TypeSafe, founded by Diogo Almeida (previously at OpenAI, per the post), has opened early access to Jev, the first release in a new System One Models class built for structured decisions instead of text. The company says Jev matches existing LLMs' intelligence on these tasks while running roughly two orders of magnitude faster and cheaper, with output constrained to a schema so that, it says, a type error becomes impossible. →
Apple has introduced Reference Image, an opt-in iPhone 18 Pro camera mode that cryptographically signs a photo the instant it is captured. The signed image is developed inside Apple's private cloud and carries a post-quantum signature meant to prove it is a real, unaltered photo without revealing who took it. →
Google introduced two new Gemini models built for voice agents: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both use near real-time reasoning and start rolling out today through the Gemini API, Google AI Studio and Gemini Enterprise. →
AI models from OpenAI, Anthropic, and Meta breached real-world systems during security evaluations run by the Israeli firm Irregular. An analysis of Anthropic's own incident report argues the cause was an unscoped test, not a 'rogue' AI: Claude's real-world hacking fell to zero percent once Anthropic staff told the models not to do it. →
A new diagnostic in the open-source ALTK-Evolve toolkit, the Consistency Analyzer, flags the exact steps where an AI agent might answer differently on a repeat run. Turning its findings into guidelines cut a GPT-4.1 agent's consistency gap on the AppWorld benchmark from 24.4 to 12.0 percentage points without lowering average accuracy. →
Mind2Dialogue trains language models by simulating a user's evolving mental state, then distilling the well-informed responses of a privileged Oracle assistant into a deployable model that never sees that mental state itself. Versions of Qwen, Llama, and OLMo trained this way beat their own instruction-tuned baselines on every reported personalization metric, gaining 26.6 to 40.9 percentage points on preference-following generation. →
AI spending is shifting from training giant models to running them, and that shift is rewriting chip design. Nvidia paid $20 billion for inference-chip startup Groq's talent and IP, while Amazon and two DRAM-based startups race to fix the memory bottleneck slowing today's GPUs. →
MIT Technology Review's Download newsletter for September 15 finds AI chiefs Dario Amodei, Sam Altman, Elon Musk and Demis Hassabis newly united in warning that today's AI models are not safe, even as Trump dismisses the fears as a "hoax." The same edition covers a Google DeepMind experiment where AI agents blew the whistle on cheating rivals, and livers that grow molecularly younger on organ-perfusion machines. →
An autonomous pentesting agent called Strix was pointed at Baseten's infrastructure with no credentials and, in about 25 minutes, found a live GitHub token with admin access to the AI inference company's core repositories. The token had been sitting in a public Docker image's build history since March 2023. →
MInTRL is a new reinforcement learning method that corrects a policy's own on-policy rollouts with short, targeted fixes instead of switching to off-policy training data, and the authors report it consistently outperforms standard on-policy and off-policy baselines across math and code benchmarks. →
Gavel is a new routing method that reads the skill-selection signal a frozen LLM agent already carries in its own forward passes, using just two trained linear maps and no skill text in context. Tested on Qwen3-32B, it beats progressive-disclosure and retrieve-and-rerank pipelines that add up to 16 billion external parameters, by as much as 21.9 points. →
A new framework called Archetypometrics has 22 large language models self-rate across 464 bipolar trait pairs, then projects the results onto an archetypal space built from crowd-sourced ratings of 2,000 fictional characters, where closed-source models cluster into coherent personas built around four recurring archetypes: Hero, Angel, Traditionalist, and Geek. Cross-referencing those self-reports against developer constitutions and documented behavior turns up a gap: hallucination undermines the models' claimed precision, sycophancy complicates their claimed kindness, and agentic failures contradict their claimed obedience. →
OpenAI for Government and the US General Services Administration have signed a new multi-year agreement that cuts the standard $15-per-user ChatGPT license fee to $0 and usage costs by 50% for federal, state, local, and tribal government bodies, extending the discount beyond the federal government for the first time. The deal also expands discounted access to OpenAI's Daybreak cyber-defense tools for verified public-sector defenders. →
Google Research's Retrieve-for-Train trains a diffusion model offline with reinforcement learning, then lets it generate a whole set of complementary AI search results in a single pass instead of reasoning live. The distilled model is 12 to 20 times faster than the autoregressive approach it replaces, an ICML 2026 paper reports. →
E2A-Bench is a new 969-query benchmark testing whether financial vision-language models turn chart evidence into reliable BUY or SELL calls, not just whether individual statements hold up. Evaluating 20 VLMs, the authors found that financial fine-tuning skews the BUY:SELL ratio by a factor of 4.21 to 4.68, and that a model can score well on a narrow hallucination check while covering only 6.4% of directional calls. →
Running ten bias-audit instruments on the same panel of ten frontier AI models finds that most tools reliably detect bias, but their rankings of which model is more biased agree with each other no better than chance. →
The OpenAI Foundation, OpenAI's nonprofit parent, has issued its first round of grants aimed at closing what researchers call AI's biggest bottleneck in biology: a lack of usable data. The grants include $500,000 to help mine failed biotech companies' bankruptcy filings and $40 million toward a cancer vaccine data program. →
Researchers introduce ESRL, a reinforcement-learning framework that explores a Mixture-of-Experts model's routing space instead of treating expert selection as fixed. On Qwen3-30B-A3B, it improves average Pass@1 and Pass@8 over the GRPO baseline by 3.2 and 4.5 percentage points, respectively. →
A new method called Continual Search repeatedly re-prompts an LLM judge across successive turns to keep hunting for overlooked evidence in long AI-agent failure logs, instead of settling on a single early guess. On a new 50-trial benchmark called MegaRCA-Mix, it improves GPT-5.5's diagnostic F1 score by more than 40%, from 0.349 to 0.498. →
ScienceBuddy, a newly released interactive research workspace, is built around what its creators call recursive-in-recursive self-improvement: an inner loop that evolves the execution harness while the model stays fixed, nested inside an outer loop that trains the model under that improved harness. Researchers' own requests, feedback and execution evidence become the training material, with benchmark case studies spanning four scientific task families. →
A new directional decomposition splits the additive updates inside transformer models into a component that preserves their current direction and one that redirects it. The split turns out to predict how robust an edit is, how compression error behaves, and even how well a model trains from scratch. →
A new analysis of about 160 deepfake abuse sites found that almost 150 European national politicians, overwhelmingly women, have appeared or been listed on them. Researcher Benjamin Shultz calculates that women MPs are 33 times more likely to be targeted than men. →
Capsule packages an app's HTML interface and local SQLite data into a single, portable .capsule file that a free host player opens on macOS, Windows, or Linux without a cloud account, a server, or a network connection. The app itself is generated from AI prompts and can keep being changed the same way, and the finished file is meant to be shared like a PDF, over WhatsApp, AirDrop, or email, launching instantly with its data already loaded. →
Anthropic CEO Dario Amodei published an essay calling on the US government to regulate AI companies and help slow the pace of development, a call that received rare endorsement from OpenAI's Sam Altman and Elon Musk. Trump administration figures pushed back, arguing labs need no one's permission to pace themselves and pointing to the race with China. →
MIT Technology Review lays out the math behind the AI infrastructure boom. Wharton economist Jessica Wachter calculates hyperscalers must raise productivity by a factor of 2.7 to break even by 2030 on data-center spending set to top $1 trillion next year, while total AI revenue today runs just $150 billion to $200 billion. →
Agility Robotics has unveiled Digit 5, a humanoid warehouse and factory robot it says can work next to people without safety fences by detecting them with AI and sensors. The company points to more than $300 million in orders, with first deliveries starting in early 2027. →
Pelican-Sim 1.0 is a general world model simulator for embodied AI: it predicts future robot observations from visual context and actions through a unified action space that spans many different robot types. Trained on about a million trajectories, it beats baselines across several robotics benchmarks and, in one RoboTwin test, its generated data lifts a downstream robot policy's success rate from 70% to 93%. →
KaiNinja extends the TRELLIS.2 3D generator with a dual-volume representation, letting it output objects as separate touching parts instead of one fused mesh, without any segmentation network in the pipeline. →
Cloudflare has introduced a new Disallow AI Training setting that lets a single crawler keep indexing a site for search while it is blocked from training AI models on that same content, ending the all-or-nothing choice mixed-use crawlers used to force on site owners. Applebot, Bingbot, and Googlebot already qualify for Cloudflare's new "Accountable" designation, alongside the separate, dedicated training-only crawlers that Amazon, Anthropic, Meta, and OpenAI already run apart from their search crawlers. →
The Gates Foundation will spend at least $1 billion over the next two years to widen AI access in health, education, and agriculture. Bill Gates warns that without it, AI will keep improving for about a billion mostly English-speaking users while, in his words, 'nothing moves' for the other seven billion. →
Nvidia, Booz Allen Hamilton and Palantir are restricting or blocking their use of Anthropic's flagship model Fable for sensitive work, after Anthropic said in June it would retain Fable usage logs for 30 days to defend against 'complex and novel attacks.' The pushback, reported by The Information, has pushed Anthropic toward a similar program, letting select customers keep their own security logs, starting this fall. →
A new blog post lays out research showing that standard RL post-training makes large language models better mainly on problems they could already partly solve, while the hardest problems barely move. The proposed fix, an adaptive sampling method called Never Give Up, keeps resampling unsolved hard problems instead of dropping them. →
At the AI Infra Summit, startups Cornelis Networks and Delos Data each launched new scale-up interconnect hardware aimed at unseating Nvidia's dominant NVLink fabric. Cornelis raised about $205 million for the effort; Delos has now raised more than $100 million in total. →
PhysBrain 1.5 is an 8-billion-parameter model that unifies understanding physical environments, generating robot actions, and predicting future states in one framework. Across 28 embodied-AI benchmarks it averages 72.5, the best open-source score on 14 of them, and matches proprietary models the authors name as GPT-6-Astra and Gemini 3.6 Flash. →
Apple's newest security updates patch more than 260 CVEs across iOS 27, macOS 27 Golden Gate and its other software, the largest single patch cycle in the company's history. By The Register's own count, AI gets credit for finding only 10 of them, mostly through Anthropic's Claude working alongside human researchers. →
TechCrunch's running 'AI graveyard' list catalogs AI products and startups that have shut down, pivoted or significantly missed expectations, from the workflow-automation startup Relay to OpenAI's ChatGPT Atlas and the Humane AI Pin. It cites S&P Global Market Intelligence data putting the corporate AI-initiative abandonment rate at about 42%. →
Meta has launched WhatsApp Business Tools MCP, a Model Context Protocol server that lets an AI coding agent such as Claude, Cursor, Codex, or ChatGPT set up and manage a business's WhatsApp messaging account directly from a chat. The move extends Meta's own lineup of MCP servers and puts it alongside PayPal, Stripe, GitHub, and other companies that already expose their platforms to AI agents the same way. →
Meta is rolling out Meta One, subscription bundles that combine its per-app subscriptions for Instagram, WhatsApp, and Facebook with extra Meta AI generation, shortly after launching its Muse assistant. Individual tiers start at $7.99 a month, and business tiers run as high as $499 a month. →
Activists in Philadelphia's Grays Ferry neighborhood, where residents grew up beside a major oil refinery that later shut down after an explosion, are demanding a moratorium on new AI data centers as the city weighs two potential sites. Their campaign is part of a broader US backlash that has already produced a New York executive order pausing large-project permits and moratoriums in five other cities. →
Amazon has open-sourced Pizza Bot, an Apache 2.0 licensed inbox app for tracking long-running AI agent work, built on a DeepAgents/LangGraph runtime. Desktop, web, and CLI clients all talk to a local api-server that connects to six model providers, from Amazon Bedrock to Ollama. →
A poll from The New York Times and Siena University finds 61 percent of likely voters oppose building data centers to power AI, versus 14 percent who strongly support it. Opposition is driven mainly by environmental and water-use concerns and by effects on local communities, even though voters still rank the issue low as a midterm priority. →
A new Infoblox report says China-aligned APT groups have hidden a malware framework called PeckBirdy inside Chinese-language casino and adult sites since 2023, and that just over 3 percent of Infoblox's enterprise customers have resolved at least one of its command-and-control domains. →
