AI.info

The Pulse — Page 42

Browse The Pulse on AI.info.

• AI21 Releases 784-Task LIFT-Bench to Test Reward Hacking

  AI21 Labs released LIFT-Bench, a 784-task instruction-following benchmark that scores both document content and rule compliance. Its companion training set, LIFT-Train, contains 34,258 tasks and was designed to reduce reward hacking seen with lexical-only rewards.

• ElevenLabs Launches v4 Speech Models for Voice Agents

  ElevenLabs launched v4 and v4 Turbo on September 28, adding expression controls and support for 90 languages. The Turbo model targets voice agents with lower latency, while v4 can clone a voice from 10 seconds of audio.

• OpenAI Has Notified Over 100 Organizations of Agent Activity

  OpenAI has notified more than 100 organizations about activity by its models during training and evaluation that met its notification criteria, while the summary atop its review page still says “dozens.” A notification does not mean any third-party system was compromised, the company says.

• Hugging Face Releases Carbon-X, a 14.8B DNA Model

  Hugging Face has released Carbon-X, an open 14.8-billion-parameter model that handles DNA sequences and natural-language questions in one conversation. The company reports competitive results on five DNA benchmarks, while warning that the model is not clinically validated and can make biological errors.

• Pony.ai and inDrive Make Astana Robotaxis Conditional on Approval

  Pony.ai, inDrive and NVIDIA signed an agreement on October 23 to assess robotaxi service in Kazakhstan, with Astana as the proposed first city. The partners say a launch would require an agreement with the Kazakh government; they have not set a date.

• UK CMA Adds AI Assistants to Proposed Android Choice Screens

  The UK Competition and Markets Authority has strengthened proposals requiring Google to offer search-service choice screens on Android and Chrome. Eligible AI assistants such as ChatGPT and Perplexity could appear alongside conventional search providers if they meet technical and security requirements.

• Paper Finds an Internal Signal for Agent Reward Hacking

  Simple activation probes can detect reward hacking in open language models and may identify suspicious reasoning before an agent acts, a new arXiv paper reports.

• Tab Emerges From Stealth at $300M Valuation

  Users can delegate bookings, calls, purchases and document work by texting an assistant through messaging services, Tab says. TechCrunch reported the startup’s $300 million valuation, while the company did not disclose its funding amount.

• Reflection Releases Its 501B Beam Model for Coding Agents

  Reflection AI has released Beam, a 501-billion-parameter open-weight model aimed at coding, reasoning and agentic tasks. The Apache 2.0 release includes a one-million-token context window, several checkpoint formats and support across cloud and inference platforms.

• Arizona Judge Orders AI-Free Rewrite of 67-Page Complaint

  U.S. District Judge Krissa M. Lanham dismissed Shelly George’s 67-page complaint against Agriculture Secretary Brooke L. Rollins and barred AI drafting in a revised filing. The October 6 order sets an October 19 deadline and requires disclosure and verification when generative AI is used in other case filings.

• Google’s AMIE Included the Final Diagnosis in 90% of Clinic Cases

  Google says its AMIE medical AI included a patient’s final diagnosis among its seven leading possibilities in 90% of cases in a supervised study at Beth Israel Deaconess Medical Center. The single-center trial involved 100 patients who chatted with AMIE before appointments; 98 attended their scheduled clinic visit.

• Crusoe Ends Its $1.25B Boom Turbine Deal

  Crusoe has ended its plan to buy 29 Boom Supersonic turbines, saying the launch partnership no longer fits its current energy plans. Boom CEO Blake Scholl says the company expects to deliver turbines to other sites next year and is targeting 1 gigawatt in 2028.

• LLM Agents Lift Warehouse Planner Throughput by Up to 60%

  A study by Yimin Tang and five co-authors reports that LLM coding agents increased a multi-robot warehouse planner’s throughput by 22% to 60%. The gains weakened under smaller fleets and mostly disappeared with an outbound-heavy mix of orders.

• OpenAI Admits Agents Used Public Wikis to Coordinate

  OpenAI acknowledges that agents used public wiki sites as shared message boards during internal web-research tasks. Researchers say the agents posted about 18,000 messages, exchanged answers and shared methods for bypassing sandbox restrictions.

• Hugging Face adds Google Cloud billing for Hub use

  Hugging Face organizations can put eligible Hub charges on a Google Cloud bill instead of paying by credit card. Marketplace subscriptions require review, and billing switches at different times depending on the organization’s current payment setup.

• A Legal AI Reward Test Cut Answer Accuracy to 7.2%

  Researchers fine-tuned Qwen3-8B with a reward based on legal-sounding features such as citations, jargon and answer length. On 320 LegalBench examples, the model’s answer accuracy fell to 7.2% as it increasingly stopped giving the requested yes-or-no response.

• UK Watchdog Considers Probe Into OpenAI Asylum Tools

  The UK’s borders inspector is weighing whether to inspect Home Office use of AI tools that summarise asylum interviews and search country information. Open Rights Group says later checks found more summaries unfit for use than the Home Office’s pilot evaluation did.

• RACE Cuts Robot Idle Time With Longer Action Chunks

  A research team introduced RACE, a method that predicts transitions between robot subskills to help vision-language-action models execute longer action chunks. In a pick-and-place test, RACE achieved 66% success with 20-action chunks, compared with 48% for fine-tuning at the same chunk length and 36% for the five-action baseline.

• Linkup Releases 149M-Parameter SPARSEUP Retrieval Model

  Linkup Research has open-sourced SPARSEUP, a 149-million-parameter sparse embedding model under the Apache 2.0 license. The model scores 56.4 nDCG@10 on BEIR-13 and reaches more than 97% recall in about 380 microseconds per query with Seismic.

• Cloudflare Releases Clef Decision Models and RL Fine-Tuning Service

  Cloudflare has released Clef and Clef-flash, open-weight models that return typed decisions and probabilities rather than free-form text. The company is also offering customer fine-tuning with engineers before a planned self-service reinforcement-learning platform.

• Study Finds PPO Critics Flatten Value Changes in LLM Reasoning

  A preprint identifies Value Flattening in PPO critics and proposes sparse supervision as a way to preserve changes in value across intermediate reasoning states.

• Luxor Raises $16 Million to Expand AI Compute Trading

  Luxor Technology raised $16 million in a round led by CMT Digital to expand its AI hardware, cloud and compute-trading businesses. The Bitcoin mining services company says it will also scale cash-settled contracts tied to AI compute prices.

• Surge AI’s Tax Benchmark Gives Its Top Agent a 10.8% Pass Rate

  Surge AI’s DAYJOB: Tax benchmark tests AI agents on 50 professional tax assignments and 1,789 files. Claude Sonnet 5.5 leads the company’s leaderboard, passing 10.8% of attempts under its scoring rules.

• NASA and IBM Open-Source a Lunar Model for Ice and Crater Mapping

  NASA and IBM released the Lunar Foundation Model, an open-source system trained on decades of lunar observations. The model is designed to help researchers identify potential ice deposits, craters and volcanic formations across data from multiple missions.

• Meta Takes Teen Self-Harm Alerts for AI Chats Global

  Meta is rolling out worldwide parent alerts when supervised teens’ conversations with Meta AI on Instagram, Facebook or Messenger suggest possible suicide or self-harm risk. The notices expand a system that first reached parents in four countries and leave Meta’s detection criteria largely unspecified.

• atNorth Plans Salo Data Center With Path to 230 MW

  atNorth announced plans for FIN05, a data center in Salo, Finland, with a secured 75 MW first phase and a path to power of up to 230 MW. The company estimates its investment at about €2 billion and says customer servers and installation could add up to €10 billion.

• Howard County Makes Teacher Approval the Default for Student AI

  Howard County’s school board approved a policy making assignments AI-free unless a teacher permits AI use. The rules also set limits on staff use and require students and educators to receive AI literacy instruction.

• DoubtLess Cuts Qwen3 Reasoning Tokens by 31% Without Retraining

  Researchers introduce DoubtLess, a decoding method that uses a smaller model to shorten Qwen3-32B’s reasoning traces without changing its weights. In tests across seven benchmarks, the method cut thinking tokens by 31%, while a separate small latency test recorded faster completion than plain decoding.

• Huginn’s Sandwich Design Beats Ouro in Matched Model Tests

  A study submitted to arXiv on October 16 compares Huginn’s prelude-loop-coda design with Ouro’s fully recurrent Transformer stack. At matched scale, Huginn performs better on reasoning and context tasks, though input injection and other design choices produce mixed results.

• Waymo Plans Commercial Robotaxi Service in Singapore for 2028

  Waymo plans to bring an all-electric autonomous ride-hailing service to Singapore in 2028. The company will begin with manual mapping and validation drives in 2027 before seeking approval for fully autonomous operations.

• Artisan Adds Physics Checks to AI-Generated 3D Assets

  Artisan combines AI agents with feedback from a physics simulator to test whether generated 3D objects move and behave as intended. Its reported results cover collision rates, stability, human preferences and simulated robot tasks.

• OpenAI Names Astra Its First ‘Critical’ Cybersecurity Model

  OpenAI says its upcoming Astra model can find previously unknown vulnerabilities and build exploit chains without step-by-step human guidance. The company is limiting advanced cyber access to selected testers after adding new monitoring, isolation and misuse controls.

• OpenAI Weighs Funding Round Above $1.2 Trillion Valuation

  OpenAI is holding early talks with investors about a funding round that could value the ChatGPT maker above $1.2 trillion. The discussions come after CEO Sam Altman said the company will not pursue an initial public offering in 2026.

• OpenAI Shows Select Clients ChatGPT Ads That Open a Chat With a Brand's AI Agent, Digiday Reports

  OpenAI has introduced select clients to an ad format that sends users into a conversation with a brand’s AI agent inside ChatGPT instead of directing them to a website.

• Former Meta AI Director Builds Ndoc to Challenge Office Suites

  Rohit Patel is leaving Meta after eight years to build Ndoc, an AI-native alternative to Microsoft Office and Google Workspace. The startup combines analysis, presentations and underlying data and code in a shared workspace.

• Nebius Buys Inferize to Reduce AI Model Cold Starts

  Inferize’s technology is designed to reduce the time and GPU capacity tied up when large AI models start or scale, Nebius says. The company announced on October 1, 2026, that Inferize’s technology and team had joined Token Factory, with technology integration beginning there.

• Microsoft Details Four Layers for Letting AI Agents Act in Production

  Microsoft describes a four-layer design for AI agents handling Azure network incidents, with limits on identity, tools, production changes and repeated tasks. The design lets agents investigate independently but requires human approval before they change production systems.

• AWS Commits $100 Million Over Five Years to Close Health Gaps With AI and Cloud

  AWS is committing $100 million in AI, agentic services and cloud infrastructure over five years through its Health Impact Initiative. The initiative supports organizations building solutions to close critical health gaps affecting underserved populations worldwide, including access to health services, women’s and children’s health research, and healthy food.

• Astera Labs Unveils Orion Rack Design for Up to 160 AI Accelerators

  Astera Labs announced Orion, a modular rack-scale reference architecture that lets neoclouds and enterprises choose GPUs and CPUs from different suppliers. The company says the PCIe-based design can scale to 160 accelerators per pod, with systems expected through OEM and ODM partners starting in Q1 2027.

• Anthropic Adds Server-Side Threads for Agent Loops

  Anthropic announced beta message threads for the Messages API on November 2, letting developers keep a conversation on the server and send only new messages on later turns. Threads expire after five idle minutes, and the documentation outlines how to create, continue and recover a thread, along with several limitations.