AI Weekly Brief

AI Weekly Brief — week ending 2026-10-03

This Week in AI

  • OpenAI says Moonshot-linked actors mined its models' hidden reasoning. OpenAI reports it cut off a campaign that logged 16,000 extraction requests across 4,000 accounts, allegedly to help train Moonshot AI's Kimi models. Tom's Hardware, The Hacker News
  • The US government now says "super intelligence," not "artificial intelligence." President Trump signed an executive order making the swap and paired it with an industry accord that asks major labs to police themselves; intelligence chief Jay Clayton is expected to be named AI czar. Politico, The New York Times, The Washington Post
  • Decision-only models became a product category. TypeSafe AI shipped Jev, which returns typed probabilities instead of text; Amazon answered with the open-source Strands Decider 2B, and Databricks launched an ai_decide SQL function the same week. infoq.com, Crypto Briefing, Databricks
  • Alibaba's Qwen drew safety and censorship scrutiny. Researchers reported alarming answers on Hong Kong and other restricted topics, and Fox News reports an investigation into claims a Chinese model returned bioweapon instructions. CBS News, Fox News
  • GitHub Copilot reached the desktop and the API. Computer use entered public preview, letting Copilot operate desktop apps, and code review is now callable from REST and GraphQL APIs. GitHub changelog, GitHub changelog
  • Microsoft delayed default Copilot usage-based billing to December 1. It also added a spending cap option, which gives organizations cost control before the switch. crn.com
  • CMS projects cheaper Medicare premiums for 2027. Weighted average Medicare Advantage premiums are projected to fall more than 16%, with MA drug premiums down 38%; CMS also finalized a mandatory Part B drug payment model. CMS, CMS
  • Anthropic put $100M into engineer training. The Claude Frontier Academy aims to train 10,000 engineers. Anthropic

Vendor Dossiers

Ai2 (Allen Institute for AI)

  • Released: AstaBrief, the fast report-generation model inside Ai2's Asta system, is now open-sourced. Hugging Face
  • Released: Olmo-core 3, Ai2's open training infrastructure for large mixture-of-experts models, the plumbing behind its open model line. Hugging Face

Alibaba (Qwen)

  • Announced: Alibaba unveiled new in-house AI chips alongside plans for its Qwen model family, a full-stack push from silicon to models. Alibaba is based in China, a fact that bears on provenance review for regulated workloads. SME & Entrepreneurship Magazine, DataCenterNews Asia Pacific
  • Researchers reported Qwen gives alarming answers on Hong Kong and keeps other topics off limits, and Fox News reports an investigation after a researcher said a Chinese model supplied bioweapon instructions. A Qwen3.5 derivative (OpenJev) landed on Databricks this week, so Qwen provenance is a live question for payer data platforms. Yahoo, CBS News, Fox News
  • A modified Qwen model underperformed in cyber attack tests. Useful context for teams assessing offensive-use risk in open models. SecurityBrief New Zealand

Amazon

  • Released: Strands Decider 2B, an open-source decision-only model positioned against TypeSafe AI's Jev. It outputs decisions rather than prose, which narrows failure modes in automated pipelines. Crypto Briefing, The AI Economy

Anthropic (Claude)

  • Announced: Claude Frontier Academy, a $100M program to train 10,000 engineers. Anthropic
  • Barclays is scaling Claude across operations and client service, a regulated-bank deployment worth watching for governance patterns that transfer to payers. Anthropic
  • Anthropic published "Claude-shaped science," an essay on how Claude shows up in scientific work. Anthropic

DeepSeek

  • Released: DeepSeek and Huawei open-sourced Ascend AI programming tools, including compute and communication libraries plus TileLang support, to cut reliance on Nvidia's CUDA. The China-based pair is building an independent hardware and software stack for open models. Tom's Hardware

Google DeepMind (Gemini, Gemma)

  • Google published its September 2026 AI updates roundup, collecting the month's Gemini and product changes. No single headline release this week. Google AI Blog

Mistral

  • Announced: Mistral signed an agreement with the European Space Agency to explore AI for space programs. Electronics Weekly
  • France-based Mistral is hiring in Canada as it expands its sovereign AI business: selling governments and regulated buyers locally controlled AI stacks. The Logic

Moonshot AI (Kimi)

  • OpenAI says actors linked to China-based Moonshot AI logged 16,000 requests across 4,000 accounts to extract hidden model reasoning, allegedly to train the Kimi family, and it has cut off the accounts. Kimi models ship in GitHub Copilot and on Databricks, so the dispute touches tools this audience already uses. Tom's Hardware, The Hacker News, SC Media

OpenAI (GPT, Codex)

  • OpenAI published a model guide for the GPT-6 family covering model choice, reasoning effort, prompts, and production prep. A practical read for teams building on the OpenAI API. OpenAI
  • Chatham Financial reports Codex and GPT-5.6 cut trade validation from 30 minutes to under 4. A concrete number for what agentic coding does to a back-office workflow. OpenAI
  • Albertsons is deploying ChatGPT Enterprise and the OpenAI API across its retail operations. OpenAI

TypeSafe AI (Jev)

  • GA: Jev, a decision-only model that returns typed probabilities instead of text. Typed outputs for yes/no, choice, and scoring questions are easier to audit than free text, which matters in claims-style decisioning. infoq.com, news.lavx.hu
  • WSJ reports Jev has sparked copycats and talk of LLM alternatives, and legal-industry press is tracking the "System One Model" category it started. WSJ, Artificial Lawyer

xAI (Grok)

  • grok-imagine-video-1.5-lite appeared in xAI's API documentation; the listing is thin on detail. docs.x.ai

GitHub Copilot

  • GA: Copilot code review can be requested through REST and GraphQL APIs with a per-request effort level, and Balanced is now the default effort level. This brings reviews into existing scripts and internal tooling. GitHub changelog
  • Preview: Computer use lets Copilot operate desktop applications, reading screens, clicking controls, and entering text, from Copilot CLI and the Copilot app on macOS and Windows. Copilot asks approval before controlling an app, and organization-managed settings can disable the feature. GitHub changelog
  • GA: Dynamic workflows define multi-agent orchestration in code inside a Copilot extension, with parallel steps and structured hand-offs between agents. GitHub changelog
  • The VS Code September releases (v1.136–v1.140) add scheduled automations in preview, agent merge for landing pull requests, and a research preview of HydraFusion model coordination. GitHub changelog
  • GitHub deprecated Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 across all Copilot surfaces on October 2; suggested replacements are Gemini 3.8 Flash, Kimi K3, and Claude Opus 5.5. Update pinned models and check Enterprise model policies. GitHub changelog

Microsoft 365 Copilot

  • Microsoft delayed default Copilot usage-based billing to December 1 and added a spending cap option. Organizations get more time to set limits before metered agent costs start by default. crn.com
  • Announced: askpolly integrated with Microsoft Copilot Cowork to connect social insights with sales data. Yahoo Finance

Microsoft / Azure

  • Nothing qualifying this week.

Open-Weight Ecosystem

  • HotHardware reviewed the HP ZGX Nano G1n, a GB10-based mini PC running a local Qwen coding agent. Desktop-class hardware for running open models locally keeps getting more practical. HotHardware
  • ServiceNow published AutoSynthData, a method for generating training data for enterprise agents, on the Hugging Face hub. Hugging Face

Databricks Watch

AI / ML

GA

  • The standard mode of ai_extract now uses a checkpoint that is 2.3x faster with better quality at no added cost. Document extraction gets cheaper with no code changes. Databricks release notes
  • Genie One memory is generally available; it retains user-provided facts such as reporting conventions across conversations, private to each user. Databricks release notes
  • New Unity Gateway models: OpenAI's GPT-6.1 Sol (release notes), xAI's Grok 4.7 (release notes), and OpenJev, a Qwen3.5 4B decision model served pay-per-token behind the TypeSafe System One API (release notes). OpenJev returns structured answers and probabilities for yes/no, choice, and scoring questions.
  • The lakebase_tokenizer extension adds a configurable whole-word tokenizer for Lakebase Search keyword relevance, with stopwords, synonyms, and stemming managed in SQL. Databricks release notes

Preview

  • Public Preview: the system.ai_gateway.usage table adds coding-agent session metadata, invocation source, and token-level cache detail. Finer cost attribution for AI workloads. Databricks release notes
  • Beta: the Agent Bricks CLI scaffolds, runs, and deploys custom code agents with managed memory, sessions, and MLflow tracing from one authenticated command. Databricks release notes

Announced

  • ai_decide, a new AI Function for fast decisions on governed data. It pairs with the OpenJev model launch above. Databricks blog
  • Lakebase Search brings full-text and vector search to Postgres, aimed at AI agent workloads. Databricks blog
  • Thinking Machine Labs Inkling retires October 30, 2026; Databricks recommends GLM 5.3 or Kimi K3 as replacements. Migrate any endpoints before the cutoff. Databricks release notes

Data & Platform

GA

  • JDBC Unity Catalog connections are generally available on serverless compute, Runtime 19+, and SQL warehouses 2025.40+, with SSL certificate support; OAuth M2M stays in Beta. Databricks release notes
  • Cross-engine attribute-based access controls: supported external engines read Unity Catalog managed tables with ABAC policies, row filters, and column masks enforced. One policy layer across engines matters wherever regulated data leaves the primary engine. Databricks release notes
  • Query tags for SQL warehouses: custom key-value tags for grouping, filtering, and cost attribution, visible in system.query.history. Databricks release notes
  • LTAP Direct Writes speeds initial loads and full refreshes of synced tables into Lakebase by writing straight to storage. Databricks release notes
  • Horizontal scaling for Databricks Apps: multiple instances behind one URL, with zero-downtime deployments. Databricks release notes
  • Lakeflow pipelines can run as an account group, so a team owns an automated pipeline instead of one person's account. Databricks release notes
  • Lakeflow Jobs table update triggers now cover OpenSharing objects and system tables. Databricks release notes
  • IP Functions bring network analytics (firewall, VPN, and DNS data) to warehouse SQL. Databricks blog
  • Excel and Google Sheets writeback now requires a workspace admin to enable a security setting. Check this if analysts write back from spreadsheets. Databricks release notes

Preview

Announced

  • Salesforce Data 360 file sharing will soon be on by default for workspaces with the compliance security profile enabled. Relevant if you run the compliance profile and federate Salesforce data. Databricks release notes
  • Lakebase Postgres branch-based restores promise faster recovery at scale. Databricks blog

Regulatory Radar

US Federal

  • President Trump signed an executive order replacing "artificial intelligence" with "super intelligence" in federal usage, paired with an industry accord under which major labs police themselves. No new binding obligations for plans yet; expect terminology drift in federal guidance and contracts. Politico, The Verge, The New York Times
  • Intelligence chief Jay Clayton is expected to be named the new AI czar. The pick will shape how the federal AI agenda gets enforced. The Washington Post
  • WARNING: Threat group TA419 posed as a former White House official to phish AI policy experts. Treat unsolicited AI-policy outreach as a phishing surface. BetaNews, inkl

US States

  • California's AI law is forcing lawyers into hard disclosure decisions about AI-assisted work. Disclosure duties land on regulated professions first, and health plans operate in several states that may follow. Bloomberg Law
  • A JDSupra analysis asks whether Colorado's AI law helps or conflicts with the Equal Credit Opportunity Act. How state AI statutes mesh with existing federal anti-discrimination law is a live question for any regulated, automated decisioning. JDSupra

Healthcare-specific

  • CMS finalized a mandatory drug payment model for Original Medicare Part B. Mandatory models change payment and data requirements for affected drugs, so watch for operational guidance on claims and encounters. CMS
  • CMS projects weighted average Medicare Advantage premiums to fall more than 16% from 2026 to 2027, with MA prescription drug premiums down 38%. Program economics this favorable shape 2027 bids and benefit design. CMS
  • CMS convened certified Independent Dispute Resolution entity leaders to discuss improving the federal surprise-billing dispute program. IDR process changes flow directly into out-of-network claims and encounters workflows. CMS
  • CMS announced about $311 million in rural health grants across six states, with telehealth, telemedicine, and emergency care among the priorities: Colorado, Alabama, Texas, North Carolina, Montana, South Dakota. Telehealth expansion keeps shifting encounter volume toward virtual care.

Worth Your Time

  • OpenAI's essay "The eternal complement" argues advanced AI may matter most for the routine execution work behind breakthrough ideas. OpenAI
  • How Databricks rolls out frontier models to 14,000 employees on day one: a practical enterprise enablement writeup. Databricks
  • "Five ways the US and China clash over AI," a primer on where the two AI powers conflict. The Straits Times
  • "Why Businesses Are Moving From a Single AI Vendor to Multi-Model Strategies in 2026" lands in a week when GitHub deprecated four models at once. citybiz
  • Databricks' 2026 guide to storing Unity Catalog managed tables covers controlling where your data lives. Databricks