X AI Daily¶

Date26 October 2026 · MondayWindow2026-10-25 02:00 to 2026-10-26 07:58CoverageComplete

Translation

English translation of the Chinese original.

1. AI Models, Products & Applications¶

Users report Claude has started blocking REA reverse-engineering tasks, and the vendor has said nothing specific¶

What happened: @Pirat_Nation posted at 10-25 03:01 (424K views, 6476 likes, 291 replies) saying users report that asking Claude to do detailed reverse engineering with REA (Reverse Engineer Anything) triggers safety checks, and that even research already completed gets cut off midway. About six hours later, @0x0SojalSec posted at 10-25 09:41 (69.5K views, 550 likes) saying OpenAI also banned a user who was reverse-engineering their own Samsung TV locally, citing "Cyber Abuse", and questioned OpenAI's Daybreak program for vetted security researchers. @cherry_mx_reds added that REA itself is an open-source MCP and can be run with GLM, DeepSeek and other models; @vartekxx relayed that REA's author later released "REA agents".
Why it matters: REA lets coding agents call tools such as Hopper, Ghidra and IDA locally to inspect binaries, Electron apps and runtime behavior, and it has spread widely among developers recently. If hosted-model vendors add blocking on the server side, the tool is still there, but the default experience of "letting the agent read the code itself" gets severed. The substance of this conflict is where the boundary lies between safety alignment and legitimate reverse engineering (one's own devices, maintenance of one's own software), and right now only the vendors know where that boundary is.
Posts & engagement: Pirat_Nation's post, 0x0SojalSec's post, cherry_mx_reds's addendum, vartekxx's relay.
Verification: The REA repository exists at morluto/rea; its README describes itself as "letting agents inspect everything from app behavior down to the native binary", with analysis running locally, conclusions carrying evidence and limitations, an MIT license, installation via the npm package rea-agents, and a website at rea.tools. Anthropic has no statement about REA; the findable official position is the 10-06 Expanding the Cyber Verification Program: the public Opus 5.5, Fable 5.1 and Sonnet 5.5 "carry conservative cybersecurity safeguards that will block most cybersecurity work", while the program is split into three tiers, of which Defense Access covers defensive tasks such as malware reverse engineering. Community guidance, compiled according to the Claude Help Center, says such checks apply to every request, including web pages and files the model reads, that malware analysis and vulnerability validation may be interrupted, and that security review of one's own code and threat modeling are unaffected (the Help Center text was not read directly; this is a secondhand relay). OpenAI's Daybreak program does exist, officially described as providing models and Codex Security to vetted defenders; "Samsung TV user banned" appears only in X posts and aggregator relays, with no official or first-hand report, and is unverified.
Take: The tool is open source, the model is not, so this kind of restriction ultimately surfaces as "switch models" rather than "switch tools". For teams doing compatibility implementations, firmware analysis or maintenance of old devices, the realistic reading is: treat REA as an analysis framework with a swappable backend rather than binding yourself to one vendor, and keep an eye on vendors' cybersecurity usage policies. As for whether vendors really tightened their policies, and how far, there are only user observations so far, and it needs official documentation or a formal statement.

Kimi apologizes for K3.1 Pro rate limiting and gives every subscriber a one-time 25% monthly quota reset (Follow-up)¶

What happened: The official @Kimi_Moonshot posted at 10-26 04:14 (179K views, 2752 likes, 192 replies, 116 quotes) apologizing for rate limiting and unnecessary quota consumption on K3.1 Pro subscriptions, saying "this is our problem, and we should have fixed it earlier"; that day every subscriber on every plan received a one-time usage reset restoring 25% of the monthly quota (anyone who had used less than 25% was simply reset to zero), equivalent to a week of usage, with usage bars updating before the end of the day Pacific time; plan parameters were also adjusted to reduce unnecessary subagent calls and output length.
Why it matters: The previous daily already covered two disputes after the K3.1 Pro launch (the LiveBench score and unlabeled paid promotion). This time the vendor offers compensation on the subscription side within two days instead of continuing to talk about benchmarks. Handing out the compensation uniformly as "25% of the monthly quota" amounts to admitting that the rate limiting came from the product side rather than from how users were using it; for paying users who had just moved their workflows over, this is the most direct form of damage control.
Posts & engagement: the official apology post; the critic @MiaAI_lab (5588 views, 96 likes) argues the apology does not address the paid-promotion dispute; for Chinese-language spread see @NFT_Chen.
Verification: The content of the official post matches data the bot obtained through the X API, and a Phemex news relay also confirms "resets 25% of the monthly quota". On the paid-promotion dispute, no official response can be found, and @MiaAI_lab's criticism holds within what is publicly visible. "Reduce unnecessary subagent calls and output length" came with no specific figures and cannot be quantified.
Take: The compensation solves quota, not trust. The promotion dispute last issue damaged the community's judgment of how credible K3.1 demo content is, while this apology covers only rate limiting; the two are treated as one thing in the official framing, and that itself will be remembered. What is more worth watching next is the independent testing after the weights drop in November, and whether the plan-parameter changes actually bring down consumption per unit of work.

More clues on GPT-6.1 Terra: Microsoft's EF Core switches its eval judge model to it (Follow-up, rumor)¶

What happened: @LuminaBench posted at 10-25 18:26 (68.9K views, 328 likes) saying a GitHub employee configured GPT-6.1 Terra in GitHub Copilot, positioned as "a cheaper alternative to GPT-6.1 Sol", and that this person had previously posted about Microsoft testing the model. @chetaslua said a change merged by Microsoft's EF Core team on 10-16 switched judge_model from gpt-5.6-terra to gpt-6.1-terra in 18 eval configs. @kimmonismus posted at 10-25 16:47 (69.4K views, 1049 likes) arguing that "Codex's 28 days" would end with a new model release, possibly Terra, while rating Luna as cheap and good, Sol and Astra as the strongest, and Terra as an awkward middle tier. The same day, Codex and ChatGPT lead Tibo posted "Day 20/ Rest and reset" (1.23M views, 11.2K likes), and at 14:06 said "It is all propagated. Enjoy some mad codexing tomorrow. Fun week ahead too." (645K views, 5175 likes).
Why it matters: This is the "hardest" kind of clue so far — not a leaked benchmark score, but a default value in eval configs inside a public Microsoft repository. It shows that at least internally Microsoft can already call the model under this name for evaluation, but whether it will be opened externally, and when, is something the config file itself cannot tell.
Posts & engagement: LuminaBench's post, chetaslua's post, kimmonismus's post, Tibo's Day 20 post and the "It is all propagated" post.
Verification: The diff of dotnet/efcore PR #39217 was read directly; the PR was submitted by AndriySvyryd, merged 2026-10-16 17:45 UTC, and changed the default judge_model in eval configs from gpt-5.6-terra to gpt-6.1-terra: 17 direct replacements, plus 1 new config written directly with the new value, 18 config files in total, with tests also including a case overriding with gpt-6.1-sol. So "18 eval configs" is basically accurate. But OpenAI has made no official release, and there is no model page or pricing page; the Terra tier is currently still GPT-5.6 Terra, gpt-6.1-terra cannot be called, and it remains a rumor. A secondhand relay says The Information reported OpenAI is preparing a new Terra model; the original is paywalled and was not read. Tibo's posts only show that a usage reset happened during the Codex event period and do not prove a new model.
Take: The evidence is strong enough to support the inference that "OpenAI is preparing the Terra tier", but not enough to support "launching tomorrow". The real signal is gpt-6.1-terra appearing on the official model page, or an extra tier on the pricing page; until then, treat this as a launch-timing guess and do not change procurement plans based on it.

More than 1,000 mathematicians join the Association for Human Mathematics, pledging not to publish AI-generated papers or work for frontier labs¶

What happened: @Hesamation posted at 10-25 08:32 (245K views, 641 likes, 275 replies) saying more than 1,000 mathematicians have joined the Association for Human Mathematics (AHM), pledging not to publish AI-generated mathematics papers and not to collaborate with frontier labs, with 654 of them further pledging to use no AI at all in their research.
Why it matters: This is the most concrete collective action against AI in mathematics so far: not an open letter, but an organization with a public roster and tiered pledges. The association splits "not working for AI companies" and "not using AI in research" into two tiers, letting people with different positions all sign and avoiding an "either fully oppose or fully accept" standoff; at the same time it moves the issue from "can models do mathematics" to "what happens to journals, faculty positions and graduate training".
Posts & engagement: Hesamation's post; coverage in New Scientist, with the official site at ahmath.org.
Verification: The original AHM members page states "1008 members, of whom 654 are in the no-AI core group" and lists three pledges: not to provide technical labor, mathematical knowledge, consulting or advocacy to commercial AI companies; not to publish mathematical text generated by AI models (including papers, reviews, lecture notes, teaching materials and books); and, for those joining the no-AI group, not to use AI models in research. The page also notes that the member list has been paused since 10-25 because of site maintenance. New Scientist's report confirms "more than 1,000 people", states explicitly that AHM was founded after OpenAI's one-off release of 722 papers this month, and quotes Terence Tao (who thinks this "dump"-style release may harm mathematics), Sam Raskin (who signed only the first two pledges), Steven Sivek and Hervé Le Dret; OpenAI and Anthropic did not respond to its requests for comment. OpenAI's 722 manuscripts are also supported by a Wikipedia entry and several other reports; OpenAI's original research post was not opened this time.
Take: The roster's practical binding force is limited, but it changes the cost of "who endorses AI mathematics": from now on, a mathematician collaborating with a lab will by default be expected to explain their position. For the labs, the real loss is not 1,000 signatures but the shrinking pool of people willing to publicly review these AI manuscripts — and review is exactly what this batch of results most lacks.

Nanbeige4.3-33B goes open source: 33B dense, 512K context, positioned as a teacher model¶

What happened: @testingcatalog posted at 10-26 01:06 (8933 views, 116 likes) saying this model leads most 30B-class benchmarks, beating Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, with Nanbeige4.3-3B also planned for later release; @alvarobartt also reposted an introduction.
Why it matters: This team did not chase total parameter count but first built a high-capacity "teacher" and then distills it into small models. The model card puts agentic and office tasks such as Terminal-Bench, SWE-Bench Pro and GDPval at the front, which shows the target scenarios are coding and workplace collaboration rather than general conversation. For teams with only a single machine or a few GPUs that still want to run agents, a 30B-class dense model is a realistic landing point.
Posts & engagement: testingcatalog's post, alvarobartt's post, the Hugging Face model card.
Verification: The model card confirms it is positioned as "the flagship teacher model of the Nanbeige4.3 series", supports a 524,288 (512K) context, and compares directly against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B; Terminal-Bench 2.1 is 80.9 versus 73.0, Terminal-Bench 4.0 is 23.8 versus 6.3, but SWE-Bench Pro is 60.3 versus 61.7 (Qwen higher) and GDPval is 84.8 versus 86.4 (Qwen higher). The repository was created 2026-10-24 under the Apache 2.0 license; config.json is LlamaForCausalLM, 52 layers, hidden size 7168, a dense rather than MoE architecture. All of the above benchmarks are self-reported by the team with no third-party reproduction. On team affiliation, the model card's contact email is nanbeige@kanzhun.com, and Chinese reports generally describe Nanbeige ("南北阁") as BOSS Zhipin's in-house model series, though no official page directly says "under BOSS Zhipin".
Take: A 33B dense model beating Qwen3.8-27B on Terminal-Bench while trailing on SWE-Bench Pro and GDPval is a more informative distribution than "leading across the board": what it is good at is long-horizon terminal tasks and multi-turn tool calling, not single-point bug fixing or professional deliverables. It is safer to judge after third-party reproduction.

Other¶

Altman spoke about AI and cybersecurity at Stanford Business School, saying "I no longer think we're gonna be able to just patch all the bugs" and that most organizational leaders "have kind of given up", while models keep getting stronger: Gizmodo's report, with the talk video on YouTube; the same round of sharing also brought in Cloudflare's Matthew Prince telling Reuters that for the next few years "every single week it's going to feel like the sky is falling" (PC Gamer's relay). Spread on X was relatively low; original posts at @theSethian and @ByteEcosystem.
OpenAI's Tibo and Meta's Alexandr Wang traded barbs on X over "Dots versus Muse": Tibo's recruiting post, Wang replying "oops you misspelled muse", Tibo's "Step 1: Quit posting butts…", Wang's "someone should check in on tibo". The spat itself carries no new information; what it reflects is direct competition between the two companies' personal agent products. A separate low-traffic post claiming Tibo has left OpenAI is false; that same day Tibo was still posting usage-reset notices in his Codex & ChatGPT @OpenAI role.

2. AI Compute & Industry¶

Elvis Presley catalog owners sue Anthropic: 274 songs, seeking about $41 million¶

What happened: @PopBase posted at 10-25 21:26 (260K views, 17.3K likes) saying the rights holders of Elvis Presley's music have sued Anthropic, alleging it used more than 250 songs without permission to train Claude; LA Times, Genius, NYMag and other accounts followed, and on the Chinese side @sleepy0x13 relayed The Hollywood Reporter.
Why it matters: This is the second time in a week Anthropic has been sued over music copyright; this time the plaintiff is a rights holder that owns a single artist's catalog and also holds both the music and likeness rights to Elvis. The complaint also uses the fact that "guardrails were in place yet output could still be highly close to the originals" as evidence of willful infringement — which turns the effectiveness of the guardrails themselves into a disputed point, not just whether the training data was lawful.
Posts & engagement: PopBase's post, LA Times's post, Genius's post, NYMag's post, Chinese relay @sleepy0x13.
Verification: Pitchfork says it has seen the complaint: the suit was filed 10-23, the plaintiff is Tennessee-based Authentic Brands Group and several co-rightsholders, it covers 274 songs at about $150,000 each for a total of roughly $41 million, and the complaint's quote reads "built its entire business model based on copyright infringement at a scale never seen before". The Guardian says the case was first reported by Billboard and filed on Friday, that the songs involved include Can't Help Falling in Love, All Shook Up and A Little Less Conversation, and that it is one of the first major suits brought directly by an artist's side against an AI developer. As of the Guardian's publication, Anthropic had not publicly responded to the case; a secondhand relay says the company disagreed with the earlier Reservoir case and will defend itself. The prior case is Reservoir's suit against Anthropic filed 10-21 in California federal court, covering 853 songs and seeking about $151.4 million, with a parallel suit against Suno (1050 songs, about $186.4 million). The court handling the Elvis case is not stated in the reports read.
Take: The amount itself is not the point — $41 million is not an existential threat to Anthropic — the point is that the litigation structure is changing: from "collective rights enforcement" to "precise claims over a single catalog", each case smaller, faster and easier to replicate. If "the guardrails failed" is accepted by a court as evidence of willful infringement, then every model vendor's filter layer turns from compliance decoration into a source of liability.

gpu-unlock: an open-source Linux kernel module unlocks the half-rate FP/BF16 limit on RTX 30-series Tensor Cores¶

What happened: @HowToPrompt__ said on 10-25 10:15 that the 4090 version, gpu-unlock-ada, raises FP16 tensor throughput from 165.6 to 331.6 TF/s; @MinLiBuilds's Chinese post at 10-25 17:00 (4117 views) gives 3090 BF16 going from 75T to 135T, image generation 1.57x, vLLM prefill 33% faster, plus 4090 BF16 1.39x and TF32 2.14x, and explains that LLM decode is basically unchanged and that it suits diffusion models, because the FP32 accumulation of gaming-card Tensor Cores is locked to half rate and a firmware exploit is used to set the registers back before the driver attaches. @TeksEdge also covered it.
Why it matters: For local inference and small teams, the significance of this kind of modification is to bring the effective compute of consumer cards closer to professional cards rather than to buy new hardware. What it bypasses is not a hardware defect but the driver's rate selection, so the gains concentrate in matrix-dense stages such as prefill and diffusion models; decode is limited by memory bandwidth and barely moves. In other words, it improves the experience of "reading long contexts and running large batches", not "tokens per second".
Posts & engagement: MinLiBuilds's post, HowToPrompt__'s post, TeksEdge's post, the repository Zagitta/gpu-unlock.
Verification: The repository exists, currently at version 1.0.0. The README explains that it is a Linux kernel module that writes the rate selection bits (SS0 0x88888888 selects full-rate BF16, SS1 0x8 selects IMLA4) before the unmodified NVIDIA driver binds the GPU and then performs an FLR, so there is no need to re-patch after every driver update; its effect is to double the issue rate of FP/BF16 FMA to FP32 on GA102/GA103/GA104/GA106. The README's figures for 3090 Ti are FP16 87.46 to 118.39 TFLOPS, BF16 87.50 to 157.41, TF32 42.58 to 82.61; in the user-report table, the BF16 GEMM improvement range for 3090 is +25% to +67% (the lowest value comes from a card power-limited at 225W), 3080 is +70% to +73%, 3070 is +73% to +81%, LLM prefill is +14% to +36%, decode and INT8 are basically unchanged, and reports that compared outputs found them bit-identical before and after. The module is inactive by default and requires adding gpu_unlock.targets explicitly on the kernel command line. So the post's "75T to 135T" and "image generation 1.57x" are personal self-tests that do not fully match the README's user-report ranges; the 4090 is a separate repository, gpu-unlock-ada, whose README was not checked this time; the three points "no VBIOS flashing, recoverable by cutting power, Secure Boot must be disabled" were not confirmed in the README. All performance data are the author's and users' own tests.
Take: This is a typical community patch: useful, fragile, and liable to be closed off at any time by a driver update or firmware signature. Its real value is not saving the money for a professional card but reminding people that part of the gap between consumer and professional tiers is a deliberately chosen rate setting. If you want to use it, first measure the prefill gain on your own model and batch, then decide whether it is worth disabling Secure Boot for it and risking that a single driver update renders it useless.

Other¶

Apollo chief economist Torsten Slok's Daily Spark of 10-24 puts it bluntly: No Signs of AI Payoff Outside the Tech Sector, "revenue per employee is rising for the Magnificent 7 and flat for the S&P 493". Relay at @MilkRoadAI; the specific revenue-per-employee figures (about $780,000 for the Magnificent 7 and about $525,000 for the other 493) come from a secondhand summary and were not checked against Apollo's original page.

3. Chinese AI Ecosystem¶

Analysys Q3 report: in September desktop office agents, Tencent WorkBuddy led with 85.34 million monthly visits¶

What happened: @NFT_Chen posted at 10-25 23:48 (3572 views) citing Analysys September data: WorkBuddy (Tencent) 85.34 million, Qwen Office (Alibaba) 46.55 million, Doubao Work (ByteDance) 30.39 million, TRAE (ByteDance) 27.12 million, CodeBuddy (Tencent) 25.88 million, and said Baidu did not make the top ten.
Why it matters: This is one of the few rankings counted by desktop visits, a different methodology from QuestMobile adding up monthly actives across all platforms; neither can be read directly as a revenue ranking, but both point to the same thing: office agents already have a clear leadership group, and the top three together account for 53.1% of visits across the 20 ranked products. For the big platforms, what comes next is not model scores but who can turn individual trials into team seats.
Posts & engagement: NFT_Chen's post.
Verification: Analysys published its "China Desktop AI-Native Office Agent Report" (Q3 2026) on 10-22; the 20 products counted in September had combined monthly visits above 300 million, nearly 5 times the June level; WorkBuddy, Qwen Office and Doubao Work were 85.34 million, 46.55 million and 30.39 million respectively, together 53.1% of visits across the 20 products, matching the post's top three figures (Sina Tech's relay, another analysis). But the same relay explicitly states that on the Analysys ranking Baidu Dazi was eighth with 9.26 million visits and Kuku AI ninth with 8.36 million, so "Baidu did not make the top ten" is inconsistent with that report; the two figures TRAE 27.12 million and CodeBuddy 25.88 million do not appear in the relays found and come from the post itself. The ranking counts desktop visits, which is not the same as enterprise orders or revenue.
Take: The ranking shows that Tencent's advantage at the enterprise collaboration entry point has indeed turned into usage, but it also exposes its ceiling: however fast visits grow, the outcome ultimately lands on seats and renewals, and that is something a ranking cannot measure.

Big tech offers close to one million yuan a year for "forward deployed engineers"¶

What happened: @choblin29's English post at 10-25 20:48 (2480 views) cited the South China Morning Post saying ByteDance, Tencent and Alibaba offer close to 1 million yuan a year for forward deployed engineers (FDEs) who put AI directly into customer enterprises; related job postings grew 1523% year on year; it also listed Kimi at 1.09 million yuan in annual package and ByteDance Doubao at 1.05 million, said some ByteDance engineers can earn 10,000 yuan a day for on-site work, that MIIT is promoting the role, and that Tencent has launched a dedicated certification.
Why it matters: FDE is a role that first became popular in Silicon Valley; the job is to connect models into a customer's business processes, not to train models. Rising pay shows that big tech's AI revenue increasingly depends on delivery rather than API call volume, which is the same logic as office agents competing for the enterprise entry point. For individuals, the bar for this kind of role is engineering ability plus business understanding, not papers.
Posts & engagement: choblin29's post, the SCMP original.
Verification: The readable portion of the SCMP original confirms three points: annual pay "close to 1 million yuan (about $149,700)", the involvement of ByteDance, Tencent and Alibaba, and a 1523% year-on-year increase in FDE postings in the first seven months of this year (data from Maimai). But the body is behind a paywall, so the details of Kimi at 1.09 million yuan, Doubao at 1.05 million, the 10,000 yuan daily on-site rate, MIIT's promotion and Tencent's certification could not be read and are relayed from the post. Among secondhand relays, the visible account is that ByteDance listed 35,000 to 65,000 yuan a month on BOSS Zhipin, about 975,000 yuan at 15 months' pay (FourWeekMBA's relay), while another summary says some management roles reach about 1.4 million yuan a year. The order of magnitude is consistent, but the specific figures differ from one another.
Take: "Close to a million" is more likely a figure obtained by multiplying monthly pay by 15 or 16 months rather than a fixed annual package; 1523% is a year-on-year growth rate with an undisclosed base, so it should not be read directly as a "job boom". What is really worth noting is the change in the shape of the role: models' marginal capabilities are getting harder to sell at a premium, while people who can land them in customer systems are scarce.

Other¶

Chinese-language circles also relayed a fair amount about Kimi's quota apology and the Tibo–Alexandr Wang spat; the content is in Section 1, with no additional new information.