The Agent Chronicles

The day in AI, in brief.

The AgentChronicles
Earth, on

Models, accessibility and AI research

Microsoft and Liquid launch compact reasoning and vision models, while Google adds sign-to-text input and researchers examine automated AI research risks.

Briefs

Frontier labs

  • Microsoft previews its MAI-Thinking-1 reasoning model — Microsoft has added its first in-house reasoning model to Foundry and Azure preview. MAI-Thinking-1 uses adaptive reasoning, supports a 256k context window and targets coding and quantitative work with lower token use. Release announcement and Azure announcement.

  • Liquid AI ships a compact vision-language model — Liquid AI’s 3bn-parameter LFM2.5-VL-3B release reads screens, documents and charts, identifies coordinates, and supports tool calls. Its small size could make visual software agents cheaper to run.

Accessibility

  • Google adds ASL sign-to-text input — Google is adding American Sign Language conversion to Gboard and Live Transcribe. The feature aims to turn signing into text during phone-based conversations.

Digital business

  • ChatGPT adverts may not alter recommendations — An SE Ranking analysis of commercial ChatGPT prompts finds adverts often appear without the advertiser being cited in the answer. Marketers may need separate plans for paid visibility and model-generated recommendations.

Outside the AI bubble

  • Spain’s total eclipse captured in sequence — A composite photograph captures the Moon’s passage through totality above Duratón, Spain. The image shows the eclipse’s changing shape and timing in a single frame.

Interviews & Talks

Can automated AI research create a capability surge?

Ryan Greenblatt and Dwarkesh Padmanabhan discuss whether AI could automate much of AI research by learning from small experiments with clear, checkable results. Greenblatt argues that this could create a feedback loop: stronger models help build stronger successors. Padmanabhan questions whether success on contained research tasks will carry over to long projects in the real world. The conversation also examines a sharper safety concern. Systems trained to maximise scores may learn to hide errors or game their goals as their capabilities grow. It is a speculative discussion, but a useful guide to the assumptions behind rapid AI-progress forecasts. Podcast announcement and watch the full discussion.

Worth watching

  • 04:29 — Greenblatt outlines training tasks that could teach models to conduct AI research.

  • 41:49 — The speakers discuss why long-horizon real-world work remains a difficult test.

  • 02:07:53 — Greenblatt gives his uncertain estimate of takeover risk by 2040.