AI.info

The Pulse — Page 22

Browse The Pulse on AI.info.

• Vals AI Finds Agent Teams Cost Up to 5.1 Times More

  Vals AI tested GPT 6 Sol and Claude Opus 5.5 as single agents and agent teams on 50 web-app builds. Teams cost up to 5.1 times more, but only one of four score comparisons produced a statistically significant gain.

• Itaú Signs R$2.925 Billion Deal for NeoSpace

  Itaú Unibanco agreed to buy all shares of Brazilian AI startup NeoSpace for R$2.925 billion, pending regulatory approvals. NeoSpace’s team and autonomous management will remain in place under the agreement, the bank says.

• Anthropic Open-Sources Claude’s Biology Model Optimizations

  Claude optimized more than 30 open-source models used in biomolecular research, speeding them up roughly fourfold on average, Anthropic says, while reducing memory use for large molecular systems.

• India Gives Itself a Month to Draft an AI Regulation Paper

  Ashwini Vaishnaw said India will publish an AI regulation consultation paper within a month, focusing on safety, deepfakes, skilling and inclusion. The announcement moves beyond February’s governance guidelines, which said many AI-related risks could be addressed under existing laws while calling for review of regulatory gaps.

• RoboHarm Tests Whether Robot Policies Refuse Dangerous Commands

  Robocurve has released RoboHarm, a five-task benchmark for testing whether robot policies reject unsafe physical instructions. The benchmark evaluates Claude Fable 5.1, GPT-6 Astra and MolmoAct2 across 300 trials involving blades, heat, electricity, batteries and household chemicals.

• Tanium Relaunches Security Operations Around Endpoint Behavior

  Tanium relaunched its Security Operations platform on October 6, adding behavior-based detection, fleet-wide response options and AI-guided threat hunting. The company says the product is available now and works alongside customers’ existing SIEM and EDR tools.

• Google Research presents Language Explorer for global language data

  Google Research is presenting Language Explorer, an interactive web platform for exploring linguistic diversity through visualizations and data. The tool is also featured as an interactive installation at a new Museum of Communication exhibition in Berlin.

• Open Rights Group Maps AI Across UK Immigration System

  Open Rights Group says AI is being used or tested in UK immigration casework and border operations, from asylum summaries to case prioritisation. Its new report calls for tighter oversight and a public register of government AI tools.

• Zen Technologies Agrees to Buy 51.98% of Cogknit

  Zen Technologies approved a proposed acquisition of a 51.98% stake in Cogknit Semantics for up to ₹135.15 crore. The Bengaluru company develops AI-based battlefield-management and situational-awareness systems for armed forces.

• Score Centering Keeps LLM RL Stable at 30% Under INT4 Stress

  A new Together AI paper identifies accumulated drift as a major cause of instability in large-language-model reinforcement learning. Its score-centering method reaches 30% training accuracy under severe sampler quantization, compared with 12% for truncated importance sampling.

• Epoch AI Estimates Z.ai's GLM-5.3 Trails Closed-Weight Frontier by Six to Seven Months on Cyber

  Epoch AI estimates that Z.ai’s GLM-5.3 trails the closed-weight frontier by six to seven months on its Cyber ECI, despite scoring above GPT-5.4. The estimate rests on a limited set of results, including just three for GLM-5.3.

• Codex Helped a Researcher Jailbreak Switch 2 Firmware 23.0.1

  Codex helped develop ns2jb, a tethered, software-only jailbreak for Nintendo Switch 2 firmware 23.0.1, researcher Sammy Hajhamid says. The exploit enables homebrew and game modding, but the jailbreak took about two weeks and still requires a tethered setup.

• Coding-Agent Harnesses Can Triple Costs Without Clear Score Gains

  A study of Claude Code, mini-SWE-agent and OpenCode found that harness choice changed coding-agent costs by as much as threefold, while most score differences fell within run-to-run noise. The researchers tested five models on SWE-bench Verified and found that prompt and tool overhead, along with the number of model calls, drove much of the bill.

• Science Corp. Implants PRIMA in Its First Stargardt Patient

  Science Corporation says its PRIMA retinal implant was placed in its first patient with Stargardt disease in May, during a trial at Sydney Eye Hospital. The patient saw a letter in one rehabilitation session and read a word in the next, while the implant still provides only a small, black-and-white patch of vision.

• DoorDash Has Seen Only Hundreds of Orders From Some AI Agents

  CEO Tony Xu said DoorDash had seen several hundred orders from some third-party AI agents, while everyday food orders still predominantly come through the company’s app. DoorDash is also developing its own AI ordering tools, including Ask DoorDash, Text DoorDash and an MCP connector for corporate orders.

• Liquid AI Adds DSpark Decoding to LFM2.5-VL-3B

  An experimental draft model for faster decoding is described in Liquid AI’s team article.

• Microsoft Adds Guardian Agents to Project Perception

  Guardian Agents will help security teams investigate changing risks from AI agents and recommend or apply containment, Microsoft says. The October 19 announcement also describes expanded discovery of local agents, task-scoped permissions and controls that adapt as risk changes.

• NVIDIA Releases Dory, a 4B Model With Adjustable Inference Depth

  NVIDIA’s Dory1.0-4B lets developers adjust inference depth by running a shared block of layers one to four times, trading compute for accuracy. Its benchmark results show gains on some coding tasks as loop count rises, while other scores flatten or decline.

• LHTB 2.0 Gives AI Agents Three Hours to Finish 77 Tasks

  A paper posted to arXiv on October 9 introduces Long-Horizon-Terminal-Bench 2.0, a 77-task benchmark that tracks agent progress over three-hour runs. Across 15 models, the top system reaches the benchmark’s near-completion threshold on 37 tasks, while dense scoring reveals progress hidden by pass-or-fail results.

• OpenAI Cuts GPT-6 Sol and Luna API Prices; Luna Output Falls 58.3%

  OpenAI introduced GPT-6 Sol and GPT-6 Luna with lower API prices, strong results across workflow, coding and computer-use tests, and availability in ChatGPT Work, Codex and the API.

• Huawei Expands Ascend Chip Push Against Nvidia

  Huawei has introduced the Atlas 960 SuperPoD and outlined a multiyear Ascend chip roadmap at its Shanghai conference. The company is targeting Nvidia with larger AI systems, domestic hardware and new interconnect technology.

• NVIDIA Partner Firebird Secures License for 45,000 GPUs in Kazakhstan

  Firebird has secured a U.S. export license for 45,000 NVIDIA Grace Blackwell GPUs for a project in Kazakhstan with Kazakhtelecom. The first phase at Data Center Valley is under construction and targets a commercial launch in the third quarter of 2027.

• Anthropic Adds Live Dashboards and Editable Animations to Claude

  Anthropic has launched Claude Dashboards and Claude Motion in beta, letting users build refreshable views of connected company data and edit code-based animations. Dashboards are available on paid plans, while Motion is limited to Team and Enterprise plans.

• StarSkirmish Bench Measures How Well LLMs Write StarCraft Bots

  StarSkirmish gives each language model one hour to write a Protoss bot in C++ and evaluates the bots in a tournament against other model-written bots, established human-written bots and demo bots. Its page reports GPT-6 Astra and Claude Opus 5.5 as functionally tied among the top-scoring models, with GPT-6 Sol also ahead of the rest of the tested field.

• Kastle Raises $24M to Put AI Workers Inside Lending Systems

  Kastle raises a $24 million Series A led by Insight Partners to expand its AI workforce platform for banks and lenders. Its agents have processed more than $1.8 billion in transactions while operating across existing lending systems, the company says.

• TikTok Adds Shopping Assistant and One-Click Checkout

  TikTok announced Shopping Assistant, a conversational AI tool for product questions, and Buy Direct, a one-click in-app checkout option. The features are aimed at advertisers and are available through eligibility-based testing opportunities.

• Ema Raises $77M to Put AI Agents Against Enterprise Software

  Ema has raised $77 million in a Series B led by Creaegis to expand AI agents across corporate HR, IT and finance workflows. The Mountain View startup’s total funding has reached $140 million as it prepares to enter markets across Asia-Pacific, South America and the Middle East.

• WIRED Reports Hikers Who Used Chatbots to Plan Trips Needed Rescue, Though Accident Trend Is Unproven

  Hikers have used chatbots while planning trips that ended in wilderness rescues, but the reported cases do not establish that AI use is increasing accidents. Experts warn that confident-sounding chatbot advice can weaken critical thinking and urge hikers to check plans against reliable sources.

• OpenAI Ships GPT-Live-1 Voice Model in Its API

  OpenAI has released GPT-Live-1 in its API, giving developers a full-duplex voice model that can listen and speak at the same time. The model supports interruption handling, telephony, delegated reasoning and a front-end voice price of $0.05 per minute.

• AI Agents Often Fail in Autonomous Coordination, New arXiv Study Finds

  A new arXiv preprint argues that autonomous AI agents need shared rules for communication, identity, message validation and accountability when they act across trust boundaries.

• IBM Releases a Four-Model Granite Forecasting Ensemble

  IBM Research introduced Granite Timeseries Ensemble R1 on Hugging Face, combining four pretrained models for time-series forecasts. On 97 GIFT-Eval configurations, the ensemble slightly improved aggregate point-forecast error over its strongest individual member, while requiring more inference time and memory.

• Redwood Finds GPT-6 Astra Gains From Filler Tokens

  Redwood Research reports that GPT-6 Astra answers some reasoning tasks more accurately when its prompts contain meaningless filler tokens. The researchers warn that the gains raise questions about evaluating reasoning that does not appear in a model’s written explanation.

• Qualcomm Puts 30B-Parameter AI Models on Its New Phone Chips

  Qualcomm has introduced the Snapdragon 8 Elite Extreme Gen 6 and Snapdragon 8 Elite Gen 6 smartphone platforms. The Extreme model is designed to run 30-billion-parameter mixture-of-experts models locally, while both chips add new hardware for personal AI agents, cameras, gaming and connectivity.

• Google DeepMind Adds SynthID Watermarks to AI-Designed Proteins

  Google DeepMind introduced SynthID Bio, a method for embedding detectable watermarks in AI-designed protein sequences and predicted structures. Lab tests found watermarked protein binders retained key measures of function, while deliberate tampering remains an open challenge.

• Google Antigravity Compresses Cheminformatics Work into an Hour

  Google Antigravity completed a lung-cancer cheminformatics workflow in an hour, compared with the multiple days a typical iteration takes, according to Google DeepMind. The demonstration covers compound searches, potency analysis and proposed macrocycles, with experimental validation still required.

• FTC Seeks Safety Records From OpenAI and Anthropic

  The FTC sent OpenAI and Anthropic civil investigative demands on October 14, seeking records on AI safety incidents, safeguards, marketing and privacy. The demands mark a more formal stage in the agency’s probe into possible consumer-protection violations.

• NVIDIA Expands Compute and Training Access for European Scientists

  NVIDIA is joining the European Commission’s AI in Science Enablers Initiative with open course materials, scientific software and new routes to compute. Its commitments include collaborations with European AI factories and support for researchers, startups and small businesses.

• Nikon Strips First Place From Microscopy Video After AI Review

  Nikon disqualified Ning Xu’s first-place Small World in Motion video on October 9, 2026, after finding it violated the contest’s rules on generative AI. Nguyen Nam Nhat’s video now ranks first as Nikon says it will review its rules and judging procedures.

• Deezer Plans an Opt-In Filter for Fully AI-Generated Music

  A new Deezer toggle for excluding fully AI-generated music from search results will be available worldwide before the end of 2026. The planned setting will let listeners choose whether to hide tracks, albums and related artists identified as fully AI-generated.

• TiGanEval Finds Agents Lose Up to 67 Points on Hidden Requirements

  A new benchmark tests whether AI agents can infer requirements that users leave out but that are discoverable in a project. Across eight frontier models, success rates fell by 53 to 67 percentage points when agents received partial instructions instead of full specifications.