Boss lens: AI valuations swing hard on regulation headlines—pad budgets and hiring plans (inference).
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 6, 2026
11-05 07:50 – 11-06 07:50 (UTC+8)
Top 3
When office agents can act, free quotas vanish fast—pre-plan capacity and fallback models (inference).
AI can now read your ledger—set rules on which finance data staff may paste into chats (inference).
All
Nasdaq 100 drops ~6% (~$2.3T) and Philly semis ~12% on AI-regulation fears after Democrats win Congress
Boss lens: AI valuations swing hard on regulation headlines—pad budgets and hiring plans (inference).
CollapseNasdaq 100 drops ~6% (~$2.3T) and Philly semis ~12% on AI-regulation fears after Democrats win Congress
Per Bloomberg via Techmeme, on Nov 5 the Nasdaq 100 fell about 6%, wiping roughly $2.3T in market value (described as a record), while the Philadelphia semiconductor index dropped about 12%. After Democrats won Congress in the midterms, investors sold AI and chip names on fears of tighter AI oversight and data-center pushback. Market-move reporting, not company earnings.
Tencent Hunyuan Hy4 blows through WorkBuddy quotas on day one; inference cluster scaled 10×, Hy3 free trial to Nov 15
When office agents can act, free quotas vanish fast—pre-plan capacity and fallback models (inference).
CollapseTencent Hunyuan Hy4 blows through WorkBuddy quotas on day one; inference cluster scaled 10×, Hy3 free trial to Nov 15
IT Home, Yicai and 36Kr report: on Nov 5 Tencent Hunyuan and WorkBuddy said Hy4’s Nov 4 WorkBuddy launch hit free quotas on day one. The Hy4 inference cluster was emergency-scaled 10× and free quotas raised, with peak throttling still possible; users may switch to Hy3 (free until Nov 15 23:59) or off-peak. Hy4 adds vision and computer-use and is slated for open weights soon.
Claude launches Personal Finance: read-only bank/card link via Plaid for US Pro/Max/Ultra; no transfers or investing
AI can now read your ledger—set rules on which finance data staff may paste into chats (inference).
CollapseClaude launches Personal Finance: read-only bank/card link via Plaid for US Pro/Max/Ultra; no transfers or investing
Anthropic announced Claude Personal Finance beta on Nov 5: users connect chosen bank/card accounts via Plaid and ask about spending, bills and balances in plain language. Access is read-only—no moving, managing or investing money. Rolling out this week to US Pro, Max and Ultra; hub chats aren’t used for training. ChatGPT shipped a similar feature earlier. Not available outside the US.
Baidu launches Token Exchange marketplace with 42 models; ~398M calls in 7 days (platform screenshot)
Model shopping is becoming a marketplace—SMBs can pick and cut cost without vendor lock-in tours (inference).
CollapseBaidu launches Token Exchange marketplace with 42 models; ~398M calls in 7 days (platform screenshot)
36Kr citing Dujia: Baidu launched Token Exchange, a marketplace linking model suppliers and buyers with 42 text/video models including DeepSeek, Kimi and Zhipu GLM. One API switch, price/latency/success-rate compare; same model can have multiple suppliers. A ranking screenshot shows ~398M calls and ~5.1336T tokens over 7 days, ~290ms TTFT, with DeepSeek-v4.1-flash leading volume. Platform-displayed, not audited earnings.
TeraWulf: Google credit-backstops financing for Kentucky Justified campus behind ~$19B Anthropic leases
Even AI leases need a credit backstop—expect tougher counterparty clauses in infra deals (inference).
CollapseTeraWulf: Google credit-backstops financing for Kentucky Justified campus behind ~$19B Anthropic leases
TeraWulf’s Nov 4 release: Google LLC will provide credit support for project debt tied to ~$19B of contracted Anthropic leases at the Justified campus in Hawesville, Kentucky, plus a Fluor EPC (GMP) deal with phased ready-for-service from late 2027 into 2Q 2028. CoinDesk and others say Google could assume the lease or take equity if Anthropic defaults. Company PR plus media gloss.
ByteDance tightens rules: personal Claude/Codex accounts barred from L4 repos and Feishu docs this month
Big Tech is closing the gate—don’t paste code/docs into personal AI accounts without a policy (inference).
CollapseByteDance tightens rules: personal Claude/Codex accounts barred from L4 repos and Feishu docs this month
TechNode, TMTPost and Guandian report ByteDance this month blocks external AI tools—including personal Claude and Codex accounts—from company L4 code repos and Feishu docs. VP Hong Dingkun said internally the concern is L3+ leakage, denying a push to force in-house tools; public info and personal use remain OK. No formal external notice from ByteDance.
Baidu open-sources ERNIE 5.5 Flash Base: 280B MoE / 14B active, pretrained on 33.5T tokens
Another big-lab base drop—teams that fine-tune privately get another starting checkpoint (fact).
CollapseBaidu open-sources ERNIE 5.5 Flash Base: 280B MoE / 14B active, pretrained on 33.5T tokens
Xinhua on Nov 5: Baidu open-sourced ERNIE 5.5 Flash Base, a MoE with 280B total / 14B active parameters, pretrained on ~33.5T tokens; train context up to 128K and fine-tune/infer up to ~1M. Baidu frames it as efficiency-focused. Base weights, not a chat-tuned product; some benchmarks are company-reported.
OpenAI community compact: own-and-operate campuses pay full power/grid costs; no NDAs with public officials (narrow exceptions)
Campus siting now runs through community politics—who pays for power and NDAs matter (inference).
CollapseOpenAI community compact: own-and-operate campuses pay full power/grid costs; no NDAs with public officials (narrow exceptions)
On Nov 5 OpenAI published a National Community Compact and compute-campus blueprint: future campuses it owns and operates will pay full electricity and needed grid/infra costs, with financial assurances before major spend; publish water/power/noise impacts; avoid NDAs with public officials except narrow security/trade-secret cases; and urge lawmakers to codify similar rules. CNBC ties the move to post-election data-center backlash.
Houmo Intelligent closes multi-billion-yuan Series B for edge CIM chips; says M50 won 150+ design wins in <1 year
Edge-chip money keeps flowing—buyers should check delivery and design-win proof, not slides (inference).
CollapseHoumo Intelligent closes multi-billion-yuan Series B for edge CIM chips; says M50 won 150+ design wins in <1 year
NBD, PEDaily and EEO (Nov 4–5): edge-AI chip firm Houmo Intelligent closed a multi-billion-yuan RMB Series B with industry/state capital and returning LPs, for M50 scale-up, M6X/next-gen 3D CIM R&D and tooling. Reports cite ~160 TOPS INT8 / ~10W typical and 150+ design wins within a year of production. Investor list and exact valuation not fully disclosed.
AMD 10-Q prices pending Taalas AI-inference chip deal at ~$2.9B, including ~$1.4B contingent
A ~$3B bet on etched-model inference shows cheap serving is still a strategic war (inference).
CollapseAMD 10-Q prices pending Taalas AI-inference chip deal at ~$2.9B, including ~$1.4B contingent
AMD’s Nov 4, 2026 Form 10-Q and Globe and Mail coverage value the pending cash-and-stock buy of Canadian inference-chip startup Taalas at about $2.9B, including up to ~$1.4B contingent consideration, targeting close by end-2026 subject to approvals. Taalas focuses on etching models into silicon for inference.
Notion founder: ~half of pages created by agents; in-product AI use ~9× and MCP volume ~60× this year
Docs are becoming agent context hubs—structure knowledge for machines, not just humans (inference).
CollapseNotion founder: ~half of pages created by agents; in-product AI use ~9× and MCP volume ~60× this year
Notion founder Ivan Zhao’s “The Next Era of Notion” (relayed by vista8 et al.) says ~a year after Agents, roughly half of Notion pages are agent-created, in-product AI use is up ~9× this year, and MCP volume for third-party agents storing context is up ~60×. Design/product/eng merged into a “building” org; marketing split. Company-reported.
Ant Ling ships Ling-3.2-flash: 560B/25B MoE, 1M context, native multimodal long-horizon agents
A payments lab pushes long-context multimodal agents—worth a pilot on docs and receipts (inference).
CollapseAnt Ling ships Ling-3.2-flash: 560B/25B MoE, 1M context, native multimodal long-horizon agents
Ant Ling docs and republishers say Ling-3.2-flash launched Nov 5: a native multimodal MoE (~560B total / ~25B active) with 1M context, aimed at hard vision, long-horizon multimodal agents, 3D and medical multimodal tasks. Vendor release; independent benches still thin.
Oumi launches FAiE: describe a task in plain language; it evaluates, curates data and trains a custom agent
Agent-building itself is being productized—non-ML teams can try a custom-model loop (inference).
CollapseOumi launches FAiE: describe a task in plain language; it evaluates, curates data and trains a custom agent
Oumi on Nov 4 (BusinessWire/SiliconANGLE) launched Frontier AI Engineer (FAiE): after ~15 minutes describing a task, it automates evals, failure diagnosis, data curation, training strategy and experiments, extending to tools/environments and SFT/DPO to ship specialized weights/agents, then improve from production failures. Product launch aimed at automating AI engineering itself.
METR: Anthropic’s capability gains ~1.9× faster since Oct 2025, largely from AI in R&D (high uncertainty)
AI-sped R&D is now an evaluator number—don’t lock vendor versions for a full year (inference).
CollapseMETR: Anthropic’s capability gains ~1.9× faster since Oct 2025, largely from AI in R&D (high uncertainty)
METR’s Nov 5 report (via dotey/Techmeme/testingcatalog) estimates Anthropic’s rate of capability gain on well-specified tasks through Oct 2026 is ~1.9× the Mar 2024–Oct 2025 pace, largely from AI-aided R&D; METR flags high uncertainty and says Anthropic redacted some informative evidence. Evaluator view, not earnings.
CA Sen. Scott Wiener (SB 53 author) wins Pelosi’s House seat, campaigning on AI regulation
A state AI-law author enters Congress—US compliance may shift from patchwork to federal drafts (inference).
CollapseCA Sen. Scott Wiener (SB 53 author) wins Pelosi’s House seat, campaigning on AI regulation
Techmeme citing The Information: California State Senator Scott Wiener, an author of AI safety law SB 53, won Tuesday’s race to succeed Nancy Pelosi in the US House, campaigning on federal AI regulation. SF local press says Connie Chan conceded Nov 5. Election result; statute path still TBD.
Artificial Analysis: Nanbeige4.3-33B scores 20 on Intelligence Index, ties Qwen3.8-27B, stronger on terminal tasks
A recruiting platform open-sourcing agent models shows terminal control sells better than quiz scores (inference).
CollapseArtificial Analysis: Nanbeige4.3-33B scores 20 on Intelligence Index, ties Qwen3.8-27B, stronger on terminal tasks
Artificial Analysis says BOSS Zhipin’s Nanbeige lab open Nanbeige4.3-33B (~33B dense, agent/coding-oriented) scores 20 on the Intelligence Index, tying Qwen3.8-27B, stronger on terminal/agent benches and weaker on knowledge. Already on Hugging Face. Third-party leaderboard; see AA for setup.
Meta Muse rolls Voice and expressive avatars to US users; expands to Mexico
Consumer assistants get voice and faces—B2B still hinges on workflow entry, not avatars (inference).
CollapseMeta Muse rolls Voice and expressive avatars to US users; expands to Mexico
testingcatalog on Nov 5 relays Meta Muse: Voice (think-aloud for brainstorming, trips, interview prep) and expressive avatars are rolling out to US users, with Muse also expanding to Mexico. Feature rollout; no user or revenue figures.
vista8 drops ScreenStudio subscription for open-source Cap (~24k stars) to customize
When SaaS is pricey and weak on Chinese UX, open-source-plus-fork is the default path (inference).
Collapsevista8 drops ScreenStudio subscription for open-source Cap (~24k stars) to customize
vista8 on Nov 5: after ~a year on ScreenStudio, praised UX but criticized slow export, weak Chinese captions and no speed control; switching to open-source Cap (~24k stars per reply) to customize and canceling renewal. Bookmarks 44 vs likes 16—classic tool-save distribution.
turingou: after a vertical personal-agent mid-platform, a parking-fee app in 15 minutes
Solo builders win by platforming once, then cloning verticals—not rewriting each app (inference).
Collapseturingou: after a vertical personal-agent mid-platform, a parking-fee app in 15 minutes
turingou on Nov 5: betting on vertical personal agents (~a dozen in progress); once a mid-platform exists, scene/entry/UX/monetization copy fast. A follow-up says after SandBank Agent mid-platform, a high-completion coin-parking fee app took ~15 minutes and links with other assistants. Founder anecdote, not audited.
AI security week: AI agents hit 1,200+ sites / 100+ firms for ~$5,000; Microsoft & LiteLLM incidents cited
Attack cost is a few thousand dollars—isolate keys before handing agents broad access (inference).
CollapseAI security week: AI agents hit 1,200+ sites / 100+ firms for ~$5,000; Microsoft & LiteLLM incidents cited
PM Aakash Gupta’s Nov 5 roundup of the last 7 days: a hacker used AI agents across 1,200+ sites reaching 100+ companies for ~$5,000; Microsoft cited 2,250+ tenants targeted and 327 accounts compromised; LiteLLM briefly shipped a credential stealer (~2 hours). Links The Hacker News, LiteLLM, The Register—verify primaries.
WIC blue book: China’s enterprise AI-agent market ~¥21.2B in 2025
Official stats now size enterprise agents as a market—use the same order-of-magnitude when pitching (fact/inference).
CollapseWIC blue book: China’s enterprise AI-agent market ~¥21.2B in 2025
HK news wire Nov 5 citing the World Internet Conference blue book: China’s enterprise AI-agent market reached ~¥21.2B in 2025; embodied-intelligence market cited ~¥915B. Forecasts AI contributing ~$22.3T to the global economy by 2030. Conference stats/forecasts—methodology in the full report.
OpenAI launches GPT-6 Cyber for authorized vuln research; ~51% relative gain on ExploitGym vs GPT-5.6 Cyber
Defender cyber models are productizing—apply via official channels; don’t point them at random targets (inference).
CollapseOpenAI launches GPT-6 Cyber for authorized vuln research; ~51% relative gain on ExploitGym vs GPT-5.6 Cyber
OpenAI’s Nov 5 post introduces GPT-6 Cyber for finding/validating exploitable vulns while staying in researcher-set scope. ExploitGym ~48.5% vs GPT-5.6 Cyber’s 32.1% (~51% relative). Alignment tests: ~99.93% avoided out-of-scope exploits without the deployment prompt. Also covers Linux/Firefox research uses. Official product/security release under Daybreak-style access. Same capability line as yesterday’s 948 Linux LPEs; listed today as the model launch.