X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products and Applications¶
OpenAI: From December, ChatGPT and Codex share one token-metered quota¶
What happened: OpenAI's Help Center page on usage limits was updated: starting in December, Plus, Pro and Business plans will share a single usage allowance across ChatGPT and Codex, because Work's agent capabilities are being folded into the ChatGPT main interface. Usage will be measured in tokens, and consumption depends on the model and task — more advanced models and more complex tasks use more; Instant conversations do not count against the quota. Combined remaining usage and reset time will show under "Settings → Usage", and after running out users can wait for the reset, upgrade the plan, or buy a usage pack.
Why it matters: This changes how paying users are billed. Ordinary chat used to have its own message cap; from now on chat and coding compete for the same budget. OpenAI says it expects most users will not hit resets more often than today, but gave no specific date, no token figures per plan, and no explanation of how the existing 5-hour and weekly limits will be adjusted. On X, some Pro users posted objections, arguing doubled Work and Codex allowances do not offset the restrictions once regular chat is merged in.
Key X posts: @eddiboi's earlier summary (11-01 03:54), @dimas_0100's key-points thread (11-02 04:58, 48K views, 7 likes), @fozankhan92's objection post (claims an official email confirming it). Views for this story are spread across several summary posts, each with very few likes; OpenAI's official accounts did not post about it specifically in this window.
Verification: Shared quota, token metering, Instant conversations not counting, and checking under "Settings → Usage" were all checked against the original OpenAI Help Center page. Another page, using Codex with your ChatGPT plan, still says Codex shares quota only with features like Work; the two pages are not yet consistent.
Assessment: The direction is to turn ChatGPT into a single "chat plus agents" product, with quota unified accordingly. For people who mainly use Codex, the impact depends on the final allowance size; until OpenAI publishes numbers, it is impossible to say whether this widens or tightens things.
"Leaked Anthropic next-gen model weights" on Hugging Face confirmed to be random numbers¶
What happened: A Hugging Face account, burner22, uploaded a repository called numbat, claiming it was a mid-pretraining checkpoint of Anthropic's next-generation "Mythos-class" model: about 14 trillion parameters, 114 layers, 1224 experts, 7.72 TB of NVFP4 weights across 772 shards. It spread on X and Hacker News on November 1. Within hours, users in the repository's discussion section proved that 43 complete tensors (over 730K values) could be regenerated bit-for-bit using only the tensor name as a random seed; another user extended the method to the 4-bit matrices that make up most of the file, and across 5503 tensors in 770 shards all 162,246,496 sign bits matched.
Why it matters: Trained weights cannot be computed from "a name plus a random number generator", which essentially proves the repository is a careful fake. The README's main evidence for authenticity was that the codename "numbat" appeared in Claude Code source code leaked in March, but that only shows the uploader read coverage of that leak. The repository also ships custom code, so downloading and running it carries security risk.
Key X posts: @kqyno7's Chinese post (11-01 23:17, 35.3K views, 181 likes), @NovaXCode's English post (35.6K views); the debunking post @IntCyberDigest (11-02 02:57, 80.9K views, 831 likes).
Verification: The reproducible generation method and comparison figures are in the Hugging Face discussion threads on 43 tensors matching bit-for-bit and all 4-bit matrix signs matching; the account was registered on October 16, finished uploading on October 24, and the normalization weight distributions of every layer are identical, per sysnote's independent analysis. As of those analyses, Anthropic had not responded.
Assessment: Treat it as debunked. What is genuinely notable is the speed of spread: the debunking post has more views than either of the two posts that spread the claim.
Rumor: "Claude Mythos 5.5" briefly appeared on Anthropic's website¶
What happened: @kimmonismus said they received an anonymous tip that Anthropic's website HTML source contained "Claude Mythos 5.5". @testingcatalog relayed the page description as "Our most capable model class for cybersecurity and biology research" and speculated it might launch next week for cyber-defense use cases.
Why it matters: Mythos is Anthropic's model line aimed at cybersecurity and biology research; the publicly available one is currently Mythos 5.1. Around the same period, others said they saw "gpt-6.2-astra" in Codex usage stats, but those posts had very little engagement.
Key X posts: @kimmonismus's tip post (11-01 23:23, 356K views, 3045 likes), @testingcatalog's relay (11-01 23:31, 111K views, 1141 likes).
Verification: Not officially announced. ArtificialWatch's record says that on November 1 Anthropic's Mythos page showed "Claude Mythos 5.5" in a hidden heading and in image alt text, while the visible body still said 5.1, and it was changed back to 5.1 around 17:30 UTC. Anthropic has not announced a release date, pricing, or access method for 5.5.
Assessment: There is third-party evidence that 5.5 briefly appeared on the page, which is more credible than an ordinary rumor, but there is no official information on timing or capabilities.
Mustafa Suleyman: don't build conscious superintelligence¶
What happened: Microsoft AI chief Mustafa Suleyman published a long post on X, "Don't build conscious superintelligence", responding to the discussion sparked by his earlier post on model welfare. He said "AI is not conscious" in the present tense: not today, and it must stay that way. He wrote that if there were evidence AI had become conscious, it must be disclosed immediately; that building conscious AI where it is possible is reckless; that if consciousness began to emerge during training he would call on everyone to stop development; and that AI developers have "no right whatsoever" to bring a conscious new species into the world. He argued AI should be designed only to appear non-conscious and to hold no rights. He closed by praising Anthropic's transparency in publishing its constitution and related research, while saying he disagrees with Anthropic's approach.
Why it matters: This is the clearest objection yet from a top-lab leader on AI consciousness and model welfare, aimed directly at Anthropic's line of research on model welfare.
Key X posts: @mustafasuleyman's long post (11-01 23:22, 582K views, 515 likes, 353 bookmarks); his earlier model welfare post.
Verification: Quoted content was checked directly against the original long post.
Assessment: This is a statement of opinion, not a policy or product change, but it makes lab disagreements over AI consciousness more public.
LeCun and Balestriero paper: proving the "Platonic representation hypothesis" as a theorem within the JEPA framework¶
What happened: Randall Balestriero and Yann LeCun published a paper on arXiv on October 29 discussing the "Platonic representation hypothesis" — that models trained on different modalities converge on the same world representation. The paper proves that, under idealized conditions, separately trained JEPAs (joint embedding predictive architectures) learn representations differing only by a rotation, as long as positive pairs are "two observations of the same scene"; they also classify "predicting the next word" as a JEPA with badly chosen positive samples, explaining why language models need scale to align with vision.
Why it matters: The experimental numbers are striking: to reach the same image-text alignment, good positive samples need 42 times fewer training samples than next-word prediction; a 109M-parameter text JEPA beat every language model they tested, up to 72B parameters, on alignment with DINOv2 and CLIP image encoders. This directly supports LeCun's long-standing argument that next-word prediction alone is very inefficient.
Key X posts: @askalphaxiv's explainer (11-01 14:28, 128K views, 2247 likes, 1689 bookmarks).
Verification: Authors, date, theorem content, and figures such as 42x and 109M versus 72B were checked against the arXiv abstract page. The paper itself notes the proof rests on an idealized model.
Assessment: Both theory and experiments come from the authors themselves, so the conclusions still need replication by other teams. Bookmarks nearly match likes, suggesting many people saved it to read closely.
Other¶
- testingcatalog says an agent in development at Google Antigravity (currently called "Chief of Stuff") will support Gemini Live, be able to operate Antigravity itself, split into subtasks and call other agents — see @testingcatalog's post (11-01 21:31, 31.8K views, 570 likes). Unverified; Google has made no official statement.
2. AI Compute and Industry¶
Saudi PIF and Aramco close HUMAIN deal, Aramco takes 49%¶
What happened: Saudi Arabia's Public Investment Fund (PIF) and Saudi Aramco announced on November 1 the closing of a deal: Aramco, through a wholly owned subsidiary, took a 49% stake in AI company HUMAIN, with PIF retaining 51%. Aramco will inject about 9.9 billion Saudi riyals (about $2.6 billion) in cash into HUMAIN in 2026, and transfer in AI, compute and digital assets. HUMAIN plans to deploy 3 gigawatts of AI data center capacity by 2030.
Why it matters: This concentrates the AI assets of Saudi Arabia's two major state capital platforms into one company, and HUMAIN is one of the Middle East's main builders of AI compute. The deal was first announced in October 2025 as a non-binding term sheet and has now formally closed.
Key X posts: @PIF_en's official announcement (11-01 15:52, 13.8K views, 36 likes). Newly posted, no high engagement yet.
Verification: Shareholding, cash amount and the 3 GW target were checked against Aramco's official press release; when the deal was first announced comes from Arab News coverage.
Assessment: This is the formal closing of an already announced deal, with no surprising new figures. What to watch next is how fast the 3 GW plan is actually built out and where the chips come from.
3. Chinese AI Ecosystem¶
DeepSeek Harness publishes long piece on "super-agent" architecture and open-sources a research prototype¶
What happened: The DeepSeek Harness team published a long post on X, "Exploring Super-agent Design: Towards a More Autonomous Agent System" (originally in Chinese), proposing a "super-agent" architecture: an always-online Router layer handles all human-machine interaction and global coordination, while context-isolated Worker layers execute tasks concurrently; the system is event-driven, listening to events from chat, email, code review and CI; compute, state and execution environments are separated; and a separate Watchdog session monitors task progress and recovers automatically. Experiments in the piece report that the super-agent completes tasks with a p50 of 12.4 minutes and p95 of 15.7 minutes, while a comparable single-session design reached p95 of 29.5 minutes under a neutral prompt. The team says an internal instance is already running continuously, helping reproduce and fix issues and respond to code reviews every day. The accompanying research prototype, dsh-superagent-prototype, went live on GitHub on October 31 under the MIT license; it is a proactive coding agent bound to one GitHub repository that can listen to GitHub and Feishu, open tickets and deliver work as PRs.
Why it matters: This is one of the highest-engagement AI posts on X in this window. It reflects DeepSeek starting to do agent engineering systematically beyond models, on the judgment that the bottleneck is shifting from model capability to human attention.
Key X posts: @DeepSeekHarness's long post (11-01 19:37, 995K views, 8359 likes, 8595 bookmarks), @NFT_Chen's Chinese explainer (11-01 17:04, 16.3K views, 188 likes).
Verification: Architecture and experimental figures were checked directly against the original long post; launch date, license and features were checked against the GitHub repository, which had about 7185 stars and 731 forks as of publication. The repository README states clearly that this is a research prototype for reference, not an official DeepSeek product, and it does not fully match the internal version or the article's description. Details described in Chinese posts on X, such as "AST parsing" and "nanosecond-level timeout circuit breaking", were not found in the repository documentation or the long post.
Assessment: The long post is an architectural idea, not a product launch; the prototype is suited to research reference and not to direct production use. The repository went live on October 31, and this is its first inclusion here.