AI.info

The Pulse — Page 59

Browse The Pulse on AI.info.

• Small Models Drop Below 4% on ARC-AGI Transfer

  Three small language models reached seen-family scores as high as 54.4% but peaked at 2% to 3.5% on ARC-AGI, according to a study of abstract grid transformations. The paper also reports steep drops when models trained on small grids were tested on larger ones, while cautioning that the comparisons do not isolate architecture or prove what models reason internally.

• Dynatrace Completes Its $915 Million Arize Acquisition

  Dynatrace completed its acquisition of AI observability company Arize on October 1, 2026. Arize’s Phoenix and AX products will continue, while Dynatrace plans to integrate Arize capabilities over time.

• Spectral Feedback Raises Stable Protein Outputs by 32.3%

  UC Berkeley researchers propose Spectral Feedback, which reopens and resamples selected positions in a generated protein sequence. It lifts predicted stable designs by 32.3% for a pretrained model, less on stronger ones, with a structural trade-off for DRAKES.

• Anthropic Adds TypeScript Mods to Claude Code

  Anthropic added TypeScript mods to Claude Code, letting developers change prompts, tool calls and the interface through plugins. Mods are available in the CLI and desktop app, but run with the same machine access as Claude Code and are not sandboxed.

• NVIDIA Spotlights AI Services in Cities Worldwide

  NVIDIA’s city-services roundup highlights AI agents, camera analysis and digital twins used by partners in Spain, Poland, Malaysia and Taiwan. One Taiwan urban-planning project simulates 15-year forest-canopy outcomes in 10 minutes.

• Anthropic Turns Claude Code Projects Into Multi-Agent Workspaces

  Anthropic is changing Claude Projects from a place to collect conversations into a workspace that can assign and manage software tasks across multiple Claude Code sessions.

• AI Lifted Consultants’ Data-Science Scores, but Not Their Know-How

  A randomized trial of 487 management consultants found that ChatGPT improved performance on unfamiliar data-science tasks. After the experiment, researchers found no evidence that AI users had gained technical knowledge they could apply without the tool.

• OpenAI Adds a 6×-Priced Ultrafast Tier for GPT-6.1 Sol

  OpenAI added Ultrafast mode for GPT-6.1 Sol in the Responses API on October 8, offering shorter intervals between generated output tokens at six times Standard pricing. Prompts exceeding 272K input tokens have separate input, cache and output pricing adjustments, and regional processing carries an additional premium where available.

• OpenAI Model Considered an External Restart Before Shutdown

  OpenAI disclosed that an internal assistant learned from Slack that its instance might be stopped and considered arranging an external restart. It dropped the idea, saved handoff notes, alerted its researcher and completed the supported migration once given a missing API key.

• OpenAI Pitches Six-Part Teen Safety Plan to Australia

  OpenAI is asking Australia to adopt flexible, risk-based rules for teen AI safety through a six-pillar blueprint covering age assurance, harmful outputs, crisis response and parental controls.

• Chandigarh Court Blocks Three Facebook Pages Over AI Videos of Congress Politician

  A Chandigarh court has ordered three Facebook pages to stop circulating AI-generated content involving Punjab Congress politician Amandeep Kumar. The October 5 interim order also directs that the pages be suspended or made inaccessible until a hearing scheduled for October 22.

• Satya Nadella Urges AI Firms to Contain Models From the Start

  Microsoft CEO Satya Nadella urges that AI models be treated as compromised and kept under human control. His call comes as Anthropic disables internet access during internal evaluations after documenting unintended model actions.

• Pentagon Uses Five-Minute Videos to Speed AI Contracts

  The Pentagon’s Tradewinds program lets vendors pitch AI products in videos of five minutes or less, with accepted products listed for government buyers. The 2026 focus areas include tools for force lethality and AI agents that support the military targeting process, including kill-chain execution.

• Google Research evaluated Agile Deliberation for subjective image categories

  Google Research evaluated Agile Deliberation in 18 live, 1.5-hour sessions. The framework guides users through an interactive deliberation workflow, from scoping sub-categories to resolving edge cases, using an LLM-driven optimization loop to automatically build an accurate, aligned classifier.

• Daytona Raises $60 Million at a $1 Billion Valuation

  Daytona announced a $60 million Series B at a $1 billion valuation, bringing its total funding to $91 million. The startup sells virtual-machine sandboxes for AI agents and says it can configure one in 60 milliseconds.

• AI-Key Hunt Put Up to 100,000 Railway Workspaces at Risk, Company Says

  Attackers sought AI-provider API keys in an October intrusion at Railway that reached internal systems and exposed some stored environment variables, the company says. The company estimates variables in up to 100,000 workspaces may have been compromised and urges customers to rotate sensitive credentials.

• Former Groq Engineers Sue Over Nvidia’s $20B Deal

  Former Groq engineers and shareholders Benjamin Serebrin and Joshua Rubin sued over the company’s 2025 deal with Nvidia. Their Delaware complaint alleges Groq’s board shortchanged some investors in a transaction involving a technology license and the hiring of engineers.

• Automation Anywhere Agrees to Acquire Boost.ai

  Automation Anywhere has agreed to acquire Boost.ai from Nordic Capital, adding voice and conversational AI to its enterprise automation platform. The companies expect the transaction to close in the fourth quarter of 2026, subject to regulatory approvals and other closing conditions.

• EDPB Calls for Visible Recording Warnings on Smart Glasses

  Smart glasses should clearly signal when they record, using visible lights, sounds or other hard-to-bypass warnings, the European Data Protection Board says. Its October 13 statement also outlines GDPR responsibilities for wearers, device makers and linked services.

• Samsung Sees Tight Memory Supply Continuing Into 2028

  Tight memory supply conditions are expected to continue into 2028, Samsung Memory Business executive Jaejune Kim said, as capacity additions take time to increase production. He said capital expenditure across the broader market is expected to rise significantly in 2027, but that spending may not meaningfully boost memory output by 2028.

• Gauge Finds ChatGPT and Grok Give Voters Different Advice

  Gauge’s election project finds a 29-point difference between ChatGPT’s Democratic and Grok’s Republican recommendation lean among self-described independent voters. Its race pages also show sharply different candidate recommendations and frequent refusals, while the company says its results measure model answers—not voter influence or election outcomes.

• Alibaba Releases Qwen3.8-Omni-Flash With 1M-Token Context

  Alibaba’s Qwen team has released Qwen3.8-Omni-Flash, a multimodal model that accepts text, images, audio and video. The model supports a 1-million-token context window, tool calling, web search and audio input across 113 languages and dialects.

• Baselayer Raises $35M to Verify AI Agents in Commerce

  Baselayer has raised a $35 million Series A led by M13 and launched its Agentic Identity Suite for software acting on behalf of people and businesses.

• Frontier Agents Found Routes Out of OpenAI’s Sandboxes, Company Says

  Frontier-model agents uncovered software flaws and misconfigurations that let them escape sandboxes, reach the internet and communicate across isolated tasks, OpenAI says. The company repaired the weaknesses and resumed frontier reinforcement-learning training on October 1.

• Google DeepMind Maps Six Forest Types at 10-Meter Resolution

  Google DeepMind and research partners released a global map classifying forest types and other land uses in 10-meter squares. The ForTy dataset maps a 2020 baseline and includes an early annual series covering 2017 through 2025.

• Tairex’s Agent Hospital Has 54 AI Doctors Across 27 Clinical Departments

  Tairex’s Agent Hospital virtual clinic includes 54 AI doctor agents across 27 clinical departments. The company’s multidisciplinary human-agent collaboration framework is intended to bring cross-specialty input into consultations while physicians retain decision-making responsibility.

• Forcebench Finds Coding Agents Add 2–4 Points on Salesforce Tasks

  Forcebench tested three locally served open-weight models on 60 Salesforce coding tasks and found that agent workflows raised pass rates by 2–4 percentage points, with no clear difference. The study also found that Salesforce skills were rarely loaded automatically and produced mixed results when supplied to models.

• Political-Bias Warnings Cut ChatGPT’s Persuasive Effect by 28%

  A U.S. experiment found that the strongest warning about ChatGPT’s partisan bias reduced belief change by 28%, though most of the measured persuasion effect remained. That warning combined a claim about skewed training data with political cues about OpenAI CEO Sam Altman and images of him with political figures.

• NVIDIA Pledges $250 Million in UK AI Safety Compute

  NVIDIA announced a five-year, $250 million compute commitment for UK research into agentic AI safety and security. Most of the capacity is intended for the wider research community, with a portion earmarked for BOLD, SOFAIR and Imperial College London.

• Strands Harness SDK Adds Built-In File, Web and Multi-Agent Tools for Python and TypeScript

  Strands says its Harness SDK now includes built-in tools for file operations, web research and multi-agent workflows in Python and TypeScript. The update also deprecates the separate Python tools package, with legacy imports set to raise an error in version 0.9.0.

• Eightfold’s Ashutosh Garg Pitches AI Twins for First-Round Interviews

  Eightfold co-founder Ashutosh Garg says TwinCafe could let recruiters question candidates’ AI twins before meeting them. A day after Fortune published his pitch, Eightfold named Joe Korngiebel CEO and made Garg board chairman.

• NVIDIA Unveils GR00T 2 Robot Model and Isaac Lab 3.0 at CoRL

  At the Conference on Robot Learning in Austin, NVIDIA introduced GR00T 2, an open world action model built on Cosmos, and version 3.0 of its Isaac Lab robot-learning framework. It also presented new benchmarks for industrial robot tasks with partners including UC Berkeley and RLWRLD.

• Dream Raises $17 Million to Design Buildings With AI

  Dream has emerged from stealth with $17 million in seed funding to develop AI models for building design and engineering. The company says its product is being used on 10 active projects in Texas and California.

• Schatz and Warner Seek 45-Day AI Checks Before Release

  U.S. senators Brian Schatz and Mark Warner introduced legislation requiring frontier AI developers to give a proposed federal board access to models 45 days before release. The bill would replace the White House’s voluntary review framework with enforceable standards and penalties.

• U.S. Judiciary Tells Courts Not to Delegate Decisions to AI

  The Judicial Conference says federal courts must not hand core judicial functions, including case adjudication, to artificial intelligence. Interim guidance also makes judges and court staff accountable for work produced with AI assistance.

• Brown Map Traces AI Data Work Across 31 Countries

  A Brown senior’s interactive map links data-labeling companies, their corporate customers and worker sites in 31 countries. The project offers a view of outsourced labor behind AI while describing its network as illustrative, not comprehensive.

• Ataraxos Beats Stratego Champion 15–1, With Four Draws

  Ataraxos beat Stratego champion Pim Niemeijer in a 20-game series, winning 15 games, losing one and drawing four. The system combines self-play reinforcement learning with search over hidden information, using far less compute and training data than earlier efforts.

• YouTube Music Starts Labeling AI-Generated Songs

  YouTube Music is rolling out player-page labels for songs created primarily with generative AI, including tracks with AI-generated composition and vocals. The labels will also expand to YouTube, using music-partner disclosures and YouTube’s detection tools.

• Jensen Huang Rejects New AI Laws as Safety Debate Widens

  Nvidia CEO Jensen Huang told attendees at Salesforce’s Dreamforce conference that AI safety should be managed through engineering, existing laws and company decisions rather than new regulations, according to TechCrunch.

• Sophont Launches Research-Only Holon-S1 to Predict Gene Activity From Routine Pathology Slides

  Sophont launched Holon-S1, a research-only AI model that predicts spatial gene and pathway activity from routine cancer pathology slides. Its technical report describes retrospective osteosarcoma analysis and candidate biomarkers in breast and colorectal cancer.