AI Weekly Brief — week ending 2026-09-20

This Week in AI

Federal Register pulls Alibaba's Qwen. The US Federal Register site ran an AI search tool built on Alibaba's Qwen days after the FBI accused the model line of copying Anthropic's technology, and removed it after backlash. Reuters, Ars Technica
California orders AI oversight. Governor Newsom issued an executive order to accelerate independent oversight and advance an AI "kill switch." The New York Times, CA.gov
GitHub Copilot retires six models on October 19. Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5 all go away; workflows pinned to them need to move to the suggested replacements. GitHub Changelog
Kimi K3 lands on Amazon Bedrock with a 1M-token context. Databricks will retire Kimi K2.7 on October 30 and names K3 the replacement, so the K2-to-K3 shift is now platform-wide. AWS, Databricks
Databricks ships Lakebase Search. Vector and BM25 full-text search are now GA inside Lakebase Postgres, so hybrid search runs in one database without a separate search stack. Databricks
Alibaba's Qwen 3.8 wave continues. Qwen3.8-Omni-Flash cuts audio pricing 98% and ships open weights, and Qwen3.8-LiveTranslate interprets across 60 languages with 2.3 seconds of average lag. Startup Fortune, MarkTechPost
CMS keeps funding technology-supported care. The ACCESS Model extends to Original Medicare in spring 2027, a Medicaid drug payment model named participants, and rural health awards fund telehealth and remote monitoring in five states. CMS
White House weighs "limited safeguards" on AI. A report says the administration may impose narrow federal AI rules. New York Post

Vendor Dossiers

Alibaba (Qwen)

Released: Qwen3.8-Omni-Flash, an omni-modal model with a 1M-token context built for agentic audio-video understanding and tool use; Alibaba cut audio pricing 98% and published open weights. MarkTechPost, Startup Fortune
Released: Qwen3.8-LiveTranslate, a real-time interpretation model that cuts average lag to 2.3 seconds across 60 languages. MarkTechPost
Released: Qwen-Image-2.1, a 7-billion-parameter open-weight image model that Alibaba claims beats closed models. the-decoder.com
The US Federal Register site ran a Qwen-based search tool days after the FBI accused the models of copying Anthropic's technology; the site removed the tool after backlash. Alibaba is a China-based company, which is central to the provenance concern. Reuters, Ars Technica, Mezha

Anthropic (Claude)

Announced: a partnership with Accenture on embedded evaluation. Anthropic
Announced: the Life Sciences Verification Program. Anthropic
Anthropic published research on measuring the pace of AI development inside frontier labs. Anthropic, plus a post on how Claude supports biomolecular modeling. Anthropic
Background on the distillation allegations behind the FBI's claim that Qwen copied Claude; distillation means training a model on another model's outputs. NeoTeo

DeepSeek

Preview: DeepSeek's voice chat gray test added four selectable text-to-speech voices. Intelligent Living

Google DeepMind (Gemini, Gemma)

No Gemini or Gemma model news this week. Google's AI news was institutional:

GA: the UN System Data Commons, an open platform from Google and the UN system for searching global statistics. Google
Google added academic advisors, fellows, and internal researchers to its AI & Economy team. Google
Google custom-built Flow video tools with designers Jane Wade and Sergio Hudson for New York Fashion Week. Google

Mistral

Firefox's AI assistant added a Mistral model, with access opening in France first; Mistral is a French lab, which gives EU users a home-jurisdiction option. Cybernews, It's FOSS
Announced: applications are open for the Mistral AI Accelerator Bootcamp for African innovators. Capital FM Africa

Moonshot AI (Kimi)

GA: Kimi K3 is available on Amazon Bedrock with a 1M-token context window. Moonshot AI is a Beijing-based lab, a fact that belongs in any provenance review before production use. AWS, Unite.AI
GA: a Kimi financial services product for banks. qz.com

OpenAI (GPT, Codex)

No GPT or Codex model news this week.

Announced: the Australian Youth Safety Blueprint, a six-pillar roadmap for safer AI experiences for young people. OpenAI

StepFun (Step)

Announced: a preview of Step 5, billed as China's cheapest frontier model yet; StepFun is a China-based lab. Intelligent Living

GitHub Copilot

GA: Copilot code review now tracks how a review changes over time, groups findings as open, resolved, or previously missed, and writes commit messages for batch-accepted suggestions. GitHub Changelog
The September 14 weekly release adds efficiency, balance, and intelligence tiers to auto model selection, Sentry integration in the Copilot app, and shell-tool validation during reviews. GitHub Changelog
Announced: six models retire on October 19, 2026 (Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, and Grok 4.5). Suggested replacements auto-enable for Business and Enterprise unless an admin turned off default model enablement. GitHub Changelog
GA: admin telemetry grew on two fronts: the impact dashboard shows 28-day feature engagement, and the usage metrics API adds agentic CLI activity for skills, custom agents, MCP servers, slash commands, and plugins. Impact dashboard, Usage metrics API

Microsoft 365 Copilot

Microsoft Copilot had an outage, with error messages traced to an API issue. tbreak.com
GA: Quartr's financial research content is available inside Microsoft 365 Copilot via MCP, the Model Context Protocol for connecting external data. Quartr
Microsoft published a Moody's case study on running risk intelligence inside Copilot. Microsoft
A meeting-tools price comparison puts a $30 gap between Otter.ai, Microsoft Copilot, and Gemini. tech-insider.org

Microsoft / Azure

Microsoft Agent 365 adds centralized governance for AI agents. Petri
GA: Copilot Studio agent governance launched in the UAE. Telecompaper

Open-Weight Ecosystem

A Rhodium Group report estimates China's top AI models earn about 10% of the revenue OpenAI and Anthropic collect, a gap between download share and paying customers. South China Morning Post, ForkLog
Released: Bonsai 2 27B claims 98% of frontier-class quality in a 5.9 GB download, sized for local hardware. Intelligent Living
A head-to-head compares KoboldCpp and Ollama for running local models. Geeky Gadgets
An analysis asks whether China can win global adoption with open-weight models as the US raises the stakes on AI. The Business Times

Databricks Watch

AI / ML

Lakebase Search brings vector search (pgvector-compatible, scales past 1 billion vectors per index) and BM25 full-text search to Lakebase Postgres, with hybrid ranked results in one database. Databricks
The Unity Gateway API for model, model provider, and MCP services is GA across the Terraform provider, CLI, and all SDKs; bundle support stays in Beta. Databricks
Unity Gateway supports all compliance security profile standards and is on by default for compliance-profile workspaces. Databricks
Genie One skills can now include multiple files and sub-folders, up to three folders deep and 50 files per skill. Databricks

Preview

Genie responses in Slack now include chart images (Public Preview); reinstall the Genie Slack app to enable it. Databricks
Account-only users can reach account-level Genie One without a workspace assignment, with Unity Catalog still governing what they see (Beta). Databricks

Announced

Moonshot AI Kimi K2.7 retires on October 30, 2026; Kimi K3 is the recommended replacement. Databricks
Anthropic models were set for removal from AWS GovCloud IL5 workspaces on September 19, 2026; affected endpoints must migrate to another supported model. Databricks

Data & Platform

Sharing foreign Delta tables with OpenSharing is GA: providers can share tables federated from OneLake, a Hive metastore, or AWS Glue without copying data. Databricks
Databricks Apps telemetry is GA, persisting OpenTelemetry traces, logs, and metrics to Unity Catalog tables. Apps is also now on by default for compliance-profile workspaces. Telemetry, Apps default-on
Service principal OAuth client secrets now auto-delete after 90 days without use. Databricks
The counter_diff window function computes differences between cumulative counter values and handles resets, for time-series work in Databricks SQL and Runtime 19+. Databricks
Metric view window measures support numeric index columns, covering fiscal weeks and 4-4-5 periods. Databricks

Preview

Attribute-based access control (ABAC) grew on two fronts, both Beta: metastore-level policies that apply across every catalog, and row filters and column masks on views. Metastore-level, ABAC on views
Data Classification can scan Unity Catalog views for sensitive data, not just tables (Beta). Useful anywhere views expose member-level fields. Databricks
Base environments manage Python dependencies on classic compute (Beta). Databricks
The Lakebase snapshots API creates, lists, deletes, and restores point-in-time branch snapshots (Beta). Databricks

Announced

Automatic change data feed reaches all supported regions by end of October; it computes row-level changes at query time, with no per-table CDF setup, on Runtime 19+ with row tracking. Databricks
The READ METADATA privilege arrives by default for compliance-profile workspaces in mid-October, giving auditors read-only visibility into grants, row filters, and masks. Databricks
JAR tasks on serverless compute auto-enable for compliance-profile workspaces in early October. Databricks
Databricks announced on-demand state repartitioning for Spark Structured Streaming on Databricks. Databricks

Regulatory Radar

US Federal

A report says the White House may impose "limited safeguards" on AI. Any federal framework could set the ceiling or floor for the state AI laws plans already track, so this bears watching. New York Post
NIST finalized guidelines on protecting online identity and access tokens from misuse. Token theft is a live risk for portals and APIs that touch protected health information, so security teams can apply this directly. NIST

US States

California's governor ordered faster independent AI oversight and work toward an AI "kill switch." California keeps writing the template other states copy, including for AI used in coverage and utilization decisions. The New York Times, CA.gov
A new California law requires disclosure when ads use AI performers. Marketing teams that advertise in California will need a disclosure step for AI-generated talent. ET BrandEquity
Colorado's updated AI law keeps AI hiring tools in scope even when a human reviews the output. Employers that screen candidates with AI, health plans included, cannot rely on a human-in-the-loop exemption. JD Supra

Healthcare-specific

CMS named participants in a Medicaid drug payment model aimed at lowering drug costs. Payment model changes flow into pharmacy encounter and rebate reporting for Medicaid plans. CMS
Starting spring 2027, the ACCESS Model extends technology-supported care to Original Medicare beneficiaries with heart failure, COPD, substance use disorders, and nicotine dependence. Remote care models produce new encounter patterns that plans will need to capture and submit. CMS
CMS announced Rural Health Transformation Program awards: $167 million for South Carolina, $104 million for Mississippi, $50 million for Connecticut, nearly $17 million for Kansas, and school-based care expansion in North Carolina, much of it for telehealth and remote monitoring. Expect telehealth encounter volume to grow in these states. SC, MS, CT, KS, NC
CMS, the SBA, and Georgia promoted CHOICE arrangements as a coverage option for small businesses, a signal on where small-group membership mix may move. CMS

Worth Your Time

Databricks shows how MATCH_RECOGNIZE finds row-sequence patterns in plain SQL. Directly useful for spotting event sequences across claim and encounter rows. Databricks
A mid-year map of US state AI statutes, useful as a single compliance reference. The National Law Review
How freely available AI helped attackers reach a TikTok user's camera. A concrete case for security awareness material. The Washington Post
A survey of open-source agent frameworks that run against local models, for anyone testing off-cloud setups. MarkTechPost
A profile of the $50 billion rise of Moonshot AI, the lab behind Kimi. Background for the K3-on-Bedrock news above. ThinkChina