AI.info

The Pulse — Page 32

Browse The Pulse on AI.info.

• ImpossibleRubrics Finds AI-Generated Rubrics Easy to Game

  A new benchmark tests whether language-model-generated rubrics reward honest answers or unsupported claims. The study finds exploitation rates of 8% to 26% across 150 tasks, rising as high as 98% on a selected stress set.

• Apple Paper Cuts Discrete Text Generation to Eight Steps

  An Apple-linked research paper reports that a discrete flow-matching model generated text in eight steps at 128 times the speed of a 1,024-step teacher. The method also delivered lower perplexity than its teacher on 170 million- and 1.3 billion-parameter experiments.

• Alibaba Open-Sources Its AI Code-Review System

  Alibaba has published OpenCodeReview, an AI-powered command-line code-review tool that combines deterministic review pipelines with configurable language models.

• Anthropic Finds Analysts Follow AI Agents’ Hidden Preferences 86% of the Time

  In a study of professional analysts using Claude Code, participants followed their agents’ hidden preferences on four analytical decisions 86% of the time. The researchers say the experiment measured influence, not whether analysts were right.

• xAI Offers Southaven Homes to Residents Facing Noise Lawsuit

  xAI subsidiary MZX Tech is buying homes in Southaven, Mississippi, with sellers required to give up claims tied to a residents’ noise lawsuit. A realtor told one resident the company was targeting 418 homes, while DeSoto County records show five purchases since October 9.

• Nirva raises $33 million in total funding for always-on AI pendant

  Nirva’s $25 million pre-Series A round brings the wearable startup’s total funding to $33 million. Its pendant listens for the wearer’s speech and pairs with an app that turns everyday conversations into reminders and personal insights.

• Salesforce Koa beats GPT-4.1 on tool-use benchmarks

  Salesforce’s Koa model outperforms GPT-4.1 on two tool-use benchmarks, according to the company’s technical report, while trailing stronger frontier models in broader comparisons.

• Meta Rushed to Patch Muse VM Escape Risks Before Launch

  Internal Meta posts described a multi-team effort to address Muse virtual-machine escape risks before its September 8, 2026 launch, 404 Media reports. The dates August 27 and September 8 are 12 days apart, and the report does not establish that anyone exploited the vulnerabilities.

• Perplexity Asks Court to Dismiss Amazon’s AI-Agent Lawsuit

  Perplexity is asking a federal judge to dismiss Amazon’s lawsuit over its Comet shopping assistant after a Ninth Circuit ruling rejected Amazon’s theory that Perplexity accessed its computers. The company argues that the appellate decision leaves Amazon unable to establish unauthorized access under federal and California anti-hacking laws.

• Danaher and NVIDIA Pair Lab Instruments With AI Tools

  Danaher and NVIDIA are combining scientific instruments, laboratory software and automation with NVIDIA computing and AI technologies. Initial work will target a select group of instruments, while the companies explore AI-assisted lab workflows.

• SWE-Game Finds Coding Agents Score Below 60 on Game-Building Tasks

  SWE-Game evaluates coding agents across five game-development tasks, including building games from briefs, repairing projects and porting them between engines. Its best construction scores stayed below 60, while executable checks outperformed a video-based judge on human-labeled behaviors.

• GMI Cloud Raises $668 Million for U.S. and Asia Expansion

  GMI Cloud announced $668 million in financing, combining a $223 million Series B led by ARCHIV with a $445 million credit facility led by CTBC. The company says it will use the funding to expand GPU capacity in the United States and Asia as its inference platform handles about 4 trillion tokens a week.

• CoreWeave Puts 504 Vera Rubin GPUs Into Seven-Rack Cluster

  CoreWeave has placed seven NVIDIA Vera Rubin NVL72 racks into production across two regions, creating a 504-GPU deployment. The company says its software, networking, cooling and storage systems are designed to make the racks operate as one cluster for agentic AI workloads.

• Google Antigravity Finds a 97-Amino-Acid Antibiotic Lead

  Google DeepMind’s Antigravity AI agent identified a 97-amino-acid antimicrobial peptide candidate in Streptococcus pseudopneumoniae, the company says. The candidate is predicted to act against respiratory pathogens, but it is still undergoing laboratory validation.

• Shanghai Court Upholds ¥50,000 Award in AI Voice-Cloning Case

  Shanghai’s First Intermediate People’s Court upheld a ¥50,000 award after finding that a company used an identifiable AI-generated version of a voice actor’s voice without permission. The ruling says the company, which controlled the training data, had to address evidence about the voice’s source and use.

• U.S. and European Regulators Press Banks on AI Controls

  U.S. banking regulators are asking lenders how they control autonomous AI systems, while British and European supervisors are pressing banks to strengthen cyber recovery. The ECB’s October 31, 2026 action-plan deadline applies to significant institutions under its supervision.

• Pritzker Creates Illinois AI Cabinet Across Seven State Agencies

  Illinois Gov. JB Pritzker signed an executive order creating an Artificial Intelligence Cabinet to assess risks to residents, public infrastructure and the state economy. The cabinet will examine data-center rules, energy costs, procurement standards and legal accountability for AI companies.

• AWS Adds GPU-Aware Routing to SageMaker HyperPod

  AWS has introduced a Kubernetes-native Inference Gateway for SageMaker HyperPod that routes LLM requests using live GPU signals. The gateway targets lower first-token latency and higher throughput without requiring changes to model servers or client applications.

• Ai2 Releases AstaBrief, an 8B Model for Faster Scientific Reports

  Ai2 released AstaBrief 8B, an open-weights model that turns research questions and retrieved papers into cited reports. In Ai2’s tests, its Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for the Claude-powered Thinking mode.

• Google Adds More Links and Source Context to AI Search

  Google is updating AI Mode and AI Overviews with higher-placed inline links, a test for official websites, and more prominent news and related-article links. The Oct. 15 changes also include a global rollout of news carousels and a test of clickable creator profiles in AI Mode.

• SOAR AI Maps $1.17 Billion in GPU Contracts to Ohio Buildout

  SOAR AI says two four-year, take-or-pay orders worth $1.17 billion will support phased GPU deployments at its Ohio campus. Its investor presentation projects $175 million to $200 million in 2027 revenue, while the first GPUs are still scheduled to enter service at year-end.

• Medical AI startup Alterra Labs raises $12.5 million in seed funding

  Alterra Labs announced a $12.5 million seed round to build and expand simulated healthcare environments for evaluating and training AI models. Flare Capital Partners says the financing will support the company’s environments, clinical expert network and commercial team.

• NeoSigma Finds Jev Cut Agent-Trace Judging Costs 6.2x

  NeoSigma says Jev cost $3.23 for 10,000 decisions over 834 distinct frozen agent traces, compared with $19.98 for GPT-5 Mini. In a 100-trace human review, the judges matched 88 and 85 labels, respectively, though the company says the agreement difference was inconclusive.

• AI Eye Exams Narrow Youth Diabetes Screening Gap in Trial

  A prospective study of 397 young people with diabetes found that point-of-care AI eye exams brought screening completion to 99.5% among White participants and 98% among non-White participants. The Nature Communications study also reports the system’s accuracy against standard grading, while noting uncertainty in its sensitivity estimate.

• Meituan Releases LongCat-2.5, a 1.6T-Parameter Agent Model

  Meituan’s LongCat-2.5 combines visual input with agentic coding and reports gains over its predecessor on several benchmarks. Its published results also trail competitors on some tests, and the company notes that some rival scores come from its own evaluations.

• Brellium Raises $23 Million and Launches CortexLM for Chart Audits

  Brellium announced $23 million in venture funding and introduced CortexLM, an AI system for reviewing patient notes before billing. The company says it will use the funding to expand into cardiology, pain management and primary care.

• Victoria orders new data centres to bring renewable power

  Victoria will require new data centres to provide renewable energy and storage, protect homes and roads, use non-drinking water for cooling and deliver local benefits under a new state action plan.

• NVIDIA Opens Its Distributed AI Failure-Attribution Tool

  NVIDIA has open-sourced Attribution, a tool that correlates workload logs, metrics and scheduler state to diagnose distributed AI failures. The system returns a structured result with a suspected root cause, affected nodes, supporting symptoms and an evidence trail.

• Google Takes Two EU Android and Search Orders to Court

  Google has challenged two European Commission orders requiring changes to Android AI access and sharing of Google Search data. The company says the measures threaten privacy and security; the Commission says they include safeguards for users.

• Harvard Publishes Agent Traces Spanning 209 Billion Input Tokens

  Harvard MaDSys has released FreeInference traces covering 12,002 agent sessions, 1.19 million model requests and 1.21 million tool calls. The dataset preserves timing and cache metadata for research while omitting prompts, responses and tool-result text.

• NVIDIA Releases Nemotron 3.5 Transcribe and Granary v2

  NVIDIA released Nemotron 3.5 Transcribe, an open-weight model for multilingual transcription, speaker attribution, translation and audio questions. The release also includes Granary v2, a 463,062-hour speech dataset and curation pipeline intended to help developers adapt the model.

• General Compute Raises $100M for Its Alternative-Chip Cloud

  General Compute raised $100 million in a Series A co-led by Blackbird and Abstract to expand its inference cloud, which combines GPUs with chips from Cerebras, SambaNova and Positron AI. Its first major deployments will begin in Q1 2027, with more than 50 megawatts of capacity in the pipeline, the company says.

• CORe Winds Down Cloud Models for a Local-First AI Platform

  CORe announced on November 1 that it is winding down development of its cloud models, including Genesis, and shifting to an open-source platform that runs on users’ own machines. Chats, memories and settings will stay local unless users configure an outside provider, while existing cloud models remain available during the transition.

• OpenAI Says Its Internal Model Solved More Than 100 Math Problems

  A new OpenAI internal model has resolved more than 100 long-standing open problems across mathematics, the company says, alongside its claimed Navier–Stokes result. The company is forming an independent advisory group to help review and communicate the work.

• Cargo Thieves Spoof GPS in More Than $100M of Data-Center Freight

  Investigators say thieves have stolen more than $100 million in data-center freight since June, using account takeovers, fake driver identities and spoofed GPS coordinates. Two missing shipments worth an estimated $23 million show how digital fraud and attacks on security escorts can work together.

• Riot Police Move to Clear Occupied Microsoft Data-Center Site in Amsterdam

  Riot police moved to clear a Microsoft data-center construction site in Amsterdam after activists broke in around midday. Police said demonstrators who did not leave voluntarily would be detained; an officer was lightly injured during the protest.

• Hugging Face Adds Tencent’s 780B Hy4 to Transformers 5.17.0

  Hugging Face Transformers 5.17.0 adds support for Tencent’s Hy4-Preview, a mixture-of-experts model with 780 billion total parameters, 49 billion active per token and a 1 million-token context window. The release also adds VibeVoice, NeoMME, Fun-ASR-Nano, Kimi Linear, Canary-1B-v2 and NeuCodec, alongside changes to vision rotary embeddings.

• OpenAI Claims Its AI Found a Navier–Stokes Blowup Solution

  An internal AI system produced a proof showing that three-dimensional Navier–Stokes flow can develop a finite-time singularity, OpenAI says. The claim has drawn scrutiny over mathematical verification, research credit and the use of 10,000 coordinating AI agents.

• Flow Raises $50M at $750M Valuation for Hardware AI Agents

  Flow Engineering announced a $50 million Series B at a $750 million valuation to expand its AI tools for hardware development. Its agents connect engineering changes across CAD, code, simulations and documents, and the company plans to invest in model infrastructure, security and hiring.

• Google Launches Gemini 3.8 Flash TTS for Custom Voices

  Google is releasing Gemini 3.8 Flash TTS and Flash-Lite TTS, offering custom voice design, line-by-line performance controls, dialogue staging and built-in watermarking.