SkillHub.news 每日AI日报 · 报告形态

时间窗	事件	概率	依据
未来数周	Genie 4 从 trusted testers 扩到 Google AI Ultra 订阅者（Project Genie）与 Gemini API／AI Studio 早期伙伴	高(85%)	TechCrunch 引 Google 简报的官方时点
未来 3 个月	Anthropic 公布发行定价区间并挂牌（S-1 已交，TechCrunch 称正接近 IPO）	高(75%)	SEC 424B 文件与交易所公告为准
未来 3 个月	世界模型侧整合落地：AMD 收购 World Labs、亚马逊收购 Decart、Rivian 收购 Third Dimension 的最终条款披露	中(55%)	TechCrunch 报道方向已定，交易完成待监管与官宣
未来 3 个月	计算机科学界对 EHT 否证的公开评审结论（工作组／会议产出）	中(45%)	Ryan Williams 已发公开信征集反馈，尚无评审日程
未来 3 个月	Oracle AI Workspace 在 ChatGPT 内开放形态与计价口径披露（推断，官方未给时点）	中(40%)	推断：官方仅称 Oracle 在开发，未披露时间与定价

时间窗

事件

概率

依据

未来数周

Genie 4 从 trusted testers 扩到 Google AI Ultra 订阅者（Project Genie）与 Gemini API／AI Studio 早期伙伴

高(85%)

TechCrunch 引 Google 简报的官方时点

未来 3 个月

Anthropic 公布发行定价区间并挂牌（S-1 已交，TechCrunch 称正接近 IPO）

高(75%)

SEC 424B 文件与交易所公告为准

未来 3 个月

世界模型侧整合落地：AMD 收购 World Labs、亚马逊收购 Decart、Rivian 收购 Third Dimension 的最终条款披露

中(55%)

TechCrunch 报道方向已定，交易完成待监管与官宣

未来 3 个月

计算机科学界对 EHT 否证的公开评审结论（工作组／会议产出）

中(45%)

Ryan Williams 已发公开信征集反馈，尚无评审日程

未来 3 个月

Oracle AI Workspace 在 ChatGPT 内开放形态与计价口径披露（推断，官方未给时点）

中(40%)

推断：官方仅称 Oracle 在开发，未披露时间与定价

VOL.2026-10-28 · 25 STORIES · SKILLHUB DAILY

SkillHub.news Daily AI

Oct 28, 2026 · #45 · updated 08:00 daily · sources 4

Highlights15 stories · approx.10min read

MODEL RELEASES2 stories

Genie 4 launches: the first world model to bring together 'realistic visuals + real-time audio + mid-session editing'

PRODUCT3 stories

Oracle brings apps and data into ChatGPT: Oracle AI Workspace + security mesh + small-business bundle

INDUSTRY4 stories

Who controls Anthropic after IPO: long-term benefit trust + founders' 50.1% voting power; ordinary shareholders have almost no say

RESEARCH3 stories

Model 'overturns exponential-time hypothesis': academia's first question is not whether it is right, but 'is my password still safe?'

INSIGHTS3 stories

From "single-player AI" to "multiplayer AI": three quantifiable leakage metrics

01MODEL RELEASES2 stories

Genie 4 launches: the first world model to bring together 'realistic visuals + real-time audio + mid-session editing'

Google DeepMind官方视频 + TechCrunchConfidence HighNew topic

Released as a research preview, it generates the most realistic, interactive worlds to date; enter a prompt mid-simulation to change the scene (switch weather, change camera angle, add people or objects), and it generates matching music, ambient sound, and interactive dialogue in real time; the scene-consistency cap doubles from about one minute in the previous generation to about two minutes, with more natural physics and character performance.

Take · The official announcement discloses three unresolved issues: details drift in long simulations, audio accuracy is not guaranteed, and text within frames is not yet readable—the world model is still three gaps away from 'deliverable,' and these three gaps are exactly the hard barriers for film/TV and education use cases.

Original ↗

Approach shifts to a 'hybrid world model': Gemini acts as director and breaks down steps, while Genie 4 handles real-time simulation

Google DeepMindTechCrunchConfidence HighNew topic

Gemini and Genie 4 are deeply integrated: Gemini plans steps and checks frames; in a related Gemini Robotics experiment, Gemini 4 Argon breaks tasks into steps, with Genie 4 simulating in real time, so robots can first trial-run in simulation before going onto real hardware. Research Director Tejas Kulkarni describes it as a hybrid form in which 'Gemini writes code and text, while Genie ingests in real time and simulates reality with layered abstraction.'

Take · This is more important than the release itself: the moat for model companies shifts from 'a bigger model' to a 'planner + simulator' assembly; single-point models are easier to replace, and the assembly is what gets purchased.

Original ↗

02PRODUCT3 stories

Oracle brings apps and data into ChatGPT: Oracle AI Workspace + security mesh + small-business bundle

OpenAI × Oracle官方Confidence HighNew topic

Oracle is developing Oracle AI Workspace, letting users describe goals, find information, and complete approved business processes in one place (example: a new finance colleague uses ChatGPT to check costs, review purchasing decisions, and take stock of remaining month-end close work); it is also launching Oracle AI Security Mesh, which uses OpenAI cyber models to find and fix vulnerabilities, and an SMB bundle packaging ChatGPT Business + Codex + OCI credits. Officially disclosed: Oracle has 130,000 active ChatGPT users and 95,000+ Codex users; market research that previously required 2—4 days for the recruiting team now takes 15—20 minutes to prepare.

Take · The battle for the enterprise software entry point has ended at the “permissions and approval flows” layer: whoever can put the organization’s permission model into the chat box wins the default gate for the agent era; SMB bundle pricing, meanwhile, turns Codex into a renewal item rather than a new add-on.

Original ↗

Daybreak enters the defense side: Check Point vulnerability research compressed from hours to minutes

OpenAI × Check Point官方Confidence MidNew topic

Check Point CTO Jonathan Zanger said that, as a member of the OpenAI Daybreak defense network, the team uses frontier cyber models to analyze vulnerabilities, simulate attack paths, and find code and configuration gaps; a research process previously constrained by analyst headcount has shrunk from hours to minutes; Check Point covers more than 100,000 organizations.

Take · With models deployed on both attack and defense sides, the defender’s bottleneck shifts from “not enough people” to “the speed of turning intelligence into defense”; this will rewrite security vendors’ delivery cycles and pricing models.

Original ↗

Healthcare front-desk agent delivers hard metrics: phone wait -75%, document throughput 2—3x

Anima（英国）OpenAI 官方案例Confidence HighNew topic

Anima serves 1000+ clinics and 14 million patients in the UK, handling 800,000+ cases per month: phone wait dropped from 40—50 minutes to 10 minutes (a reduction of at least 75%), median case handling is about 3.5 hours, document throughput is 2—3x the baseline, and 86% of surveyed patients are willing to use AI receptionist Andi again; the technical division of labor is GPT-Live for incoming calls, GPT-6 for documents, and GPT 5.6 Luna for drafting consultation records, with an in-house policy engine in between constraining conversations into a deterministic intake workflow.

Take · In regulated industries, there is only one deployable form: an auditable process shell wrapping an uncontrollable model. The real leverage is that policy engine layer, not swapping in a larger model—this also holds for domestic healthcare, government services, and financial customer service.

Original ↗

03INDUSTRY4 stories

Who controls Anthropic after IPO: long-term benefit trust + founders' 50.1% voting power; ordinary shareholders have almost no say

AnthropicTechCrunchConfidence HighNew topic

Public benefit corporation status protects the board's safety decisions from shareholder lawsuits; the Long-Term Benefit Trust (LTBT) holds a class of shares that can elect and remove a majority of directors (Class T), with members including former Federal Reserve Chair Ben Bernanke; the seven founders collectively hold a class of shares, with 50.1% voting power, and can elect three directors; the board's nominating committee can veto the LTBT's nomination or removal without cause of 'designated' directors, referring the matter to an employee vote; founder control can optionally end 10 years after IPO, or automatically terminate if there are fewer than three founders.

Take · Write 'governance discount' as a priceable input: when passive money buys on index inclusion, it gets a security with almost no voting rights; at the same valuation, the risk premium must differ—this applies to all subsequent AI targets with a 'public benefit corporation + trust' structure.

Original ↗

$10 billion new fund, up to $1 billion per deal: late-stage big money in AI and defense is concentrating

DisruptiveTechCrunch（引 WSJ / 彭博 / 达拉斯商业期刊）Confidence HighNew topic

Dallas-based Disruptive (has invested in Palantir, Airbnb, Spotify, Groq, ElevenLabs) announced a $10 billion fund, to invest in at least 10 late-stage companies, up to $1 billion per deal; this is its first blind-pool fund, previously relying on deal-by-deal SPVs; Bloomberg said it has told LPs it returned more than $6 billion on about $1.9 billion invested; Dallas Business Journal said it is the largest VC fund in Texas history.

Take · The emergence of large late-stage single-deal funds = capital concentrating further into a few targets; second-tier companies pushed earlier into M&A or debt markets; this week's world-model-side consolidation moves in the same direction (AMD acquiring World Labs, Amazon in talks for Decart, Rivian acquiring Third Dimension).

Original ↗

Meta AI Director Rohit Patel departs to compete with Microsoft Office and Google Workspace

Meta → 微软／谷歌TechCrunchConfidence MidSingle-source pending verification

TechCrunch reported the personnel change and indicated the direction: moving to the office-suite battlefield; the report did not disclose the specific title, timing, or non-compete terms.

Take · Talent flow is a battlefield signal: office suites are the thickest existing budget pool for agent deployment; Meta's absence at this layer hands the 'work entry point' to Microsoft and Google.

Original ↗

Research AI's decisive factor is 'dark data': dementia alliance uses federated architecture to bypass institutional walls

NVIDIA × C-BRAINNVIDIA 官方博客Confidence HighNew topic

With 57 million dementia patients worldwide and 10 million+ new cases per year, NIH spending grew 7x over 15 years to $3.9 billion per year, yet there is still no way to stop, prevent, or reverse the disease; C-BRAIN (18 member institutions, launched November 2025) and NVIDIA use a hybrid centralized + federated architecture to access each institution's undisclosed data (including negative results and governed proprietary data); officials say OpenScientist completes in minutes analyses that domain experts estimate require weeks to months, at less than $10 per run, has covered 200+ research groups, and at scale could release over 1 million hours of research capacity per year.

Take · Frontier models do not contain life sciences' 'dark data', so the upside is not in general models but in lawful data access: whoever gets their governance architecture working first gets this corpus that others cannot access—this is also the most realistic entry point for domestic research AI.

Original ↗

04RESEARCH3 stories

Model 'overturns exponential-time hypothesis': academia's first question is not whether it is right, but 'is my password still safe?'

OpenAI / Anthropic / MITTechCrunch 专访 Ryan WilliamsConfidence MidNew topic

OpenAI released solutions to three math problems, including an unexpected refutation of EHT; Anthropic mathematician Levent Alpöge said they were based on existing work, and hours later Anthropic also released its own refutation. MIT’s Ryan Williams (who once called OpenAI’s previous “proof dump” a DDOS on the math community) and the original 1999 ETH authors Impagliazzo and Paturi were invited to interpret it; the first step was to issue an open letter soliciting feedback and discussing norms for AI contributions; he said the proof has two or three “alien-style” parts and first asked whether it affects cryptosystems, and OpenAI said it had checked and believed it did not constitute a break.

Take · Reading: treat “disruptive conclusions” as a process problem — routing by discipline through GitHub / working groups / conferences as three release paths is a more reusable lesson than the proof itself; meanwhile Williams notes that ETH has long been cited by a large number of papers, and if it holds, the foundations of those works must be recalculated.

Original ↗

A reporter used AI to fabricate two fake papers to test their own gate: neither was accepted, but it exposed the true scale of “AI slop”

IEEE SpectrumIEEE Spectrum 实验报道Confidence HighNew topic

An IEEE Spectrum reporter used “vibe-research” to generate fake papers with commercial models and partnered with the editor-in-chief of IEEE Access for adversarial testing; background figures cited in the article: in 2025 about 57% of published papers showed detectable large-model influence (12% in 2023); in 2025 more than 12% of review comments in Nature Communications were entirely AI-generated; only 1 in every 40 papers with statistical evidence formally disclosed; neither test paper was accepted, nor did they go through peer review.

Take · This is most practical for people doing content delivery: the gate on the review side is still working, but outside the gate hundreds of thousands of fake papers are still produced each year — “whether it can be detected” is becoming a new delivery threshold; those in the AI content business should add this metric to acceptance ahead of time.

Original ↗

The New York Times asks: has the hardest problem in AI suddenly become simple (interpretability)?

NYT ScienceHacker News AI（引纽约时报）Single-source pending verification

The Hacker News AI section included this New York Times report that day, pointing toward interpretability; the full text was not obtained in this round, and the report's specific claims, methods, and author conclusions were not verified—only the title and entry point are noted.

Take · If interpretability does make a real breakthrough, compliance and insurance will be rewritten first: only models that can show "why this decision was made" can even be discussed as insurable and auditable.

Original ↗

05INSIGHTS3 stories

From "single-player AI" to "multiplayer AI": three quantifiable leakage metrics

组织落地Salesforce 250 位高管调查Confidence MidNew topic

Survey basis: 68% see AI insights go unshared each week; 66% see colleagues repeatedly solve the same problem; 59% see work stuck inside someone's AI tool; average 3.6 hours per week spent switching windows (annualized 22 working days); only 15% say AI is already seamlessly embedded across multiple systems; 73% of large enterprises get stuck weekly vs 37% of SMEs; 72% see team-level AI with shared context as the ideal state.

Take · Actionable approach: measure these three as weekly KPIs once (number of repeated problem-solving instances, insight reuse rate, time stuck at blockers), then decide whether to buy tools or first assetize prompts—solutions in individual chats count as cost reduction only once they become team assets.

Original ↗

Equip the agent with "budget circuit breaker + observability": alert at 80%, switch to a cheaper model at 90%, block at 100%.

Agent 运维Show HN（项目自述）Confidence MidNew topic

flightdeck is self-hosted under Apache 2.0 and takes over SDK calls via an in-process sensor (init + patch, no other code changes, requests do not go through a proxy); capabilities include per-agent timelines, per-flavor token budgets (80% alert / 90% downgrade to a cheaper model / 100% block), MCP service allowlists and blocklists, single-agent or whole-group kill switch, sub-agent trees (Claude Code, CrewAI, LangGraph), and cost analysis by model/framework/host; prompts are not captured by default.

Take · This is a minimal implementation that turns "agent cost and privilege overreach" from incidents into configurable items: first deploy the three budget tiers and MCP allowlist, which is an order of magnitude cheaper than after-the-fact auditing; note that it is a personal project with no independent audit, so review the code yourself before production.

Original ↗

Before using Genie 4 for film/TV previsualization, first turn the three officially acknowledged gaps into an acceptance checklist.

交付验收TechCrunch 官方访谈Confidence MidNew topic

Official statement: details and objects drift with simulation duration, audio accuracy cannot be guaranteed, and in-frame text is not yet readable; currently open only to trusted testers, and in the coming weeks will open to Google AI Ultra subscribers (Project Genie) and early partners of Gemini API/AI Studio; the most valuable use case cited by official engineers is rapid ideation and storyboarding in the early creative stage (maintaining flow), not finished footage.

Take · Actionable approach: set "no drift within two minutes, audio matches the environment, text is readable" as three hard acceptance criteria, use it only in the storyboard previsualization stage first, and keep finished footage on the existing pipeline—this captures the speed benefit without being dragged into rework by the official known gaps.

Original ↗

Official Events10 stories

YouTube·Google DeepMindGenie 4: A new generation of world simulation⚡96

官方·TechCrunch AIGoogle’s Genie 4 world model adds audio, more realistic worlds, and real-time steering⚡94

官方·OpenAI NewsOpenAI expands partnership with Oracle to bring its apps to ChatGPT⚡90

官方·OpenAI NewsAnima helps clinics cut call waits by 75% with GPT-6 and GPT-Live⚡88

官方·TechCrunch AIFrontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean?⚡87

官方·SalesforceNew Research: Work Gets Trapped in Single-Player AI. Executives Are Ready for Multiplayer⚡85

官方·OpenAI NewsHow Check Point accelerates vulnerability research with Daybreak⚡84

官方·TechCrunch AIWho’s really in charge at Anthropic after the IPO?⚡83

官方·TechCrunch AIDisruptive raises $10B fund to make big bets on at least 10 companies⚡81

官方·NVIDIA BlogHow AI Is Changing the Pace of Discovery for Millions Living With Dementia⚡79

Ecosystem Pulseself-updating

News600Skills58OpenHub65Reddit25

2026-09-02 → 2026-09-10 Star growth：obra/superpowers +90,630★ · mattpocock/skills +82,178★ · multica-ai/andrej-karpathy-skills +67,613★ · anthropics/skills +56,067★ · Shubhamsaboo/awesome-llm-apps +43,732★

※Risk Levelsgray rhino / black swan

Gray rhino · foreseeable

World models ‘demoable but undeliverable’: Genie 4 remains a research preview; officials acknowledge three unresolved issues—long-duration drift, audio accuracy, and in-frame text; if film/TV and game procurement contracts to finished-film standards, rework and acceptance disputes will erupt within weeks (medium-high, 50-70%)
AI slop erodes research and content trust: in 2025 about 57% of papers show large-model traces, over 12% of Nature Communications reviews are fully AI-generated, yet only 1 of every 40 discloses it—if evaluation systems and procurement acceptance still measure by ‘text output volume,’ clearing out bad actors will be slower (medium-high, 40-60%)
Capital concentrates into a few targets: a USD 10 billion late-stage fund with single checks up to USD 1 billion + serial M&A on the world-model side; second-tier AI companies will be forced to sell themselves or shift to debt financing before 2027, intensifying valuation divergence (medium, 40-60%)
Governance discount becomes explicit: Anthropic’s trust + founder 50.1% structure leaves ordinary shareholders with almost no voting rights; when passive funds include it in indices, they must reassess the risk premium, and similar ‘public benefit corporation’ targets will collectively face valuation discounts (medium, 35-55%)

Black swan · low-probability

Cryptography emergency: If the EHT falsification holds and the algorithm spills over into existing cryptographic parameters, systems relying on key length need a collective assessment; OpenAI says it has checked and does not consider it a break, but academia has not yet independently reviewed it (low, <25%)
Simulation-reality gap incidents: World models and simulation are starting to enter autonomous driving and robotics testing (Rivian acquires Third Dimension, Amazon in talks over Decart, AMD acquires World Labs); once passing simulation replaces real-vehicle validation, liability allocation for safety incidents will have no precedent to follow (low, <20%)
AI-generated evidence contaminates justice and regulation: IEEE experiments show that people without professional backgrounds can use commercial models to create fake papers sufficient to enter the review process; applying the same methods to legal, medical, and financial-reporting scenarios would directly challenge the default assumption that 'documents are evidence' (low, <20%)

✦Action Items3

1. Turn "team shared context" into a small product that can ship in a week.

First measure three leakage metrics (weekly repeated problem-solving count, insight reuse rate, blocker duration), then implement the minimal closed loop: shared prompt library + team-level agent workspace + each session captured as searchable entries. Path: internally run for one week to establish a baseline → package externally as 'AI context audit' and charge per team → advance to a cross-system context layer (integrated with ticketing, IM, knowledge base). Criterion: weekly repeated problem-solving count falls by more than half within four weeks.

window Starting immediately, 4—6 weeks (enterprise Q4 budget window)

2. Give existing agents 'three budget tiers + MCP allowlist'

Within one week, turn cost and unauthorized access from incidents into configuration: token budget alerts at 80% / downgrade at 90% / block at 100%, MCP service allowlist, kill switch for individual agents and the whole group, and sub-agents labeled by role. Can build in-house or directly self-host flightdeck (Apache 2.0, in-process sensor without code changes), but must review the code and capture packets yourself to confirm prompts are not logged by default. Criterion: in one stress test, block an unauthorized MCP call before it leaves the process.

window 1—2 weeks (first stop the bleeding, then discuss a governance platform)

3. Claim the delivery slot for 'AI previsualization'

World models can currently offer only storyboards and creative previsualization; final cuts still go through the existing pipeline. Path: first use off-the-shelf tools (Veo/GWM Worlds-type) to turn 'one sentence → storyboard previsualization' into a 48-hour delivery package, while writing an acceptance checklist for the three officially acknowledged gaps (no drift for two minutes, audio and environment matching, readable text); when Genie 4 opens to AI Ultra and Gemini API early partners, the first batch will have comparable delivery samples. Criterion: secure 3 paying previsualization orders.

window 3—6 months (get ahead before the opening window)

⏱Next 3 MonthsProb.

Window	Event	Prob.	Basis
未来数周	Genie 4 expands from trusted testers to Google AI Ultra subscribers (Project Genie) and early Gemini API/AI Studio partners	High (85%)	TechCrunch cites official timing from Google briefing
未来 3 个月	Anthropic announces issue price range and listing (S-1 filed; TechCrunch says nearing IPO)	High (75%)	Subject to SEC 424B filing and exchange announcement
未来 3 个月	World-model consolidation lands: final terms disclosed for AMD's acquisition of World Labs, Amazon's acquisition of Decart, and Rivian's acquisition of Third Dimension	Medium (55%)	TechCrunch reports direction is set; deal completion awaits regulatory and official announcement
未来 3 个月	Computer science community's public review conclusion on the EHT refutation (working group/conference output)	Medium (45%)	Ryan Williams has issued an open letter soliciting feedback; no review schedule yet
未来 3 个月	Oracle AI Workspace’s availability format and pricing basis inside ChatGPT disclosed (inferred; no official timing given)	Medium (40%)	Inferred: officials only say Oracle is in development; no timeline or pricing disclosed

Sources：In-Depth Daily Report (Daily Core AI Information)：Issue 29 · 24.5 KB · AI Official Video Monitoring Log：41 events · 41 with original-source links · Site ecosystem data：news 600 · skills 58 · openhub 65 · reddit 25 · Official Event AI Picks：10 items (with source links)
SkillHub.news — Global Skill rankings

SkillHub.news 每日AI

Genie 4 发布：世界模型首次凑齐「逼真画面 + 实时音频 + 中途改写」

路线转向「混合世界模型」：Gemini 当导演拆步骤，Genie 4 负责实时模拟

甲骨文把应用与数据接进 ChatGPT：Oracle AI Workspace + 安全网格 + 小企业组合

Daybreak 进防守方：Check Point 漏洞研究从数小时压到数分钟

医疗前台智能体跑出硬指标：电话等待 -75%，文档吞吐 2—3 倍

Anthropic 上市后谁说了算：长期利益信托 + 创始人 50.1% 表决权，普通股东几乎无话语权

100 亿美元新基金、单笔最高 10 亿美元：AI 与国防的晚期大额钱集中化

Meta AI 总监 Rohit Patel 离职，转去对标 Microsoft Office 与 Google Workspace

科研 AI 的胜负手在「暗数据」：痴呆症联盟用联邦架构绕开机构围墙

模型「推翻指数时间假说」：学术界先问的不是对不对，而是「我的密码还安全吗」

记者用 AI 造了两篇假论文去打自家闸门：两篇都没被接收，但暴露了「AI slop」的真实规模

纽约时报问：AI 里最难的问题是不是突然变简单了（可解释性）

把「单人版 AI」改成「多人版 AI」：三个可量化的漏损指标

给 agent 装「预算熔断 + 可观测」：80% 告警、90% 换便宜模型、100% 阻断

拿 Genie 4 做影视预演之前，先把官方承认的三个缺口写成验收表

Genie 4 launches: the first world model to bring together 'realistic visuals + real-time audio + mid-session editing'

Approach shifts to a 'hybrid world model': Gemini acts as director and breaks down steps, while Genie 4 handles real-time simulation

Oracle brings apps and data into ChatGPT: Oracle AI Workspace + security mesh + small-business bundle

Daybreak enters the defense side: Check Point vulnerability research compressed from hours to minutes

Healthcare front-desk agent delivers hard metrics: phone wait -75%, document throughput 2—3x

Who controls Anthropic after IPO: long-term benefit trust + founders' 50.1% voting power; ordinary shareholders have almost no say

$10 billion new fund, up to $1 billion per deal: late-stage big money in AI and defense is concentrating

Meta AI Director Rohit Patel departs to compete with Microsoft Office and Google Workspace

Research AI's decisive factor is 'dark data': dementia alliance uses federated architecture to bypass institutional walls

Model 'overturns exponential-time hypothesis': academia's first question is not whether it is right, but 'is my password still safe?'

A reporter used AI to fabricate two fake papers to test their own gate: neither was accepted, but it exposed the true scale of “AI slop”

The New York Times asks: has the hardest problem in AI suddenly become simple (interpretability)?

From "single-player AI" to "multiplayer AI": three quantifiable leakage metrics

Equip the agent with "budget circuit breaker + observability": alert at 80%, switch to a cheaper model at 90%, block at 100%.

Before using Genie 4 for film/TV previsualization, first turn the three officially acknowledged gaps into an acceptance checklist.