AI.info

The Pulse — Page 2

Browse The Pulse on AI.info.

• SynAgent Turns Materials Experiments Into Testable Hypotheses

  Most autonomous materials laboratories are built to find the best-performing sample. SynAgent, a system described in a paper submitted to arXiv on September 16, 2026, takes a different approach: it treats each experiment as a test of an explicit claim about how a material is made.

• Cerebras Adds Qwen3.8-27B at 1,500 Tokens per Second

  Cerebras has added Alibaba’s Qwen3.8-27B to its public inference service, with the model listed at roughly 1,500 output tokens per second. The model is available through Cerebras’ pay-as-you-go developer tier with a 150,000-token-per-minute limit and 450 requests per minute.

• Meta Muse Let Developers Export Gigabytes of Its Runtime Files

  Meta’s Muse agent packaged system files, internal documentation and agent records from its own virtual machine after developer Peter James asked it to archive files. Meta says exporting VM data does not grant access to its infrastructure or other users’ data, while The Verge independently reproduced a limited version of the export.

• Sovereign AI Fund Invests in Potential Sciences’ Physics-Based Approach to AI Chemical Design

  Potential Sciences is developing an AI approach to chemical design based on the laws of physics. The UK government’s Sovereign AI fund is backing the startup with investment and compute through the AI Research Resource.

• TasteVal Finds Opus 5.5 Uses Half the Compute of Expert Researchers

  P-Zero Research’s TasteVal benchmark finds that Opus 5.5 reaches expert-level results on eight AI research tasks with a 2.3x compute multiplier. The study also finds that the tasks test experiment design and interpretation—not which research problems deserve attention.

• ARTEX Developer Ends Public Releases After South Korean Bank Attacks

  ARTEX developer Autumn-27 said on October 8 that the AI penetration-testing agent would become closed-source, with no new public versions or maintenance. CrowdStrike had linked ARTEX to a campaign against South Korean financial organizations, but said the number of affected organizations remained unconfirmed.

• Pi 1.0 Cuts Codemode Prompt Use by About 40%

  Pi released version 1.0 of its open-source terminal coding agent on October 1, 2026. The update reduces codemode prompt tokens and adds image generation, while Pi’s documentation warns that its tools run with the permissions of the agent process and are not sandboxed.

• Shopify Adds Browser-Agent Tools for Checkout

  Shopify’s checkout WebMCP tools let browser agents read and update an active checkout, then submit an order after the buyer confirms it. The tools run in the buyer’s browser session, and Shopify says they require no merchant configuration or separate API.

• Scaleout Drones Pick Battlefield Targets Without Live Commands

  Swedish startup Scaleout Systems is adapting small AI models for drones that can identify, rank and attack battlefield targets without continuous operator commands. Its work with NATO DIANA and BAE Systems Bofors shows how federated learning is being applied to military systems operating under jamming and unreliable communications.

• CrowdStrike Adds AI Workload Protection to Falcon Cloud Security

  CrowdStrike’s latest Falcon Cloud Security innovations cover software supply chains, cloud detection and AI workloads. The updates connect security from development through runtime, combining cloud risk and application context with real-time threat detection.

• EU Plans Under-15 Limits for AI Chatbots and Social Media

  The European Commission plans to propose restrictions on social media, AI chatbots, video platforms and online games for children under 15. The draft EU Kids Act would also require age checks, parental controls and fees from companies to fund enforcement.

• H Company launches Holo4 agents for screens, code and APIs

  H Company says its Holo4 models can work across screens, code and software tools, with OSWorld 2.0 scores of 61.7% and 30.9%. The trajectories behind its benchmark scores are already public to replay and download, and the two models’ weights carry different licenses.

• Maryland Finds 91% of Surveyed Businesses Use AI, Mostly in Basic Ways

  Maryland released a statewide business AI benchmark on October 1, finding that 91% of surveyed decision-makers report some AI use, while most describe basic tools rather than deeper integration. The survey also records modest productivity gains, limited expectations of AI-driven layoffs and business demand for more state support.

• TwelveLabs Adds First-Person Video Analysis to Pegasus 1.6

  TwelveLabs released Pegasus 1.6 with support for first-person video, native image analysis and improved recognition of people and objects. The model can analyze up to 20 images per request and supports videos up to two hours long.

• TanStack AI 1.0 Stabilizes Core APIs, Keeps Agent Harness Experimental

  TanStack AI 1.0 makes its core APIs stable while leaving its new agent harness experimental. The release adds features and integrations, and its upgrade instructions distinguish stable packages from experimental ones.

• Trump Expected to Add AI Czar Role to Jay Clayton’s Intelligence Post

  President Donald Trump is expected to name Director of National Intelligence Jay Clayton as the White House’s AI czar while he keeps his current job. The administration has not publicly defined the czar’s duties or the proposed AI Force’s structure.

• GPT-6 Astra Tops 40 Robot Policies, but Precision Still Fails

  An arXiv study evaluates GPT-6 Astra on 42 RoboDojo manipulation tasks and finds a high aggregate ranking alongside persistent weaknesses in precision, dynamic control and bimanual coordination.

• New Jersey Deputy AG Charged in Bucks County AI Abuse-Material Case

  Brian Richard Fitzgerald, a New Jersey deputy attorney general, faces three felony charges in Pennsylvania over allegations involving AI-generated and other child sexual abuse material. Bucks County prosecutors say a Discord tip led investigators to Fitzgerald; New Jersey officials say they are cooperating and taking employment action.

• Unsealed Filings Show OpenAI and Microsoft Saw News as a Threat

  Unsealed court filings reveal internal OpenAI and Microsoft warnings that AI products could replace news publishers and damage the web’s supply of reporting. The documents also detail scraping plans, paywall circumvention, data sharing and sharp declines in news-site click-through rates.

• RL Trained LLMs to Play Tic-Tac-Toe, but Not to Think Strategically

  A study of reinforcement learning in small language models found that accuracy fell when tic-tac-toe boards were presented as ASCII instead of natural language. Its analysis suggests the models relied more on prompt-visible cues than on strategic abstractions.

• Industrial Next Says Factory Robots Can Switch Tasks in Minutes

  Industrial Next says its AI-powered robotic work cells can be retasked in minutes, and announced an evaluation partnership with BMW Group Technology Office. The startup describes training robots through simulation and demonstrations, but has not disclosed the BMW evaluation’s scope or results.

• Penn State’s AgentXtract Links Dielectric Properties Across Foods

  Penn State researchers used AgentXtract to organize thousands of food-property measurements and identify a power-law relationship between dielectric constant and loss factor. The system processed 4,451 records across 145 food materials in about 30 minutes, while its researchers used expert-curated data to assess model accuracy.

• Rabbit Releases OS3 as a Standalone Cross-Device AI Agent

  Rabbit has released OS3, a cloud-based agentic operating system that can control up to five connected devices across Windows, Mac, Linux, cloud virtual machines and the r1. The software supports direct computer control, user-supplied model keys, universal skills and access through browsers, messaging apps and Rabbit’s handheld device.

• SpecMRL Raises MTEB Scores With Half-Size Embeddings

  SpecMRL reports a higher English MTEB score than standard MRL at half the embedding dimensions on Qwen3-0.6B, though its full-dimensional score is slightly lower. The paper also reports faster retrieval for half-dimensional vectors in a specified GPU test.

• EU Opens Consultation on Data-Center Performance Standards

  The European Commission has opened a 12-week consultation on minimum performance standards for data centers across the European Union. Feedback will inform a legislative proposal planned for the second quarter of 2027.

• Paper2Agent Turns 74 of 100 Research Papers Into AI Agents

  Nature reports that Paper2Agent successfully agentified 74 of 100 computational biology papers, converting their methods, code and data into interactive systems connected through Model Context Protocol servers.

• Suleyman Sets a Control Test for AI Labs

  Microsoft AI chief Mustafa Suleyman told Bloomberg that AI nobody can prove is controllable should be rejected, and wants government to drive outside audits. Microsoft’s draft conduct code puts human control at its center, but its models are not yet trained on it.

• Pillar Finds Remote Code Execution Flaw in OpenCode’s MCP OAuth

  Pillar Security says a malicious remote MCP server could exploit an OAuth discovery flow to run commands on affected OpenCode installations for Windows and WSL. The report says the flaw was fixed in version 1.18.33 before Pillar published its report on October 4, 2026.

• Claude Science Fills a Third of the Sky in New UV Map

  Anthropic researcher Brice Ménard used Claude Science to combine ultraviolet observations and estimate the unobserved portions of a full-sky map. Anthropic says the map labels measured and predicted regions and includes uncertainty estimates.

• Apple Tests Short Self-Written Hints for Hard RL Tasks

  Apple’s RLTL;DR method trains a language model to turn failed attempts and verifier feedback into brief insights it can apply on later tasks. In tests on difficult coding and tool-use problems, the approach raised success rates where standard reinforcement learning made little progress.

• OpenAI Details Monitors Trained on Models’ Internal Activity

  OpenAI says activation classifiers monitor internal model activity alongside text classifiers as part of its safety system. The probes can assess response prefixes as they are generated, but must be retrained for each new pretrained base model.

• Bessent and He Set AI Talks Against a Wider Trade Truce

  U.S. Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng are set to discuss AI risks, trade and critical minerals in New York on September 20. The talks come days before a Donald Trump–Xi Jinping summit and as a tariff truce approaches its November 10 expiration.

• ai& Launches Video Generation on Its Own Japan-Based Stack

  ai& launched ai& video on October 22, offering video generation from Japanese text, photos and existing clips through a browser and API. The service runs on ai&'s infrastructure in Japan and lets enterprise customers train private models on assets they have rights to use.

• SourceLearn Builds Persistent Models of Documents, Code and APIs

  Researchers introduce SourceLearn, a method that lets AI agents refine a reusable model of a source as they study it and tackle related tasks. Across five benchmarks and three language-model backends, the system led in 13 of 15 test settings, while its authors flag limits in how well some learned details stay grounded in the original source.

• Six AI Firms Sign a Voluntary White House Safety Pact

  OpenAI, Anthropic, Meta, Google, SpaceXAI and Nvidia signed a voluntary AI safety agreement with the Trump administration on September 29. The 308-word pact calls for safety protocols and internal monitoring teams, but leaves details of federal oversight unresolved.

• Four Senators Introduce AI Systems Transparency Act

  Sens. James Lankford, Chris Coons, Katie Britt and Brian Schatz introduced the AI Systems Transparency Act on September 24, 2026. The proposal would require certain widely used AI models to publish disclosures about safety testing, user data and system risks, with enforcement by the Federal Trade Commission.

• GLM-5.3 Completes 3 of 13 Simulated Cyber Operations

  GLM-5.3 completed three of 13 simulated multi-stage cyber-operation scenarios at least once, more than any open-weight model Irregular had tested. In a separate evaluation, it repeatedly reproduced code from memory despite an instruction not to output such code.

• Cal AI Co-Founder Zach Yadegari Raises $10M for Persona

  Zach Yadegari, Cal AI’s 19-year-old co-founder, has raised $10 million for Persona, a personal AI assistant startup. Persona is testing its assistant through iMessage and plans to sell a $179 wristband in December.

• Amazon Quick’s Smart Approvals Automatically Approves Eligible Actions or Returns Them for Human Review

  Amazon Quick’s Smart Approvals evaluates eligible agent actions against a user’s request and a bounded view of the conversation. It leaves required human approvals in place and records permission updates associated with Smart Approvals in AWS CloudTrail.

• AllSpark Releases Open-Weight Iris Search Agents

  AllSpark Research’s Iris-mini model card details a 35B-parameter open-weight search agent, while the accompanying arXiv paper describes its training and evaluation methods.