X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products & Applications¶
OpenAI Announces Three AI-Produced Scientific Findings in Mathematics, Computer Science and Physics¶
What happened: On 2026-10-26 OpenAI posted that its internal models produced three findings: first, a proof of the local smoothing conjecture in all dimensions, from which results related to Kakeya and Stein restriction follow; second, a refutation of the Exponential Time Hypothesis (ETH), together with a Lean formalization; third, the existence of an interacting quantum field theory in four dimensions including gauge fields and fermions (including QCD), with subproblems touching on pure Yang–Mills existence and thus related to the Millennium Problem. OpenAI calls this batch of material "raw research data" and has invited outside scholars to serve as shepherds; the computer science group has published the natural-language proof and Lean code, while the physics group is not yet releasing the AI outputs.
Why it matters: This pushes the craft from "AI-assisted derivation" to "AI produces results on open problems, which domain scholars then take up and verify," and the verifier for the ETH item includes the person who originally posed the problem. The spread of this item was the largest of any event that day.
Original post & engagement: @gdb, posted around 05:50 Shanghai time on 2026-10-27, about 234,000 views, 4555 likes, 226 replies, 190 quotes, 1055 bookmarks (the retweet field was not returned by the API).
Verification: OpenAI official blog (2026-10-26). Shepherds on the CS side include Impagliazzo, Paturi and Williams; on the physics side, David Gross and others. The conclusions still await independent verification by the academic community; Lean only verifies the formalized statement itself.
Assessment: Major, but it should wait for the outside shepherds' conclusions to settle, and should not be written up as "the problem has been definitively solved."
GPT-6.1 Terra Goes Live: Focused on Visual Reasoning and Computer Use, Price Halved (Follow-up)¶
What happened: OpenAI officially released GPT-6.1 Terra, emphasizing visual reasoning, computer use and more natural writing. Standard API pricing is 50% lower than GPT-5.6 Terra / GPT-6.1 Sol: $1 per million input tokens, $0.05 for cached input, $5 for output. OpenAI says that at the medium reasoning setting on Agents' Last Exam it scores 15 points above 5.6 Terra at roughly one third of the cost; on Terminal-Bench 4.0 it reaches 50.3% at the maximum reasoning setting (5.6 Terra: 21.5%); and the internal hallucination-related error rate drops from 53.6% to 16.7% at the low reasoning setting. The model is available in ChatGPT Work and Codex (Plus/Pro/Business/Enterprise/Edu, subject to region and workspace settings); the API name is gpt-6.1-terra, it does not appear on the regular Chat page, and it does not support UAE inference residency.
Why it matters: Yesterday it was only a name in configuration clues; today it is a callable product. It is positioned as the high-value computer-use tier rather than a replacement for top tiers like Astra / Sol.
Original post & engagement: @OpenAI, about 2,270,000 views, 19,800 likes, 1648 reposts, 1028 replies, 1399 quotes, 3074 bookmarks.
Verification: Introducing GPT-6.1 Terra; ChatGPT release notes.
Assessment: It turns rumor into one slot in the product line; the pricing and Terminal-Bench figures are the publisher's own numbers, and the real-world feel still needs validation in use.
Mirendil Releases Fern 0.1: Claimed Context of Up to 10 Million Tokens¶
What happened: Mirendil, made up of members with former Anthropic / Google backgrounds, released its first model, Fern. Fern 0.1 is built on Mirendil-Base with short post-training, with context of up to 10 million tokens, roughly equivalent to about 23MB of code or about 500,000 lines (Megatron-LM scale). It is positioned for the dirty work of AI R&D: auditing training data, debugging from logs, reading entire agent traces. The larger Base-Large / XL are still in training, and access is labeled "coming soon."
Why it matters: Ultra-long context is one of the hardest narratives right now, and offering ten-million-token scale in one go while aiming squarely at R&D workflows clearly distinguishes it from general chat models.
Original post & engagement: @bneyshabur, about 117,000 views, 1576 likes; @mirendil, about 30,000 views, 549 likes.
Verification: Introducing Fern. All benchmarks are vendor self-tests and await third-party reproduction.
Assessment: Strong narrative, but the product is not yet publicly accessible, and the truly usable context and effective recall still need verification.
Cursor Agent SDK Generally Available¶
What happened: Cursor announced that the Agent SDK is available: define agents in code, write prompts in Markdown, write tools in TypeScript, hosted or self-hosted, with built-in evals. On the same day Cursor also announced that DeepSeek V4.1 Flash is coming to Cursor.
Why it matters: Coding agents move from "used inside a chat box" to "versioned, evaluated and hosted inside the codebase" — a step along the engineering path.
Original post & engagement: @cursor_ai, about 176,000 views, 2422 likes, 144 reposts, 225 replies, 127 quotes, 1262 bookmarks.
Verification: The party's own post; details are subject to the official docs (the docs were not fetched item by item in this issue).
Assessment: Clear direction; the quality of execution depends on whether the SDK's eval and self-hosting capabilities are as described.
Replit Proposes Agent Habitat and the "Self-Driving Company"¶
What happened: Replit CEO Amjad Masad published "The Rise of the Self-Driving Company," proposing Agent Habitat: giving agents controlled access to company knowledge and systems, clear approval boundaries and auditable ways of collaborating. Marc Andreessen and a16z published a long post in echo the same day, arguing for letting chaos loose first and then reining that chaos into the habitat.
Why it matters: This is a product concept and narrative battle that pushes agents from point tools toward organizational form, and it also explains why "permissions, approvals, auditing" are becoming the keywords of the next stage.
Original post & engagement: @amasad, about 58,000 views, 838 likes, 966 bookmarks; @pmarca, about 201,000 views, 1200 likes (links to the a16z long post); @a16z, about 23,000 views.
Verification: Replit blog; a16z article. The customer efficiency figures in the text are the party's own claims.
Assessment: Concept and narrative come first; this is not a new model release. Real usability depends on how the Habitat permission model is implemented.
Riverside Launches Agents: An AI Team for Video Production¶
What happened: Riverside's CEO launched four agents: Producer, Editor, Clipper and Marketer, covering the production and distribution stages of video.
Why it matters: Breaking "one model" into "a pipeline of job roles" in a vertical scenario is the most typical application-layer form this year.
Original post & engagement: @NadavKeyson, about 132,000 views, 282 likes, 108 replies, 92 quotes.
Verification: The party's own announcement; the scope of capability is subject to the official demo, and this issue has not independently tested it.
Assessment: Engagement is above threshold and this is a noteworthy application deployment, but there is still a gap between demos and actual output.
2. AI Compute & Industry¶
Anthropic Files S-1, Plans Nasdaq Listing Under Ticker ANT¶
What happened: On 2026-10-26 Anthropic announced it has submitted Form S-1 to the SEC, intending to list Class A common stock on the Nasdaq Global Select Market under the ticker ANT; the number of shares offered and the price range have not yet been determined, and the statement makes clear the filing is not yet effective. The S-1 disclosure table (in millions of USD) shows: revenue of 16,286 for the first half of 2026 (about $16.29 billion), versus 1,163 for the first half of 2025; operating loss of 693; net loss of 80,748, of which about 72,388 is a non-cash item from changes in fair value; compute and infrastructure spending of 8,945. In addition, media have relayed a very large long-term compute commitment, which is "media relay" rather than a confirmed item in the announcement.
Why it matters: This is a milestone as a leading AI lab moves toward the public markets, and it also lets outsiders see, for the first time in prospectus terms, the relative scale of revenue, losses and compute spending.
Original post & engagement: @AnthropicAI, about 5,150,000 views, 15,400 likes, 1924 reposts, 1055 replies, 1936 quotes, 2526 bookmarks. Third-party spread such as @MarioNawfal, about 41,000 views.
Verification: Anthropic announcement; S-1 on SEC EDGAR.
Assessment: Figures are taken from the S-1; valuation rumors are not accepted. Non-cash fair-value changes should not be read as operating cash burn.
Jensen Huang's Example: NVIDIA Software for Running Agents on an AI Factory Is About $2,000/GPU/Year¶
What happened: At GTC Berlin, Jensen Huang gave the example that the NVIDIA software needed to run agents on an AI factory is roughly $2,000 per GPU per year. This is not a confirmed price list; existing AI Enterprise list pricing is about $4,500 per GPU per year, and he did not explain the relationship between the two.
Why it matters: Layering a "software subscription" on top of the compute business is a continuing shift in NVIDIA's narrative; once the market treats the number 2000 as pricing, it will directly affect the unit economics of agents.
Original post & engagement: Spread post @Pirat_Nation, about 184,000 views, 2003 likes, 233 replies. Media follow-ups include Tom's Hardware, TechSpot and others.
Verification: The on-stage remark has been relayed by multiple media outlets; no official price list has been updated in step.
Assessment: The number is an "example" rather than an announced price, and should be read as a directional signal, not quoted as a price.
3. Chinese AI Ecosystem¶
Poor Box Office for "Sanxingdui: Future Past," the First Domestic AI Theatrical Film, Sparks Discussion¶
What happened: Chinese-language social discussion says the film, released on 10-23, took in roughly 1.25 million RMB at the box office over about four days, with a rumored production budget of over 10 million; the discussion compares it with the high box office of low-cost animated films such as "Niu Lai."
Why it matters: This is the first time "full AI pipeline production + theatrical distribution" in China has been examined under real box-office terms, and the result directly undercuts the earlier imagination that "AI cost reduction equals commercial success."
Original post & engagement: @pianhangx, about 16,000 views, with a relatively low like count; other related Chinese-language posts have even lower engagement.
Verification: Both box office and production cost come from social media relay; no authoritative box-office figure was confirmed in this issue, so this is marked as pending verification.
Assessment: Usable as a record of a phenomenon in the Chinese ecosystem, but not to be cited as precise box-office data.