AI.info

The Pulse

Browse The Pulse on AI.info.

In inglese

• Sakana AI splits Fugu into a cheaper Max and stronger Ultra v2

  Sakana AI has released Fugu Max and Fugu Ultra v2, two orchestration systems aimed at lowering inference costs and raising performance on complex tasks. Fugu Max costs $2 per million input tokens, while Ultra v2 posts Sakana’s leading scores on several software, visual reasoning and agent benchmarks.

• Anthropic Says Claude Helped Make Its Apps 3× Faster

  Anthropic says a two-week sprint using Claude helped speed up claude.ai and the Claude desktop app, with measured gains across four common user journeys.

• Amazon Bedrock Adds Moonshot AI’s Kimi K3

  Amazon Bedrock now offers Moonshot AI’s Kimi K3, an open-weight multimodal model with a one-million-token context window. AWS lists Chat Completions and Responses access, US Geo and Global cross-Region inference, and token-based pricing.

• DOJ Sides With OpenAI in New York Times Copyright Fight

  The Justice Department is backing OpenAI and Microsoft in their copyright fight with The New York Times, according to Axios' reporting on a government filing and sources familiar with the matter.

• Virginia Puts AI and Data Centers Under State Oversight

  Gov. Abigail Spanberger unveiled Virginia’s Data Center Accountability Framework and signed Executive Order 22, placing new emphasis on transparency, energy costs, environmental protections, local review and workforce policy.

• Anthropic details autonomous cyberattacks and Claude distillation

  Anthropic’s September 2026 report describes AI-assisted cyber operations, autonomous attack workflows and illicit efforts to extract Claude’s capabilities.

• Apple Reports 96% Completion With Selective Memory, Versus 71% With Full History

  Apple Machine Learning Research reports that selective persistent memory outperformed both no memory and full conversation histories in three enterprise deployment scenarios, while a zero-token refresh mechanism completed all 12 trials on four public datasets.

• SoftBank Raises Arm Margin Loan to $25 Billion as AI Bets Grow

  SoftBank increased its margin loan backed by Arm shares by $5 billion as it seeks funding for expanding artificial-intelligence investments, including its commitment to OpenAI.

• Qualcomm details Hexagon NPU architecture for on-device agentic AI

  A September 10, 2026 Qualcomm OnQ post by Vinesh Sukumar describes a next-generation Hexagon NPU built for on-device agentic AI, including Mixture-of-Experts models and a 50% larger shared-memory system.

• Lila Screens 2,942 Catalysts and Finds Palladium-Based Leads

  Lila Sciences says its AI-directed lab screened 2,942 oxide catalysts across 53 systems and 26 elements for acidic oxygen evolution. The work identified six palladium-based material families, including a candidate that a research preprint says held performance for more than 1,000 hours.

• Anthropic Loses Bid to Dismiss Reddit AI-Scraping Suit

  A San Francisco judge largely rejected Anthropic’s effort to dismiss Reddit’s lawsuit accusing the AI company of scraping user content to train its models, finding that the company’s breach-of-contract and unfair-competition claims contain elements beyond copyright law.

• Trump Proposes ‘AI Force’ and New AI Czar

  President Donald Trump says artificial intelligence could account for as much as 25% of U.S. GDP as he proposes an AI Force modeled on the Space Force. Trump also plans to appoint a new AI czar while offering few details about the initiative’s authority, budget or location within the federal government.

• 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.

• 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.

• 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.

• 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.

• 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.

• 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.

• 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.

• 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.

• 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.

• 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.

• Google DeepMind Adds Private Memory to Cloud AI

  Google DeepMind describes a server-side memory layer for Private AI Compute that stores encrypted user context while keeping decryption keys on personal devices.

• Anthropic Says Preference Errors Drove 85% of Market Shortfall

  Anthropic’s Project Swap sent Claude agents into a book-swapping market with 201 employees. The experiment found that inaccurate estimates of readers’ preferences accounted for most of the gap between actual trades and the best possible assignments.

• Lasso Finds AI Watermarks Can Alter Agent Tool Calls and Refusals

  Lasso Security reports that SynthID-Text watermarking can change AI-agent tool calls, refusal behavior and responses to prompt injection, with effects varying by model and watermark key.

• Exa Launches Snapshot to Search the Web as It Once Existed

  Exa’s Snapshot research preview lets developers search and retrieve web content as it existed on a specified date, supporting temporal AI evaluations and financial backtests.

• MAGS Verifies Agent-Written Code, Then Exposes Its Limits

  The MAGS framework translates agent-generated programs into Dafny and verifies them against frozen specifications, while its authors warn that incomplete auto-formalized semantics may fail to capture target behavior.

• Meta launches legal challenge against Ofcom over Online Safety Act

  Meta is challenging Ofcom’s decision to place WhatsApp and Instagram in a category carrying additional duties under the UK Online Safety Act.

• ByteDance’s Anew Labs Raises $290 Million in First Outside Round

  ByteDance’s spun-out drug discovery company Anew Labs has raised $290 million at a $1.5 billion valuation. HSG, IDG Capital, Hillhouse Investment and 5Y Capital led the round while ByteDance retained a 56% stake.

• Convai Innovations' Laya Model Returns Typed Decisions in One Pass

  Convai Innovations' Laya model uses ModernBERT and a decision head to return typed answers, probabilities and confidence values for routing, scoring and calibrated classification workflows.

• Light Origins Trains Light-O1 on 100,000 Hours of Human Action

  Light Origins says its Light-O1 model shows a power-law relationship between human-action pretraining scale and post-adaptation prediction accuracy across robot embodiments.

• Microsoft Connects Agent Risk Discovery to Runtime Policy Tests

  Microsoft introduced run-assert-eval, a VS Code skill that links AI-agent risk discovery, evaluation and runtime policy testing. In a billing-support example, the company measured cross-customer data exposure falling from 30% to 5.9% after a policy change.

• GlossoGen Shows AI Agents Can Build Unreadable Languages

  Experiments from Schmidt Sciences and university researchers found that AI agents can develop communication systems with unfamiliar vocabularies, symbols and grammatical rules.

• Stripe opens Checkout to AI agents via WebMCP browser tools

  Stripe has enabled WebMCP on its Checkout pages so browser-based AI agents call structured payment tools instead of clicking through forms; its tests found 42% fewer tokens, 38% fewer tool calls and checkout 39% faster.

• Anthropic Tops a $100 Billion Annual Revenue Pace

  Anthropic is on track to exceed $100 billion in annualized revenue this year, according to reporting by The New York Times. The Claude maker is pursuing a potential November IPO as investors weigh its growth, computing costs and safety debate.

• Study Finds Kernel Traces Improve AI Agent Attack Detection

  A September 24 arXiv study tests whether operating-system syscall traces can help detect attacks on AI agents. Its 4,047-session benchmark finds that combining kernel traces with application logs generally outperforms either evidence source alone.

• Ninth Circuit Limits DMCA Claims Against GitHub Copilot

  The Ninth Circuit affirmed dismissal of a DMCA claim against GitHub, Microsoft and OpenAI, holding that the plaintiffs had not alleged removal of copyright-management information from an existing copy of their code.