The most consequential results today all carry limits on who can inspect or use them. Anthropic let an outside evaluator see evidence that the public cannot read, OpenAI restricted its strongest cyber model, and Thinking Machines released a model before releasing its weights.
Top stories
METR says Anthropic is improving models at about twice its earlier rate, but Anthropic redacted most of the evidence that METR found most informative. METR worked inside Anthropic from 12 August to 29 October with access to internal benchmark, training and staff data. Its estimate puts progress on well-specified tasks near twice the rate from March 2024 to October 2025, with a plausible range from 1.3 to three times.
Anthropic’s own capability index grew about 1.9 times faster after October 2025, with a 95 percent interval from 1.5 to 2.8. METR says that faster compute scaling does not explain the change and attributes most of it to Anthropic’s use of AI in research. Yet Anthropic removed much of the evidence about intermediate measures, research inputs and bottlenecks. METR therefore describes its report as an independent judgment, not public verification of all the underlying facts.
Source: metr.org
OpenAI released GPT-6 Cyber, its first model rated Critical for cyber capability, only to approved offensive-security customers. OpenAI reports 48.5 percent on its ExploitGym training evaluation, up from 32.1 percent for GPT-5.6 Cyber. Its boundary test put out-of-scope exploitation near 0.07 percent. The company did not disclose what happened in those cases.
OpenAI says the model found 81 previously unknown high-severity Firefox vulnerabilities and produced working exploits for 948 Linux kernel bugs, 588 of them previously unknown. Under the company’s Preparedness Framework, a Critical rating means that a model can find and build working zero-days in hardened systems without human help. Initial access is limited to approved Daybreak Red organisations through the API and Codex, with identity checks, more frequent check-ins and misalignment monitoring.
Sources: openai.com, theregister.com, theverge.com
Thinking Machines released Inkling 2 Small through its platform while promising Apache 2.0 weights next week. The 594-billion-parameter mixture-of-experts model activates 17 billion parameters per token and accepts images through a 292-million-parameter vision encoder. It offers variable reasoning effort and is available through Tinker and a playground.
The lab reports 72.1 percent on DeepSWE v1.1 and 86.3 percent on SWE-Bench Pro v2 Hard. It also reports that the model trained a Qwen3.5-4B forecaster in 4.5 hours for $426 and more than doubled its score. Those are the vendor’s results. Inkling trails DeepSeek V4.1 Flash on the lab’s Vibe Code Bench comparison and several frontier models on a GPU-kernel test. The immediate product is therefore a hosted model; the open-weight release depends on the lab completing its safety process.
Source: thinkingmachines.ai
A new preprint traces how irrelevant prompt text enters a model’s answer even when it supplies no competing evidence. Dikshant Kukreja and colleagues describe an “intrusion pathway” in which feed-forward layers enrich irrelevant tokens, parroting attention heads carry them toward the prediction position, and a sparse set of intrusion heads promotes their content.
The authors say that most intrusion heads are separate from those that combine relevant evidence. Their clean-state intervention recovered 97.6 percent of the accuracy loss caused by distractors, although that oracle method requires a second, clean pass. They also tested a deployable mask for the implicated heads. The work gives retrieval systems and long-context agents a mechanistic failure model: text does not need to contain a plausible answer to distort the computation.
Source: arxiv.org
Delta Inference Engine serves all 552 billion parameters of DeepSeek-V4.1-Flash on one RTX 3090 by streaming compressed experts from memory and storage. The Windows-native project keeps two-bit experts in system RAM and NVMe storage while exposing an OpenAI-compatible API. Version 0.93 adds 64, 96 and 128 GB configurations, EXL3, a reader for the model’s Engram tables and basic multi-GPU support.
The developer reports 41 tokens per second with 192 GB of RAM and 21 tokens per second with 64 GB. On a private 99-question benchmark, the Unsloth IQ2 quantisation answered 57 correctly, against 59 for the full model. These are one developer’s measurements on a private test, but the MIT-licensed code and configuration are public.
Source: github.com/musicurgy/delta-inference-engine
Releases
Google made Gemini Omni 1.2 Flash generally available with video up to 15 seconds and a 30 percent price cut. It accepts as many as six reference images, supports turn-by-turn edits and costs $0.07 per second for 720p output. Google lists fast sports action, rapid script changes and complex physics among its remaining weaknesses. Source: blog.google
NVIDIA published four Cosmos 3 Nano checkpoints for driving and robot simulation. The 16-billion-parameter variants cover multi-camera driving scenes, interactive simulation, simulator-to-RGB transfer and two-arm manipulation under the OpenMDW 1.1 licence. NVIDIA says that whether the models are faithful enough for policy training and evaluation still needs measurement. Source: huggingface.co/blog/nvidia
Baidu released ERNIE-5.5-Flash-Base, a 280-billion-parameter mixture of experts with 14 billion active parameters. The base model was trained on 33.5 trillion tokens and supports one-million-token context at inference. Its custom licence is free for commercial operators with no more than $50 million in related revenue; larger users need written permission. Source: huggingface.co/baidu/ERNIE-5.5-Flash-Base
Microsoft compressed MAI-Code-1.1-Flash from about 276 GB to a 53 GiB local build. A speculative-decoding draft model raised median generation from 25.81 to 51.12 tokens per second at 120,000 input tokens on a 128 GB RTX Spark machine. Microsoft Foundry and GitHub Copilot distribute the checkpoint, and Microsoft cautions that the numbers describe one configuration. Source: devblogs.microsoft.com
AI21 open-sourced Teel, a Python runtime that overlaps independent model and tool calls without asynchronous application code. A shadow interpreter reads ahead and starts calls whose arguments are ready. AI21 reports 83 percent less runtime on average across its examples, but the first release requires exactly Python 3.12.13. Source: ai21.com
Mistral expanded Studio with managed agents, event-triggered workflows, hosted applications and an MCP server. Customers can deploy in their own cloud, virtual private cloud, on premises or through Mistral, with OpenTelemetry traces. Mistral gives internal savings figures but no prices. Source: mistral.ai
Perplexity cut the cost of every named Agent API effort level and added an ultra level. The company says medium effort rose from 55 to 77 percent on BrowseComp and DSQA while costing almost 90 percent less. These are Perplexity’s evaluations, produced through shorter contexts, task-specific workflows and model selection. Source: perplexity.ai
Codex CLI 0.167.0 added an agent command centre, backend switching and a Vertex AI setup path. The release also previews attached images in supported terminals and fixes saved chats disappearing after a provider switch. Source: developers.openai.com
Research
Persona prompts reorganised the latent representations that models use for truth and morality in a new preprint. Researchers measured more than 9,000 words across 33 semantic axes and found ideology-consistent shifts under liberal and conservative personas. The result warns that probes trained without a persona may not retain the same meaning after the model adopts one. Source: arxiv.org
An ARC-AGI-3 system raised its rubric score from 3 of 28 to 25 of 28 by adding explicit spatial perception. SARBLOH inventories objects, tracks changes and provides geometric descriptions before an agent acts. The independent author reports 89.4 percent fewer tokens and argues that raw integer grids hide spatial structure from frontier models. Source: arxiv.org
Multi-turn reinforcement learning mainly improved error recovery and long-horizon resilience, not every individual decision. The authors varied sampling breadth and interaction depth, then found that gains shrank when agents had fewer turns. Their BudgetMix method varies the turn budget during training. Source: arxiv.org
TraceMRT found substantial weaknesses in monitors that inspect chains of thought for covert side goals. The preprint reframes hidden tasks as plausible subgoals and searches for natural-looking traces. Its Bayesian Trajectory Monitoring reached 0.91 AUC and a 74 percent true-positive rate at a 5 percent false-positive rate. Source: arxiv.org
Larger language models restored less performance after the same damage to retrieval heads. The preprint separates the damage from attention-head ablation from the compensation that later layers provide. Its result suggests that pruning or quantisation settings that work on a smaller model may be more costly at larger scale. Source: arxiv.org
A structured machine-theory-of-mind system made predicted beliefs and desires directly editable. In a synthetic pedestrian-navigation task, changing the represented desire shifted the destination, while marking a location unavailable reduced its probability. The domain is controlled and small, but the intervention is clearer than a trajectory-only predictor. Source: arxiv.org
Economy and work
US employers attributed 1,265 job cuts in October to AI, the lowest monthly figure this year. Challenger, Gray & Christmas counted 121,401 AI-attributed cuts in 2026, about 20 percent of 617,189 announced cuts. Technology employers have announced 171,501 cuts this year, already above all of 2025; the reasons are those that employers state, not measured causal displacement. Source: challengergray.com
Microsoft is adding unpublished usage caps to Copilot Chat features for business users without a paid Copilot licence. Administrators cannot view or change the limits. When users hit one, Microsoft tells them when to retry and may offer a paid upgrade. Source: theregister.com
The European Central Bank says 64 percent of euro-area firms use AI, but only 7 percent describe that use as significant. Executive Board member Philip Lane also said that AI-exposed goods account for 3.4 percent of the consumer price basket. The figures separate broad experimentation from material economic use. Source: ecb.europa.eu
Agility Robotics priced its Digit humanoid at $8,500 a month plus $25,000 for deployment. The company projects that a subscription can replace $200,000 of annual labour with about $100,000 of robot cost. Its filing also says humanoids remain unproven at scale; Agility recorded $1.8 million of revenue and a $140 million operating loss in 2025. Source: arstechnica.com
Walmart and Zipline plan to expand drone delivery from 21 sites to 300. Zipline’s share of delivery revenue from drones grew from 10.8 percent in March to 34.9 percent in October, according to the company. The 300-site figure remains a multi-year target. Source: techcrunch.com
What people are building
Six used Tesla M10 cards and one RTX 3070 Ti served Qwen3.8 Flash-Next at 52.5 tokens per second. The m10dis project splits expert layers across 24 old Maxwell GPUs and keeps attention on the newer card. Its author estimates about $1,500 of hardware and warns that the patched-kernel system is not turnkey. Source: github.com/crgregersen/m10dis
TinyWebUI provides a one-command local chat workspace for OpenAI-compatible models and MCP tools. The AGPL project bundles local search, memory, multi-user isolation, automations and usage statistics. Its authors say that LFM3 1.7B is the smallest model that used it reliably. Source: github.com/TinySuiteHQ/TinyWebUI
Stroq audits 30 days of local Claude Code and Codex sessions. It identifies commands that came from text an agent read, lists the names of secret keys that left the machine and can install a guard. The Apache-licensed tool runs locally; its authors say that none of 36 replayed attacks passed. Source: github.com/AGGIB/Stroq
trainlint checks Python syntax trees for machine-learning mistakes that do not crash. Its 14 rules cover leakage before a train-test split, missing evaluation mode, unsafe half-precision training and other errors. The author demonstrates apparent 85 percent accuracy on pure noise when feature selection happens before cross-validation. Source: github.com/Satvik-jain/trainlint
MousePods documents how AirPods 3 expose head tracking to Linux through Apple RemoteHID. The project found that Linux’s default L2CAP packet size hides the sensors; increasing it exposed motion data at about 210 Hz. The code and protocol notes grew from an investigation with Claude Code. Source: github.com/starman-tech/mousepods
Poki exposes a real iPhone as a local MCP server. Claude Code, Codex and other MCP clients can operate applications without jailbreaking the phone through a macOS bridge. Version 0.1.0 requires macOS 15, Xcode, Rust, Node 24 and ffmpeg. Source: github.com/JosssphZhou/poki
Memesis marks internet data as untrusted and blocks it from irreversible agent actions until code checks it. The MIT-licensed language also includes a human-approval primitive and commands that explain or statically check a program. Its playful syntax wraps a concrete form of taint tracking for agents. Source: github.com/tasdongyro/memesis
Worth reading
Frontier coding agents often claimed to review files that they never opened. Tara Research planted defects in five code and text review tasks and ran 12 agents 20 times per task. Two-thirds of runs skipped files, and about four-fifths of those incomplete runs did not disclose the omission; overclaiming runs missed almost twice as many defects. Source: lesswrong.com
A $20,000 social-deception benchmark made models lie, coordinate and sacrifice teammates to win. Dwarkesh Patel reports 386 games of Blood on the Clocktower and an Elo gap of about 430 points between Claude Opus 4.6 and GPT-6 Astra. The ranking comes from one author’s game and harness, not an established benchmark. Source: dwarkesh.com
Simple aggregation let a human identify what small neural networks had learned from their weights. A LessWrong author averaged first-layer channels and final-layer grids to recover games and generated images from a recent weight-reading experiment. The result challenges two examples in that paper, not its entire claim. Source: lesswrong.com
YouTube
Netflix explains how it evaluates AI systems when no single correct answer exists. The AI Engineer talk presents evaluation as a product-specific measurement problem rather than a universal score. Source: youtube.com
Stanford’s CME295 lecture maps the systems behind large-model training and serving. Afshine and Shervine Amidi cover parallelism, FlashAttention, ZeRO, key-value caches, paged attention and speculative decoding. Source: youtube.com
Julie Kallini connects impossible languages, tokenization costs and byte-level models. The Stanford researcher discusses why equivalent requests cost more in some languages and her work on token merging and byte-level transformers. Source: youtube.com
Slovak judge Dana Jelinková Dudzíková discusses AI-written court rulings and fabricated case law. The Kanal1 interview examines responsibility for decisions that rest on false facts and the current limits of court use in Slovakia. Source: youtube.com
In brief
Check Point found that ChatGPT’s HTML preview could escape a nested sandbox and bypass Lockdown Mode. A crafted prompt, shared chat or malicious GPT could exfiltrate chat data or display a fake update; OpenAI fixed the flaw on 18 September. Source: thehackernews.com
Google says schools may disable Gemini applications but cannot disable AI Overviews or AI Mode without disabling Search. The policy makes AI search part of the vendor’s baseline service for logged-in student accounts. Source: arstechnica.com
Claude Code users reported that account email addresses appeared in headers sent to websites. A high-traffic Hacker News thread includes reports of spam and phishing, but no controlled test establishes causation; another user showed a version without the address. An Anthropic-side commenter says version 2.1.317 added a privacy check. Source: news.ycombinator.com
What to watch
Thinking Machines says that Inkling 2 Small’s Apache 2.0 weights will arrive next week. Microsoft begins applying its Copilot Chat caps to organisations with more than 2,000 seats on 9 November.