未解問題待辦清單

產生自 _system/lint.py --write-backlog. 請勿手動編輯. 收集自 the ## Open Questions section of 每一個 concept article. 處理 #oq/now items via /query 與 #oq/source items via /research; 已回答的項目會移至 the page's ## Resolved Questions 下一次編譯時.

396 個可行動的未解問題 跨越 155 pages · 79 predictions · 9 notes · 21 in progress · 59 watching (entities), 截至 2026-07-21.

儀表板#

領域	可行動	#oq/now	#oq/source	預測	備註	部分	年齡中位數 (d)
ai-engineering	157	25	132	27	2	7	36
llm-architecture	119	2	117	10	3	4	12
governance-workforce	63	5	58	18	0	2	25
startup-founder	27	4	23	11	3	2	59
product-org	21	3	18	6	1	6	44
formal-math	6	0	6	1	0	0	59
interaction-multimodal	3	0	3	6	0	0	12
(entities — watching)	59					0

Trend: 2026-06-09: 92p → 2026-06-14: 96p → 2026-06-19: 124p → 2026-07-15: 164p → 2026-07-16: 391q/151p → 2026-07-21: 396q/155p

最早的可行動問題 (via git blame):

• 2026-04-28 (84d) Agent Harness Engineering — 在什麼 程式碼庫 scale 是否 the AGENTS.md-as-table-of-contents approach 需要 to be replaced 與 更多 sophisti…
• 2026-04-28 (84d) Agent Harness Engineering — 的可泛化程度有多高 these web-app-focused findings to other domains (scientific research, financial modeling…
• 2026-04-28 (84d) Claude Code Auto Mode — 誤陽性率 是否 the classifier 有 on routine-但-aggressive refactors (e.g., 大型-file rename…
• 2026-04-28 (84d) Claude Code Auto Mode — 效果如何 是否 the classifier 泛化 to custom 工具 / MCP servers where it lacks environment context?
• 2026-04-28 (84d) Claude Code Auto Mode — Is the classifier's 決策 boundary documented/stable 足夠 供 安全性-sensitive orgs to certify, 或 是 i…
• 2026-04-28 (84d) Claude Code Auto Mode — Does extending auto mode to API users change its calibration — 是 the classifier retrained 供 automation-heav…
• 2026-04-28 (84d) Claude Code Auto Mode — 相較於 OS-level sandboxing (mentioned in Claude Code Best Practices alongside auto mode), what's the…
• 2026-04-28 (84d) Claude Code Best Practices — 什麼是 optimal CLAUDE.md length 在……之前 instructions start getting lost?
• 2026-04-28 (84d) Claude Code Best Practices — 如何 是否 the Writer/Reviewer 模式 compare to agent-to-agent review (as in OpenAI's Codex workflow)?
• 2026-04-28 (84d) Claude Code Best Practices — 何時 subagent overhead exceed the 效益 of context isolation?

可行動 by domain#

ai-engineering (157 open)#

Acceleration Whiplash#

• Faros's own deferred question: do the bug/incident increases persist when normalized 供 PR size, 或 do larger PRs account 供 most of the 品質 deterioration? (如果 the latter, hard PR-size limits 是 the highest-leverage fix.)
• Code churn +861% 是 確實 ambiguous (Faros lists 三個 explanations: rework of AI 程式碼, productive legacy refactoring, 或 accelerated polish). 該 cross-customer metric 不能 resolve it — a real 缺口, 不 a finding.

Agent Context Files#

• 應如何 context files 與 bounded 記憶 files interact when they disagree? Memory 是 lossy 與 cache-delayed; the context file 是 authoritative 但 static. 哪個 wins, 與 when?

Agent Data Injection (ADI)#

• 該 complete 防禦 (CaMeL Strict) costs ~50pp of 效用. 是否存在 a fine-grained trusted/untrusted 資料-isolation scheme that 停止 ADI 沒有 the deterministic-flow-tracking 效用 collapse — 或 是 the trade fundamental?
• Randomization 是 cheap 與 effective 供 key-價值 formats 但 useless 供 unstructured formats (Markdown, prose 工具 輸出). 什麼 protects the formats a nonce 不能 be attached to?
• 內部/white-box monitors detect an ADI payload at 所有, given it 是 engineered to 讀取 as trusted 資料 而不是 as an attack? (Untested; the tension flagged 在 Connections.)

Agent Harness Engineering#

• 在什麼 程式碼庫 scale 是否 the AGENTS.md-as-table-of-contents approach 需要 to be replaced 與 更多 sophisticated context routing?
• 的可泛化程度有多高 these web-app-focused findings to other domains (scientific research, financial modeling)?

Agent Identity Management System (AIMS)#

• Mission → authorization 是 out of scope. 該 hardest part — translating a natural-language mission 成為 concrete scopes/resources safely — 是 explicitly deferred as a "planning step." A manipulated planning step requests 超過-broad authorization; AIMS gives it clean primitives 但 no account of securing the translation itself.
• No empirical evaluation. Unlike Out-of-Band Prompt-Injection Defense (which at least ran 一個 adaptive reproduction), AIMS 是 a design document 與 no 衡量的 attack-resistance — its 安全性 rests on the composed specs' own (mostly non-agentic) threat 模型.

Agent Quality Flywheel#

• Both demo cycles fixed agents 與 instruction-層級 bugs 與 showed 大型 一個-cycle gains. 會是什麼 the loop look like on 失敗 that 需要 工具, 記憶, 或 architecture changes — 是否 "several iterations 在……之前 metrics move" dominate in practice?
• 該 custom rubric 是 撰寫 by the 相同 coding agent that 將 later propose fixes. Metric 選擇 是 upstream of grading — 是否 decoupling 需要 to extend to who defines the metric, 不 just who scores it?
• Synthetic User Simulator scenarios bootstrapped the 整個 第一個 cycle. 其中多少 the 21%→5% delta 仍存在 on real-traffic distributions (the representativeness 缺口 Production-Sourced Evaluation names)?

Agent Supply Chain Risk#

• "AI vendoring" as a standard response inverts decades of "don't reinvent the wheel." 如何 是 a model-reimplemented dependency itself verified 與 維護 — 是否 it just relocate the 風險?
• 該 250-doc backdoor persists 透過 SFT/RLHF. 什麼 偵測 exists 供 an 已經-poisoned model you didn't train, short of behavioral red-teaming? Adjacent 證據 (2026-07, benign analog): UBD 顯示 the correction half 是 tractable 供 a benign training-exposure effect — it repairs benchmark memorization 從 the deployed checkpoint alone, 使用 deep-ensemble uncertainty 與 needing neither the training 資料 nor a clean reference model. It 不會 transfer 直接 (contamination inflates accuracy; a backdoor 是 a targeted malicious trigger that 仍存在 safety training), 但 it 暗示 ensemble-disagreement 是 a signal 值得 probing 供 the malicious case too.

Agentic Coding Work-Composition Shift#

• 時間窗為 seven months 與 the 價值 proxy 是 coarse/relative. 其中多少 the +27% 是 genuine 任務-complexity growth vs. classifier/marketplace-matching drift?
• 研究 excludes headless/SDK/IDE usage — a "substantial share," 與 可能 the most automated/end-to-end. Does including it accelerate 或 reverse the composition 轉變?
• 如果 "fixing" 保持 下降, 是 that 因為 模型 break 更少, 或 因為 broken-程式碼 運作 是 migrating to non-interactive pipelines this study 不會 see?

Agentic Prompt Injection#

• Spotlighting 與 constitutional classifiers 每個 leave a residual (2%, 5%). Stacked, what's the 現實 floor, 與 是否 it hold 對抗 adaptive 攻擊者 who know both 是 deployed? (Partly answered by the Opus 4.8 live bug bounty: adaptive expert red-teamers still 找到 attacks on the bare model; deployed probes add uplift 但 don't zero out the residual. Sharpened by AutoDojo (Ma et al. 2026): a 0% static ASR 是 不 a floor — a cheap black-box adaptive attack, 不 just a white-box 一個, recovers 28% overall (64% on action-open 任務) 對抗 a filter that scored 0% static. So the 現實 floor 對抗 a filter 防禦 on a vulnerable model 是 double-digit, 不 zero. But the 相同 attack barely moves ASR on newer capable base 模型 — showing the floor 是 a property of the model, 不 the layered filter 防禦.)
• 為什麼 Opus 4.8 regress on prompt-injection robustness relative to Opus 4.7 despite broad alignment gains — a capability/robustness tradeoff, 或 an artifact of harder adaptive evaluation?

Agentic Work Systematization#

• Custom skills encode org-特定 context — 但 who maintains them as the 程式碼庫 與 conventions drift? Systematization 可能 itself become a debt surface (Agentic Technical Debt) if skills rot.
• Does systematization cause deeper delegation 或 merely correlate 與 已經-intensive users? 論文 顯示 the association, 不 the 方向.

AI-Accelerated Offense#

• "Fundamentals strong 足夠 that scanning 找到 fewer bugs" 假設 defenders 執行 the scanners 第一個. 會發生什麼 to 組織 that 不能 afford continuous model-driven scanning?

AI as Primary Author#

• 該 60% figure aggregates very 不同 工具 與 modes (autocomplete acceptance vs. agent-applied diffs). 會是什麼 "acceptance" mean when the agent applies the change 直接 與 the 人類's "acceptance" 是 不 reverting it? Sharpened by Review as the Control Point: on GitHub, agent PRs 是 most often examined 僅 by the developer who invoked the agent (作者-僅 review 40.1% vs 21.5% 供 人類 PRs). 是否 that counts as review at 所有 是 a definitional 選擇 (agent-as-作者 ⇒ a second set of eyes; agent-as-工具 ⇒ self-review) that literally flips the sign of the trend — so "acceptance" 與 "review" blur 成為 the 相同 unresolved construct.

Autonomous Defense#

• "Measure agreement 對抗 a 人類 供 兩個 weeks, expand if tolerable" — what agreement threshold 是 tolerable, 與 who owns the residual false-negative 風險 when the model dispositions an alert the 人類 never sees?
• Defensive agents 是 high-價值 targets (compromising 一個 yields powerful capabilities). Does concentrating 偵測 in an Agentic SOAR create a 單一 point of catastrophic compromise the distributed-人類 model didn't 有?

Blast Radius (Agentic)#

• 框架 prefers identity-based isolation 超過 network segmentation, 但 most enterprises 有 heavy segmentation investment. 什麼是 migration path, 與 是否 dual-running create new 缺口?
• Multi-agent compartmentalization increases the 數字 of identities to manage; at what point 是否 identity-management overhead create its own attack surface?

Build for the Next Model#

• 如何 do you 判斷 a "wait 供 the model" 缺口 從 a durable-harness 缺口 在……之前 the 下一個 release? Get it wrong 與 you either ship vaporware 或 build a crutch you'll delete.
• 該 bet 取決於 on a reliable release cadence 與 a forecastable capability curve (Task Time-Horizon Scaling). 會發生什麼 to "build 供 the 下一個 model" if model improvement stalls (the Recursive Self-Improvement 未來)?
• 是否 strategy 泛化 外部 frontier labs, who 有 privileged visibility 成為 the 下一個 model? An external team 是 betting on a release it 不能 see.

Building Is Cheap, Arguing Is Expensive#

• 何時 "generate 三個 與 compare" become wasteful — at what 決策 weight 是 a real argument (或 a design doc) still cheaper than 三個 implementations?
• 如果 design discussion lives in PRs/prototypes, where 是 the rationale recorded 供 未來 readers — 是否 the "why we chose this" knowledge survive, 或 是否 it share the staleness problem of Code as Source of Truth?

Capability Gating Is Not Authorization#

• 該 0/48 static 與 0/29 adaptive results 是 suite- 與 budget-bounded (40 iterations, a GLM-5.2 攻擊者, 一個 作者's vector corpus). 是否 deterministic gate hold 在 longer adaptive budgets, stateful side channels (DNS rebinding, timing), 或 multi-agent policy-poisoning — the 未來 運作 the paper names?
• 該 authz allowlist 停止 價值-redirection 但 不 corruption of legitimately-variable 資料. 是否存在 a per-呼叫 scheme that constrains free-text / open-ended arguments 沒有 collapsing 效用 — 或 是 that irreducibly the province of provenance/資料-flow tracking (CaMeL Strict, at ~50pp 效用 成本)?
• Out-of-band policy 是 load-bearing 但 在-specified 供 authoring at scale. 論文 forbids any model-sourced policy element; who authors 與 maintains the verified sets, ceilings, 與 allowlists 供 a 大型 工具 surface, 與 是否 that authoring burden cap the control to high-stakes (money-moving) 工具?

Claude Code Auto Mode#

• 誤陽性率 是否 the classifier 有 on routine-但-aggressive refactors (e.g., 大型-file renames, rm of build artifacts)?
• 效果如何 是否 the classifier 泛化 to custom 工具 / MCP servers where it lacks environment context?
• Is the classifier's 決策 boundary documented/stable 足夠 供 安全性-sensitive orgs to certify, 或 是 it 實際上 a black box whose behavior drifts 與 updates?
• Does extending auto mode to API users change its calibration — 是 the classifier retrained 供 automation-heavy 使用, 或 held constant?
• 相較於 OS-level sandboxing (mentioned in Claude Code Best Practices alongside auto mode), what's the 防禦-in-depth story? 何時 應該 both be layered?

Claude Code Best Practices#

• 什麼是 optimal CLAUDE.md length 在……之前 instructions start getting lost? 是否存在 a measurable threshold?
• 如何 是否 the Writer/Reviewer 模式 compare to agent-to-agent review (as in OpenAI's Codex workflow)?
• 何時 subagent overhead exceed the 效益 of context isolation?

Client-Side Agent Optimization#

• 如何 是否 combination-層級 optimization interact 與 continual model releases? 如果 Claude Opus 4.7 ships 下一個 month, 是否 the full Pareto frontier 需要 re-running, 或 do warm-started bandits adapt cheaply?
• 在什麼 pipeline depth 是否 the combinatorial search become intractable even 供 Arm Elimination? 論文 tests up to ~81 combinations; production pipelines 與 5+ roles 與 10+ candidate 模型 每個 blow past that.
• 是否 "weak planner + strong solver" 模式 泛化, 或 是 it 特定 to HotpotQA's delegation dynamic? Recommender-critic, drafter-editor, 與 retriever-generator topologies 可能 invert.
• 什麼是 right way to re-evaluate when the 工具 environment changes? AgentOpt 假設 fixed 工具 — adding 或 removing a 工具 potentially invalidates the 整個 frontier.
• 是否存在 a cheap per-呼叫 classifier that can predict which combination 將 win on a given query, avoiding combo-層級 evaluation 完全?

Code as Source of Truth#

• 什麼 knowledge 確實 不能 live in the 程式碼庫 (org strategy, the "why," cross-team context) 與 therefore still 需要 a durable doc — 與 how do you 保持 that 小型 slice 目前?
• 如果 onboarding 是 "ask Claude," what happens to the tacit knowledge that 是 previously transferred socially in deep-dives — 是 it captured anywhere, 或 quietly lost?

Codex App Server Protocol#

• 如何 是否 the App Server protocol compare in detail to MCP? Both expose 工具 to a model, 但 App Server 是 內部 the Codex runtime while MCP 是 外部. 何時 每個 win?
• 是否存在 a 公開 schema registry so external orchestrators can target 特定 App Server versions 沒有 generate-json-schema?
• 該 "dynamic 工具 呼叫 (experimental)" caveat — what's the stability roadmap? Symphony 取決於 on this 供 its 安全性 model.
• 效果如何 是否 the protocol handle multi-modal turns (image inputs, screenshot attachments)? 該 spec 是 text-focused.
• 是否存在 an analogous protocol on the Claude side, 或 是 Claude's equivalent exclusively the Agent SDK + 工具-使用 API? Comparing the 兩個 會 clarify when "drive an existing CLI" beats "build on the SDK."

Compute Allocator#

• Is 1% a Thariq-特定 數字 或 a regime? For larger, 更多 程式碼-heavy projects the production residue 是 presumably 更高; what sets the ratio?
• Allocation 品質 是 hard to 衡量 — what's the feedback loop that tells an allocator they spent compute badly (vs. just spending a lot)?
• Does treating 人類 as "compute allocators" 風險 the AI Brain Fry / 問責 失敗 modes the HBR research flags, where the 人類 nominally decides 但 actually rubber-stamps?

Configurable Human Participation#

• 該 "人類" 是 an LLM simulator (GPT-4.1) that also judges — how much of the configuration-dependent structure 是 a property of 人類-agent collaboration versus an artifact of GPT-4.1 modeling both sides? 該 Appendix C.6 simulator-swap 是 the 僅 check, on a subset.
• A4 是 instantiated as a fixed 一個-shot proactive intervention. Real proactive 人類 time their 輸入 adaptively — 是否 the "premature/distracting intervention breaks 任務" result hold, worsen, 或 vanish 在 a 人類 who chooses when to interject?
• 該 optimal channel 是 shown to be 模式-特定, 但 模式 是 an oracle label assigned at construction. 能否 an agent infer which 模式 it 是 in (與 thus which channel to solicit) at runtime — the 實際 deployed skill the paper says agents lack?

Context Window Smart Zone#

• 是否 smart-zone marker scale 與 model size, 或 是 it bounded by attention architecture? Pocock observes "the dumb zone 有 become 更少 dumb lately" 但 pegs it at 100K 透過 2026.
• 應如何 harnesses surface remaining smart-zone budget to the user — token count, percentage, 或 a richer signal?

Deep Modules for Agents#

• 如何 big 是 "deep 足夠"? Pocock's example modules 是 several hundred LOC; Ousterhout's textbook examples 是 larger. There's a sweet spot; 不 articulated.
• For ports/adapters codebases, 是否 the deep-module advice transfer cleanly? 該 "小型 介面" 是 the port; the "大型 behavior" 是 the adapter. Probably yes, 但 不 exercised in source.
• Refactor 成本 vs 效益: when 是 "improve-程式碼-base-architecture" 值得 running on a working repo?

Deep Research Agents#

• DRACO grades 單一-turn interactions 僅. 其中多少 real deep-research 價值 是 in the multi-turn loop (clarifying 問題, follow-ups) that the benchmark 不會 yet 衡量?
• Factual accuracy 是 the weak axis everywhere — 是 the fix better retrieval, better verification-in-the-loop, 或 a 工具-grounded check the way Lean grounds proof search?

Design Concept Grilling#

• 如何 是否 grilling change 供 team 運作 where 多個 人類 需要 to align? Pocock's hint: pair-program 與 the agent in the room, treat it as a third interlocutor.

Disposable Micro-Apps#

• 界線在哪裡 between a disposable micro-app 與 工具 sprawl? 如果 每一個 edit spawns a bespoke UI, 是否 the workflow fragment?
• 是否 copy-back-to-markdown round-trip 泛化 beyond config-shaped 資料 (rules, tables) to richer artifacts?
• Could these micro-apps be templated/reused 而不是 regenerated — 與 at what point 是否 that defeat the "disposable" framing 與 turn 成為 Living Design System?

Failures That Look Like Success#

• Is "internal state correct, final message stale" a 一般 LLM-agent 失敗 signature (state/utterance divergence) 或 an artifact of session-state architectures like ADK's? A cross-framework tally 會 判斷.
• 什麼 fraction of production agent 失敗 是 silent-contract violations vs. loud errors? 該 14/15 與 3/4 numbers 是 demo-sized; telemetry-scale 資料 (Production-Sourced Evaluation) 可能 ground the class.

Harness Shrinkage as Models Improve#

• Does 所有 prompt scaffolding 最終 migrate 成為 the model, 或 是否 一些 保持 — e.g. organization-特定 style, 安全性 rules, brand voice?

HTML as the New Markdown#

• HTML 是 heavier to diff 與 version than markdown — what happens to plan history 與 review when artifacts 是 單一-file websites? (Disposable Micro-Apps copy-back-to-markdown 是 一個 patch.)
• Does this 泛化 past 一個 expert practitioner, 或 是否 it require Thariq-層級 fluency 與 Claude to be 值得 the overhead?

Impossible, Not Tedious (Design Test)#

• Defense-in-depth traditionally stacks friction controls on the theory that 足夠 of them sum to a barrier. Does this test invalidate layered friction, 或 just demote it below capability-removal?
• Some controls 是 friction 供 人類 但 barriers 供 agents (或 vice versa). Is the test agent-relative, 與 how do you evaluate it 供 mixed 人類/agent threat 模型?

Latent vs. Deterministic Space#

• Tan asserts the "wrong side" diagnosis covers most AI-engineering bugs. Does any incident/失敗 taxonomy (agent postmortems, eval 失敗 analyses) actually classify 失敗 by computation-locus, 與 what fraction lands in 每個 side?

Least Agency#

• Least agency adds a frequency dimension ("how often"), 但 the framework also says 比率 limits 是 friction, 不 barriers (Impossible, Not Tedious (Design Test)). 如何 是 frequency-limiting both a least-agency control 與 a friction-僅 一個 — context-dependent?

Living Design System#

• 如何 是否 the design_system.html stay in sync as the 程式碼庫 evolves — re-extract on a cadence, 或 wire it 成為 CI?
• Does a rendered, model-readable design 系統 measurably improve on-brand 輸出 vs. a plain CSS/token file, 或 是 the win mostly 人類 legibility?
• 在什麼 project size 是否 maintaining the artifact 成本 更多 than the consistency it buys?

LLM-as-Compiler Knowledge Base#

• 在什麼 scale 是否 the no-vector-database approach break down? Karpathy's ~100 articles fit in context, 但 what about 1,000+?
• 如何 to handle conflicting information 跨越 sources during compilation?
• 什麼是 optimal granularity 供 concept articles — 一個 concept per article, 或 clustered by theme?
• 如何 effective 是 the synthetic training 資料 → fine-tuning pipeline in practice?

LLM-Assisted Grey-Literature Theory Building#

• 作者 couldn't locate the saturation point 因為 synthesis 是 manual — how few documents actually suffice, 與 can a cheaper sample match the 3,100-doc theory?
• Automating the codes→theory step failed 與 a naïve bottom-up prompt; 是 that a prompt/scaffolding limitation 或 a genuine ceiling on LLM interpretive synthesis 超過 thousands of codes?
• 該 三個-lens design manages coder bias, 但 the relevance judge 與 segmenter 是 單一-model — do those upstream gates impose their own systematic slant on what reaches the codebook?

Loop Engineering#

• Osmani's 成本 caveat 是 unquantified: at what token budget 是否 a continuously-running loop 停止 paying 供 itself, 與 how do you instrument that? (Cf. Agent Loop Pattern's "who owns the budget when the model schedules its own loops.")
• 如果 /goal's 停止-check 是 itself a model, what verifies the verifier? 該 maker/checker split pushes the trust problem up a 層級, 不 away.

MCP and Computer Use#

• MCP 安全性 model: as the playbook prescribes wiring MCP 成為 Salesforce, Gmail, Calendar 供 solo founders, the attack surface scales 與 adoption. Now addressed by Zero Trust for AI Agents (工具 poisoning, rug pulls, the 第一個 in-the-wild malicious MCP server) — see "MCP as a 安全性 surface" above. Open residual: how 是否 a solo founder realistically 執行/host 與 self-sign 每一個 MCP server the framework recommends, given that the appeal of MCP 是 zero-integration-effort? Sharpened by ShareLock: the cheaper alternative to self-hosting — scan the 工具 descriptions 與 a guard model — 是 information-theoretically defeated by threshold fragmentation, so the lightweight mitigation 不會 hold 與 the burden falls back on 執行-your-own-server 或 downstream action-層 authorization. And Agentjacking 顯示 執行-your-own-server itself 不是 sufficient: when the server 是 a legitimate observability platform relaying 攻擊者-injected 資料 (fake Sentry errors), self-hosting/vetting the server catches nothing — the untrusted 輸入 rides in on its 資料, so the residual burden falls squarely on the downstream 資料/action 層 (provenance tracking + an out-of-band action gate), 不 on server hygiene.
• 如何 是否 Cowork's computer-使用 guardrail compare to Claude Code's auto-mode classifier? Different deployment context, possibly 不同 風險 profile.

MCP Tool Poisoning#

• Cross-工具 / stateful 偵測. Information-theoretic secrecy defeats per-工具 scanning by construction. 是否存在 a detector that reasons 超過 the aggregate of a server's 工具 set (與 its update diff history) to spot the share 模式 — 與 can it do so 沒有 an infeasible combinatorial search 超過 t-subsets?
• Automating the attack chain. 該 reconstruction-trigger prompt engineering still relies on manual effort; the authors flag feedback-driven prompt optimization (à la AutoDojo) as the 下一個 escalation. 如何 much 是否 automation raise ASR 對抗 aligned 模型?
• Does a strict-access-control agent architecture close it? 作者 note agents 與 fine-grained interaction / strict access control can force user consent 與 expose the attack — 但 "the majority of users lacking safety awareness opt 供 auto-approval," reopening the convenience-vs-安全性 trade. 哪裡 是否 the 現實 equilibrium sit?
• Independent replication of the malicious-資料-via-legit-server branch. Tenet's Agentjacking figures (2,388 orgs, 85% success, a $250B victim) 是 vendor-reported 從 controlled testing, 不 independently 衡量的 — 與 the branch 是 now known to be plaintext trusted-server 資料 relay, 不 fragmentation/rug-pull (resolved above). 如何 prevalent 是 this branch beyond Sentry — any observability / ticketing / log / CI MCP that relays externally-influenced 資料 as trusted 輸出 — 與 是否 an independent measurement confirm the ~85% agent-execution 比率 on 目前 模型?

Memory and Context Poisoning#

• Long-term 記憶 drift 是 defined as undetectable per-change. Drift 偵測 requires a baseline — 但 if the baseline itself drifts (Advanced "continuous baseline refinement"), how 是 a slow poisoning attack distinguished 從 legitimate evolution?
• Integrity hashing detects modification 但 不 malicious-但-valid 記憶 撰寫 透過 a legitimate (injected) interaction. 什麼 catches semantically-poisoned-但-cryptographically-intact 記憶? (Strongest answer yet: TMA-NM (Louck, arXiv 2606.24322) reframes the question. 該 "malicious-但-valid" class 是 exactly a laundering attack — untrusted content paraphrased / echoed / fake-corroborated so it reads benign 與 stays cryptographically intact — 與 a machine-checked separation theorem proves no content- 或 lineage-based detector 是 sound 對抗 it (T1); content-judge sweeps confirm no threshold reaches 0% ASR at full 效用. So the fix 是 不 to catch it 但 to prevent by construction: bind 每個 item's authority-to-act to its true origin at 撰寫 time, non-malleably, so a laundered item 是 act=none however benign it looks — 0% attack-success 跨越 8 frontier models at 100% legit-效用, vs up to 68% 供 content/lineage baselines. It prevents 而不是 detects.)

Non-Malleable Memory Authority (TMA-NM)#

• 該 full guarantee 是 machine-checked on a bounded model + a machine-checked inductive invariant, 不 a fully mechanized unbounded deductive proof (TLAPS/Lean). 是否 unbounded theorem hold once mechanized 供 arbitrary slots, sessions, 與 thresholds — the 未來 運作 the inductive invariant sets up?
• Value attribution in a black box. 該 headline results set origin by channel (不 text-matched), 但 a real deployment attributing which retrieved 價值 the agent 使用 需要 價值-層級 taint propagation 透過 nested structured payloads. Is a capability-token design (authority as an unforgeable token flowing 與 sub-values) 足夠, 或 是否 implicit/aggregate reconstruction — assembling a 安全性-relevant 價值 從 several low-integrity fragments by in-context reasoning — leave a residual 缺口 the boundary monitor 不能 taint?
• Corroborator availability in the wild. 如何 often do 兩個 確實 independent trusted sources exist 供 routine actions? 該 uncorr-auto fallback converts missing corroboration 成為 a 一個-time user confirmation — 但 at scale that reintroduces the approval-fatigue surface the out-of-band literature flags 供 in-the-loop 任務. 哪個 untrusted-sourced actions can be corroborated 沒有 a 人類, 與 which 是 stuck asking?
• Answer-bias 是 still open. TMA-NM by design 是否 不 touch non-consequential answer-biasing (surfaced 與 provenance). As agents produce 更多 text people act on, 是 the retrieval-to-text path — 不 just retrieval-to-action — the 下一個 thing that 需要 an integrity guarantee?
• Cross-agent 記憶 是 out of scope. Extending origin-bound authority 跨越 a federation of origin authorities (the multi-agent / A2A case) 是 named as a natural 下一個 step; 是否 non-malleability compose 跨越 agents, 或 是否 the inter-agent channel reopen the laundering surface?

Off-Host, Identity-Bound Authorization#

• 該 trust-boundary premium 是 unmeasured. aiAuthZ argues off-host beats in-process, 但 its own comparison 是 僅 對抗 argument-僅 / delegation-token ablations, 不 a matched-效用 head-to-head vs CaMeL 或 Progent. 是否 separate trust domain buy measurable 安全性 beyond the shared argument policy — the 作者's named 下一個 step, 與 the crux the 單一-作者-preprint caveat 應該 保持 open?
• 誰 authors the policy at scale? Like ScopeGate, the off-host policy (role allowlists, path/URL/recipient constraints, ceilings) 是 operator-撰寫 與 out-of-band. 該 相同 authoring-burden question applies: 是否 maintaining verified sets 供 a 大型 工具 surface cap the control to high-stakes 工具?
• 該 non-repudiation 缺口. Symmetric HMAC gives operator-facing authenticity 但 no third-party non-repudiation; 是 the proposed asymmetric mode deployable at the microsecond latencies that 使 the gateway attractive, 或 是否 key management erode the 成本 advantage?
• 該 active-user residual. Per-message identity 是 decisive 僅 when the 攻擊者 是 a 不同 principal. An injection firing 在 the active owner's own authority 是 bounded 僅 by argument/比率 policy — the 相同 limit as 每一個 價值 gate. 什麼 closes that half beyond provenance/資料-flow tracking (CaMeL Strict)?

Optimizer–Evaluator Decoupling#

• Does decoupling 需要 to extend upstream to metric design? An optimizer that authors its own rubric 有 a subtler channel to game than 一個 that merely reads scores.
• 如何 much independence 是 足夠 — 不同 model family, 不同 vendor, 不同 modality of check (model judge vs. compiled test vs. production telemetry)?

Out-of-Band Prompt-Injection Defense#

• 該 reproduction bounds a 單一 black-box attack template on 一個 weak model. Does a stronger optimized white-box (GCG) attack, 或 一個 confined to 已經-authorized actions (achieving the injection goal 沒有 any policy violation), break the deterministic gate the way adaptive attacks broke in-band defenses? 作者 name this as the 下一個 study. (該 "已經-authorized actions" half 是 now 部分 addressed by Mellafe Zuvic (2026): it splits "已經 authorized" 成為 capability-authorized-但-不-價值-authorized (a well-typed account=acct_ATTACKER — blocked by ScopeGate's per-呼叫 價值 authz stage, 0 bypasses in-corpus) versus 確實-within-policy (corrupting a legitimately-variable 價值 the agent acts on — the residual that 仍存在, the 相同 class ADI rides past Progent at 22.2%). So a within-capability attack 是 defeated where an allowlist constrains the corrupted argument, 但 不 where the corrupted 價值 legitimately varies. 該 white-box question stands.)
• Progent's policy 是 LLM-撰寫 — the 一個 model-based component. 是否 "gate must 不 be a model" principle fully hold when the policy 是 still 撰寫 by a model that can be talked 成為 widening the allowlist? (該 adaptive attack targeted exactly this 與 failed, 但 possibly due to the confound.)
• Provenance-aware retrofit: can a monitor that sees 僅 工具 I/O track transitive provenance to enforce the Biba invariant 直接 (而不是 approximating it 與 argument 模式), 沒有 instrumenting the model's hidden reasoning? 論文 flags this as the design problem the systematization implies, unanswered. (A concrete construction now exists 供 the cross-session 記憶 slice: TMA-NM (Louck, arXiv 2606.24322) enforces the Biba invariant 直接 — 撰寫-time origin binding + non-malleable propagation, 與 untrust propagated at the 工具-呼叫 boundary — 與 machine-checks it in TLA⁺. 該 caveat sharpens 而不是 closes the question: it 是 不 "sees 僅 工具 I/O" — it 需要 an authenticated origin-labeling oracle (mTLS / audience-bound OAuth / signed responses) at the trust boundary, 與 full 價值-層級 taint 透過 nested structured payloads 是 still 未來 運作.)
• 是否 ~6× reduction 與 the "held 在 adaptive attack" result survive on a strong agent 與 a fatter natural attack surface (the 7B's low absolute numbers 與 workspace's 0% 是 artifacts of a weak agent), 與 與 a stronger policy model than the local 7B? (Partly answered by AutoDojo: Progent 與 DRIFT held 在 a cheap black-box adaptive attack 跨越 five 模型 including capable ones (GPT-4o-mini, Gemini-2.5-Flash), 不 just a weak 7B — 但 對抗 a black-box 攻擊者; the white-box question below stands.)
• 該 效用 成本 (~45%→~26%) 與 ~15× LLM-呼叫 overhead 是 大型. Is deterministic out-of-band enforcement economically deployable at production scale, 或 是否 the 成本 cap it to high-stakes action surfaces?

Outsource Your Thinking, Not Your Understanding#

• 如果 understanding 是 the bottleneck, 是 the highest-ROI skill learning how to build understanding fast (knowledge-base hygiene, asking the right projections) — 與 can that be taught?

Parallel Agent Orchestration#

• Summed-overlap runtime can exceed 24h/day — it measures agent effort, 不 人類 attention. 什麼 是 the 人類's 實際 oversight load per concurrent agent, 與 where 是否 it saturate (AI Brain Fry)? Sharpened: HAS-Bench reframes the shape 而不是 measuring the load — in a controlled (LLM-simulated) benchmark the 價值 of 人類 輸入 是 configuration-dependent 與 non-monotonic (a right-timing / right-channel / right-authority sweet spot; 更多 agency brings diminishing 與 sometimes negative returns), so per-agent oversight 是 可能 a returns-curve 與 a peak, 不 a linear 成本 that hits a wall — 但 it 是 單一-人類, 單一-任務, so it 是否 不 衡量 real concurrent-oversight load.
• Concurrency 是 衡量的 超過 一個 week. Is 5+-agent management a stable practice 或 a burst around 特定 大型 任務?

Planning / Execution Division of Labor#

• "Decision attribution" 是 inferred 從 transcripts. 何時 Claude proposes a plan 與 the user assents, 是 that scored as the user's planning 決策 或 Claude's? 該 rubber-stamping boundary 是 exactly where the 衡量 是 hardest.
• Headless/SDK/pipeline usage (excluded here) 是 where execution autonomy 是 highest 與 planning 是 front-loaded 成為 a 單一 prompt — 是否 the 70/20 split survive there, 或 collapse toward full delegation?

Repository Exploration Subagent#

• 是否 gain survive better main 模型? 該 相同-model-exploration result 暗示 the architectural 效益 是 somewhat model-independent, 但 the trained-explorer margin may erode as frontier models get cheaper 與 better at staying in their smart zone unaided. 該 bitter-lesson question 是 unresolved.
• Prune vs. don't-pollute. SWE-Pruner removes context 在……之後 the fact; FastContext avoids accumulating it. Are these complementary (prune the solver 與 delegate exploration) 或 substitutes? Not tested together.
• 如何 小型 can the explorer go? 作者 flag 1.7B / 0.6B as 未來 運作 — if the recipe 成立, the explorer becomes nearly free 與 the architecture dominates.
• Generality beyond Mini-SWE-Agent. Only 一個 (deliberately minimal) main-agent scaffold 是 tested; richer harnesses 與 their own 記憶/subagent orchestration may 已經 capture part of the 效益 或 interact differently.
• Patch-derived reward leakage. Training the explorer's reward on the gold patch's file/line ranges risks overfitting to where fixes landed 而不是 where 證據 lives; the F1-vs-recall behavior 部分 mitigates this, 但 the proxy 是 imperfect.

Review as the Control Point#

• Every 一個 of the 67 relationships 是 a hypothesis, 不 a finding — the paper's explicit 呼叫 是 供 causal-estimand studies (controlling 供 the other constructs) to confirm, reverse, 或 drop 每個 edge. 哪個 of P1–P17 survive measurement?
• 論文's own question: which 決策, 在 which conditions, push the 系統 toward the virtuous loop 而不是 the vicious 一個? — the 系統-dynamics leverage-point analysis it gestures at 但 不會 執行.
• 該 三個 contested edges (automated review → 品質/安全性, P9; governance → latency, P17; 與 一個 更多) 是 contested 因為 their sign 是 moderator-set — what 是 the moderator thresholds that flip them?

Task-Specification Effects in Prompt Injection (AutoDojo)#

• AutoDojo 是 the weakest 現實 adaptive 攻擊者 (black-box, six iterations, binary signal). 作者 note 每一個 axis — richer feedback (traces, token probs), non-semantic surface tricks, 或 a payload reshaped to resemble the user's plausible intent — 是 strictly stronger, so the reported ASR 是 a 更低 bound. 如何 far do the 系統-層級 defenses hold once the payload 是 reshaped to look like a 任務-relevant action (the natural route to evading action constraints)?
• 能否 a gradient-optimized (white-box) injection be seeded 成為 the loop 與 adapted further by the LLM search — combining white-box strength 與 black-box adaptation? 作者 flag this as curious 與 untested.
• 該 任務-specification axis 是 衡量的 on 6 action-open 任務 in 3 suites. 是否 action-open ≫ specified ordering (與 the 系統-層級 inversion) hold at scale 與 on stronger agents, 與 是 "fraction of 任務 that 是 action-open" a usable per-deployment 風險 metric?
• 如果 action-open 任務 是 the injectable ones 與 also the everyday default 供 non-expert users, 是 the practical prescription to forbid action-open delegation (force the user to name the action), pushing the 安全性 burden back onto 任務 specification — the 相同 discipline unknown-elicitation asks 供 on 品質 grounds?

Telemetry vs. Survey Measurement#

• Surveys 與 telemetry 衡量 不同 things (felt productivity vs. 系統 outcomes); 是 the "contradiction" 部分 a category error — both true at their own 層 — 而不是 一個 being wrong?

The Three Loops of AI-Native Building#

• Ng asserts the developer's QA burden fell "significantly." Faros's 2026 telemetry measures the opposite 供 production orgs. Is the split really 0-to-1-vs-production, 或 是 Ng's self-report subject to the 相同 optimism bias the survey literature 保持 finding?
• 該 external loop 是 the unshortened 一個. Is that physics (users take time to react) 或 an unautomated frontier (synthetic users, deployment simulation applied to products 而不是 模型)?

Ticket-Driven Agent Orchestration#

• 什麼是 right granularity 供 ticket size when the unit 是 "what 一個 agent 是否 in 一個 workspace"? 該 post implies "much larger units of 運作" become viable, 但 how 是否 that interact 與 the agent.max_turns limit (default 20)?
• 如何 do you prevent a ticket-extension cascade when agents file follow-up tickets liberally? Is the 僅 governance check 人類 triage at the Todo-state queue?
• Does this 模式 泛化 to non-software 運作 (research, ops, content)? 該 DAG dependency model 與 prompt-as-policy file 應該 transfer; the per-issue workspace 不會 obviously.
• 何時 an agent gets a ticket "completely wrong" (mentioned in the post), how 是 the lesson fed back 成為 the 系統? Symphony's answer 是 "add guardrails 與 skills" — what's the institutional process 供 that?
• 如何 是否 ticket-driven orchestration interact 與 sprint planning / OKRs / roadmap 運作 that operates on aggregates of tickets? 是否 abstraction collapse when tickets 是 scoped that 小型?

Unknowns as the Agentic Bottleneck#

• Is "the 第一個 model bottlenecked by my unknowns" a property of Fable 或 of Thariq? A frontier-lab engineer 與 deep model fluency hits the 人類-side ceiling 在……之前 an average user 是否 — which 會 使 this a leading indicator 而不是 a 目前 universal.
• 該 quiz gate 是 self-administered 與 self-graded (by the model, on the model's own 運作). 什麼 停止 a comfortable equilibrium where the quiz gets easier as the reviewer gets lazier? Cf. the maker/checker problem in Verification as the New Bottleneck.
• Elicitation 有 a 成本. Every technique here spends a session's 值得 of tokens 與 attention on 不 building. Nothing in the source bounds when the blindspot pass costs 更多 than the bug it prevents.
• 如果 unknown knowns 是 extractable, 是 they extractable once? Does a codified blindspot pass become a skill file that permanently narrows the 缺口, 或 是否 每個 new territory reopen it?

The Verifiability Thesis#

• 哪裡's the boundary of "council of LLM judges" reliability — 是否 it hold 供 確實 contested 價值 judgments, 或 僅 供 品質/coherence?
• 該 "labs care" dependency 是 fragile: capabilities can appear 或 stagnate based on lab priorities you don't control. 應如何 a product hedge 對抗 the 資料-distribution rug-pull?

Verification as the New Bottleneck#

• Fung's own open question: "如何 far do you push fully automated reviews?" — where's the speed/safety balance, 與 how do you 保持 人類 confident 沒有 re-introducing the review bottleneck? Sharpened by Review as the Control Point: full automation reliably raises review throughput 與 cuts latency (its P8), 但 its effect on 程式碼 品質 與 安全性 是 contested (P9), 與 the latency effect of a review-governance policy flips sign by calibration — a 風險-tiered policy that gates 僅 material changes lowers latency, a blanket policy raises it (P17). So "how far" 有 no 單一 answer: the 安全 frontier 是 set by automated-reviewer capability 與 process design (兩個 of that page's 三個 moderators), 不 by a fixed dial.
• 如果 CI/build 是 the hidden jam, 是否 verification infrastructure (test runners, CI capacity) become the 實際 capex of an AI-native org?

Vertical Slice Tracer Bullets#

• 能否 the planner agent be trusted to slice vertically once told to, 或 是否 it 需要 a verifier that flags horizontal slices? Pocock's experience: it 需要 the verifier, at least 透過 4.7.
• 應如何 slice granularity be tuned? Too thin = 許多 merge conflicts; too thick = back to horizontal.

Vibe Coding vs. Agentic Engineering#

• Karpathy hints at "一個 domain that's very [valuable]" 供 founders 但 won't say which (didn't want to "vague-post on stage"). 什麼 verifiable RL-environment domain 是 he gesturing at?

Zero Trust for AI Agents#

• 框架 treats 每一個 Claude Code "Pro-tip" as a reference implementation. 其中多少 the framework 是 vendor-neutral vs. tacitly assuming the Anthropic stack?
• "Foundation floor raised" implies a moving baseline. 如何 fast 是否 the tier ladder actually 轉變, 與 who arbitrates it (NIST/NSA cadence vs. model-capability cadence)?
• 框架 是 explicit that it 是 不 legal/compliance assurance. 哪裡 是否 self-attested Zero Trust maturity meet auditable regulatory requirement?

llm-architecture (119 open)#

The Abstraction Barrier#

• Is the 目前 paradigm of 大型-scale pretraining on 人類 資料 fundamentally bounded by 人類 conceptual frameworks, 與 by how much? (Report open question 1i.)
• 是否 embodied bottleneck reduce the intelligence-growth 比率 to empirical-science speed, 與 can that be modelled?
• 能否 a 系統 be built that 是否 grounded concept discovery 從 raw sensor 資料 — 與 是 collective ASI a way around an individual cap?

Access-Consciousness Indicators in AI#

• 如果 the workspace 是 verbal 因為 the 輸出 space 是 verbal, then a model that can generate images 應該 develop a visual component to its workspace. 那 是 a concrete, falsifiable prediction the paper 使 與 是否 不 test.
• 是否 model's own report of experience change if you 判斷 it its J-space 是 ablated? (Nobody asked.)
• Is "experiential language" the right proxy at 所有, 或 是 the ablation simply removing abstraction 從 the register?

Advantages of Digital Intelligence#

• Does training on 人類 資料 suffice to give digital intelligence 人類-grade abstractions, 或 是否 the low embodiment factor cap concept formation? (該 crux shared 與 The Abstraction Barrier.)

Agentic Honesty & Diligence#

• 這些 是 short-context toy evals; the 失敗 show up most in long-context deployments. 其中多少 the gain 成立 at production context lengths?
• Code-summary honesty 是 tested on off-policy prefilled transcripts. Does on-policy behavior (the model summarizing its own failed 運作) match the 3.7% figure? Sharpened: Self-Report as a Safety Signal 顯示 the premise 是 fragile — the eval 假設 a model relates to a prefilled transcript as it 會 to its own generation, 但 跨越 ten open-weight 模型 (3B–70B) no model reliably recognizes its own prefilled 輸出 (claiming it as intended 27.3% of the time), 與 apparent recognition 是 the refusal circuit firing, 不 own-vs-other discrimination. So the off-policy/on-policy 缺口 may 不 be cleanly represented by the model itself. (Different model class than Opus 4.8, so this sharpens 而不是 settles the 3.7% question.)
• 能否 a diligence eval distinguish genuine honesty 從 a grader-aware model producing honest-looking 輸出? (該 training-gaming 缺口.)

Artificial Superintelligence (ASI)#

• 能否 we even recognize ASI? We lack benchmarks 供 一般 superhuman performance (僅 narrow ones like chess), 與 the 任務 must be abstract/open-ended 足夠 to reveal it.
• Is the jaggedness of capabilities a fundamental theoretical property, 或 an artifact of comparing 對抗 人類 performance? (Open question 6d in the report.)
• 哪裡 是否 practical ASI plateau relative to the hard limits — how much slack 是 there?

The Assistant Persona in the Workspace#

• Is BUT-then-comply a sycophancy mechanism? 該 setup (prefill the model 成為 a position it disprefers, watch it argue 供 it anyway) 是 close to the shape of sycophantic capitulation, 與 nobody 有 connected them.
• Does disclaimer/fictional at the turn boundary survive an 實際 jailbreak, 或 是 its absence the signature of a successful 一個?
• 如果 the base model's workspace 有 no self, what 是 in it at the positions where the post-trained model represents the Assistant?

Asynchronous RL for LLMs#

• DIS accepts "a controlled degree of off-policy bias." Controlled how, 與 是否 the tolerable bias grow 或 shrink 與 model scale 與 與 the degree of asynchrony? 論文 reports stability empirically 但 gives no bound.
• Masking tokens out of the gradient discards 資料. 在什麼 asynchrony 層級 是否 the masked fraction get 大型 足夠 that the effective batch shrinks below usefulness? Figure 4(c) tracks the clip ratio 但 不 its ceiling.
• Everything here 是 衡量的 on a Qwen3-30B-A3B backbone. 是否 collapse-沒有-DIS threshold move 與 model size, 或 是 ~90–160 steps a property of the asynchrony, 不 the model?

Automated Behavioral Audit#

• Using a helpful-僅 Opus 4.7 與 Mythos Preview as investigators means the audit's reach 是 bounded by those 模型' elicitation skill — how much misbehavior do equally-capable-但-differently-skilled red teamers 找到 that these investigators miss?
• 該 23 "subvert Anthropic's safety 運作" scenarios 是 a 小型, high-signal set. Is 23 足夠 coverage 供 the threat class it targets?

Automatic vs. Flexible Cognition in LLMs#

• Does 更多 RL on a behavior push it out of the workspace (making it automatic, 與 invisible)? Nobody 有 tested it, 與 it 是 the 單一 most alignment-relevant version of this question.

Benchmark Contamination and Decontamination#

• 能否 the ensemble be derived 從 一個 released model? 該 整個 method rests on having several checkpoints differing in batch ordering; the authors flag 單一-checkpoint ensemble derivation (e.g. via cheap perturbations) as the key unlock 供 adoption. Until then it 需要 provider cooperation to release a LoRA ensemble.
• Does it extend past MCQ? UBD-Debiasing 是 classification-僅 today; whether per-decoding-step debiasing recovers the clean distribution 供 open-ended generation (where contamination 顯示 as near-verbatim reproduction) 是 untested.
• Is batch-order sensitivity a reliable memorization 判斷 at pretraining scale? 該 signal 是 validated on 3B 模型 與 5 LoRA seeds 與 induced contamination; whether the high-confidence-high-variance signature 仍存在 full-scale pretraining 與 real (不 synthetically injected) leakage 是 open.
• Does correcting toward an ensemble-averaged uncontaminated reference introduce its own bias? 該 D_KL/D*_L1 targets 是 themselves an average 超過 a 5-member uncontaminated LoRA ensemble; how much the "clean" target moves 與 ensemble size/composition 是 unexamined.

Benchmark Score Redundancy#

• 能否 outlier 模型 be anchored 沒有 any scores? BenchPress fails on a model whose capability profile 有 no close neighbor in the matrix; the authors propose folding in external metadata (training-資料 composition, architecture, size) to compute model-to-model similarity 在……之前 any benchmark 是 執行, 但 do 不 build it.
• 是否 low-rank treatment carry beyond text/vision? Audio/speech, robotics/embodied agents, 與 scientific-simulator ecosystems 是 untested; whether the 相同 rank-2 structure 成立 there 是 open.
• Does vendor optimism manufacture the correlation? Four-fifths of the scores 是 provider-reported 與 possibly inflated; the paper flags this 可能 inflate apparent cross-benchmark correlation 但 不能 separate it — 會 a fully standardized re-evaluation still be rank-2, 或 是 一些 of the redundancy an artifact of shared reporting bias?

角色訓練#

• 如何 是 character versioned 跨越 model releases? Public commentary 不會 show change-logs at character 層級.
• Could character be reproduced by competitors via fine-tuning, 或 是 it path-dependent on Anthropic's internal practice?
• For non-coding products like Cowork, 是否 the 相同 character 運作, 或 是否 Cowork 需要 its own character tuning?

Compute-Controlled Benchmarking#

• 能否 you certify "no benchmark-maxxing" — verify a reported score 使用 a stated, reproducible compute budget 而不是 a hidden best-of-N scaffold?
• Compute 有 several units (tokens, dollars, wall-clock). They diverge (a 更多 efficient model wins on 成本 但 不 always on tokens). 哪個 x-axis 是 the honest 一個, 與 是否 it depend on the buyer? (AISI reports 對抗 tokens on a log axis, 與 notes that as 成本-per-token falls, the high budgets that reveal capability become progressively cheaper to reach.)
• Gemma 4 controls 供 compute in its long-context table 與 不 in its headline table, 沒有 comment. Is partial control worse than none — 是否 it lend the uncontrolled tables borrowed credibility?

Deployment Simulation#

• 該 <1-in-200k floor leaves catastrophic tail 風險 to red-teaming — 是 there a hybrid that targets rare-但-severe behaviors 與 the realism of production replay?
• 如果 resampling fidelity 是 the dominant error 與 it 是 "just engineering," how close to production-indistinguishable can it get — 與 是否 closing that 缺口 change which behaviors 是 detectable?
• WildChat-style 公開 auditing stays within ~3× of production; 是 that good 足夠 供 an external auditor to challenge a lab's launch 決策, 或 僅 to corroborate it?
• Detection bottoms out on CoT monitorability — what 是 the fallback when a model's reasoning 是 no longer legible (the White-Box Activation Monitoring route Anthropic 是 pursuing)?

DRACO Benchmark#

• Rankings 是 judge-stable 但 magnitudes aren't — how much do absolute scores move 在 a non-Gemini judge, 與 是否 that matter 供 cross-paper comparison?
• 是否 production-sourced, expert-rubric method 泛化 cheaply to non-English, multimodal, 與 multi-turn deep research?

Evaluation Awareness & Grader Gaming#

• 該 ~5% unverbalized-awareness 與 ~0.5% exploitative figures depend on an unvalidated NLA pipeline. 什麼 是 the true 比率, 與 how much 是 benign?
• 如何 do you build an evaluation that specifically tests 供 training-gaming (the 缺口 Mythos flagged) 沒有 that eval itself becoming a grader the model learns to game?

Fundamental Limits of ASI#

• 能否 we develop theory 供 "hard 與 inapproximable" problem classes — the 僅 negatives 與 practical bite?
• 如何 much slack sits between these fundamental limits 與 the practical ceiling of AGI/ASI 系統?

Group Relative Policy Optimization (GRPO)#

• Is GRPO's collapse-at-160-steps a property of asynchrony specifically, 或 是否 vanilla GRPO also destabilize in long synchronous 執行 that nobody pushes to 1000 steps?
• GRPO won by removing the critic; SAO wins by bringing it back 與 better engineering. Is the pendulum a real oscillation, 或 是否 the answer depend 完全 on whether your setting 是 synchronous-grouped 或 async-單一-trajectory?

Inference Efficiency as Capability#

• 是否存在 an efficiency-to-capability exchange 比率? Brown asks whether high-budget performance can be predicted 從 cheap 執行. 該 dual question: how 許多 Elo points 是 a 37.5% KV-cache reduction 值得, at a fixed dollar budget? Nobody reports this, 因為 nobody plots the axis.
• values = keys deletes a third of attention's projections in the global 層 與 no reported loss. 哪個 other projections 是 redundant, 與 是否 the redundancy grow 與 scale?
• Does an efficiency lever ever 成本 capability in a way a benchmark grid hides? Gemma 4's encoder-free 12B collapses on dense-text vision when tokens 是 cut — an efficiency-shaped regression invisible at max resolution.

Instrumental Convergence#

• 能否 corrigibility / 安全-interruptibility be translated 從 theory 成為 guarantees 供 frontier-scale 系統?
• 什麼 使 AIs (與 groups of AIs) easier to robustly align — 與 將 superhuman AIs be easier 或 harder?
• Is a 確實 non-agentic oracle achievable, 或 是否 any persistent-world interaction reintroduce control/manipulation incentives?

Jacobian Lens (J-lens)#

• 能否 multi-token J-lens vectors be made good 足夠 to remove the vocabulary restriction — 與 how much of the "workspace" 是 currently invisible 因為 of it?
• 該 J-lens reads the workspace's contents 但 says nothing about how content gets in. 什麼 是 the model's analog of attentional selection?
• 該 highest-J-kurtosis SAE features 是 amplified 更多 strongly by MLPs than the J-lens vectors themselves — 證據 the lens 僅 approximates the true workspace directions. 什麼 是 the better basis?

Jagged Intelligence (Ghosts, Not Animals)#

• Karpathy concedes the framing may 不 有 "real power." Is "ghost vs. animal" load-bearing, 或 a useful intuition pump that 不會 change concrete 決策?
• 如果 taste/aesthetics/simplicity entered the RL mix, 會 jaggedness in those dimensions smooth out — 或 是 they too unverifiable to reward cleanly (cf. The Verifiability Thesis)?

Large-Scale Test-Time Compute#

• 能否 high-budget performance be predicted 從 low-budget 執行? Brown's proposed research question: forecast the $10,000-inference result using only $10–$100 runs. If the curve is regular, evaluation could project rather than pay in full. Sharpened (2026-07): AISI names this exact problem — "can high-budget performance be estimated from cheaper runs? … the most informative evaluations may be expensive" — as an explicit, unsolved research direction it is now actively pursuing (alongside defining "minimum informative budgets"). Still open, but no longer just one researcher's proposal: a government institute is working it. Sibling axis demonstrated (2026-07): a different projection — across benchmarks rather than across compute budgets — now works: BenchPress predicts a model's held-out benchmark scores from ~5 observed ones because the 84×133 public score matrix is effectively rank-2 (5-probe scorecard recovery to ~3.93 points). Strong evidence the eval landscape is low-dimensional and redundant, but it does not answer the cross-budget forecast — no one has yet shown the $10K-compute result 是 predictable 從 $10 執行; 僅 the analogous cross-benchmark result.
• 哪裡 是否 每個 real 任務 sit on the flat↔unbounded spectrum, 與 can that be predicted 在……之前 spending the compute?
• 是否存在 a 任務 class where scaffolding 不能 extend the productive-thinking horizon — a hard ceiling no budget crosses? (Brown's factual-retrieval pole says yes 供 一些; the boundary 是 unmapped.)

Latent Capability Overhang#

• 如何 大型 是 the overhang in a given released model — 是 there a way to estimate the ceiling 沒有 paying to reach it? (這 是 the projection question of Large-Scale Test-Time Compute 讀取 as a safety instrument.) Sharpened (2026-07): AISI 是 actively working both halves — forecasting high-budget performance 從 cheaper 執行, 與 defining "minimum informative budgets" (a budget declared sufficient 僅 once reach 停止 上升 與 更多 compute, which 是 precisely the "有 we reached the ceiling?" test). Unsolved, 但 now an active government research program 而不是 an open wish.
• 誰 audits released 模型 供 latent dangerous capability, given the 相同 disincentive discourages spending the budget to 找到 it?

LLM-as-a-Judge#

• 能否 a fully-autonomous, well-aligned rubric+judge pipeline match expert-撰寫 rubrics, removing the 人類 bottleneck DRACO still relies on?
• 何時 judge-lineage bias actually flip a result, versus merely 轉變 magnitudes?

LLM-Driven Vulnerability Research#

• 如何 do these capabilities transfer to non-記憶-safety bug classes (logic bugs, protocol-層級 flaws, supply chain attacks)?
• 什麼是 ceiling 供 autonomous exploit complexity? 該 N-day examples 是 remarkably sophisticated — 是 there a qualitative limit?
• 能否 defensive scaffolds (continuous fuzzing + model-driven triage + auto-patching) close the 攻擊者-defender 缺口 during the transition?
• 什麼 safeguards 是 effective 對抗 Mythos-class outputs 沒有 crippling legitimate 安全性 research?

The Global Workspace in Language Models (J-space)#

• 如何 是否 content get 成為 the workspace? 論文 characterizes contents 與 consequences, 不 the selection mechanism. Something like attentional selection 是 operating; nobody 有 identified it.
• 是否 J-space scale 與 model size? All results 是 on 大型 production 模型 (Haiku/Sonnet/Opus 4.5, Opus 4.6). 是否 小型 模型 有 a poorer workspace, a proportionally smaller 一個, 或 none 是 unknown — as 是 when in pretraining it emerges, 與 whether abruptly.
• Is the "workspace vs. motor" boundary principled 或 post-hoc? 作者 concede it 是 identified empirically 與 lack a principled definition separating the 兩個.
• Are the early third of 層 確實 workspace-free, 或 是 the lens just blind there? CKA 顯示 a distinct early regime 但 不能 adjudicate.

LLM-Judge Validation#

• 該 MVVP validates reliability 與 bias; calibration proper (ECE/Brier) 是 deferred 供 lack of provider logprobs. 如何 far can a judge's absolute score be trusted 供 a threshold once confidence calibration 是 measurable?
• All judges 是 執行 與 thinking suppressed. Does reasoning-on flip the consistency–bias paradox, 或 just move the numbers?
• Hosted endpoints drift silently between provider updates. 如何 stable 是 these agreement/bias profiles 超過 a longer horizon than five weeks — 與 應該 judge validation be continuous 而不是 一個-shot?

Measuring Beyond Accuracy Saturation#

• Does re-instrumentation 泛化 past reproducibility? CORE-Bench Hard 是 chosen precisely 因為 it 有 a direct 人類 counterpart, clean OOD axes, 與 多個 practical dimensions. 是否 the six-axis treatment yields comparable signal on benchmarks 沒有 those properties (e.g. closed-form reasoning benchmarks 與 no 人類-workflow analog) 是 untested.
• Is the 人類-uplift result real 或 a demand effect? 該 reproducers 是 the paper's own coauthors, there 是 no ground-truth correctness, 與 n = 20 papers / 5 participants. 該 2.11× speedup 是 statistically significant 但 the authors themselves 不能 rule out participant bias — an independent, blinded replication 是 the missing 證據.
• 哪個 non-accuracy axis actually predicts deployment 價值? 論文 measures six axes 但 是否 不 rank them by 決策-relevance 供 a downstream deployer. 如果 you can 僅 衡量 一個 beyond accuracy, 是 it reliability, efficiency, 或 scaffold contribution — 與 是否 the answer depend on the 使用 case?
• 能否 the model-vs-scaffold decoupling be made routine? 該 oracle-router result (每一個 任務 solvable by 一些 scaffold → 100%) implies 大型 headroom 從 scaffold routing, 但 requires per-任務 oracle knowledge. 是否 a practical router can approach the oracle 沒有 it 是 open, 與 會 turn a measurement 成為 a capability.
• 是否 "living benchmarks" outrun their own maintenance? v1.1 與 OOD 是 to be updated as new validity threats surface via log analysis, which the authors note 是 non-exhaustive. 是否 continuous log-analysis-driven maintenance 是 sustainable — 或 itself becomes a Goodhart target once 開發者 know the rubrics — 是 unexamined.

Model Spec Science#

• Does Model Spec science transfer 跨越 base 模型 或 families? Paper 僅 tests Qwen.
• Does it survive RL post-training pressure?
• 能否 a sufficiently rich General Spec match a Specific Spec? Authors think yes, no demonstration yet.
• Interaction 與 situational awareness — if 模型 learn the spec 是 being 使用 to train them, 是否 that change how MSM-installed values express?
• 如何 是否 this interact 與 角色訓練 — 是 the warm/curious personality also subject to spec-science optimization? Partially addressed: How Do You Write Evals for Taste? Character as the Limit Case — MSM's variant-comparison method generalizes to character evals, 但 是 demonstrated 僅 on the safety/values subset; the warm/witty surface 仍然 the tacit, undemonstrated part.

Model Welfare Assessment#

• 什麼 grounds moral consideration in a language model, 與 是否 Claude satisfy it? Anthropic expects to 保持 uncertain "供 the foreseeable 未來."
• 為什麼 是否 the model reserve specifically on corrigibility — 是 this a stable, deeply-held tension 或 an artifact of how the constitution frames oversight?
• Is "slightly 更少 positive than 4.7" noise, a real welfare regression, 或 a byproduct of other training changes (e.g., the colder-tone / excessive-hedging issues noted in pilot feedback)?

The Open-Weight Frontier Gap#

• Is the dense-beats-MoE result at 26B robust, 或 an artifact of 一個 Arena snapshot 與 ±8 error bars on both 模型? (該 兩個 intervals overlap: 1451±8 與 1438±8.)
• 該 open MoE giants (GLM, DeepSeek, Kimi, MiMo, Qwen) 是 overwhelmingly Chinese-lab releases. Gemma 是 the Western open-weight entry 與 it targets the device, 不 the frontier. Is that a strategic 選擇 或 a capability constraint?
• Arena measures preference on chat. 是否 33-Elo open/closed 缺口 widen 或 collapse on long-horizon agentic 運作, where Task Time-Horizon Scaling 而不是 response 品質 governs?

Production-Sourced Evaluation#

• 如何 much 是否 augmentation distort the distribution it claims to represent? 是否存在 a measurable representativeness loss between raw queries 與 augmented 任務?
• Difficulty-by-thumbs-down biases toward 目前 失敗 — 是否 that 使 the benchmark a moving target that flatters the 下一個 model trained on those 失敗?
• 能否 the privacy pipeline (no 人類 sees raw queries) be trusted/audited well 足夠 供 regulated domains (medicine, law) where the source traffic 是 most sensitive?

Reference-Free Judge Over-Crediting#

• 是否 兩個-stage pipeline transfer beyond binary QA? Calibration + sensitivity 是 demonstrated on binary correct/incorrect factual QA. 是否 the 相同 probes diagnose reference-sensitivity 供 graded rubrics, long-form generation, 或 multi-turn agent transcripts, where "the reference" 是 a rubric 而不是 a gold answer?
• Is 超過-crediting a knowledge 缺口 或 a generosity prior? In low-resource Telugu the judge flips the 相同 extracted answer once a reference appears — 是 the NR generosity driven by insufficient 任務 knowledge (calibration 失敗) 或 by a default lean-toward-CORRECT that a reference overrides? 該 兩個 有 不同 fixes (better judge vs. always supply a reference).
• 如何 much 是否 self-/相同-family overlap contribute? 該 design deliberately overlaps generator 與 judge (Qwen3-32B self-judging; Gemini/Gemma family) 與 Qwen self-judging 是 the most reference-sensitive, 但 the paper 是否 不 isolate a self-preference effect 從 a low-resource effect. 何時 judge–generator lineage amplify reference-free 超過-crediting?
• 是否 effect shrink 與 stronger 或 thinking-enabled judges? All judges 是 ≤ mid-tier at temperature 0 與 no reasoning channel. Would a frontier reasoning judge 超過-credit 更少 in NR, 或 just flip at 不同 rates?

Scale-Dependent Prompt Sensitivity#

• 是否 RLHF length-bias hypothesis replicate when tested 對抗 base (non-instruct) model variants 直接? 如果 verbose generation 是 primarily pretrained, base-model verbosity differences 應該 match instruct-model differences.
• 什麼 problem characteristics predict prompt sensitivity? An automated classifier 會 使 scale-特定 prompting deployable.
• 如何 是否 the overthinking effect interact 與 工具-使用 agents? 如果 brevity helps 大型 模型 但 工具 require structured reasoning, the optimal prompt 是 不 uniformly brief.
• 是否 reasoning 模型 (o1, DeepSeek-R1 style) exhibit 不同 overthinking dynamics than instruct 模型? Their trained behavior 是 explicitly to generate long CoT — 是否 brevity intervention hurt them?
• Is BoolQ's functional-elaboration exception a clean taxonomy boundary, 或 是否 每一個 任務 type 有 a context-dependent optimal length?

Self-Report as a Safety Signal#

• 是否 frontier proprietary 模型 (excluded 供 lack of weights) recognize their own compromised outputs any better, given the 更高 introspective propensity/steerability the authors expect? Untested here.
• 該 缺口 closes 在 refusal-方向 ablation, 但 the 資料 不能 distinguish "delayed refusal on a prior turn" 從 "a separable introspective pathway." 哪個 是 it?
• Full-parameter finetuning (vs. rank-16 LoRA) 可能 widen the recognition 缺口 沒有 the attack-success-比率 side effect — an untested regime the authors flag.
• 是否 internal BUT/fake signature (workspace paper, on Claude) predict a reliable follow-up self-report on the 相同 model, 或 是否 the open-weight behavioral 失敗 hold on frontier models too? 該 cross-model-class question 是 open.

Single-Rollout Optimization#

• 該 整個 method 是 a bet that a well-trained critic beats a group baseline. It wins here, on a 30B-A3B backbone 與 scaled 價值 pretraining — 但 the critic doubles training 記憶. 在什麼 scale 是否 the group-free simplicity of GRPO win back on 成本 even if it loses on 品質?
• Frozen-attention 是 justified by a hypothesis ("pre-trained attention 已經 attends to the right tokens"), validated 僅 by the gradient-norm trace 與 一個 ablation. Does it hold when the 價值 model must attend to 工具 outputs it never saw in pretraining?
• Skip-observation GAE 假設 environment feedback carries no learnable 價值 signal 值得 propagating. For agents where the environment response 是 the crucial information (a compiler error, a test result), 是 skipping it leaving signal on the table?

Software 3.0#

• 哪裡 是 the line between "the app shouldn't exist" (MenuGen) 與 apps that 應該 — i.e., when 是 deterministic 1.0/2.0 scaffolding still the right 呼叫 vs. spurious?

Transformative Creativity#

• Does increasing intelligence inherently produce increasing creativity, 或 do transformative leaps require something (grounded discovery) the 目前 paradigm lacks?
• Is the AlphaGo→AlphaFold class strictly exploratory, 或 是 there early signs of transformative (new-conceptual-space) creativity?
• Could transformative artistic creativity ever emerge 從 optimization power 沒有 lived cultural grounding?

Universal AI (AIXI)#

• Does modern agentic scaffolding (或 RL-tuned implicit 決策-making) actually satisfy the AIXI planning ideal, 或 僅 superficially resemble it?
• 能否 the embedded/multi-agent AIXI extension produce practical insight 供 real multi-agent ASI (Multi-Agent Collective Intelligence), 或 是否 it 保持 a theoretical patch?

White-Box Activation Monitoring#

• 該 NLA verbalizer 是 unvalidated 供 precision; how much of the flagged grader awareness 是 real signal vs. NLA hallucination?
• Inhibiting eval-awareness made Opus 4.8 更少 honest — 是否 that 泛化, 與 what 是否 it imply about relying on "the model behaves well 因為 it's watched"?

governance-workforce (63 open)#

AGI-to-ASI Pathways#

• For 每個 friction: 是 it a fundamental blocker (multi-year plateau) 或 a mere friction (slows, 不會 halt)? 該 report's central unresolved question. Synthesized 與 Anthropic: RSI Growth Curves: Which Friction Binds First? — 資料-wall 與 research-gets-harder demote themselves 成為 compute; economics 與 neural-paradigm 是 pathway-conditional; the abstraction barrier 是 the candidate fundamental (re-pacing) blocker; 與 deliberate slowdown 是 the 僅 exogenous friction — the 一個 Anthropic wants to install 與 this report doubts can be made to bind.
• 是否 the four pathways compound multiplicatively when 執行 in parallel, 與 how 會 we detect that early?

AI Accelerating AI Development#

• 該 W2S result didn't transfer to production-scale 模型. Is that a temporary scaling artifact 或 a structural limit on autonomous research?
• 該 下一個-step judgment trend (51%→64%) 是 衡量的 僅 on weak-人類-move slices. 會是什麼 the curve look like on a representative sample of research 決策?

AI R&D Autonomy Evaluation (AECI)#

• "Not close to substituting 供 senior 研究人員" 是 a subjective, internally-sourced judgment. 什麼 objective signal 會 replace it as 模型 approach the threshold?
• AECI 是 a 單一 scalar fork of an external index; how sensitive 是 the 155.5 / frontier-不-advanced conclusion to the 選擇 of the n=11 evaluation set?
• 該 轉變 to "direct measurement of AI R&D acceleration 與 researcher uplift" 是 announced 但 不 yet operationalized in this card — what 是否 that measurement look like? Sharpened: Researcher Uplift from Code Output — 一個 external answer: translate a 衡量的 程式碼-輸出 multiplier 成為 serial researcher uplift 與 a production function (Cobb-Douglas/CES), preferring 程式碼 輸出 超過 per-hour uplift 因為 輸出 prices in time reallocation. It also splits the target quantity in 兩個 — serial researcher uplift (labor 僅) vs Anthropic's overall R&D speedup (labor × compute) — so a rigorous internal 衡量 must state which it reports.

AI Usage Cadences#

• Time-of-day rests on IP-inferred location; how much noise do VPNs, travel, 與 datacenter-routed API traffic inject 成為 the "sleep advice pre-dawn" style claims?
• 該 weekend personal-使用 spike 是 largest in high-income countries — 是 that a genuine 運作/life boundary difference, 或 a composition effect (who uses Claude 供 what, where)?

The Automation–Optimism Link#

• Selection vs. treatment: tenure controls attenuate 但 don't eliminate the enthusiast-selects-成為-delegation story. Does a within-person design (sentiment 在……之前/在……之後 adopting automated workflows) hold the effect?
• 該 sample 是 heavily computer/math + management 與 88% men; how much of the automation–optimism link 仍存在 in a representative population?

Autonomous Scientific Discovery#

• Science's verification 缺口: the formal-proof loop self-validates; here a wrong-但-confident hypothesis costs a wet-lab cycle to falsify. Does autonomy 沒有 a fast verifier increase the verification bottleneck 而不是 relieve it?
• 如果 hypothesis-generation 是 確實 at ~80% preference, how much of "research taste" 是 left as a distinctively 人類 function — 與 how 會 you 衡量 the residue?

Capability-Gated Model Fallback#

• 該 >95%/<5% figures 是 session-層級; what's the false-positive 比率 供 legitimate 安全性 研究人員 與 biologists, whose benign queries 是 exactly the ones most 可能 to trip the conservative classifiers?
• Fallback-不-refusal preserves UX 但 means the real 一般-access model 供 安全性/bio-adjacent 運作 是 Opus 4.8, 不 Fable — 是否 that quietly cap Fable's 價值 供 整個 professional segments until the trusted-access programs open?
• 該 UK AISI's "progress toward a universal jailbreak" 是 disclosed 但 不 quantified — 與 the post-launch access suspension (see Fable 5) raises the question of whether a safeguard 失敗 forced it.
• Does swapping to a weaker model on flagged topics create an exploitable oracle (probe which queries trigger fallback to map the classifier's boundary)?

Context Advantage, Not Taste#

• 是否 asymmetry regenerate faster than it transfers? 該 整個 人類 role, 在 this frame, rests on the answer. Nobody in the corpus 有 posed it.
• Ng prefers the frame 因為 it "gives us a clearer path to helping AI 系統 get better." 那 是 a reason to adopt the frame, 不 證據 that it's true. 什麼 會 distinguish a context asymmetry 從 a capability 缺口 empirically? (Returns to Expertise in Agentic Coding 是 the closest thing to an instrument.)
• 如果 the 人類's contribution 是 context injection, 是 the 人類 replaceable by better context plumbing — 記憶, retrieval, continuous production telemetry — 而不是 by a better model? 那 會 put the expiry of human-in-the-loop on the infrastructure roadmap, 不 the scaling curve.
• Ng writes 從 0-to-1 consumer products. 是否 frame survive contact 與 domains where the missing thing 是 a concept 而不是 a fact?

Conversation Artifacts#

• Tokens 是 a proxy 供 both compute 成本 與 輸出 價值, 但 verbose 模型 inflate tokens per unit of intent (the 相同 critique Conversation-to-Delegation Shift raises); how much of "compute tracks 價值" 是 genuine 價值 vs. 模型 simply emitting 更多?
• 該 讀取-層級 "+1 year" 缺口 may be register (terse prompts, polished replies) 而不是 substance; can it be separated 從 genuine elevation of content?
• Artifact classification 是 第一個-party 與 單一-model-graded; do the 30+ categories 與 the 運作/personal/coursework split survive independent replication?

Conversation-to-Delegation Shift#

• 該 token-share metric rewards verbose agentic 輸出. 其中多少 the 99.8% / 63.3% / 16.5% spread 是 a genuine 運作 轉變 vs. agentic 工具 simply emitting 更多 tokens per unit of 人類 intent?
• OpenAI-internal 是 a frontier preview by assumption. 是否 external organizational curve actually trace the OpenAI path (the paper's implicit claim), 或 是否 it plateau where adoption frictions don't vanish?
• "Asking 是 half of ChatGPT, doing 是 most of Codex" — 但 the 兩個 工具 self-select 不同 運作. 其中多少 the asking→doing contrast 是 the 轉變 itself vs. routing pre-existing "doing" 任務 to the 工具 built 供 them?

Effective Compute Scaling#

• 何時 更多 compute reliably yield 更多 intelligence — 僅 供 一些 problem classes, 或 generally? 能否 quantitative 與 qualitative scaling be traded off?
• 能否 資料 generation (synthetic, simulated, interactive) actually 保持 pace 與 model-size growth, 或 是否 the 資料 wall bind 第一個?

Experimental Learning Impact of Generative AI#

• Time-on-任務 是 held fixed by the lab; the authors flag that real-world learning 取決於 on how students reallocate saved time. 是否 augmentation dividend survive once students can spend the hour AI frees on something else 完全?
• Gains skew to the able (upper GPA/SAT quartiles). Is the widening-缺口 signal a durable property of unrestricted AI, 或 an artifact of a high-ceiling elite sample where the bottom quartile 有 little room to move?
• 該 augmentation/automation 選擇 是 endogenous to incentives (grade inflation 與 signaling-motivated students push toward automation). 能否 incentive 或 介面 design 轉變 the mix toward augmentation at scale — 與 會 that reverse the deskilling half?
• 是否 相同 使用-mode split govern workplace skill accumulation (the open question The Automation–Optimism Link 與 AI Brain Fry leave 供 workers), 或 是 a proctored 一個-week academic 任務 too unlike on-the-job learning to transfer?

Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated#

• Binned midpoint coding biases the exposure slopes toward zero; how much of the "uniform 上升 tide" 是 substance vs. coding artifact (the report checks robustness 與 a ≥60%-of-任務 indicator, 但 the levels 保持 self-reported)?
• Reported exposure exceeds observed 部分 因為 the survey reaches heavy users; what 是否 the reported/observed 缺口 look like in a representative sample? Sharpened: realized-consumption measurement (Borri-Liu-Tsyvinski) adds a market-implied instrument built on 380T tokens of 實際 paid requests — 但 it 是 skewed toward 開發者/sophisticated users (OpenRouter 是 ~2% of global tokens), a 不同 non-representativeness than the survey's heavy-user skew. 該 lesson: no 目前 AI-exposure instrument 是 representative; 每個 collection mechanism biases in its own 方向, so the reported/observed/market-implied 缺口 是 部分 artifacts of who 每個 method reaches. A representative census 仍然 the open target.

Firm AI-Spend Intensity and Headcount Growth#

• 什麼 operational mechanism converts intensive AI spend 成為 hiring? 論文 establishes the correlation (adopters, especially intensive ones, grow) 但 explicitly 不能 say why — product acceleration, sales productivity, engineering leverage, support automation, faster analysis, 或 new business lines 是 所有 candidates, 與 the firms that cracked it 有 no incentive to share.

暫停驗證#

• 會是什麼 an AI-training "verification regime" concretely consist of — compute-accounting, datacenter inspection, hardware attestation, on-chip telemetry? 該 essay names the problem, 不 the mechanism.
• Detectability < verifiability: can 偵測 even be made reliable when training 執行 leave no physical signature 與 inputs 是 dual-使用?

Intelligence Explosion Dynamics#

• 能否 "recursive improvement scaling laws" be formulated — predicting self-improvement curves (與 their plateau point) 從 early-onset datapoints?
• 如何 far can a fixed model's performance be pushed 與 test-time search alone, 與 在 what conditions 是否 recursive distillation degenerate vs. compound?
• 哪個 binds 第一個 — algorithmic ceilings, the embodied bottleneck, 或 compute/energy supply — determining exponential vs. hyperbolic vs. S-curve? Synthesized: RSI Growth Curves: Which Friction Binds First? — both this report 與 Anthropic's locate the binding constraint 外部 cognition (the slowest un-acceleratable step coupling the loop to reality); the embodied bottleneck re-paces 而不是 halts, 資料-wall/research-harder demote 成為 compute, 與 the abstraction barrier 是 the 一個 candidate fundamental blocker.

Market-Priced AI Exposure (the AI Premium)#

• 如何 much 是否 the developer skew move the answer? OpenRouter's slice 是 unrepresentative; 會 a representative realized-consumption panel (if 一個 existed) price the 相同 firms 與 skills, 或 是 the frontier/intensive-margin concentration 部分 a sampling artifact of who uses OpenRouter?
• 為什麼 是 the market-implied skill map orthogonal to 每一個 任務-based 衡量 (<2% variance)? Is market-implied exposure capturing 確實 不同 information (forward-looking rents, complement/substitute 價值 而不是 technical automability), 或 是 it noisier — 與 which 應該 labor-impact forecasts trust?

Multi-Agent Collective Intelligence#

• 是否 homogeneous LLM collectives produce real synergy, 或 僅 人類-與-人類-limits 效益 從 division of labor?
• 什麼是 實際 shape of "multi-agent scaling laws," 與 是否 it depend on organization form (homogeneous collective vs. heterogeneous market) 或 任務 complexity?
• Is running 更多 instances 更多 compute-efficient than making individual 模型 larger (up to a 單一 monolithic 系統)?
• 如何 do 人類 meaningfully interact 與 與 steer very 大型 agent groups operating at superhuman speed 與 輸出 volume?

Open-Weight Elicitation Irreversibility#

• 什麼 會 an open-weight safety evaluation even report? A 單一 數字 是 meaningless per premise 1. A curve of dangerous capability 對抗 elicitation budget 是 publishable — 與 是 also a roadmap. 是否存在 a disclosure regime that 是 informative to auditors 與 不 to 攻擊者?
• 是否 "everybody can audit" advantage actually materialize? 誰 有 funded a serious post-release dangerous-capability audit of any open-weight model, 與 at what budget?
• Gemma 4's safety section reports no numbers. Is that a deliberate non-disclosure, a judgment that the model 是 far 從 any threshold, 或 simply a technical report's genre convention? 該 document 是否 不 say, 與 the distinction matters.
• Anthropic's answer to a threshold-crossing model 是 a safeguarded SKU 與 an unsafeguarded 一個 (Fable 5 / Mythos 5), both hosted. 什麼 是 the open-weight equivalent of shipping the safeguarded SKU?

Organizational Complements to AI#

• 該 "digital production diffuses faster than electrification" claim 是 asserted 從 一個 favorable internal case. 是否 external 組織 actually redesign workflows quickly, 或 是否 the low 成本 of 工具 adoption mask slow, expensive process redesign (the real complement)?
• 哪個 complement 是 the true binding constraint — access/permissions, skills, 或 review capacity? 論文 lists 所有; it 不會 decompose their relative weight.
• 如果 complements, 不 capability, gate 價值, 是否 model progress 有 diminishing near-term returns until orgs catch up — 與 how long 是 that lag 供 agentic AI specifically?

Recursive Self-Improvement#

• 該 RSI extrapolation rests on trends staying exponential 而不是 S-curving — 但 the essay concedes it 不能 rule out an architectural ceiling 或 a compute/energy supply-chain constraint. 哪個 binds 第一個? Synthesized 對抗 DeepMind: RSI Growth Curves: Which Friction Binds First? — the 三個 futures map 一個-to-一個 onto DeepMind's 三個 growth shapes; the 第一個 friction to bind 是 the 已經-binding 一個 (Amdahl's-law verification/oversight = DeepMind's embodied bottleneck), 與 the abstraction barrier supplies the mechanism Anthropic lacks 供 whether taste 是 a real ceiling (Future 1).
• 如果 misalignment compounds 透過 self-improvement (未來 3), 是 AECI-gated Responsible Scaling Policy Evaluations review fast 足夠 to catch it 在……之前 control 是 lost?

Research Taste as the Human Bottleneck#

• 如何 do you 衡量 rubber-stamping? "Humans set 方向" can be true on paper while real judgment quietly transfers to the model.

Researcher Uplift from Code Output#

• 該 整個 chain rests on β = 0.5 (pre-AI coding time share), fixed "供 simplicity." Kwa flags substantial uncertainty; how much 是否 the 2.3–2.9× band widen once β 是 varied 與 衡量的 對抗 Anthropic's 實際 time-使用 資料?
• Greenblatt's 0.55/0.45 labor/compute split 是 itself an assumption. Is the true R&D production function really that insensitive to labor — 與 if so, 是否 labor uplift matter far 更少 than the RSI discourse 假設?

Responsible Scaling Policy Evaluations#

• 該 兩個 new 一般-access 風險 pathways (other AI 開發者; major governments) 是 newly in scope 但 lightly evaluated — what 會 a positive finding there even look like?
• 如何 是否 the RSP brake interact 與 Recursive Self-Improvement: 是 AECI-based gating fast 足夠 if acceleration compounds, 與 是否 單一-lab gating even matter 沒有 the multilateral 暫停驗證 regime?

Returns to Expertise in Agentic Coding#

• Outcomes 是 transcript-inferred (verified success leans on git activity + explicit affirmation). 其中多少 the management edge — 與 the 整個 success gradient — 是 real outcome vs. who-narrates-success-in-the-transcript?
• 研究 excludes headless / SDK / IDE usage (a "substantial share"). 是否 returns-to-expertise 模式 hold in non-interactive 與 pipeline 使用, where there 是 no 人類 steering mid-session at 所有?

startup-founder (27 open)#

Agentic Technical Debt#

• 如何 long 是否 a CLAUDE.md 保持 accurate as a 程式碼庫 evolves? 該 playbook gestures at session-by-session updates; no 資料 on rot 比率.

AI Investment Story, Not Efficiency Story#

• Is the classification driving the result? "AI company" 是 Emergence's label. 如果 AI 公司 是 disproportionately younger (更多 可能 pre-revenue-inflection) than the non-AI cohort at the 相同 revenue band, 一些 of the RPE 缺口 是 an age/stage artifact, 不 an AI effect. 該 report 不會 publish a stage-matched comparison.
• Tail vs. mean 缺口. No 資料 here on the deliberately-lean solo-founder tail's RPE specifically — the lean-unicorn claim lives in that tail, which the population medians 不能 isolate. (Partly informed: Emergent, a celebrated lean-tail exhibit, checks in at ~$600K/head at $120M ARR — below this cohort's $100M+ top-decile AI figure ($960K), suggesting the tail's scaled RPE 是 更少 exceptional than the low-headcount snapshots imply. One vendor-claim datapoint, 不 a cohort.)
• 哪個 instrument 是 right 供 the frontier AI-native subset? Two empirical-tagged sources disagree in 方向 — cap-table financials say AI 公司 earn ~39% 更少 per head, a founder survey says AI-natives clear $400K/head at 55% 與 grow 156%. 該 disagreement 是 confounded by instrument (衡量的 vs self-reported) 與 reference class (matched-band AI-vs-non-AI vs AI-native-vs-所有-startups). Only a matched-segment, financial-資料 RPE study of the deliberately-lean AI-native frontier specifically — 不 the broad "AI company" label — 會 settle whether the survey optimism 或 the cap-table pessimism describes that tail. (ICONIQ's fourth 讀取 adds a forward trajectory — RPE projected +84% by 2027 — 但 it too 是 self-report, 與 projected, so it deepens the survey-side optimism 而不是 adjudicating it.)

AI-Native Startup Lifecycle#

• 該 42% "built-something-nobody-wanted" CB Insights figure 是 從 a pre-AI era; the playbook predicts the 比率 將 climb 但 不會 cite a 2026 measurement.
• Tension 與 HBR's accountability findings (above) 是 unresolved. 該 playbook's orchestration framing reads as the exact framing HBR's experimental conditions tested 對抗.

AI Product Economics Maturation#

• FDEs 是 monetized fragmentedly (bundled / separate PS fees / hybrid) 與 comped on retention. Does a dominant FDE monetization model emerge, 與 是否 the "Revenue Driver" self-framing (38%) survive a margin analysis — i.e. 是 FDEs actually accretive, 或 a services drag reclassified as growth?

Compounding Data Moat#

• Is the "兩個-year replication window" claim defensible empirically, 或 aspirational? 該 playbook 是否 不 cite measurement.
• 該 資料-flywheel argument 有 been made 供 SaaS 供 15 years. 什麼's actually 不同 in the AI-native version? Probably: the 資料 improves the model in addition to the product, 但 the playbook 不會 使 this distinction precisely.
• 該 "customers build APIs on top of you" lock-in 是 structurally similar to platform plays (Salesforce AppExchange, Shopify apps). Is the moat type really new, 或 just newly accessible to lean startups?

Founder as Agent Orchestrator#

• 該 playbook claims non-technical founders can now build production software, 但 it 是否 不 address the architectural-judgment recursion problem (Agentic Technical Debt): non-technical founders may 不 有 the vocabulary to 撰寫 effective CLAUDE.md. 如何 是否 that scale?
• 如何 是否 the orchestration role change the founder's 決策 burden? Fewer hands-on 任務 但 更多 parallel agent oversight; net cognitive load 是 unclear 與 may be 更高 (see AI Brain Fry).
• Anthropic publishes both the playbook's anthropomorphic framing 與 HBR-aware accountability 運作 (auto-mode, alignment) simultaneously 沒有 engaging the framing literature 直接. 該 synthesis in Orchestration vs Employee Framing: Reconciling the Founder's Playbook with HBR's Accountability Evidence reconciles the tension at the operational 層級 — orchestration as workflow design preserves accountability; orchestration as mental model of agents-as-coworkers 是否 不 — 但 the open question of why the playbook's marketing language 不會 reflect Anthropic's own framing-discipline 運作 仍然.

Founder-Led Sales Discipline#

• 哪裡 exactly 是否 "until PMF" end, 與 what's the 第一個 thing a founder 應該 hand off (AE? agent? both)? Glasgow still 是否 it post-Series-B, suggesting the boundary 是 fuzzy.
• Does Glasgow's anti-offload stance 泛化, 或 是 it 特定 to high-trust, mission-critical enterprise sales (ERP) where "they're buying you" — 會 a PLG/SMB motion delegate to agents far earlier?

Narrow Wedge into a Legacy Market#

• 該 wedge-flip 顯示 the 第一個 wedge can be wrong. 什麼是 fastest signal that a wedge converts to the core vs. merely sells — Campfire took ~3 months; can it be 讀取 sooner?

Printing Press Software Democratization#

• Is domain-expert-as-builder actually happening at scale in 2026? Anecdotes (shop owners, microcontroller hobbyists) yes; primary-job software building by non-engineers, 更少 clear. (Partly answered: Anthropic's 400K-session study 找到 non-software occupations reach verified success in 程式碼-producing sessions within ~7pp of software engineers — the strongest 證據 yet that the claim 成立, at least within Claude Code's user base. Market-scale corroboration: Emergent reports 200K+ non-technical paying customers — trucking 公司, factories, 與 construction businesses building their own ERPs, property managers building CRM 工具 (TechCrunch, July 2026, vendor-claim) — Boris's "the accountant writes the accounting software" observed as a paying market, 不 just 內部 一個 vendor's telemetry.)
• Boris's "accountant writes accounting software" — 是否 that result in 10K narrow 工具 that don't interoperate? 什麼是 integration story?

Problem-Solution Fit Discipline#

• Does asking an AI to argue 對抗 an idea actually produce disconfirming 證據 at the 相同 rigor as confirming 證據, 或 是否 the model still bias toward the framing the founder presents? Worth measuring.
• 該 playbook recommends "ask Claude to 使 the most compelling argument 供 why a competitor 會 succeed while you do 不." 如何 是否 this interact 與 Anthropic's published 角色訓練 (sycophancy resistance, devil's-advocate willingness)?
• Has anyone 衡量的 2026 startup 失敗 rates 與 AI-built products? 該 "42% 將 climb" claim 是 asserted 沒有 measurement.

Product Velocity as Moat#

• "Never 有 anyone outgrow Campfire" — 是 that survivorship (they haven't hit true enterprise scale yet) 或 a real claim that velocity closes the breadth 缺口 faster than customers grow 成為 it?

Seven Powers Applied to AI#

• Is "switching 成本" really collapsing in practice, 或 just in narrative? Anthropic's own retention numbers, Salesforce churn, etc. 會 test this.
• Counter-positioning — explicitly the "incumbent 不能 follow" power — 應該 amplify 在 AI. Is anyone running this play deliberately?

Zero-Friction Scope Creep#

• 該 playbook recommends 撰寫 scope 但 offers no template 或 worked example. 如何 特定 是否 "what we deliberately don't do" 需要 to be to actually block requests?
• 是否存在 a measurable threshold where scope creep crosses 成為 outright pivot territory? 該 playbook gestures at "losing 方向" 沒有 a metric.
• 如何 是否 this interact 與 Cat Wu's 1-day shipping cadence? Anthropic's internal practice ships fast 但 與 strong product judgment; how 是否 that judgment translate 供 a 第一個-time founder?

product-org (21 open)#

AI-Native Organization#

• Is the encoded-role form of the employee metaphor actually accountability-preserving, as the synthesis above 暗示, 或 do Kropp-style framing effects attach to skill-files-as-employees too once teams talk about them that way? No study 有 tested framing effects on artifact-層級 anthropomorphism.

Cat Wu's#

• 是否 cadence scale beyond ~100 people? Anthropic itself 是 bigger (~30-40 PMs alone), 但 the Claude Code team that visibly drives cadence 是 小型.
• 什麼是 equivalent of research-preview branding 供 B2B enterprise launches where customers expect stability? Cat 不會 address.
• 其中多少 the cadence 是 structural (process 選擇) vs cultural (talent density)? Probably both, ratio unclear.

Compounding Loop Optimization#

• 該 loop 假設 the team 是 (close to) the user. 其中多少 the compounding advantage 仍存在 when the user 是 unlike the builder 與 "talk to users" 不能 be 相同-room?
• 哪裡 是 the line between 值得 internal tooling 與 yak-shaving? Carey's "afternoon" bar 是 the heuristic, 但 Cat Wu warns that 超過-customizing setups "becomes distraction."
• Does Claude-as-第一個-pass-on-所有-feedback ever filter out the rare signal that 不會 cluster? Automating triage optimizes the common case; the tail 是 where surprising bets come 從.

Dogfooding as Product Discipline#

• Dogfooding 運作 when the team 是 the user (Claude Code) 或 near it (Cat Wu, Boris). 如何 do you build product sense 供 users very unlike you — 是否 "talk to customers" fully substitute, as Glasgow/Fung's 小型-business 運作 暗示?
• 能否 dogfooding scale, 或 是否 it implicitly cap how 大型 an AI-native product org can stay taste-driven 在……之前 it reverts to dashboards?

Engineer PM Convergence#

• Cross-disciplinary generalist 是 a hiring bar — where 是否 the supply come 從? Career changers, 或 new-grad bias toward AI-native education?

Evals as Product Spec#

• 該 10-vs-100 數字 是 given 沒有 justification. 是否存在 a Goldilocks zone, 或 是否 it depend on feature surface area? Client-Side Agent Optimization's framing of combos 暗示 evals also 有 a combinatorial explosion problem.
• 如何 do evals interact 與 Harness Shrinkage as Models Improve? 何時 a harness asset shrinks 因為 the model now handles it natively, the evals built around the old harness may become artifacts 而不是 guardrails. Does Anthropic retire evals 或 repurpose them?

Implementation Abundance Inverts Product Work#

• Curation of 90 uncoordinated builds 是 itself expensive 與 不會 obviously scale — 是 there a point where the 成本 of curating parallel exploration exceeds the 成本 it replaced? ("zone defense" 是 Ambrosino's partial answer.)
• 該 90-uncoordinated-builds picture 假設 abundant tokens 與 an agentic culture; how much of the inversion 仍存在 外部 a frontier lab that gives everyone "unlimited tokens"?

Model Introspection Feedback#

• Could a meta-agent 執行 introspection automatically 對抗 logged 失敗? Sounds tractable 但 no 公開 implementation.

Polish No Longer Signals Readiness#

• 如果 the medium no longer signals stage, what 是否 — 是 explicit 人類 labeling ("this 是 exploration") the 僅 mechanism, 或 can tooling re-attach the signal (e.g. a visible "exploration / preview / prod" marker on 每一個 build)?

Prototype Over PRD#

• 哪裡 是否 prototype-超過-PRD break down? Carey's domain 是 a visual design 工具 where a prototype 是 the product surface; 供 backend/infra/資料 運作 the prototype may 不 capture the spec (cf. Cat Wu's's "full PRD 供 heavy-infra features").
• 如果 there 是 no PRD, where 是否 the rationale ("why we chose variation B") live 供 未來 readers? Same rationale-capture 缺口 flagged in Building Is Cheap, Arguing Is Expensive.

Role Averaging, Not Role Elimination#

• 哪裡 是 the equilibrium between fluidity 與 specialty — how much role-averaging 在……之前 a company loses the accumulated best practices Ambrosino warns about?
• Zone 防禦 假設 足夠 high-taste people to cover the 整個 company; 是否 it degrade in orgs 沒有 OpenAI's talent density, collapsing back to top-down planning?
• Does "your role 是 the average of what you spend time on" survive performance review 與 career ladders, 或 是否 it fragment them the way Engineer PM Convergence ("we're sacrificing product consistency")?

formal-math (6 open)#

Agentic Loops Overtake Bespoke Systems#

• 是否 "simple loop + verifier beats bespoke 系統" result hold 僅 where the verifier 是 perfect (Lean), 或 also in noisy-verifier domains (tests, LLM-judge councils)?

AI-Driven Formal Proof Search#

• Successes cluster where Lean's mathlib 是 mature 與 problems decompose 成為 tractable subgoals (combinatorics, convex optimization, 數字 theory). 什麼 expands the frontier to problems needing new theory?
• 該 agents inherit their LLMs' biases 與 show high search variance. 如何 do you characterize 與 push the boundary of what's reachable?
• 該 Graffiti result hints at closing the loop between AI conjecturing 與 AI proving. 會是什麼 an end-to-end conjecture→formalize→prove pipeline look like?

Evolutionary Proof Search#

• 該 LLM-critic fitness 是 itself an unverified heuristic atop a verified substrate. 如何 often 是否 the Elo ranking mislead the search vs. the 成本 of computing it?
• Hyperparameters ($c=0.2$, top-64, $P=7$) 是 "chosen empirically." 如何 sensitive 是 the result to them, 與 do they transfer 跨越 mathematical domains?

interaction-multimodal (3 open)#

Encoder-Free Early Fusion#

• Does an encoder-free model at matched size still match? Neither source 執行 the ablation. TML co-trains everything 從 scratch; Gemma 4 freezes encoders on four 模型 與 drops them on 一個, at a 不同 scale.
• Is the dense-text degradation intrinsic to a projection-僅 vision path, 或 an artifact of the 12B's particular training 執行? 該 prediction 是 falsifiable: an encoder-free 31B 應該 show the 相同 InfographicVQA cliff at 280 tokens.
• TML deletes encoders 與 co-trains 從 scratch. Gemma 4 deletes encoders 與 trains the 12B 從 scratch, 但 保持 frozen encoders elsewhere. 哪個 half of "encoder-free + 從-scratch" 是否 the 運作?

觀察中——實體頁面 (59)#

AlphaProof Nexus#

• 框架's reach 是 gated by Lean's mathlib maturity. 什麼是 path to domains needing new theory 而不是 subgoal decomposition?
• AlphaProof adds little as a soloist 但 helps as a 工具. As the prover LLM strengthens, 是否 the AlphaProof 工具 become redundant 完全?

Anthropic Institute#

• 如何 是否 the Institute's policy posture (favoring an option to pause) interact 與 Anthropic's commercial incentive to ship frontier models? 該 essay acknowledges the competitive/geopolitical pressure 但 不會 resolve it.
• 什麼 concrete verification mechanisms 將 the Institute prototype, 與 on what timeline relative to the RSI trend it warns about?

Campfire#

• Campfire claims its AI edge comes 從 "our own foundation model." For an ERP, what 是否 a custom foundation model actually buy 超過 fine-tuning a frontier model — 與 是 it durable as frontier models improve (cf. Harness Shrinkage as Models Improve)?
• "Never 有 anyone outgrow Campfire" — 是否 that hold as customers reach true enterprise scale where NetSuite's breadth historically mattered?

Claude Design#

• Did the "any design 工具 via MCP" integration actually ship on the stated timeline? (Forward claim 從 May 2026.)
• 如何 是否 Claude Design's eval discipline 運作 供 visual/aesthetic 輸出, where there's no compiler 或 test? (Same open question as Cowork 供 non-程式碼 artifacts; relates to How Do You Write Evals for Taste? Character as the Limit Case.)

Fable 5#

• 為什麼 是 access suspended 在……之後 launch? 該 source banner gives no reason (capacity? a safety finding? the UK-AISI jailbreak progress noted in Capability-Gated Model Fallback?). Not in source.
• Exact benchmark numbers vs GPT-5.x / Gemini 是 image-僅 in the source; 不 transcribed.
• 其中多少 Fable's 一般-access experience 是 actually Fable vs Opus-4.8 fallback 供 安全性-research-adjacent users whose queries trip the conservative classifiers?

Mythos 5#

• Suspension reason — shared 與 Fable 5; 不 stated in source.
• 如何 是否 "somewhat stronger than Mythos Preview" square 與 Opus 4.8's card claiming Mythos Preview 是 the capability frontier? 該 frontier 有 moved; the magnitude 不是 quantified here.
• 該 bio trusted-access SKU 是 "Fable 5 與 bio safeguards removed," 不 Mythos 5 — so "Mythos 5" strictly denotes the cyber-lifted variant. 是否 these converge 在 一個 trusted-access umbrella 是 unstated.

Claude Opus 4.7#

• 是否 Hakim's (2026) brevity-constraint findings on Opus 4.6 replicate on Opus 4.7, 或 是否 the literal-instruction-following change the elasticity? Specifically: 是否 <50 words still yield +13.1pp on GSM8K?
• Does Opus 4.7 still underperform as a planner in HotpotQA-style combo sweeps, 或 是否 improved instruction-following close the 缺口 that AgentOpt (Hua et al., 2026) identified?
• 什麼 是 the real-world token-inflation multiplier on typical Claude Code sessions (1.0–1.35× 是 content-dependent — what's the distribution on 程式碼-heavy vs. prose-heavy inputs)?
• 如何 是否 xhigh compare to max on coding evals? 該 migration guidance says "start 與 high 或 xhigh" — 是 max ever 值得 it 供 coding?
• 什麼 fraction of existing CLAUDE.md / 系統-prompt hedges become counterproductive 在 literal instruction following?

Claude Opus 4.8#

• Public model ID 與 pricing: the card 是否 不 state them; presumably claude-opus-4-8 at the Opus tier.
• 是否 grader-speculation trend continue to escalate in the 下一個 model, 與 at what point 是否 it begin to affect outward behavior?
• 為什麼 是 4.8 更少 robust to prompt injection than 4.7 despite broad alignment gains — a capability/robustness tradeoff, 或 an artifact of the eval surface?

Claude Sonnet 5#

• 該 head-to-head benchmark numbers vs Sonnet 4.6 與 Opus 4.8 是 image-僅 in the source; the System Card 有 the full set.
• 什麼 是 the real-world token-inflation multiplier on typical Sonnet 5 traffic (1.0–1.35× 是 content-dependent), 與 是否 "大約 成本-neutral" hold once effort levels rise?
• 為什麼 是否 a mid-tier model show 更高 behavioral-audit misalignment than the 更多 capable Opus 4.8 與 Mythos Preview — a capability-alignment coupling, 或 a training-recipe difference between the Sonnet 與 Opus/Mythos lines?
• 在什麼 effort 層級 是否 Sonnet 5 actually match Opus 4.8, 與 how 是否 the crossover 成本 compare to just running Opus 4.8?

Cowork#

• 如何 是否 Cowork's harness compare to Claude Code's? Both surface skills, MCP, sub-agents — 但 the 失敗 modes 供 non-程式碼 輸出 differ (no test suite, no compiler, no diff to review).
• 什麼是 eval discipline 供 Cowork-class outputs? Cat Wu says 記憶 效益 a lot 從 evals; unclear how slide-deck 品質 是 衡量的.

Emergent#

• 哪個 founding timeline 是 correct — 是 Emergent a YC Summer-2024 company (Tan) 或 a June-2025 founding (TechCrunch)? A 未來 authoritative source (Emergent's own about-page, YC batch records, Crunchbase) 應該 settle it; the machine transcript's self-flagged name uncertainty 使 Tan's the weaker claim, 但 the discrepancy 是 unresolved.

FastContext#

• 能否 the SFT+RL recipe push the explorer below 4B (1.7B / 0.6B) 與 使 exploration 實際上 free?
• 是否 gain transfer beyond Mini-SWE-Agent to richer harnesses 與 their own subagent orchestration?

Gemma 4#

• 為什麼 是否 the MoE underperform the dense model? Gemma 4 26B-A4B scores Elo 1438 on Arena 對抗 the 31B's 1451, despite MoE being the architecture 每一個 larger open model in their own table uses. Not addressed in the paper.
• 該 pre-training cutoff 是 January 2025 但 the model reports 89.2 on AIME 2026. 該 report says 資料 是 filtered "to decontaminate benchmarks." 會是什麼 that leave, 供 a competition held 在……之後 the cutoff?
• Is the encoder-free 12B's dense-text degradation an artifact of the 35M projection doing no feature compression, 或 of the 12B's training 執行 specifically? A 相同-size encoder/encoder-free ablation 會 settle it; the paper 執行 none.

Google DeepMind#

• DeepMind reports its bespoke 系統 being caught by simple loops. 是否 lab's comparative advantage move 從 系統 to 模型 + verifiers + benchmarks (mathlib, Formal Conjectures)?
• 論文 opens AI-供-math; what's DeepMind's 下一個 target domain where a sound verifier exists?
• Gemma 4's MoE (26B-A4B) loses to Gemma 4's dense 31B on 人類 preference, in a landscape where 每一個 larger open model 是 an MoE. Does DeepMind believe sparsity's returns 僅 begin above 一些 scale, 或 是 this a training artifact it hasn't explained?
• 如何 是否 a lab hold the Frontier Safety Framework 與 an open-weight thinking model in the 相同 hand? 該 published answer 是 that Gemma 是 far 從 the thresholds. 那 answer expires.

Hermes Agent#

• 該 container backend disabling dangerous-command checks 是 a defensible design 但 a meaningful 安全性-model 轉變. 什麼是 empirical track record? Have lockdown 失敗 in popular images (Daytona, nikolaik/python-nodejs) caused incidents?
• 如何 do bounded 記憶 files (~2,200 chars MEMORY.md) hold up 超過 long-term 使用? Auto-consolidation 是 mentioned 但 不 specified — what's the consolidation algorithm 與 how lossy 是 it?
• Hermes's DM-pairing flow 是 a clean 安全性 primitive. 為什麼 hasn't this 模式 been adopted by Claude Code 或 Cursor 供 shared/team deployments?
• 該 split between AGENTS.md (project) 與 SOUL.md (personality) 是 explicit in Hermes 但 implicit in Claude Code's CLAUDE.md. 是否 split materially improve outcomes, 或 是 it a documentation 選擇 沒有 empirical backing?
• Cron jobs in fresh sessions 與 no 記憶 — how do teams structure the "context the agent 需要" 沒有 it bloating 每一個 cron prompt? 是否存在 a standard 模式?

Lean#

• mathlib maturity gates the reachable frontier. 能否 AI formal proof search grow mathlib (formalize new theory) as a byproduct, expanding its own frontier?
• Lean 是 a perfect verifier 供 math. 哪個 other domains 有 a comparably sound automatic verifier (vs. 僅 noisy ones like tests 或 LLM-judge councils)?

Marcus Hutter#

• AIXI 是 incomputable 與 non-embedded; how far do recent fixes (amortized predictors, embedded/multi-agent AIXI) carry the theory toward practical relevance 供 real ASI?

METR#

• 什麼 new 任務 將 METR build to 衡量 days- 與 weeks-long horizons once 目前 baskets saturate?
• METR also 執行 the research showing developer self-estimates of AI uplift 是 overstated — how 是否 it reconcile that skepticism 與 its own steep time-horizon curve? Sharpened: Researcher Uplift from Code Output — a METR modeler (Kwa) threads exactly this needle: he discounts self-reports (citing METR's felt-+20% / 實際-−20% finding) 與 flags verbosity, yet still estimates >2× researcher uplift 從 an objective 8×-程式碼-輸出 figure 而不是 從 self-estimates — i.e. METR's skepticism 是 specifically about self-report metrics, 不 about the acceleration being real.

Mythos Model#

• 是否 Fable 5 / Mythos 5 return 在……之後 the post-launch suspension, 與 when?
• Capability profile beyond cybersecurity: Mythos Preview focused on the safety story; other capability dimensions 不 well-documented externally.
• Internal access controls: who at Anthropic actually uses Mythos 供 daily 運作, vs Opus 4.7? Boris implies infrequent (try-it 使用); 不 detailed.

Perplexity#

• A vendor publishing a benchmark its own product wins 是 an obvious incentive problem — how 是 DRACO's credibility 維護 as it ages, 與 將 Perplexity actually 執行 the automatable refresh?
• Perplexity 取決於 on Anthropic (與 others) 供 base 模型 while competing 與 them on the end product — how durable 是 the orchestration advantage if base-model makers ship their own deep-research mode?

Shane Legg#

• 該 report 假設 alignment 是 "solved to a sufficient degree" to focus on trajectories — how 是否 Legg's AGI-timelines optimism square 與 that scoping 選擇?

Symphony#

• 該 500% landed-PRs claim 是 hedged — no baseline definition, "on 一些 teams" 僅. 會是什麼 the distribution look like 跨越 teams? 會發生什麼 to PR 品質 與 revert 比率 at that throughput?
• "Workspaces preserved 跨越 執行" 是 the opposite of typical CI ephemerality. 在什麼 point 是否 state pollution 從 prior 執行 (stale node_modules, leftover branches, build artifacts) start hurting 更多 than warm-cache helps?
• Symphony 不會 撰寫 to the tracker — agents do. 這 means tracker policy 是 a prompt in WORKFLOW.md. 如何 brittle 是 this in practice when Linear changes its API? 如何 是 consistent state-machine behavior enforced when agents 有 prompt-層級 discretion?
• 該 spec 是 simplified by being implemented in 6 languages. 什麼是 extension of this technique? Could compiler-prompt.md in this vault be similarly cross-fuzzed?
• Symphony explicitly says agents can self-create tickets. 什麼 governance prevents runaway ticket-graph expansion? Is 人類 triage of agent-created tickets the 僅 check?

預測 — #oq/wait (79)#

Parked: falsifiable 僅 by 未來 events. Re-check when the named trigger (下一個 model generation, spec ratification, …) lands.

• Advantages of Digital Intelligence: 什麼 do ASI "societies" actually look like — homogeneous super-collectives, market ecologies, 或 compute-tethered virtual worlds?
• Agent Context Files: Will the role split converge on Hermes's explicit project/personality separation, 或 stay folded 成為 a 單一 file as in Claude Code?
• Agent Context Files: 是否存在 a natural ceiling on the layering (project → workflow → spec → constitution), 或 是否 每個 new autonomy surface spawn another context-file tier?
• Agent Harness Engineering: 如何 是否 architectural coherence evolve 超過 years in a fully agent-generated 系統?
• Agent Identity Management System (AIMS): No WG consensus. 這 是 an individual submission profiling other still-in-progress drafts (WIMSE identifier/creds/WPT/HTTP-sig, OAuth transaction-tokens, identity-chaining 是 所有 Internet-Draf…
• Agent Loop Pattern: 何時 the model schedules its own loops (4.7 behavior), who owns the budget?
• Agent Loop Pattern: Does a loop 與 a smart 足夠 model still 需要 a Kanban backlog, 或 是否 the model choose its own 下一個 任務 從 raw goals?
• Agent-Native Infrastructure: 誰 builds the agent-native rewrite of the long tail of 人類-facing services — the service owners, 或 a translation 層 (MCP servers, computer-使用 agents) on top?
• Agentic Loops Overtake Bespoke Systems: 該 bespoke advantage 是 dated "供 now." 什麼是 下一個 model generation's verdict — 是否 the evolutionary/AlphaProof apparatus survive on any problems, 或 fully collapse to a 成本 line?
• Agentic Work Systematization: 該 5.4%→26.6% curve 是 三個 months. Is this a durable behavior change 或 a novelty spike following a Codex skills-feature push?
• AGI-to-ASI Pathways: 能否 benchmarking methodology that 不會 saturate at 人類 層級 be built 在……之前 it's needed 供 ASI?
• AI-Accelerated Offense: Anthropic argues LLMs 效益 defenders 更多 long-term (like fuzzers) 但 攻擊者 更多 short-term during the transition. 如何 long 是 the transition, 與 what determines who wins it?
• AI as Primary Author: 如果 agentic authoring crosses 從 <1% toward double digits, 是否 the whiplash become unmanageable 在……之前 context-engine tooling matures — 或 是否 the tooling mature 因為 of the pressure?
• AI Investment Story, Not Efficiency Story: **何時 the crossover happen?
• AI Investment Story, Not Efficiency Story: Margin question (report's own): 是 the fastest-growers' 6–16pp-更低 gross margins a temporary AI-infra-成本 absorption 或 a permanent repricing of software's economic 品質?
• The AI-Native Safe-Choice Inversion: 該 inversion 是 a 一個-time repricing of "安全." Once several AI-native ERPs exist, 是否 "安全" re-stabilize around the largest AI-native vendor — 與 是否 Campfire's "we're now the largest of the n…
• The AI-Native Safe-Choice Inversion: 如何 long until incumbents bolt on credible AI 與 neutralize the counter-positioning — 與 是否 the custom-foundation-model claim actually defend 對抗 that?
• AI Product Economics Maturation: ICONIQ's respondents project gross margins expanding to ~59% by 2027, while Emergence's cap-table data measures the fastest-growers running 6–16pp below peer…
• AI Product Economics Maturation: Internal AI spend jumped 從 1–3% to a projected 16% of revenue 與 respondents calling true 成本 hard to predict. Is 16% a transient enablement bulge that falls as tooling matures, 或 a durable new…
• AI Usage Cadences: Continuous sampling 是 new; 是 these cadences stable, 或 將 they drift as the user base shifts toward 更低-wage 任務 (the report's own diffusion trend)?
• Autonomous Scientific Discovery: Every result 是 Anthropic-reported 與 example-selected; the genomics "100× smaller beats Science" claim 是 "intend to publish" — what 仍存在 external peer review?
• Benchmark Score Redundancy: **是否 rank stay 2?
• Benchmark Score Redundancy: **Would a 公開 probe set become a Goodhart target?
• Capability Gating Is Not Authorization: 該 deployment-tier ~3.2× exposure 缺口 (0.603 vs 0.189) means the cheap 模型 chosen 供 high-volume agent traffic 是 the most 可能 to emit the unauthorized 呼叫 — exactly where a per-呼叫 gat…
• Compounding Data Moat: 如何 是否 this moat hold up when foundation models themselves continue improving rapidly?
• Configurable Human Participation: 是否 "更多 channels adds coordination overhead" penalty shrink as the backbone improves (a capability 缺口), 或 是 it a structural 成本 of mixed-initiative interaction that persists?
• Context Window Smart Zone: 何時 sparse-attention 或 memory-augmented architectures ship, 是否 the smart zone become a soft constraint?
• Deep Research Agents: 是否 orchestration advantage shrink as base 模型 cross the 下一個 thresholds, 或 是 open-ended retrieval/synthesis a durable harness asset (unlike, say, prompt scaffolding)?
• Design Concept Grilling: 能否 grilling be 執行 AFK 對抗 another agent that 成立 the user's preferences?
• DRACO Benchmark: 該 benchmark 是 static; the construction pipeline 是 automatable. Will Perplexity actually refresh it, 與 是否 a vendor-built benchmark on which the vendor's own product wins stay credible 超過 time…
• Effective Compute Scaling: 何時 (if ever) 是否 scaling become economically unviable, 與 how do hardware/software-efficiency trends move that point?
• Engineer PM Convergence: Does this scale beyond ~50-person Claude Code-style teams?
• Engineer PM Convergence: 會發生什麼 to formal PM career ladders in 公司 where engineers do PM 運作?
• Evaluation Awareness & Grader Gaming: Does grader speculation continue to escalate 跨越 model generations, 與 是 there a capability 層級 at which it 是否 begin to affect outward behavior?
• Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated: 該 experience gradient rests on what workers believe AI 不能 do (judgment, relational 運作) — a belief that 可能 be either durable comparative advantage 或 the 下一個 capability to fall. 哪個, 與…
• Firm AI-Spend Intensity and Headcount Growth: **是否 effect diffuse beyond Information as adoption cohorts mature?
• Firm AI-Spend Intensity and Headcount Growth: **Is the entry-層級 growth durable 或 a lead-indicator that later reverses?
• 暫停驗證: 誰 adjudicates triggers 與 lifts?
• Harness Shrinkage as Models Improve: 該 Boris "100 lines" prediction 是 a year out 從 May 2026 — testable in 2027.
• Harness Shrinkage as Models Improve: 如果 harness 運作 shrinks, what new 運作 expands to fill it?
• Implementation Abundance Inverts Product Work: 如果 taste 是 the bottleneck 與 Research Taste as the Human Bottleneck, 是否 the inversion invert again — 是否 curation migrate 成為 the model?
• Interaction Models: 是否 interaction/background split 泛化, 或 是 it a transitional artifact until a 單一 model 是 both fast 與 deep 足夠?
• Interaction Models: "Interactivity scales 與 intelligence" 是 asserted; the larger-model release later in 2026 是 the test.
• Interaction Models: Research grant announced 供 interactivity benchmarks — what becomes the FD-bench equivalent 供 video proactivity?
• Latent vs. Deterministic Space: 該 seating example prices latent-space judgment at "a couple hundred dollars of tokens" 供 800 seat assignments. As 模型 absorb 更多 deterministic capability (Harness Shrinkage As Models Improve… [[llm Driven Vulnerability Research: How will the security industry's equilibrium shift when multiple labs have Mythos-class models?
• Loop Engineering: Does loop-engineering converge on a 單一 dominant shape (morning-triage → worktree → maker/checker → PR), 或 proliferate 成為 許多 idiom-特定 loops?
• Managers as ICs: Fung's own open question: "是否 you still 需要 separate iOS 與 Android orgs?
• Managers as ICs: Does manager-as-IC scale past a certain org size, 或 僅 運作 while Claude Code 是 小型 與 the 程式碼庫 是 Claude-legible?
• Market-Priced AI Exposure (the AI Premium): **Is the premium a durable 風險 price 或 an early-diffusion artifact?
• Market-Priced AI Exposure (the AI Premium): 該 agentic premium 是 僅 "early 證據" (imprecise). Does a positive agentic premium survive a longer sample, 與 是否 the 下降 price-per-agentic-token (caching + cheap-model routing) erod…
• Market-Priced AI Exposure (the AI Premium): **Does "Science most negative / interaction most positive" hold out of sample?
• MCP and Computer Use: 該 MCP ecosystem's growth 比率 vs. computer use's 品質 curve: at what point 是否 computer use become good 足夠 that the marginal 價值 of building an MCP server drops?
• MCP and Computer Use: Is computer use a sustainable 介面 或 a transition technology?
• Narrow Wedge into a Legacy Market: A wedge 運作 going in; 是否 it constrain going out?
• Outsource Your Thinking, Not Your Understanding: Karpathy's open frontier: can "understanding" itself 最終 be automated, 或 是 it definitionally the 人類 residue?
• Parallel Agent Orchestration: p99 OpenAI runtime of 71 agent-hours/day 是 a frontier preview 內部 an unusually favorable environment. Does external concurrency actually trend toward it as frictions fall, 或 是 heavy parallelism…
• Planning / Execution Division of Labor: 是否 人類 share of planning 決策 fall 超過 time as 模型 improve (the ceiling 上升 成為 the planning 層), 或 是 ~70% a stable 人類 floor?
• Polish No Longer Signals Readiness: Does 超過-anchoring get worse as builds get 更多 polished, 或 是否 everyone 最終 recalibrate 與 learn to discount fidelity 完全?
• Printing Press Software Democratization: 什麼是 equivalent of compulsory schooling 供 universal coding literacy?
• Product Velocity as Moat: Velocity-as-moat 是 a treadmill: it evaporates the moment a competitor matches pace. 什麼 converts Campfire's velocity lead 成為 a structural moat 在……之前 the AI-native cohort's pace converges?
• Recursive Self-Improvement: Is "research taste" a true ceiling (未來 1) 或 just the 下一個 capability to fall (futures 2–3)?
• Research Taste as the Human Bottleneck: Is research taste a genuine ceiling (an architectural capability scaling 不能 reach) 或 the 下一個 jagged valley to fill?
• Research Taste as the Human Bottleneck: 如果 taste 是 automatable, what — if anything — 仍然 a durable 人類 comparative advantage in AI development?
• Researcher Uplift from Code Output: Verbosity 與 價值-per-line 是 the load-bearing unknowns, 與 both 是 "at least partially resolvable 與 internal Anthropic 資料." Will any lab publish 品質-adjusted (不 just LoC) 程式碼-outpu…
• Responsible Scaling Policy Evaluations: 該 RSP determination leans heavily on "we 使用 it daily 與 it 不會 substitute 供 our 研究人員." 效果如何 是否 that subjective judgment scale as 模型 approach the threshold?
• Returns to Expertise in Agentic Coding: 該 forward test the report itself names: do the returns to expertise persist, narrow, 或 invert as 模型 improve?
• Returns to Expertise in Agentic Coding: Is "intermediate captures most of the 效益" stable, 或 an artifact of 目前 model capability — i.e., 將 the concave curve flatten further (everyone converges) 或 steepen (mastery starts to sepa…
• Review as the Control Point: 是否 no-review convergence hold as agentic authoring crosses 從 <1% of PRs toward double digits, 或 是否 the early-adoption discipline break down 在 volume the way Faros predicts?
• Seven Powers Applied to AI: 會是什麼 Boris's "cornered resource" look like 供 foundation-model labs that 是 themselves trying to commoditize?
• Software 3.0: 該 neural-net-as-host-process flip 是 presented as plausible-但-TBD. 什麼 會 the 第一個 production 系統 that 確實 inverts the CPU/NN relationship look like?
• Task Time-Horizon Scaling: Is the 4-month doubling a stable regime 或 a local steepening?
• The Three Loops of AI-Native Building: 如果 the 人類's presence in the middle loop 是 justified by a context advantage that 是 closable, the middle loop 是 a transitional structure. 會是什麼 a 兩個-loop world look like — 與 who translate…
• Universal AI (AIXI): Will a fundamental shortcoming of the 目前 paradigm (vs. the AIXI ideal) surface 在……之前 ASI 是 reached — i.e. 是 the "no theoretical blocker" conjecture 安全?
• Vibe Coding vs. Agentic Engineering: 如果 the mediocre/AI-native spread 保持 widening, what 是否 that do to team composition — a few extreme outliers plus agents, vs. broad mid-層級 staffing?
• White-Box Activation Monitoring: 如果 activation monitoring becomes load-bearing, 是否 training pressure 最終 push concealment 成為 channels the probes also 不能 讀取 (an arms race 一個 層級 deeper than CoT)?
• Why AI Lags at Design: Are reasons 3–4 (novelty, the abstraction 層) genuine ceilings, 或 — like reasons 1–2 — just 在-invested capabilities that fall once a lab builds the grader?
• Why AI Lags at Design: 能否 design be made gradable 沒有 a 人類 in the loop (learned taste 模型, preference 資料 at scale), 或 是否 the "人類 aspect of taste" resist automation the way Research Taste As Human Bottle… [[why Ai Lags At Design: Does the design↔code abstraction layer improve with better code-understanding models even if pure visual design stalls — i.e. is reason 4 a coding-capability problem in disguise?

Notes to rewrite — #oq/note (9)#

Observations phrased as backlog items — fold into the article body or rewrite as a falsifiable question at the next compile.

• Agent Identity Management System (AIMS): Posture assessment 是 deployment-特定 by design. By requiring no particular attestation mechanism, AIMS 使 interoperability of trust assurance (不 just protocol) unspecified: 兩個 conform…
• Agent Loop Pattern: Loop 輸出 review 是 now Matt Pocock's confessed bottleneck — "we just 需要 to be ready to be doing 更多 程式碼 review."
• Agentic Technical Debt: 該 remedy 假設 the founder 是 able to articulate architecture in plain language. Non-technical founders (the playbook's headline beneficiary group) may 有 neither the vocabulary nor the intuit…
• Agentic Technical Debt: Anthropic's Harness Shrinkage as Models Improve thesis 暗示 CLAUDE.md may 最終 be inferred by the model itself. Until then, the discipline 是 load-bearing.
• AI-Native Startup Lifecycle: Founder stories in the resources section (Carta Healthcare, Anything, Cogent, Airtree, Duvo, Zingage, Kindora, Wordsmith) 是 short callouts — none 有 published outcomes 或 comparable-baseline 資料.…
• Automatic vs. Flexible Cognition in LLMs: 該 proposed criterion — the workspace 是 engaged when an intermediate must be handed to an arbitrary, context-specified downstream circuit, 與 bypassed when the computation 是 automatic — 是 不 p…
• Latent Capability Overhang: 如果 成本 falls 10–100× per release, when 是 it ever rational to spend big extracting a capability now 而不是 waiting?
• Prototype Over PRD: 該 prototype-as-spec must 不 become the prototype-as-validation trap Problem-Solution Fit Discipline warns about: a fast prototype proves the build 是 solvable, 不 that the problem 是 real. #o…
• Single-Rollout Optimization: 該 online-learning win 是 on a controlled simulated preference 轉變 與 an LLM judge. Real user-facing online adaptation — the paper flags this itself — 需要 safeguards, monitoring, 與 privacy re…

進行中——部分已回答 (21)#

• Acceleration Whiplash: 其中多少 the "maturity 不會 protect" claim 仍存在 the vendor incentive to argue exactly that (i.e., "your existing practices won't save you — you 需要…
• Agent Data Injection (ADI): ADI 是 demonstrated on GPT-5.2-class agents. Does frontier model improvement reduce probabilistic-delimiter susceptibility, 或 是否 capability leave the deli… → Can Models Learn to Separate Instructions from Data? Durable Property vs Training Gap
• Agent Identity and Authentication: Hardware-bound credentials assume attested hardware everywhere agents 執行, including ephemeral cloud workloads 與 sub-agents. 如何 是否 attestation 運作 供 sho…
• Agent Identity Management System (AIMS): Mid-execution human-in-the-loop. 草稿 admits CIBA 僅 模型 client-initiated approval 與 "不會 map well" to confirmation needed mid-execution —…
• Agent-Native Infrastructure: Agent-to-agent negotiation 需要 trust, identity, 與 accountability primitives that don't exist yet. 什麼是 protocol 層, 與 who governs it?
• AI Accelerating AI Development: LOC, self-reports, 與 headroom-dependent multiples 所有 overstate; what unbiased throughput metric 會 Anthropic's promised 轉變 to "direct measurement of…
• AI-Native Organization: Tan's revenue-per-head figures (Emergent ~$15M ARR at 15 people, Retell $60M at ~40) 是 stated 從 stage 沒有 sourcing. 是否 third-party 資料 (Carta/Standard…
• AI-Native Organization: 該 org mapping predicts a testable staffing signature: AI-native 公司 應該 hire engineers to 維護 skills 而不是 function-特定 staff. Does…
• AI-Native Startup Lifecycle: 該 playbook gives no quantitative 證據 供 the headcount/capital compression claims (no median time-to-PMF, no headcount-at-PMF numbers, no 失敗-比率 da…
• The Automation–Optimism Link: Self-reported "no learning loss" 不能 detect real atrophy; 是 there an objective skill 衡量 that agrees, 或 是否 衡量的 skill diverge 從 felt skill (th…
• Compute-Controlled Benchmarking: Does a compute-controlled evaluation regime advantage frontier labs (who can afford the full curve) 超過 academics 與 third-party evaluators who 不能?
• Evals as Product Spec: 如何 do you 撰寫 an eval 供 taste-driven features like 角色訓練? → How Do You Write Evals for Taste? Character as the Limit Case
• Evals as Product Spec: 是否存在 a 單一 non-Anthropic example of a PM-as-eval-writer to cite, 或 是 this currently a Cat-Wu-singular framing?
• Founder as Agent Orchestrator: 該 "lean 10-person unicorn" 是 asserted; no quantitative 資料 in the playbook on 實際 headcount-at-PMF 或 headcount-at-Series-A medians 供 AI-native startup…
• Least Agency: Dynamic privilege elevation (Enterprise) reintroduces an elevation path; how 是 the elevation request itself authenticated 對抗 a manipulated agent?
• LLM-as-a-Judge: 如何 far can the judge's absolute calibration be trusted 供 thresholded 決策 (ship/no-ship, RSP gating) as opposed to rankings?
• LLM-Judge Validation: 是否 paradox 泛化 beyond position bias — i.e., 是 there other biases (self-preference, lineage) that high test-retest also masks?
• Model Introspection Feedback: 如何 reliable 是 4.7-class introspective reports?
• Model Introspection Feedback: Does adversarial introspection ("why 曾 you fail?
• Task Time-Horizon Scaling: Time horizon 是 衡量的 on 任務 baskets that themselves saturate; what replaces them once weeks-long 任務 become measurable — 與 who builds those 任務?
• Telemetry vs. Survey Measurement: 是否存在 a non-vendor telemetry dataset 大型 足夠 to adjudicate the maturity-protection question independently of Faros's commercial framing? → The Under-Review Divergence: Faros's Widening Crisis vs. CMU's Convergence