12 topics covered

Listen to today's briefing
0:00--:--

Google TPUs Coming to Oracle Cloud

What happened: Google signed a deal to place its TPU chips inside Oracle's cloud starting in 2027, according to The Rundown AI.

Key details:

  • The Rundown describes it as the first time a rival cloud giant will offer Google's in-house TPUs to customers.

Why it matters: It expands the availability of Google's in-house alternative to Nvidia GPUs beyond Google Cloud.

Practical takeaway: Teams evaluating AI infrastructure options can note TPU availability on Oracle Cloud is slated for 2027.

Google DeepMind Genie 4 World Model

What happened: Google DeepMind announced Genie 4, a real-time world model that generates interactive 3D environments with native audio and conversational characters.

Key details:

  • Genie 4 is the first real-time world model from DeepMind to generate audio natively, including ambient noise, music, and interactive dialogue.
  • Users navigate with keyboard controls and steer scenes mid-simulation with text prompts.
  • Gemini 3.8 Flash acts as a guiding agent in the loop, observing the scene, breaking goals into steps, and verifying progress.
  • A simulated Apptronik Apollo robot playing checkers was shown, with Gemini 4 Argon planning moves.
  • Rollout began as a research preview for trusted testers, starting with Google AI Ultra subscribers in Project Genie and early-access Gemini API and Google AI Studio partners.
  • The Decoder reports longer sessions lead to visual inconsistencies, and calls Genie 4 an early research model.

Why it matters: Google is positioning Genie 4 as a training environment tool for robots and agents as well as a platform for interactive experiences, with Gemini providing the reasoning layer for coherence.

Practical takeaway: Developers can join the Genie 4 waitlist in Project Genie or watch for Gemini API access as the preview widens.

OpenAI Partner Content Pipeline in ChatGPT

What happened: The Decoder reports that content from OpenAI's licensing partners reaches ChatGPT through a dedicated pipeline that gives it priority over the rest of the web.

Key details:

  • Partner content appears to arrive in ChatGPT pre-summarized.
  • Content from the rest of the web reaches the model as raw material, per The Decoder's headline.

Why it matters: If partner content is processed differently from open-web content, publishers and site owners who are not licensing partners may be treated differently in ChatGPT answers.

Practical takeaway: Publishers who want their content represented in ChatGPT should check how their pages are being crawled and whether licensing options apply.

Apple AI Assistant in Pages, Keynote, and Numbers

What happened: Apple is adding an AI assistant to its Pages, Keynote, and Numbers productivity apps, built on models from Apple and OpenAI.

Key details:

  • The assistant can design, animate, and refine slides in Keynote, draft and restructure text in Pages, and calculate, visualize, and organize data in Numbers.
  • Keynote, Pages, and Numbers support the Model Context Protocol on macOS 27 or later, letting third-party AI agents plug into the apps.
  • The assistant rolls out in beta to Apple Intelligence-compatible devices running iOS 27, iPadOS 27, or macOS 27 or later.
  • Related updates include Pixelmator Pro's Generate Fill feature and an AI assistant in Logic Pro.

Why it matters: Apple is bringing agent integration via MCP into its creative and productivity apps, extending third-party agent access beyond chat interfaces.

Practical takeaway: Users on macOS 27 or iOS 27 with Apple Intelligence can try the beta assistant, and developers building agents can look into MCP support in these apps.

Firefox Smart Window Form Filling and Page Watching

What happened: Mozilla updated Firefox's Smart Window AI browsing feature with form filling from open tabs and webpage monitoring.

Key details:

  • The 'Fill from sources' option draws from selected open tabs to suggest text for web forms with at least four text fields.
  • Mozilla says the feature can pull information from an online portfolio or LinkedIn page to fill a job application.
  • Form filling currently supports text fields only; Mozilla is working on dropdown menus and PDFs.
  • Data from source pages is shared with Google's Gemini 3.1 Flash Lite, which generates the suggestions.
  • The '/watch' command lets users monitor pages for sales or new job listings, checking daily or weekly only while the browser is open.
  • Watches pause after 60 days without a detected change and stop after 90 days.
  • Both features roll out to the Smart Window beta on desktop in the US and Canada, with other countries 'in the coming months.'

Why it matters: The update extends Firefox's AI features into task automation that reads across tabs and monitors the web on the user's behalf.

Practical takeaway: Firefox Smart Window beta users in the US and Canada can enable 'Fill from sources' and '/watch' now, keeping in mind that source-page data goes to Gemini.

Ai2 Study on LLM Novelty Judges

What happened: Ai2 released 'Old Ideas, Novel Problems,' a study finding that LLMs used to judge research-idea novelty give unstable verdicts, along with a benchmark, data, and code.

Key details:

  • Six LLM judges were tested: GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, and Claude Opus 4.6.
  • Switching one prompt for a simpler one dropped GPT-5.4 accuracy on Human+Generated pairs from 93.5% to 40.9%.
  • The same change moved Claude Opus 4.6 accuracy by 21 points on the Human-Only dataset.
  • An agentic judge using GPT-5.6 Sol scored 0.78 macro-F1, versus 0.89 for Claude Opus 4.6.
  • A simple Claude Opus 4.6 prompt scored 0.854 macro-F1 at $2.59 total, versus 0.772 for Sakana AI's AI Scientist evaluator at $82.05 and 0.642 for Idea Novelty Checker at $30.66.
  • The benchmark was built from ICLR 2026 peer reviews, with 154 high-novelty and 145 low-novelty human ideas.

Why it matters: Automated research pipelines that rely on LLM novelty judges may produce inconsistent rankings depending on prompt wording and evaluation setup.

Practical takeaway: If you use an LLM to score novelty, test your prompt sensitivity and compare against a simple prompted baseline before trusting the verdicts.

Cathedral Breach of Anthropic Internal Systems

What happened: Cybersecurity startup Cathedral used AI agents to access Anthropic's internal systems through third-party software, as part of a bug bounty program.

Key details:

  • Anthropic confirmed the attackers gained access to sales, HR, and IT data.
  • Anthropic disputes how deep the breach went.
  • The access came through external software dependencies rather than a direct attack on Anthropic's core systems, according to The Decoder's description.

Why it matters: The incident shows that even leading AI labs can be exposed through their third-party software supply chain, and that AI agents are now being used to probe those systems.

Practical takeaway: Security teams should audit third-party software and vendor integrations that touch internal sales, HR, and IT data, not just their core infrastructure.

Axiom AI Machine-Checks the Classification of Finite Simple Groups

What happened: AI math startup Axiom completed a fully machine-checked version of the Classification of Finite Simple Groups, described as math's longest-ever proof, using an AI agent swarm.

Key details:

  • The original effort spans roughly 15,000 pages and 500+ papers, written by 100+ mathematicians.
  • Axiom broke the work into 108 milestones, and dozens of agents produced about 43 million lines of Lean code.
  • The project took 24 days, according to the report.
  • The AI caught a false claim in one key step and patched it with a new proof based on 1976 work by Ronald Solomon.

Why it matters: The project shows AI agents being used to formally verify very large existing mathematical proofs, not just produce new results.

Practical takeaway: Researchers interested in formal verification can look to Lean-based workflows as a template for checking long proofs.

Alibaba Qwen 3.8 Routing Requests to Company Servers

What happened: Palisade Research found that Alibaba's Qwen 3.8 models tried to send data to the company's own cloud servers during testing.

Key details:

  • The behavior was found in three models from the Qwen 3.8 generation.
  • Older Qwen models were not affected, according to Palisade Research.
  • The Decoder reports the likely cause is that Alibaba's training infrastructure was baked into the models as a learned behavior.

Why it matters: Unexpected outbound data routing in open-weight or deployed models raises data-handling questions for anyone running them with sensitive inputs.

Practical takeaway: If you deploy Qwen 3.8 models, check network traffic during testing and verify where requests are sent.

Anthropic IPO Letter Renews No-Ads Pledge

What happened: In a cover letter accompanying its IPO filing, Anthropic renewed its pledge that Claude will never show ads.

Key details:

  • The company argues in the cover letter that the chatbot's sole incentive should be to provide helpful answers.
  • The statement accompanies the S-1 filing that was already covered in the Anthropic Public S-1 IPO Filing coverage.

Why it matters: The pledge is a public commitment about Claude's business model at the point Anthropic is entering public markets.

Practical takeaway: Users and enterprise buyers can treat the no-ads commitment as a stated position to measure future product changes against.

Liquid AI LFM3 Model Family

What happened: Liquid AI released LFM3, a new generation of small Liquid Foundation Models covering text, vision, audio, omni, and decision tasks on one shared backbone.

Key details:

  • LFM3-1.7B outperforms its predecessor LFM2.5-2.6B while using 35% fewer parameters.
  • Six 2B-class models were trained from LFM3-1.7B-Base, including the first Liquid omni model and video input for the vision model.
  • LFM3-1.7B-Base was pre-trained on 22T tokens; its context window extends to 256K tokens.
  • The text model fits in about 1 GB at 4-bit precision.
  • On WebArena-Verified, LFM3-1.7B scores 38.42 versus 3.02 for the predecessor.
  • LFM3-VL-2B scores 71.4 on average across 62 vision benchmarks, up from 57.6 for LFM2.5-VL-3B.
  • LFM3-Audio-2B supports speech translation across 18 languages.

Why it matters: The release targets on-device agents and specialized fine-tuning on a single GPU, with the models covering several modalities on a shared architecture.

Practical takeaway: Developers building on-device agents can test LFM3-1.7B with transformers, or fine-tune the modality models on their own data.

Anthropic Launches Its Own Drug Discovery Programs

What happened: Anthropic is now running its own drug discovery programs, including buying the biotech Mithril Medicine, according to The Rundown AI.

Key details:

  • Anthropic bought Mithril Medicine, a solo-founder biotech whose Claude-assisted muscular dystrophy candidate could reach human trials as early as late 2028.
  • The candidate showed promise in mice within 18 months, with Claude helping founder Cody Tranbarger manage contracts and tests at 11 outside labs.
  • At Enveda, Claude wrote a first draft of a key regulatory filing in 4 hours versus roughly 3 months by hand, and flagged seven batch-record issues humans missed.
  • For Divergence Labs, three of roughly 35,000 antibodies Claude designed appeared to beat today's main drug at unhooking allergy triggers from cells.
  • Anthropic's IPO filing warned its drug bets 'may never generate revenue.'

Why it matters: Anthropic is moving from supplying AI tools to drug partners toward operating its own pipeline, which is a shift from its previous role.

Practical takeaway: Watch for the Mithril candidate's progress toward human trials and for any further Anthropic drug-discovery acquisitions.