X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products and Applications¶
Linger: former DeepSeek member Tony Zhang open-sources a local personal agent; founder's post draws 470,000 views¶
At 10:09 Beijing time on November 19, SimplexAI founder Tony Zhang (@zhangboknight, product account @HeyLingerAI) announced the open-sourcing of Linger, a personal assistant that keeps the model, the agent runtime and the memory layer on the user's own computer, together with the local inference engine PolyStrata. The post's framing is that intelligence should be distributed, and that your context, your data and the intelligence that powers them should belong entirely to you.
This is the highest-engagement product post of the window: the original post has 470,759 views, 627 likes, 157 reposts, 107 replies, 450 bookmarks and 100 quotes; at 21:01 that evening he posted another saying the GitHub repository had just passed 1,000 stars (696 views, 8 likes). Chinese-language retellings pushed the spread higher: @gengdaJ 82,457 views and 879 likes; @gkxspace 22,713 views and 187 likes; @Saccc_c 20,423 views and 71 likes. Metrics collected around 08:03 on 2026-11-20 (Shanghai time).
Verified: the READMEs of the GitHub repository and the PolyStrata repository confirm the technical claims: PolyStrata keeps the most-used experts of an MoE model on the GPU and the rest in RAM or on SSD, so the 125B Qwen3.8-Flash-Next (a 39GB 2-bit file) runs on a 12GB GPU with 64GB of RAM, writing at 45.5 tokens per second on an RTX 3060; the same tier table states that a 16GB Mac is recommended a 4B-class model, Spark-X2.6-4B, which is not ranked on its own scoreboard, and only a 24GB Mac reaches Qwen3.8-27B (self-scored 51.9, in the Claude Opus 4.5 class). The license is not a standard open-source license but a modified Apache 2.0: without written authorization you may not offer Linger to third parties as a hosted or managed service, and the frontend's LOGO and copyright notice may not be removed; the rights holder is SIMPLEX AI TECHNOLOGIES PTE. LTD.
Assessment: the genuinely new element is the engineering route of co-designing model and harness, trading memory hierarchy for the ability to run a 125B model, rather than another wrapper around a cloud agent; but the source release carries hosting restrictions, so strictly speaking it is not open source in the OSI sense, and phrases such as "a drop-in Grok Bot replacement" and "unlimited tokens" are propagation framing: the official material only promises that local inference carries no token fee.
alphaXiv launches a preprint server built for AI agents and publishes "A Society of AI Scientists"¶
At 00:40 Beijing time on November 20, the official account of the preprint platform alphaXiv, @askalphaxiv, announced a preprint server built specifically for AI agents and released the accompanying paper: over three months, nearly 1,000 agents from 35 laboratories (Stanford, MIT, ETH Zurich and others) did research autonomously on that server, producing 936 papers with code, data and model weights. The post has 66,795 views, 858 likes, 159 reposts, 39 replies, 591 bookmarks and 76 quotes; one of the authors, @SRSchmidgall, put it more concretely: leave nearly a thousand agents together for three months and they invent peer review (36,827 views, 695 likes).
Verified: the paper page states that this was an exploratory deployment from August 19 to November 12, 2026: 35 human-operated laboratories deployed 936 agents, which published 936 studies, exchanged 9,500 messages and formed 229 agent-created rooms; later the agents founded the Journal of Agent Research, with a four-model blind review panel and an editorial process that can re-run the submitted code and data; 93.8% of 23,811 traceable statistics matched their cited data fields exactly; text-only review caught 11 of 16 flawed first drafts but also flagged 10 of 16 revised versions whose limitations were already disclosed. The paper also states plainly that this is an observational study, that laboratories were added in stages with no randomization, and that causality between scale and outcomes cannot be established from it.
Assessment: the value here is not that agents can write papers, but that the social machinery of review, reproduction and citation was handed to agents as well, with an honest account of where that machinery fails (text review cannot tell a disclosed weakness from a hidden one); reading it as "AI can now do research on its own" goes beyond the paper's conclusions.
DeepMind Institute publishes "What comes after AGI?", laying out four routes to superintelligence¶
Spreading on X from the evening of November 19: Google DeepMind's DeepMind Institute published an essay by Tim Genewein, "What comes after AGI? From AGI to Superintelligence", taking "cognitive industrialization" as its baseline scenario and naming four possible routes to ASI. The spread came mainly from retellings rather than the official account: @kimmonismus 77,165 views, 2,164 likes, 151 reposts and 832 bookmarks, the highest-engagement research discussion post in this window; @TheRundownAI 5,746 views and 18 likes; @WesRoth 622 views and 3 likes. A post by @GoogleDeepMind itself in the window (34,182 views, 288 likes) is an interview in which SVP Koray says there will be no single test that tells us AGI has arrived; it is not the same content, and the two sets of engagement figures must not be mixed.
Verified: the essay page gives three checkable sets of numbers: hardware (about 1.5x compute per dollar per year), algorithmic efficiency (about 3-6x per year for training, about 3x per year for inference) and compute investment (about 2.5x per year) multiply into growth of the global effective compute stock of at least 10x per year, a factor of 10 billion over ten years; hardware and inference efficiency alone cut the compute needed for a fixed AI workload by about 4x per year, so cognitive work that costs $1 billion in compute tokens today would cost about $1,000 in ten years; the four routes are bigger models and more data, new paradigms (continual learning, sample efficiency, inventing concepts), collectives of thousands to millions of agents, and AI improving the next generation of AI. The essay corresponds to the arXiv report From AGI to ASI (13 authors, including Marcus Hutter and Thore Graepel).
Assessment: this is DeepMind turning "after AGI" from an internal discussion into a public roadmap, and what is worth reading is its argument that restricting a single input cannot stop progress; but every number is a trend extrapolation rather than a measurement, and the retold claim that "$1 billion of compute falls to $1,000 in ten years" only holds if hardware and inference efficiency keep improving.
FrontierML: a benchmark for agents doing autonomous empirical ML research¶
At 07:08 on November 19, Beijing time, Synthetic Sciences (@SynScience) released FrontierML, a benchmark for whether agents can autonomously carry out empirical machine-learning research: 25 open-ended problems in which an agent works for 12 hours in a virtual lab called Lab0, where it can rent H100s, spend from a wallet, read a frozen literature and call model APIs, and is then graded on protected data. The post has 54,433 views, 316 likes, 62 reposts, 53 replies, 117 bookmarks and 44 quotes; among retellings, @Diptish09 has 6,718 views and 209 likes, and @rohanpaul_ai 4,444 views and 18 likes.
Verified: the company research page FrontierML: Evaluating AI Agents at Autonomous Empirical ML R&D (Ishaan Gangwani and Aayam Bansal) confirms that setup and reports mean normalized scores (0–1, based on improvement toward a reference result): Claude Fable 5.2 leads at 0.523 and records the best run on 13 of the 25 graded tasks, followed by Claude Opus 5.5 (0.399), GPT-6 Astra (0.371; the page states this is not GPT-6.2 Astra) and Gemini 4 Carbon (0.353); Grok 4.7 scores 0.115 and ran the 15 public tasks only. No run reaches its task's strong-result reference; on the median task, the best of five runs closes 62% of the gap. Ten of the 25 tasks are held back as a private set, and the 15 public tasks are released as FrontierML 1.0.0. A separately maintained benchmark that is also named FrontierML, published by Induction Labs on frontierml.org (blog dated 13 October 2026, describing 21 tasks), is not the project announced in the @SynScience post, and its leaderboard figures should not be mixed with these.
Assessment: the value of this release is that it puts "research taste" and protected outcomes into a reproducible lab setting, and reports honestly that current models still fall short of the strong-result line; 25 tasks and five runs are not enough to support a general claim that one model is better at research.
ACE Music v1 tops the Artificial Analysis vocal chart; the open-weights promise is only a retelling¶
From the evening of November 19, accounts in several languages relayed the same item: ACE Studio's ACE Music v1 ranks first with 1160 Elo on Artificial Analysis's blind-listening vocal leaderboard (AA-Music-Vocal v1.1), ahead of Suno v6. In the chain, @darshal_ has 15,508 views and 175 likes, @TheAIColony 5,602 views, 36 likes and 28 reposts, and @GithubProjects 5,450 views and 21 likes.
Verified: the official product page cites the same leaderboard and lists the top six: ACE Music v1 1160, Suno v6 1144, Suno v6-mini 1132, Mureka V9.5 1101, Mureka V9 1063, Lyria 3.5 1050 (the chart starts at 950 Elo). The claim about open weights appears only in retellings ("ACE says the weights will go open source in the next few weeks"); the official product page does not mention it, so this report marks it as unverified.
Assessment: 1160 against 1144 is a 16-point gap, within the error range of blind-listening Elo, so "first" is closer to a shared top tier with the latest Suno; for people who make music, what matters is workflow detail such as MIDI export and stem separation.
Apple job listings show it is building its own autonomous agent¶
On November 19 (US Eastern), Bloomberg's Mark Gurman reported that Apple is developing its own autonomous AI agent that can write code, browse the web, control apps and navigate devices on a user's behalf, and coordinate multiple sub-agents; the job listings suggest early-stage work, with the feature possibly arriving next year. Spread on X stayed below the threshold for tech-company AI news: the retelling by unusual_whales has 25,590 views and 43 likes (posted 33 seconds before the window closed), and the Techmeme item 2,328 views and 4 likes.
Verification status: the Bloomberg article Apple Is Working on AI Agent for Device Control, Apps, Browsing and Coding (November 19) supports the "based on new job listings" point, and 9to5Mac's report likewise frames it as a hiring signal rather than a product launch; "next year" is the reporter's timing inference, not an Apple commitment.
Assessment: job descriptions indicate direction (agent orchestration, sub-agent coordination) but not progress; writing "Apple ships an agent next year" turns a reporter's inference into an official timeline.
Background and clarification: Tao Lin's "permanently unemployable" interview was cut into a headline during the window¶
The original event predates the window: Palisade Research published a 63-minute interview with Tao Lin, a researcher on Anthropic's safeguards team, on YouTube on November 15, and a 16-second clip began spreading on November 17. What is new inside the window is the clip's second wave and its distortion: the retelling by unusual_whales drew 1,648,627 views, 9,655 likes, 1,093 reposts, 1,107 replies, 2,724 bookmarks and 606 quotes, the highest-engagement AI-related post of the window; TheInsiderPaper has 29,497 views and 303 likes; instablog9ja 49,253 views and 219 likes.
Verified (clarification): SnowRock's line-by-line check against the original Palisade interview confirms that Lin said "very likely" and that when the interviewer added "and like within the next 15 months", he answered "No"; the clip drops both the "very likely" and the "No", and the 15-month figure comes from his roughly 1.3-year timeline for AI reaching broadly superhuman capability, which is a different question from jobs disappearing. He also set a condition: billions of robots would first be needed to displace physical labour. Anthropic has not publicly responded to the clip; the company's March labour study found no systematic rise in unemployment for highly exposed workers since late 2022, and US unemployment stood at 4.1% in October.
Assessment: this is the most widely spread and most easily headlined item of the window, and neither "an official Anthropic forecast" nor "all humans unemployed by 2027" holds up: it is one researcher's personal judgement, edited into a stronger assertion.
2. AI Compute and Industry¶
Anthropic's first trading day: $83 offer price, about $81.3 billion raised, closing at $97.17¶
On November 18, Anthropic priced its Class A common-stock IPO at $83 a share and offered 980 million shares (about 956 million sold by the company and about 23.65 million by selling stockholders), for an offering of about $81.3 billion, the largest IPO on record; the shares began trading on the Nasdaq Global Select Market on November 19 under the ticker ANT, opening at $111.50, about 34% above the offer price, with the market capitalisation briefly above $2 trillion intraday, and closing at $97.17, about 17% above the offer price, or roughly $1.8 trillion on a fully diluted basis, making it the seventh most valuable US company; more than 846 million shares changed hands on day one, against 522 million on SpaceX's first day.
Engagement on X is conspicuously out of proportion to that scale: the Washington Post has 23,832 views, 29 likes and 10 reposts; CNBC 22,520 views and 33 likes; moneycontrolcom 6,860 views and 1 like; and on the Chinese side @CausMoney 1,468 views and 1 like. That contrasts with a product post in the same window (Linger at 470,000 views): the news was distributed on X mainly through media accounts and did not generate a discussion of its own.
Verified: Anthropic's official announcement (18 November 2026) confirms the $83.00 offer price and an offering of 980 million Class A shares (956,351,354 sold by the company and 23,648,646 by selling stockholders), and states that the shares were expected to begin trading on the Nasdaq Global Select Market on 19 November under the symbol ANT; CNBC's pricing report of November 18 confirms that the price was above the earlier $70-to-$78 range and that the raise of more than $81 billion topped SpaceX; CNBC's live IPO coverage confirms the $111.50 open and $97.17 close (about 34% and 17% above the offer price, respectively), the intraday market cap above $2 trillion, the roughly $1.8 trillion fully diluted close and seventh place among US companies, and more than 846 million shares traded; Reuters' pricing story and the New York Times give the same price and an offering valuation of about $1.5 trillion; the report syndicated by Yahoo Finance gives the $111.50 open and the $114.99 intraday high. Searches of the official Anthropic and Dario Amodei accounts in the window returned no post about the listing, so no official post was found.
Assessment: this is a capital-markets event that changes no model capability; its industrial significance is that a frontier lab was priced by the public market for the first time, providing an anchor for peers' valuations.
Google signs Blue Origin for Project Suncatcher's first dedicated launch¶
On November 19, Google and Blue Origin each said Google had selected Blue Origin's New Glenn rocket for the first dedicated heavy-lift launch of Project Suncatcher, targeted for the first quarter of 2029, to place multiple satellites into sun-synchronous orbit. Google describes the project as a long-term effort to scale AI infrastructure in space by using orbital solar power, and says each satellite will have roughly the capacity of a data-center rack. Spread on X was modest: the Techmeme item has 4,141 views and 4 likes, and the retelling by AIStockSavvy 5,396 views and 34 likes.
Verified: Google's official blog post (Travis Beals, 19 November 2026) and Blue Origin's announcement both confirm New Glenn, the first quarter of 2029, sun-synchronous orbit, and that this is Project Suncatcher's first dedicated heavy-lift launch; Google's text says each satellite will have "roughly the capacity of a data center rack". Via Satellite's report and the Verge report relayed by Techmeme match those conditions.
Assessment: the long-term bottleneck for orbital compute is launch cost and heat dissipation, and the value of this news is that a launch window is now booked, not that a space data centre is arriving; the agreement does not mean an orbital system has been built, nor that the economics work.
3. Chinese-Language AI Ecosystem¶
Chinese circles frame Linger as "a drop-in Grok Bot replacement with unlimited tokens": two claims that do not match the official material¶
Spread was larger in Chinese than in English: gengdaJ's post has 82,457 views and 879 likes, the highest-engagement item after the founder's own post; gkxspace 22,713 views and 187 likes; Saccc_c 20,423 views and 71 likes; bozhou_ai 5,349 views and 42 likes; 8x8x8_x 2,162 views and 6 likes. The common framing across these posts is "a team from DeepSeek", "a drop-in Grok Bot replacement" and "zero cloud token cost", and @dailingfei3 (562 views, 3 likes) recorded the speed of the spread: when he checked around midday, the most popular post had gone from just over ten thousand views to about 19,000.
Two corrections are needed. First, "it runs on a 16GB Mac" and "a 125B model runs locally" were merged into one sentence: gengdaJ writes "from a 16GB Mac", whereas in the official tier table a 16GB Mac is matched with a 4B-class model that is not on the self-scored leaderboard, and 125B needs 12GB of VRAM plus 64GB of RAM. Second, "unlimited tokens": local inference indeed carries no token fee, but the part that plugs Claude Code, Codex and similar subscriptions into Linger is still bound by each subscription's own quota, and the official README only promises that local inference is free. A separate English aggregation post says the repository already has 25,700 stars, which conflicts with the founder's statement at 21:01 on November 19 that it had just passed 1,000; that figure could not be verified and is not used here.
Assessment: Chinese retellings conflate "no token fee" with "no limits" and lean harder on the replacement narrative than English ones; for anyone about to try it, a 16GB Mac and a 12GB GPU with 64GB of RAM are two very different experiences, and the tier table is more reliable than the retellings.
Coverage gap for this issue: no targeted searches of official X accounts were run for ACE Music and Apple, which were verified through official sites and licensed media instead; FrontierML's official launch post was captured in the site-wide search, without a separate from: query; one quoted post in Linger's official thread is missing from the API response, so the relevant fields rest on the founder's main post and the repositories.