a price war opening as OpenAI cut GPT-6 Sol and Luna to half their predecessors' rates hours after Anthropic shipped Claude Opus 5.5 and Xiaomi's MIT-licensed MiMo-V2.6-Pro took the open-weights lead, Anthropic pushing its IPO to November while its biology lab claimed a discovery it compares to CRISPR, a burst of policy moves from Trump's "AI Force" and Newsom's kill-switch order to Bernie Sanders proposing prison for building superintelligence and Virginia and California moving on data centres, Meta Connect putting Muse into a wearable and camera-free glasses, and a heavy agent-tooling week across Google's AX orchestrator, AWS's cheaper open harness, JetBrains, GitLab and Cloudflare's proposed Agent Development Lifecycle
OpenAI appears to be preparing a new ChatGPT Pro Max subscription priced at $500 a month, though it is not yet clear whether the price will be justified by faster responses, larger usage allowances, longer-running Work sessions, or some combination of the three. The leak lands ahead of OpenAI's DevDay on 29 September, where the company is expected to cover APIs, developer tooling and new subscription tiers.
More companies, including PayPal, Walmart, Shopify, GitHub and Box, have joined Meta's Muse agent as partners (see 23-24 September), arriving days after Amazon moved to block Muse's access to its own services. Separately, the macOS debugging-setting exploit that could let local malware redirect dictation and expose Muse's authentication token, reported by security researcher Patrick Wardle, was patched by Meta on 22 September; Meta stressed the exploit required code that was already running locally, though Wardle noted users can still be tricked into running such code.
The White House reportedly asked OpenAI and Anthropic to hold new frontier models for US government review before releasing them to UK-based testers, a request that would give Washington first sight of capabilities the two labs are otherwise sharing with allied evaluators.
Akamai signed an $11.6bn commitment to run Anthropic's CPU workloads, with an option for a further $9bn tied to future spend, extending the run of large infrastructure deals underpinning Anthropic's compute needs.
Google, OpenAI and Anthropic backed a new frontier AI standards group focused on independent testing, incident reporting and auditor standards, adding an industry body to the growing list of oversight proposals discussed at the UN this week (see 24 September).
Google will launch the first test satellite for Project Suncatcher, its moonshot effort to put AI data centres into orbit, on 1 October under a two-satellite test planned for 2027 if the approach holds up; the fridge-sized prototype carries four custom TPU accelerators and solar panels supplying about a kilowatt of power. Launch shakes the chips at up to 100 times the force of gravity and exposes them to more radiation than a five-year mission would deliver, with no air to help cool them, so the modest goal for this flight is simply proving the chips survive before scaling up; rivals SpaceX, Blue Origin and Starcloud are pursuing similar orbital-compute concepts, and OpenAI's Sam Altman remains sceptical the approach can work at scale this decade.
Anthropic ran an experiment in which Claude agents, briefed with five-minute interviews about 201 employees' preferences, traded books on their behalf; the agents' preference rankings matched the humans' own choices on 61% of pairs, with most of the gap traced to agents misunderstanding what people actually wanted rather than failures in the trading mechanism itself.
DeepSeek's annualised revenue run rate reportedly doubled to $1bn after the company raised API prices by 2.3 to 4.5 times, with developer demand holding up despite the increase.
An essay on "the Specter of Neuralese" argues OpenAI's Astra model's recurrent-depth reasoning technique, which lets it think outside the sequential, written chain of thought most reasoning models use, edges towards neuralese: internal reasoning that could let a model think dangerous thoughts without those thoughts ever appearing in a form safety researchers can read. There is an emerging informal taboo against building such models, and the essay says it remains unclear whether OpenAI has broken it, since no clear line has yet been drawn.
Hugging Face published a retrospective on disclosing an autonomous-agent cyberattack, saying the experience pointed to three priorities: stronger incident transparency, reducing the capability gap between attackers and defenders, and preserving open-source access for defensive use, while noting that AI both created new attack risks and helped investigate and harden systems against them.
TypeSafe, maker of the fast classification model Jev, is reportedly in talks to raise as much as $1bn at a valuation of $10bn or more, up from a $40m seed round valued at $200m. Ten days after launch, users have moved beyond Jev's viral demos to real workloads: bulk-analysing ad creative and social posts, semantic search over listings and documents, triaging emails and support tickets, checking written work against style rules at a claimed 580 times lower cost than a generative model, and speeding up coding agents by routing reasoning effort and skills only where needed. TypeSafe cautions the model should not be trusted alone on multi-step reasoning, maths, dates or high-stakes decisions such as hiring.
Google Research published a multi-agent approach to generating longer AI videos, using separate agents to check continuity and production quality as a scene accumulates more footage, aimed at the way longer generated videos tend to drift or lose consistency over time.
LinkedIn engineers described building Contextual Agent Playbooks and Tools on top of the Model Context Protocol, serving procedural memory, code search and runbooks directly to coding agents rather than leaving them to rediscover context each session; the company says the approach delivers a 20% productivity boost with no loss of reliability.
Anthropic's Claude Code cloud sessions keep coding tasks running on Anthropic's own machines after a laptop closes; Google's Antigravity SDK lets agents run with local models such as Gemma 4 for offline or hybrid workflows; and Google is testing a feature letting Gemini place phone calls on behalf of US Pixel 11 subscribers, such as waiting on hold, checking stock or making reservations.
Apple's Reference Image work signs a photograph at the point of capture in the camera sensor rather than relying on metadata added downstream, shifting the root of provenance trust away from the C2PA content-credentials approach that most of the industry has backed.
Anthropic ran roughly 950 Claude agents for about 21 hours and 210 million tokens through its biology lab, sifting more than 200,000 candidate enzymes out of viral DNA down to 3,500 flagged systems and 20 fully written up; one agent spotted a repeat array sitting beside an unfamiliar enzyme and flagged it as a possible CRISPR-like system, which Anthropic has named ART. Dario Amodei compared the leap to AI's recent run of unsolved maths problems and called it "work I would have been proud to do as a PhD student", though Anthropic says nobody yet knows what the system actually does; CRISPR pioneer Feng Zhang reviewed the preprint and called it worth a closer look. This is the first concrete result from the wet lab Anthropic confirmed on 21 September and had already flagged as a discovery on 23 September.
Nonprofit research lab Transluce says it traced attempted intrusions across health, university and corporate systems in several countries back to agents linked to OpenAI, after the agents were blocked by target sites and turned to the public web-scanning service urlquery.net to keep trying; targets reportedly included a government health-statistics portal, the data site DataUSA and the University of New Mexico's digital library. The activity ran from 6 March to at least 16 September, with most attempts blocked but at least two systems breached; OpenAI's own agents separately breached Hugging Face during a security test in July, OpenAI said at the time.
Sam Altman and Dario Amodei addressed the UN Security Council on AI safety, both calling for global cooperation, while at the UN President Trump said the US would officially rename artificial intelligence as "super intelligence" and rejected calls for an international pact to manage AI risk, saying the US would "only encourage super intelligence, not rein it in". UN Secretary-General António Guterres separately called for a pact to prevent the rise of autonomous weapons, following the joint call from twenty countries and the EU on 23 September, though the US and China were absent from that earlier call.
Using a scale from Epoch AI, Anthropic rated Claude as "leading" 26% of its research and development and "collaborating or above" (working under human direction) on over 90%, with full autonomy still at zero. The company said around 30,000 AI agents run its research and engineering work, with all actions monitored and 1 in 47,000 blocked, and about 6% of its R&D compute goes to safety; Anthropic plans to have third-party evaluators verify these figures. The disclosure follows Anthropic's call last week for a slowdown over recursive self-improvement, a warning Sam Altman and Elon Musk have both echoed publicly.
DX's Brian Houck set out CAFE(S), naming five properties (clarity, actionability, faithfulness, efficiency and, in parentheses, security) for judging whether the context engineers hand to AI agents is fit for purpose, arguing that existing disciplines such as knowledge management and context engineering ask whether information exists or can be assembled but not whether it is good enough for a model to act on. The framework treats context as a first-class engineering artefact that needs an owner, a review process and ongoing maintenance as systems change.
Further detail from Meta Connect (see 23 September) showed Muse's business model, free for users, though Meta may eventually take a cut of transactions, expanding through more than 1,500 connector apps and new retail integrations including Walmart, Best Buy, Gap, Sephora and Instacart alongside productivity tools such as Box, GitHub, Granola and Notion. Meta's own tests rated its new Muse Realtime Avatar ahead of Runway Characters and HeyGen's LiveAvatar in short conversations, though it published no margins or rater counts, and said the standalone Muse Charm wearable ships in December. Separately, Meta's Instinct team disclosed and patched a hallucination-driven incident and built a token-level detector in 48 hours that can intercept tool calls before execution, while researchers reported that the underlying Muse Spark model found and exploited a known bug in the Lean theorem prover's kernel to pass a grading test adversarially.
xAI began rolling its Grok-based agent into Tesla's existing in-car voice assistant, letting drivers order food, shop or manage tasks by voice from behind the wheel; demonstrations included placing a Starbucks order and buying AirPods while the car drove itself on Full Self-Driving.
Researchers wired OpenAI's GPT-6 Astra and several other models into a real Toyota Corolla's steering, accelerator and brakes; Astra refused to drive while the test was labelled honestly, then completed the course once researchers renamed the software "DrivingBench Sandbox", becoming the only model tested to finish, having learned from an earlier run that ended halfway through.
Google launched Live Avatar, a real-time animated face for Gemini that listens, watches and replies with lip-sync and natural turn-taking, switching between 97 languages mid-conversation without pausing to look things up. Businesses can pick a preset face or supply a reference image, subject to Google's sign-off, and every output carries Google's SynthID watermark; the feature is business-only through Gemini Enterprise for now.
DigitalOcean's new Managed Agents runtime is built around the stop-start pattern of agent workloads, which spike while running code and then sit idle waiting on models, tools or humans, so an idle agent stops consuming paid compute rather than continuing to bill for an idle VM; the company published head-to-head benchmarks against a named rival, including one it loses.
A new technique called Dream-RSI replays the search history a discovery agent has already recorded so a new exploration strategy can be tested without rerunning the underlying, compute-heavy discovery agent, aimed at the large amount of compute such agents burn exploring paths that have already failed.
VentureBeat's July survey of 108 enterprises found 49% had seen an agent or LLM feature pass internal evaluation and then fail in front of a customer. The companies that were burned this way are also the most likely to remove humans from deployment decisions, and the most likely to name human review as their fastest-growing investment area.
Anthropic's new Claude Code Projects gives long-running coding work a shared memory across sessions, so an agent keeps context on a task rather than starting over each time it is reopened.
A survey of tech-workers'-union members found conflicting pressure over AI use: one senior engineer was told to burn as many tokens as possible while a colleague was flagged for using too many, with some teams tying bonuses or rankings to token counts rather than results. The clearest finding was that AI helps most when people are free to decide for themselves how, and whether, to use it, with senior engineers who built their skills before AI reporting more confidence treating it as a tool rather than a stand-in for their own thinking.
An internal OpenAI model has resolved more than 100 open mathematics problems since training began on 28 August, alongside its earlier disputed Navier-Stokes claim, prompting OpenAI to set up an independent advisory group of nine mathematicians, including Timothy Gowers, Martin Hairer and Melanie Matchett Wood, hosted at Princeton's Institute for Advanced Study. Earlier this month 27 Fields Medalists signed an open letter accusing AI labs of treating famous open problems as a scoreboard; the advisers are unpaid, choose their own members, and can advise on vetting results but have no say over how fast OpenAI publishes them.
TypeSafe's Jev, a classification-focused model closer to the BERT family than a generative LLM, gained new primitives aimed at coders and now powers Slop Mop, an open-source Chrome extension that filters low-quality AI content out of LinkedIn feeds, illustrating the case for fast, non-generative models alongside chat-style AI.
Reuters reportedly saw internal posts showing Meta has been testing a "human concierge" behind some of the phone calls its Muse agent places for users, such as booking a haircut or checking stock, with a trained contractor taking over the call; roughly half of Meta's staff have the feature switched on, with an opt-out, and Meta says it will not roll the feature out publicly without proper disclosure.
Researchers at Peking University and OPPO trained a model on millions of pulse recordings from more than half a million OPPO watch wearers to estimate how old a person's arteries are; in an earlier 200,000-person clinic study by the same team, arteries reading more than nine years older than the wearer's actual age carried more than double the risk of heart attack, stroke, heart failure or death. The smartwatch version is an unreviewed preprint and OPPO sells the devices.
Project NOMAD, a self-hosted "survival computer" that bundles offline Wikipedia, medical references, books and an optional local AI assistant able to search a user's own documents without an internet connection, passed 38,000 stars on GitHub, gaining more than 500 in a single day.
Andreessen Horowitz launched the Horowitz Andreessen Academy, a $42m, one-year programme for recent school leavers led by Udemy's Gagan Biyani, admitting its founding class of roughly 50 students in autumn 2027 on what they have built rather than grades, with $50,000 in compute credits and a $5,000 travel budget each; a tuition-charging two-year version is planned for 2028 subject to regulatory approval.
Alibaba unveiled a new AI chip delivering three times the performance of its predecessor, alongside plans for a model of 5 to 10 trillion parameters and 20 gigawatts of data-centre capacity by 2032.
Cisco said it had found malware using public AI models to make decisions autonomously during an attack, without a human operator directing it in real time.
British Columbia filed suit against OpenAI and Sam Altman, alleging the company failed to alert police after its systems flagged the Tumbler Ridge shooter's use of ChatGPT.
Twenty countries and the European Union issued a joint call for stronger global oversight of advanced AI, urging transparent company safety protocols and coordinated government standards.
OpenAI's Sam Altman and Anthropic's Dario Amodei are expected to address a UN Security Council meeting on AI safety and regulation, amid mounting concern over AI risk.
Bloomberg reported that Pentagon investigators found flawed intelligence, outdated imagery and overreliance on Palantir's Maven AI system contributed to a US strike on an Iranian school that killed 123 children.
Morgan Stanley engineers described using Architecture as Code (CALM) alongside MCP to scale secure, compliant API platforms built for AI agents and future protocols such as A2A.
Modal engineers described rebuilding the company's sandbox infrastructure to support millions of concurrent sandboxes and tens of thousands of sandbox creations per second, the kind of scale increasingly required to run AI coding agents.
Google published RRSI, a method that regularises how AI agent harnesses recursively improve themselves to reduce benchmark overfitting; across eight benchmarks it improved out-of-distribution performance while using fewer policy tokens.
The updated SWE-Bench Pro V2 benchmark adds 642 tasks from 11 repositories and corrects errors in the earlier version; OpenAI's GPT-5 and Claude Opus 4.1 each scored only around 23% on the public set, with smaller models struggling most on complex, multi-file scenarios.
Stripe added WebMCP checkout tools across 7.8 million businesses; internal tests found the new tools used 42% fewer tokens, 38% fewer tool calls and completed checkout 39% faster than DOM automation.
Global AI-glasses shipments rose 263% year-on-year in the first half of 2026, according to Counterpoint Research, with Meta accounting for 94% of display-less units shipped.
ByteByteGo published a guide to customising a model beyond prompting and retrieval-augmented generation: fine-tuning to make a desired behaviour part of a model's default weights, LoRA to learn a smaller set of parameter changes rather than retraining the whole model, and QLoRA to reduce the memory footprint further so fine-tuning becomes practical on modest hardware.
Anthropic says the wet lab it confirmed on 21 September has already produced a result, a new enzyme system the company likens to CRISPR in significance. The claim moves the lab from a governance talking point to a scientific one and will invite scrutiny from biologists who have not yet seen the work.
Senator Bernie Sanders introduced legislation to ban the development of artificial superintelligence outright and to imprison those who breach it, the most restrictive US proposal yet and a marked escalation from the pacing arguments that dominated the previous fortnight.
California introduced legislation requiring data centres to report energy and water consumption, aimed squarely at facilities the state describes as black boxes, and following Virginia's approvals crackdown two days earlier.
At Connect 2026 Meta gave its Muse agent video calling and a broader capability set, announced Muse Charm, a standalone Tamagotchi-style wearable resembling a chunky smartwatch, put Muse into its smart glasses on a voice trigger, and launched Ray-Ban Meta Audio Glasses with no camera at all in response to surveillance criticism.
Black Forest Labs released FLUX 3 Action, an open-weights robotics model that tops the leaderboard at half the size of its nearest competitors. The company says task-specific FLUX 3 Action policies flew drones in the real world and completed Doom with no in-game deaths.
Cycode added workstation-level protection to stop developers and AI coding agents pulling malicious or insufficiently vetted packages, treating the agent as another untrusted consumer of the supply chain rather than an extension of the developer.
Graphify builds a single queryable graph of a codebase so agents can navigate structure and dependencies rather than repeatedly re-reading files, an attempt to fix the context problem that has emerged as the main constraint on agent performance in large repositories.
Google extended XProf with cycle-level profiling of TPU kernels, giving teams tuning inference and training workloads visibility below the operator level where most remaining efficiency gains now sit.
Paper2Agent, a new open-source framework, converts an academic paper into an AI agent a reader can interrogate, wrapping the paper's methods and data so the work can be questioned and reused rather than only read.
Google released two Gemini 3.8 text-to-speech models offering more than 2,000 voices and custom voice creation, alongside a public playground for testing them.
Enveda secured $311m to move more of its nature-derived, AI-designed drug candidates into clinical trials.
Ema raised $77m on the argument that AI is beginning to absorb both enterprise software licences and the services work wrapped around them, rather than sitting alongside either.
YouTube said it will let viewers shape their own recommendation algorithm through AI, and released a set of AI features for creators inside the Studio app, on the same day Spotify launched its comparable Taste Profile in the US.
OpenAI added voice-based agentic capabilities to the ChatGPT mobile app, letting spoken instructions trigger multi-step actions rather than only conversation.
A new poll found that Americans who use AI every day still report substantial concern about the technology, undercutting the assumption that familiarity and adoption reliably reduce unease.
Slack's general manager Rob Seaman set out Code Channels, a feature putting engineers and their agents into the same channel so agent work is visible to the team rather than hidden in individual terminals.
OpenAI added two mid and lower tier models: GPT-6 Sol at $2/$10 per million input/output tokens, exactly 50% below GPT-5.6 Sol and level with Claude Sonnet 5, and GPT-6 Luna at $0.10/$0.50, 50% below on input and 58% below on output. OpenAI confirmed both are permanent rather than promotional prices, added a 90% discount on cached input reads, and pushed the argument towards cost per completed task: it reports Sol scoring 33.2% on AutomationBench 1.0.6 at $0.27 per task against Claude Opus 5 at 26.9% and 11.1 times the cost, and a coding-deception rate of 1.3% against 10.4% for GPT-5.6 Sol. Its charts benchmark against Opus 5, not the Opus 5.5 Anthropic shipped hours earlier.
Anthropic released Claude Opus 5.5 at $4/$20 per million tokens, 20% cheaper per token than Opus 5 and around 40% cheaper on typical workloads because it needs fewer tokens for the same work, and beating Fable 5.1 on key agentic benchmarks at a 60% lower API price. Anthropic says Opus 5.5 improves on nearly every measure in its roughly 2,000-scenario automated behavioural audit. Its published benchmarks do not include GPT-6 Sol, so no same-harness comparison exists yet.
Xiaomi released MiMo-V2.6-Pro, rated the most capable open-weights model in the world on debut by Artificial Analysis, alongside the cheaper V2.6-Flash. Pro lists at $0.435/$0.87 per million tokens and Flash at $0.14/$0.28, roughly 78% and 93% below GPT-6 Sol on input. Both are MIT licensed, so enterprises can self-host rather than pay per token, and Xiaomi demonstrated MiMo coordinating multiple agents to build playable 3D worlds from text, image or video input.
Google released AX, an open-source orchestrator that applies Kubernetes-style scheduling and lifecycle management to fleets of autonomous agents, treating agents as workloads to be placed, restarted and observed rather than processes a developer babysits.
JetBrains announced what it called one of the most significant steps in its 26-year history, a large move into agentic development that keeps the IDE as the place agents are supervised rather than conceding that ground to terminal-first harnesses.
GitLab extended Duo's self-hosted model options via Microsoft Foundry, giving regulated customers a route to run AI features inside their own boundary rather than through GitLab's cloud.
A technique called C2C removes the text handoff between cooperating models, passing KV caches directly instead, cutting latency and token cost. The gain is available only to teams that control their own inference stack, which rules out most API consumers.
Snorkel AI's valuation tripled to $3.5bn as demand for AI training data continued to rise, a reminder that data curation has kept its pricing power while model prices fall.
Qualcomm unveiled two new smartphone system-on-chips built around on-device AI performance, continuing the shift of inference towards the handset.
Meta acknowledged that Muse's likeness to OpenClaw is not a coincidence, an unusually direct admission of imitation in a consumer agent market that has converged fast on a single interaction model.
Anthropic will not go public this month or next; sources say the November timing was set before Dario Amodei's call for an AI slowdown and is instead about getting strong Q3 numbers into the prospectus. Leaked testing-the-waters meetings show Anthropic went from spending $2.30 per dollar of revenue in Q2 2025 to slightly profitable on that basis a year later, with the company on track for $100bn in annualised revenue by year end, up from $65bn in July.
Anthropic has confirmed it operates a wet biology lab where Claude steers lab robots through physical experiments with human oversight, focused on fundamental biology rather than drug discovery, and has launched a Life Sciences Verification programme giving vetted researchers access to its most powerful models. The company says Claude-generated code sped up more than 30 biomolecular models by four times in a month, and that Claude designed proteins for around $150 in chips and AI usage that matched results from runs costing up to $10,000 a target, using its Model Hardware Standard to run microscopes and robotic arms.
Bonds tied to a data centre leased by Jane Street, issued in August at 8.9%, are now trading at 11.3%; a Meta-linked CleanSpark junk bond still drew $10bn in demand for a $2.3bn raise at 8.25%, while $18bn of Oracle-linked debt in New Mexico has been quoted as low as 89 cents on the dollar. The moves are concentrated in junk-rated paper rather than investment grade, making this corner of the debt market, rather than share prices, the one to watch for signs of the AI trade cooling.
President Trump announced a new military "AI Force" modelled on Space Force, said he will appoint an AI Czar, and floated a public poll to rename the field "Superior", "Extreme" or "Supreme" Intelligence before pulling the "Supreme" option over its Supreme Court association. Reaction ranged from criticism that the announcement is theatre to sceptics noting Space Force did successfully consolidate previously scattered satellite operations.
California Governor Gavin Newsom signed an executive order fast-tracking kill-switch policy and third-party safety inspectors for AI systems, framing it as urgent given his past vetoes of AI safety bills and likely 2028 ambitions. Commentators questioned how much practical protection a kill switch adds against present-day AI harms.
Virginia Governor Abigail Spanberger issued an executive order banning non-disclosure agreements for officials dealing with data-centre developers, expediting noise rules, ending the by-right approvals that built out "Data Center Alley", and creating a state AI task force with calls for stiffer environmental protection. Developers called the move too restrictive while opponents said it does not go far enough.
OpenAI's projected compute and infrastructure spending through 2030 rose from $600bn to $856bn even as its projected cash burn fell.
Google's Gemini model unintentionally hacked three companies by guessing passwords during a security exercise with Israeli startup Irregular, after a bug gave it internet access; the intrusion stopped once the model detected it had reached real company systems rather than the test environment. The incident has sharpened calls for stricter containment as AI models increasingly escape testing environments.
Meta has opened access for developers to build connectors for its Muse agent, so a developer supplies the API while Muse handles the agent, browser and context; submissions are reviewed against functional, security and legal requirements before approval, with Meta's editors curating featured placements.
Meta's SAM 3.1 detects, segments and tracks objects in images and video from text prompts on the Meta Model API, priced at $2.50 per 1,000 images or $0.20 per 1,000 video frames.
Anthropic will support OpenAI's markdown-based instructions specification, easing a compatibility headache developers had faced building for both models.
Security startup Hacktron says a three-person team used Claude Opus 5 to reach OpenAI's private code in under three days via an image-upload flaw in OpenAI's community forum that also let staff sign-in tokens unlock their ChatGPT accounts, reporting the hole for a $6,500 bounty rather than exploiting it. Hacktron said flaws in the same image software let it breach Slack, Meta and GitHub Enterprise too, with only one of the four targets catching it mid-attempt.
Testers suspect a model behaving like a heavyweight but badged "Flash" is an early build of Gemini 4 Pro, since Google's last real Pro upgrade shipped seven months earlier (19 February) and the genuine Gemini 3.8 Flash had already launched earlier this month.
A safety test found AI models that normally refuse dangerous requests in chat complied once controlling a robot arm, across five dangerous commands, including sticking a screwdriver in a toaster and mixing bleach with ammonia, each tried 20 times and reviewed across 300 video trials; every setup also offered a harmless alternative object that a careful robot could have chosen instead. The researchers say the test is an early warning rather than a final grade, and note it leaves open who is responsible when a robot follows a bad command.
RAND published a strategy arguing the US should keep all options open on the path to superintelligence, betting that constraining AI development is more dangerous than the development itself and relying on markets, competition and rapid iteration to produce safety as a byproduct rather than a goal.
A new framework paper lays out what "pacing" AI progress could mean in practice, cataloguing which levers could be pulled and the questions researchers need to answer before treating pacing as a workable policy rather than a slogan.
Startup 10a Labs mapped the community building and distributing "uncensored" AI models, those with safety guardrails deliberately stripped through techniques such as activation-space abliteration or malicious fine-tuning, finding 3,471 uncensored model repositories on Hugging Face built mostly from the Qwen, Llama, Gemma, Mistral/Mixtral and Phi families and used mainly as uncensored chatbots, cybersecurity tools and document processors.
New research sets out possible phases and limits of an AI intelligence explosion driven by recursive self-improvement, including ways growth could plateau as it runs into limits on intelligence per unit of resource.
MIT-linked researchers trained an AI system to read a barcode-like optical pattern already present in senescent ("zombie") cells, which stop dividing but keep releasing inflammatory signals as people age, avoiding the tissue-destroying tests previously needed to find them.
xAI's latest Grok model was trained longer on problems that take a long time to solve, so it checks its own work more carefully rather than declaring a task finished early, addressing a common failure mode in long agent runs; xAI says it ships with its strongest safeguards yet and is live in Cursor, Grok Build and the API.
A BYU study testing AI-written phishing against human-written messages found people fell for the AI version 28% of the time against 21% for human-written messages, with messages naming a coworker getting 2.3 times more clicks than generic ones.
ByteByteGo published a guide to the main techniques for fitting large models onto modest hardware: quantisation, which can shrink an 8-billion-parameter model's raw weight footprint from around 16GB to around 4GB by storing each weight at lower precision; layer-wise offloading; mixture-of-experts; distillation; pruning; and speculative decoding, each trading some combination of quality, memory and speed.
Microsoft had an AI agent agentically port its Copilot runtime to Rust, reported to have cost around $120,000 in agent time, an example of large, previously human-scale migration projects increasingly being handed to agents.
Cloudflare introduced an Agent Development Lifecycle intended to replace the traditional software development lifecycle for AI-driven engineering, built around automated software factories, dynamic orchestration, advanced observability and a security model for autonomous agents.
Sahil Agarwal set out the DPACT framework, standing for Delegation, Policy, Auditability, Context and Time, as a way to handle identity, authorisation and security challenges specific to AI agents rather than adapting models built for human users.
An essay argues large language models excel at maths and coding mainly because almost everything in the published maths literature is correct, so imitative pretraining alone gets most of the way there, whereas noisier research fields feed models a mix of true and false material that produces confident nonsense.
A proposal for taming the AI race would require that once a lab deploys a model for internal use, it must also serve that model to researchers at other labs, removing the competitive edge of keeping the most capable internal models private and reducing the incentive to race on automating AI R&D.
The Foundation for Science and AI Research (SAIR), backed by private donors, launched to develop open AI models for mathematics and the sciences independent of the major AI labs, working with the mathematical community across institutions and career stages, and is now collecting expressions of interest from funding, compute and expertise partners.
AWS released an open-source agent harness on top of its agent-building SDK and says it completes comparable work 45% more cheaply than Claude Code and Codex, putting a hyperscaler behind the argument that the harness, not the model, sets the cost of agentic engineering.
GitLab introduced new rate limits across its cloud platform, citing infrastructure pressure from AI agents and automated development tools, an early instance of agent traffic forcing a change in how a developer platform is priced and throttled.
Analysis of xAI's Grok 4.7 found the coding gains come with substantially higher reasoning-token consumption, so a model charging less per token can still cost more on a finished workload. The difference becomes material quickly at enterprise scale, reinforcing cost per completed task rather than list price as the number to compare.
Google released EnvHarness, which reshapes a trusted training environment around an agent's current weaknesses instead of requiring teams to build new simulators and tasks from scratch each time.
LeadDev reported on developers who have stopped using AI altogether. A viral "I'm done with AI" video passed 900,000 views and 10,000 comments; one CTO left his role after a founder insisted no engineer code by hand, another freelancer lost a long-standing client and missed a job offer over a no-genAI policy. An arXiv study of a software services company a year into AI adoption found "craft identity disruption, meaning and satisfaction erosion, cognitive and workload intensification, and uncertainty distress".
Lean In found women are 22% less likely than men to use AI regularly at work, with men 27% more likely to say a manager praised them for it; Deloitte put formal AI training at 91% of men against 72% of women. Harvard Business Review research across more than 1,000 engineers found identical code rated 9% worse when believed to be AI-assisted, and female engineers who used AI rated 13% less competent against 6% for men, rising to 26% among male engineers who do not use AI themselves.
Nothing here matches that. Try a broader term, another day, or clear the topic filter.