AI.info

The Pulse — Page 44

Browse The Pulse on AI.info.

• AWS Rebuilds AgentCore Runtime Around Elastic Memory and Snapshots

  AWS has released a new version of Amazon Bedrock AgentCore Runtime that reclaims unused session memory and restores prepared snapshots for more predictable cold starts. The company says the runtime reaches about two seconds at the 75th percentile across container images from 200 MB to 2 GB.

• Nvidia Turns to Insurers to Cover Risk on AI Chip Loans

  Nvidia is discussing insurance structures that could protect lenders if smaller cloud providers default on loans backed by its chips. The talks are preliminary, with no confirmed policy or deal.

• BizBench Finds Rule Violations Behind Agents’ Success Rates

  BizBench tests AI agents on 50 simulated business tasks based on payment workflows, checking both task outcomes and compliance during execution. In the benchmark, Qwen3.8-27B reached its goal on 42% of runs but met all success criteria on 32.7%.

• Seven Senators Press Meta Over $6.6 Billion in R&D Credits

  Seven senators led by Ron Wyden asked Meta CEO Mark Zuckerberg to explain $6.6 billion in research tax credits claimed from 2023 to 2025. Their letter questions whether any credits covered AI data center construction, while Meta says its AI facilities require extensive testing.

• Ballistic Ventures Seeks $490M for Two Cybersecurity Funds

  Ballistic Ventures is seeking $490 million across a $360 million flagship fund and a separate $130 million vehicle. Its SEC filing covers the flagship offering, while the firm’s broader fundraising plans come amid heightened investor attention to cybersecurity and artificial intelligence.

• RadixArk team describes serving Meta’s Muse Spark Open on one TPU host

  The RadixArk team describes running Meta’s 816-billion-parameter Muse Spark Open on a single TPU v7 host using SGLang-TorchTPU. Its post details a megakernel, attention and expert-processing changes, and results on the AgentX benchmark.

• Anthropic Launches Cyber Mission for Critical Infrastructure

  Anthropic is launching a Cyber Mission that brings Claude models, engineers and threat research to companies securing critical infrastructure. It is also offering free, automated vulnerability scans to participating open-source projects, with reports that do not receive human review before delivery.

• Microsoft Tests a Morphology-Aware Tokenizer Across 104 Languages

  Microsoft researchers introduce the Latent Core Tokenizer, a 200,000-token system evaluated across 104 languages. Their paper reports better morphological alignment and lower tokens-per-word than several common tokenizers, while noting limits in the downstream tests.

• Inworld Buys Ultravox to Add Voice Agents to Its Speech Stack

  Inworld acquired Ultravox, bringing its real-time voice-agent platform together with Inworld’s speech models and inference infrastructure. Built-in Inworld voices on Ultravox now use Realtime TTS-2, with existing voice IDs retained and no additional charge.

• Supabase Acquires Turso to Give AI Agents Their Own Databases

  Supabase is acquiring Turso to build infrastructure for AI agents that need cheap, on-demand databases. Turso will keep supporting SQLite while Supabase continues to build around Postgres.

• Snorkel AI raises $350M at a $3.5B valuation

  Snorkel AI has raised $350 million in a Series E round that values the company at $3.5 billion. Its annualized revenue run-rate has exceeded $375 million, the Stanford spinout says, as customers demand specialized training data, evaluations and simulated environments.

• JFrog Finds Unauthenticated Code Execution Flaw in LMCache

  JFrog disclosed CVE-2026-105192, a critical remote-code-execution flaw in LMCache’s multiprocess ZeroMQ transport. The vulnerability affects versions 0.3.9 and later, but remote exposure depends on whether operators bind the transport to a routable address.

• EditAgent Gives AI Image Editors More Tools and Better Checks

  EditAgent combines general-purpose and specialized image-editing methods, then gives a vision-language model evidence to help judge the results. In the paper’s benchmarks, verification records improved candidate selection, while the system’s added search and checking increased runtime.

• US Agencies Accuse Six Chinese AI Firms of Model Distillation

  Six China-based AI companies extracted billions of tokens from U.S. frontier models through distributed, industrial-scale campaigns, the NSA, FBI and CISA say. The agencies urge model providers, cloud companies and API aggregators to share signals and alter responses to suspected abuse.

• OpenAI Maps 165 ChatGPT Work Activities Across Nearly 100 Countries

  OpenAI released data tracking 165 work activities in consumer ChatGPT messages across nearly 100 countries from Q1 2025 to Q3 2026. The dataset shows work use spreading across more activities, but excludes enterprise messages and Codex and leaves many US job tasks outside its published categories.

• Andromeda 2 outperforms drug-formulation baselines

  A preprint posted September 16, 2026, reports that Andromeda 2 produced more high-performing paclitaxel formulations than Andromeda 1 and a wet-lab design-of-experiments campaign.

• Synthflow Launches Vector, an AI Agent That Tracks Lending Accounts

  Synthflow launched Vector, an AI agent that follows consumer lending accounts across calls, texts and email. The company says a four-week collections trial reached borrowers more often than the lender’s human team, while keeping underwriting and hardship decisions with people.

• Gallatin AI Raises $50 Million for Military Logistics Software

  Gallatin AI announced a $50 million Series A to expand its engineering team and develop Navigator, its military logistics decision-support platform. The tool is used by Army and Air Force customers and is deployed with U.S. military units around the world, the company says.

• Boomi Acquires Elementary Data, Promising to Keep Its Open-Source Project

  Boomi’s acquisition of Elementary Data will add data observability and governance capabilities to its platform. Elementary’s founders say the open-source project and standalone service will continue, while the two announcements give different counts of data teams using the product.

• EPA alumni flag 30 actions that could raise AI data-center health risks

  A report from the Environmental Protection Network identifies 30 federal actions it says could increase pollution-related health risks as AI data centers expand.

• XConf Teaches Models to Calibrate From Their Past

  A new arXiv preprint describes XConf, a confidence-estimation method that uses a model’s graded past episodes to improve calibration and selective prediction.

• Databricks Pledges More Than $350 Million for Singapore

  Databricks plans to invest more than US$350 million in Singapore over three years, expanding its workforce, headquarters, training programs and startup support.

• Ghost Raises $11 Million for Its $3,499 Personal AI Computer

  Ghost emerged from stealth on October 5, 2026, with an $11 million seed round led by Andreessen Horowitz and its first product, Core. The $3,499 screenless computer is designed to run AI agents locally and connect with users’ apps, files and smart-home devices.

• Osavul Raises €8.5M to Sell Hybrid-Threat Intelligence to Businesses

  Luxembourg-based Osavul raised €8.5 million in a Series A led by 33N Ventures. The company plans to expand its AI threat-intelligence platform from government and defence customers to enterprises and critical-infrastructure operators.

• Dream Releases MiST Cybersecurity Models With 13.1-Point Gain

  Dream released two open-weight MiST cybersecurity models, including an 8B model that scored 61.7 against 48.6 for its Qwen3 baseline. The 8B model’s average rose 13.1 points across 13 cybersecurity tasks, the company says, while its 30B-A3B release gained 8.8 points over its base model.

• Sakana AI Names Jürgen Schmidhuber Chief Scientific Advisor

  Sakana AI announced on September 24 that Jürgen Schmidhuber will join as chief scientific advisor and help guide its Recursive Self-Improvement Lab. The company says the lab is developing its first world models and aims to build systems that can research, improve their own code and interact with the physical world.

• Z-Library Paper Contains at Least 10 Untraceable Citations

  A peer-reviewed paper about Z-Library contains at least 10 references that TorrentFreak could not locate among its 44 citations. The authors disclosed using Consensus AI for literature searches and ChatGPT for language editing; Taylor & Francis says its ethics teams are investigating.

• Fresh-Context Agent Branches Lift Scores Across 13 Tasks

  A study evaluates “fork-and-flush,” a method that gives parallel agent branches a fresh chat context while preserving their accumulated workspace. Across 13 research and engineering tasks, the authors report higher normalized average scores than single-run and best-of-four baselines.

• Common Sense Media Rates ChatGPT for Teens an Unacceptable Risk

  Common Sense Media’s Youth AI Safety Institute rated ChatGPT for Teens an “Unacceptable Risk” after testing more than 4,000 prompts. OpenAI disputed the assessment, citing the timing of parental-control setup and defending Study Hours as a flexible learning feature.

• OpenAI Adds 200 Ireland Jobs for New Dublin Applied AI Lab

  OpenAI says it will create 200 additional jobs in Ireland, including technical roles at a new Applied AI Lab at its Dublin headquarters. The new roles add to 250 announced in July, bringing the company’s stated commitment to at least 450 new jobs locally.

• Go.AI Raises $85 Million for On-Prem Enterprise AI

  Chicago-based Go.AI has raised $85 million in Series A funding led by Updata Partners. The company will expand its on-premises AI hardware, Go.OS software and sales operations for regulated industries.

• Raindrop Announces Series A, Says It Has Raised $50M in Total Funding

  Raindrop has raised $50 million in total funding, led by CRV, as the startup expands from monitoring AI agents in production to testing their behavior before new code reaches users.

• AudioShake’s Refinery Separates Overlapping Voices for AI Training

  AudioShake introduced The Refinery on October 8, 2026, to turn existing recordings into speaker-separated training data. Its system has processed more than 100 million minutes of audio, the company says, and reports fewer transcription errors in a benchmark against an open-source separator.

• Fractile filings show £577.4m Series C share issue

  Companies House filings show Fractile allotted £577.4 million in Series C preferred shares between September and October 2026. The top-priced shares imply a value of about £3.84 billion for issued shares, but the company has not publicly confirmed the round.

• Marvell Agrees to Buy Lumentum in $115B All-Stock Deal

  Marvell and Lumentum announced an all-stock acquisition agreement with an implied enterprise value of $115 billion for Lumentum. The deal calls for Lumentum shareholders to receive 4.5 Marvell shares for each share they own and is expected to close in the first half of 2028, subject to approvals.

• LightOn Releases ColGrep Agent With 34.3% Fewer Claude Code Tokens

  LightOn released ColGrep-Agent-2B, a local code-search model that sends relevant file paths and line ranges to larger coding assistants. In LightOn’s 52-question Claude Code test, the agent reduced API token use by 34.3% but increased average response time.

• Goldman Limits AI Coding Tools to Engineers

  Goldman Sachs is restricting traders and bankers from using Anthropic’s Claude Code and Microsoft’s GitHub coding tools, according to internal guidance updated in October. The change follows months of wider access and leaves an exception for employees whose roles span trading and software engineering.

• Nvidia Adds a Record $150 Billion to Its Buyback Authorization

  Nvidia’s board approved a $150 billion increase to its share repurchase authorization, bringing the remaining amount to $235 billion. The company expects to carry out the program through fiscal 2028 as it balances shareholder returns with investment in AI and accelerated computing.

• IBM Research Trains Palette, a 20B Model for Slide Decks

  IBM Research introduced Palette, a fine-tuned 20B model for building slide decks in the IBM Design Language. On PresentBench, the team reports first-place rankings in Economics and Education and a 3.5-point average gain from its slide-repair harness.

• Study Finds Strong AI Agents Need Less Harness Machinery

  A study published by Apple Machine Learning Research finds that a single coding-agent session matched or beat four specialized harnesses on tested machine-learning engineering benchmarks. The authors caution that the results are limited by benchmark coverage, small seed counts and little testing of coordination between parallel agents.