Daily editorial briefing

№ 20261025

Anthropic files its prospectus as AI finds bugs faster than humans patch them

The day's most important item is Anthropic's prospectus: for the first time an AI company has to put its revenue, its losses and its compute contracts into a public document. A…

The day’s most important item is Anthropic’s prospectus: for the first time an AI company has to put its revenue, its losses and its compute contracts into a public document. A colder thread runs through the same batch of signals — the vulnerabilities that AI helped surface over the past week are now being exploited, and the patching still depends on people racking servers. Open weights, hardware portability and model capability all kept moving; the gap between “looks stronger” and “holds up in use” did not close.

1. Anthropic files: an AI company’s books become readable

According to The Information, Anthropic could publicly file its S-1 as early as Monday, October 26, with a roadshow starting November 9, pricing the following week and a listing in mid-November. The raise is targeted at up to roughly $100 billion against a valuation of about $2 trillion, led by Morgan Stanley and Goldman Sachs. As of the morning of October 26 Beijing time, the filing was not yet visible in the SEC’s EDGAR system; last week’s talk of a filing “as early as the 23rd” also failed to materialize.

The numbers that have already leaked matter more. Per a relatively recent confidential draft seen by The Wall Street Journal, revenue for the first half of 2026 was $16.3 billion, against $1.1 billion a year earlier. Citing FT and Reuters, the company told potential investors at its San Francisco headquarters on October 14 that third-quarter revenue was about $16.8 billion, up roughly 45% from $11.6 billion in the second quarter — sequential growth slowing from about 147% to 45%, even though the single quarter now exceeds the entire first half.

The loss figure needs unpacking. The same draft shows a first-half GAAP net loss of $80.7 billion, of which about $72 billion comes from accounting adjustments on convertible instruments and similar financing, with no corresponding cash outflow. The company tells investors it has been profitable for two consecutive quarters on its own metric, without saying which costs that metric excludes.

What really shapes the long term is cloud dependence. Per a confidential draft seen by Reuters in late September, about 47% of 2025 sales — roughly $2.16 billion — were sold through the Amazon and Google Cloud marketplaces, which took about $351 million, or 16 cents on every dollar. Two unnamed customers each accounted for 12% of 2025 revenue. The draft lists roughly $450 billion in cloud compute commitments over the next decade: $111 billion to Google Cloud, $110 billion to AWS and $31 billion to Microsoft. The cloud partners are simultaneously channel, shareholder and supplier.

The timing is pointed. This is also earnings week: Microsoft and Alphabet report Wednesday, Amazon Thursday, and all three hold compute contracts with Anthropic — Amazon and Google are also major investors. Once the prospectus is public, the market can for the first time hold one set of numbers against three cloud businesses.

All of these figures come from confidential drafts seen by media and a single investor meeting, not from a public prospectus. The accounting basis may change.

Sources:

2. The week’s security ledger: discovery accelerated, patching did not

The two PHP vulnerabilities summarized by AI Valley are the hardest example. Called php2shell, they let an attacker execute code on affected servers without logging in, and they had sat in PHP for roughly 14 years. Palo Alto Networks identified 22.6 million potentially vulnerable instances, and Cloudflare reportedly saw millions of exploitation attempts within hours of the news leaking. OpenAI says GPT-5.6 Sol helped turn the suspicious behavior into a working exploit. PHP has shipped fixes; there is no complete workaround, so upgrading is the only option.

The patching side runs on money and people. OpenAI announced $1 billion for Patch the Planet to help organizations audit their systems. But finding bugs faster does not mean fixing them faster: getting tens of millions of servers updated remains a stubbornly human problem.

The same day brought a Tailscale SSH privilege issue. According to a Tailworm report, an attacker who controls one device on the tailnet may be able to reach other machines that run Tailscale SSH and allow the relevant forwarding along already-permitted SSH connections. One researcher published 11 vulnerabilities without notifying the vendor privately first. The company has responded and is fixing it; those who do not want to disable Tailscale can set TS_SSH_DISABLE_FORWARDING=1 on the target machine and restart, at the cost of dropping current sessions, so keep a backup way in.

Consumer hardware looks just as bad. A researcher who rooted a Samsung TV found an always-on microphone; the person who amplified it used “why smart TVs are so cheap” as the explanation, which places data collection inside the hardware cost model.

Trust is eroding too. At GEEKCON 2026 in Shanghai, the drone attack-and-defense demo was halted by police, and contestants and judges were taken in for questioning; the complaint was “inciting others to crack drones.” The event has explicit responsible-disclosure rules, and security expert Tan Xiaosheng and founder Wang Qi both objected publicly. Who filed the report remains unconfirmed. On the other side, a user reports that Claude simply refused and “couldn’t continue” on video-editing work that involved cybersecurity content.

Put these together: AI is speeding up the hunt for bugs, while security practitioners can be blocked by a model for describing a vulnerability and reported to police by a vendor. Patching capacity has not scaled with discovery capacity. That is the most concrete risk in this week’s record.

Sources:

3. Kimi K3.1 Pro: 2.8T open weights and a marketing accident

Kimi shipped a 3T-class open-weight model, K3.1 Pro: 2.8T total parameters with 104B active in a mixture-of-experts layout and a 1M-token context window. Marketing copy says it can write whole game projects, run Unreal and Unity, and handle Blender and CAD modeling. Those specifications come from an aggregated digest — there is no primary model card in the archive, and the same is true of the claim that pricing is high.

The marketing drew more attention than the parameters. A blogger published what appeared to be an official placement brief that, in this dispute, asked recipients not to label the post as paid partnership and to praise the model in specified terms. The blogger refused on the spot; the story reached Reddit and even prompted questions about FTC rules on deceptive advertising. Other Chinese writers then questioned the “whistleblower” — their follower-to-following ratio and posting history — arguing that every link in the chain looked strange.

What the company actually addressed afterwards was something else: it admitted that abnormal quota consumption was its own fault and returned 25% of the monthly quota to all users in one go, roughly a week’s worth of usage. Both the response time and the apology came after days of user complaints.

One capability data point got lost. Anthropic co-founder Jack Clark says Chinese models trail the West by six to twelve months; one writer cites a cybersecurity evaluation that AISI and CAISI ran when Kimi K3 launched: given a 32-step simulated attack path, Kimi K3 advanced on average to step 17, while the strongest US model of the period reached step 28.5. That comparison is second-hand and the original report is not in the archive, but it at least shows a gap between answering questions and sustaining long-horizon autonomous work.

Sources:

4. Google moves PyTorch onto TPUs

At its PyTorch conference, Google open-sourced TorchTPU, which lets PyTorch run natively on TPUs. According to the aggregated digest, the stack needs only one parameter to deploy on Hugging Face, already covers 4,917 models, and reaches 97% parity with GPUs.

Those two numbers matter for migration cost, not for a performance leaderboard. CUDA’s stickiness over the years came from “change the hardware and you rewrite the code”; once the same code runs on two kinds of accelerator, bargaining power shifts back to the buyer.

A supporting signal is MiniMax’s official post. Among new contributions to its H3 ecosystem it names the TorchTPU team alongside PTD, DMAD, SparkH3 and FlashML, which suggests the stack is wired into real training and inference workflows rather than a demo.

The 97% parity and the 4,917 models both come from a summary. There is no first-party Google announcement in the archive, and nothing says which 3% of operators are missing. The direction is clear; the exact figures wait on primary material.

Sources:

5. Block’s Dessa: a speech model that skips text

Block, the company co-founded by Twitter co-founder Jack Dorsey, released a speech model called Dessa: trained from scratch on pure human speech, with no text and no text pretraining, end-to-end speech-to-speech, able to listen and speak at once and to be interrupted at any point.

The engineering numbers the company gave are specific: the whole training run used a single NVIDIA GB200 NVL72, consumed several hundred times fewer tokens than a large language model, and runs offline on an M4 Max MacBook faster than real time. That is a different route from the three-stage “transcribe, reason, synthesize” pipeline behind current voice assistants, and skipping the text bottleneck keeps sighs, hesitations and laughter from being flattened.

Treat it as a preview, though. It is a small research preview from an internal Block lab, and the public demos were chosen by the company. How long a purely audio-trained model holds up on complex multi-step reasoning without a text knowledge base is unknown. Chinese commentary claiming “the AI architecture has been completely destroyed” is the author’s mood, not evidence.

Sources:

6. Agent infrastructure is converging: less orchestration, more constraint

Agensh, a multi-agent framework open-sourced by Microsoft Research, took a counterintuitive direction: it removes the central scheduler so that each agent grabs work and merges code on its own. The same codebase scales from one agent to 1,024, and ProgramBench pass rates rose 9.47 percentage points.

LangChain’s “company harness,” distilled from three internal agents in GTM, engineering and marketing, lands somewhere completely different: entry inside the Slack workspace people already use; identity split per tool between acting on behalf of the invoking user (OBO) and a fixed service account; memory in two layers, global and per-user; code always executed inside a sandbox that reaches authenticated services through a secret proxy; and many runs started by schedules or events rather than a person.

Thorsten Ball of AmpCode gives the same conclusion another shape: putting an agent into “the slot where the text editor used to be” does not change output. What multiplies returns is agent × codebase readability × process — letting the agent cover build, release, CI, production logs, feature flags and monitoring. His team’s preparation is roughly 15 AGENTS.md files plus skills so agents can see, start and seed the system.

The Claude blog is even more restrained: when building agents, folding sub-agents back into the main loop and using progressive disclosure through Skills often beats one sub-agent per team plus a top-level orchestrator, which tends to copy the org chart and brings context loss, latency and bloated prompts.

The interface layer is shifting too. A developer who has long worked in the terminal argues the CLI is no longer the ideal place to work with agents; interfaces like Grok Bot and the Codex app make it easier to scale parallel work, collaborate and manage the lifecycle, and his test has become “don’t use an interface that caps the model.” A shorter formulation from another practitioner: rather than hand-holding the method, supply the development workflow, toolchain and materials, and let people speak only at key nodes, choices and constraints.

Five paths point the same way: the orchestration layer thins out while engineering constraints thicken. But Agensh’s pass rate and LangChain’s division of labor are both reported second-hand, with no repository or documentation link to check.

Sources:

7. The other side of compute: sold-out capacity, squeezed middlemen

AWS says AI cloud capacity is booked into 2027, with 2028 slots filling quickly. At the same moment, Musk said publicly that the AI bubble has already popped and that the remaining 200-plus neoclouds face “guppy-level” mass death, because debt markets have shut out compute companies without stable cash flow outside AI.

The two are not contradictory. Sold-out capacity at the giants means demand is still there; what is being squeezed out is the middle layer that borrowed at high interest to buy GPUs and rent them by the slice. His own comparison is that SpaceX just closed a $40 billion GPU-related debt financing, on the strength of an investment-grade rating plus Starlink and launch as cash engines.

Chinese-language commentary fills in the mechanism: capital once treated GPUs as collateral, but as NVIDIA’s architectures iterate faster, residual value of older cards shrinks and default rates among smaller model customers rise, so lenders now demand stable operating cash flow beyond AI. Some investors already describe the outcome for 200-plus neoclouds as guppy-level death and consolidation; others watch CoreWeave’s share price as a thermometer.

The macro warning comes from Krugman: if AI spending is a bubble and it bursts, the shock would exceed the dot-com bust of the late 1990s, while the Fed now has only 4% of rate cuts of room and could end up in a liquidity trap. That is a comment Gary Marcus relayed, not data.

The most useful part for ordinary readers is not who dies but the change in criteria: for two years the question was whether you could get cards; next it is the balance sheet and real operating cash flow.

Sources:

8. Capability rises, reliability does not

A DeepMind researcher points to a numerical precision problem in the softmax implementations of PyTorch, JAX and TensorFlow: during training gradients suddenly inflate, loss spikes and learned capability degrades. The fix is to change how the computation is done; afterwards accuracy went from 94.3% back to 100%. Bugs of this kind raise no error and only make the model quietly worse, which makes them among the hardest to find.

The capability side offers three handsome cases. Google DeepMind and several universities tested agents on research-grade physics problems drawn from experts’ unpublished work; the best agent solved 49 of them at under $3 per problem, and GPT-6-Astra solved 49 with web search and 46 without. A physics professor fed nearly 500 pages of quantum gravity derivations — a year of human hand calculation — to Claude Opus 5.5, which finished in ten hours, produced 183 pages of derivation notes and caught several sign and constant errors in the paper.

The same material also holds counter-evidence. A DeepMind reinforcement-learning paper on moral dilemmas used real dilemmas from classic novels, each checked by a logic solver to confirm the conflict was genuine, and trained a policy that rates pairs of its own arguments in a self-judged tournament; its ratings agreed with human annotators. The result: answers became more stable when options were reordered, yet a user pushing back or reframing the question still changed the verdict. Order invariance — an easy thing to measure — improved while the underlying judgment did not.

The direct lesson for training: if self-reward or consistency metrics are used as a training signal, robustness under pressure has to be measured separately, or the only thing optimized is formal stability.

Set the silent softmax regression next to the high physics score and the “876x” paper reproduction and only one conclusion holds: a single benchmark or a single case demonstrates a ceiling, not reliability. When using such numbers, it is worth also asking whether they survive a different batch of inputs.

Sources:

9. Three moves on the hardware side: unlocking, assets, demos

Start with the hobbyist modding. One writer says the community has released a method that removes part of the Tensor Core compute limit at a low level: RTX 4090 from 165.6 to 331.6 TFLOPS, RTX 3090 Ti BF16 from 87.5 to 157.4, and RTX 2080 Ti FP16 throughput nearly doubling in specific workloads. Per the post, no new hardware, no BIOS flash and no modified official driver are needed — on Linux, install a kernel module, set boot parameters and trigger it on a cold start; someone has reproduced it on Windows 11 with a modified RTX 4080.

Treat this as an unverified community result: it is one post plus a few community repositories, with no independent replication and no power, stability or precision comparisons. If it holds, the biggest effect is on the effective compute pricing of used cards — depreciation models built around the original product segmentation have to be redone.

Two other items point the same way. NVIDIA released 2,575 SimReady 3D assets on Hugging Face, 472GB in total, each carrying weight, material and center of mass, ready for robot grasping and digital-twin training. AgiBot published video of 110 robots performing in sync, formations tidy, which looks more like a demonstration of cluster scheduling. Trainable assets and schedulable fleets each move a step forward, but a successful demo is not volume production.

Sources:

10. Skills are turning from prompts into deliverables

The clearest case is leanrea: an 8KB SKILL.md replaces an entire reverse-engineering MCP setup, skipping 200-plus tools and teaching the agent to use native commands like otool, jadx and Ghidra directly. The README’s measurement claims 38% faster, 35% cheaper and 60% fewer tokens at equal accuracy, with more than 2,300 stars in its first week.

A whole batch of projects is packaging one job into a reusable unit. brag lets an agent produce a 20-second film with one command after a project finishes, complete with music, motion and share copy; OpenWiki turns an entire codebase into an interlinked Markdown library where every claim traces back to source, with more than 17,000 stars; JetBrains’ open-source ThinkRail is essentially an IDE shell around Pi, whose core design is a single SPEC.md file recording project state.

Process management gets more granular. pi-subagent runs each sub-agent in its own tmux session so they can investigate in parallel and review diffs; from the parent session, /subagent drops you into that terminal to watch, and you can append instructions while a task runs.

One design pointing the other way is worth noting: for content-reading tools built for agents such as x-reader, the difference is not how much they read but that they report what they actually read — a search snippet is not the article, a tweet’s caption is not its video, a missing layer is labeled “partial”, and failure is stated as failure.

Together these show where competition is moving: from “can the agent do it” to “can this capability be reused and audited by someone else.” Beating 200 tools with 8KB is not about a stronger model; it is about more precise packaging.

Sources:

High-value briefs

  • Nanbeige4.3-33B: a 33B dense model from BOSS Zhipin’s Nanbeige lab, tuned for agent capability, scoring above Qwen3.8-27B on agent tests and weaker on coding. The architecture returns to standard Llama full attention with GQA, and the 512K context is achieved by setting rope_theta straight to 120 million; the company positions it as a teacher model. https://x.com/karminski3/status/2114545217407963524
  • 11 additions to the MiniMax H3 ecosystem: 8-step reference-conditioned video plus audio, 4-step distillation, sparse attention, reference-token compression, and serving and workflow integrations across GPU servers, TPUs and Apple Silicon. https://x.com/MiniMax_AI/status/2114504329499660419
  • Grok Bot 0.77 pre-release: adds calls on your behalf, autonomous payments, iMessage access, tailnet access after admin approval, and bot sharing — alongside a vault that can store logins and card details. Capability and authorization boundaries expand together. https://x.com/xiaohu/status/2114550091667386736
  • Initia bridge exploited: attacked on October 23, the USDC bridge reserve drained to $63, withdrawals stalled, roughly $3 million in real assets out against $7 million of unbacked balances. https://x.com/ohxiyu/status/2114500203931988037
  • Coding market share reverses: a leak claims Rippling’s coding usage went from 57% Anthropic and 16% OpenAI in May to 52% OpenAI and 18% Anthropic now, with cost as the stated reason and Replit switching after discounts above 30%.
  • Armin Ronacher’s three-platform poll: to the same question — do you still hand-write more than 20% of your code — 12.7% of 4,326 voters on X said yes, versus 26.4% on Bluesky and 75.2% on Mastodon. Self-selected samples, but the platform gap is too large to call an industry finding.
  • OpenAI explains its ad stack: the official developer account says ChatGPT’s advertising system follows an evolving conversation while keeping answers independent. Ads and answers now share one surface, and what “independent” means in practice was not detailed. https://x.com/OpenAIDevs/status/2114489740930232731
  • Qwen reportedly restricts Iranian users: discussion says Tongyi Qianwen blocks users in Iran, with commentary noting this is not something Alibaba can decide alone and that complaining to the vendor is pointless. Export controls reaching open-model distribution will produce more of this friction. https://x.com/dotey/status/2114443220432372195
  • rasbt’s reasoning-model course: an 11-hour video course starting from a pretrained Qwen3 0.6B, hand-writing a verifier, inference-time scaling (sampling, self-consistency, scoring, self-refinement), GRPO-style RLVR and distillation, all evaluated on MATH-500. https://x.com/shao__meng/status/2114533468826952112
  • Research on looped and depth-grown models: at 3B scale, models that copy layers deeper or loop a shared block use context better than a dense model of the same depth, scoring higher on GSM8K at every number of demonstrations; the mechanism shows up in attention heads, with strong final heads descending from a few early induction heads. https://x.com/omarsar0/status/2114404629064942018
  • Manus’s organization experiment: in “Rebuilding Manus with Manus,” the authors’ conclusion is summarized as “Context Not Control” — instead of tightly specifying how the agent should work, build out the process, toolchain and materials, and let people supply constraints at key nodes. https://x.com/fkysly/status/2114546068386136198
  • A local voice studio hits the trending list: a fully local project supporting 646 languages, bundling voice cloning, dubbing, dictation and audiobooks, gained 1,588 stars in a day for 85,737 total. On the speech line, open source plus local execution now covers most everyday needs.
  • A weekend price war in analytics: a paid revenue-attribution tool costs about $39 a month; someone shipped a lifetime-license clone and raised it from $29 to $49; the same day a self-hosted, MIT-licensed equivalent appeared and passed 300 stars in two days. The comment thread is more interesting than the product — “don’t sell to people who will clone it.”
  • Nathan Lambert’s assessment: in a guest lecture for Harvard CS 2881R he argued Chinese open models lead broadly while the US holds a 10x-plus compute advantage, and that distillation is not the key issue. https://x.com/shao__meng/status/2114511015371329845

🕐 Selected hourly signals

PT time Signal Why it is worth remembering
00:00 Owly implements AWDL and AirDrop in pure Rust on Linux Apple’s private transfer protocol reverse-engineered, so an iPhone can exchange files with an ordinary Linux machine
01:00 DBDelve: a native multi-engine database client in Rust and GPUI Aimed straight at DBeaver’s 5-second startup and 606MB idle footprint; memory is the wedge
04:00 The dimensional-drawing Skill Draws architecture, not copy; variables become formulas; unknowns marked [TBD] — a Skill can carry design rules
05:00 prompt2rig Generate a short clip of a person moving, then recover 3D motion onto a character, 15 joints per hand, 2–3 seconds on a 4090
07:00 Scrawl’s screen-recording workflow Bundles keyframe collages, speech transcripts and action logs for the agent, closer to the scene than describing a UI bug in words
08:00 Aravind Srinivas’s multiplayer SF 2049 Rooms and lap timing run on Cloudflare Durable Objects; the trailer was shot by three agents in parallel
14:00 Laurence Moroney’s advice Build each core piece by hand once, or you cannot review generated code or notice retrieval returning junk
19:00 SOMA hands CAD to the agent The agent writes models in Python and rebuilds on every change, with 3D-printing and CNC rules and overhang and wall-thickness checks

Editorial conclusion

The two things worth keeping from today are plain. For the first time an AI company has to open its books, and slowing growth, channel commissions and compute commitments are all written there; the day’s security signals show that AI has pulled discovery forward while patching stays tied to human labor. Open weights, hardware portability and model capability keep advancing, but none of those three lines is the same line as “holds up in use.”

Sources and method

Reviewed 20 PT hourly captures and 3 named sources in the 2026-10-25-pt directory, about 189KB in total; the signal pool is rich. hubtoday and AI Valley are aggregator and newsletter sources, so the specifications, benchmarks and financing figures they carry are not first-party confirmed; single social-media relays are treated as single-source and labeled as such in the text.

WeChat QR code for 智简 Smart&Concise

FOLLOW ON WECHAT

智简 Smart&Concise

Search in WeChat for independent development and AI updates.