OpenAI o1 got the exact or very close diagnosis in 67.1% of early ER triage cases from real patient records, compared with roughly 50% to 55% for physicians.

My goodness, Anthropic. Please fix, I'm getting these a lot atm.

Google announced four Gemini AI upgrades at The Android Show today, including the Googlebook laptop!

What are we expecting from OpenAI today?

Codex/ChatGPT super app coming this week?

Traditional software platforms are undergoing a structural reset. - Tools like Jira and Salesforce are fading into silent, headless systems of record - AI agents are becoming the new front door to everyday work - Per-user seat pricing is starting to crack

Let's try and clear this up. Is this text book real? Firstly, the image being passed around (without referring to the actual source in most cases btw!) appears (at least partly) AI generated, but this is likely from whoever added the annotations and cleaned up the text.

Thinking Machines Lab has released a research preview of interaction models that process audio, video, and text in continuous 200ms micro-turns - unlocking real-time human–AI collaboration without external turn-taking harnesses.

Did Google DeepMind move from London? 😜

Good question, do we want Codex added to the ChatGPT app or ChatGPT added to the Codex app? Or something new?

Opus 4.7 paired with Cursor CLI tops the Artificial Analysis Coding Agent Index at 61 points.

Customer service postings are now "outpacing the overall job market" - despite predictions that AI would replace this work. A longer view image than the original post: www.a16z.news/p/charts-of...

Gemini Omni video model apparently seen in the wild but many users (though I've only seen the "Zachary" shot so far) in the Gemini mobile app. www.reddit.com/r/GeminiAI/...

Happens to them all at some point. 😀 x.com/burkov/stat...

Remote access coming to Codex. Screenshot from Code app on macOS. Must be coming this week.

Did you make the AI 1000 on Digg? No idea on the metrics here btw. Cool new AI news site though. di.gg/u/x/koltreg...

I have zero interest in the 50% success rate chart, the 80% success chart is a far better representation of where we are at. People are sharing the linear scale chart, which is great, but I do like looking at the log scale for another perspective. metr.org/time-horizons/

Hermes Agent is #1 on OpenRouter today - and it’s right up there on the weekly rankings too. Daily charts swing a lot, so it’s worth waiting for the weekly (and longer-term) view before calling it a real trend. openrouter.ai/rankings?vi... (weekly is the second image)

Google DeepMind has partnered with Fenris Creations to research general-purpose AI inside EVE Online’s player-driven simulation. The work will happen only in controlled offline versions of the game that stay disconnected from the live Tranquility servers.

This is why Claude Mythos has not been released by Anthropic.

Big Codex update coming? The super app? Perhaps Thursday?

New image models can now produce forged documents that look convincingly real. - Fake council letters with spot-on logos - Realistic paper creases, shadows, and matched lighting - Believable business receipts and invoices

Grok Build seen in the wild. Includes marketplace, skills, MCPs, plugins, projects, planning, GitHub, web browsing and files support. Can sign in with xAI API or via SSH or Docker. Web and CLI versions. Looks like it might be SuperGrok Heavy plans only. Coming to Windows, Mac and Linux.

OpenAI has finally shipped a browser plugin (be it only for their coding agent). I’ll have more to say once it’s available in my country, but this is a textbook example of how backwards the rollout was.

AI agents grew 15x year-over-year in Microsoft 365 according to the 2026 Work Trend Index.

Continuous agentic loops can chew through output tokens fast. - Hidden reasoning steps drain your budget before you see any value - API calls and compute costs can spike far beyond standard chat - Output spend is hard to predict and even harder to control

Grok Computer (local computer use) release soon? Spotted in the Gro mobile app.

Keep waiting for a new Advanced Voice Mode in ChatGPT I'm afraid:

GPT 5.5 leads Blueprint-Bench 2 with a score of 0.362. AI agents get roughly 20 interior photos per apartment and must output a full floor plan with correct room connections and consistent scale across 50 apartments in sequence while using a persistent notepad to refine their approach.

Autonomous agents running code in the background create serious security risk. - AI can inadvertently execute malicious code or pull in compromised packages - Data exfiltration is a major threat to corporate networks

AI companies improvise more than they strategise. Ethan Mollick notes that what looks like exponential progress often masks a string of reactive pivots - prompted by shifting markets and surprising model behavior - not some grand master plan.

Nobody has won the personal AI agent race yet. Peter Yang tested OpenClaw, Hermes, Claude Code, Codex and Gemini against the ten must-have capabilities for a true chief of staff in his deep dive.

GPT-5.5 leads every category and modality on the new LLM Stats Index. The Index aggregates 200 plus benchmark results into a single TrueSkill rating per model spanning law healthcare coding tool calling vision and reasoning after 3.2 years of tracking every frontier model release.

Anthropic is doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans, following a new partnership with SpaceX.

SubQ 1M-Preview delivers 12M-token context at linear cost scaling. It runs on a sub-quadratic sparse-attention architecture from Subquadratic that uses content-dependent selection to route exact attention only to meaningful token pairs.

Misleading conscious AI claims should be treated as an AI safety issue due to their legal and ethical risks, according to Luiza Jarovsky's article.

AI job anxiety rises where Claude is already doing more of the work.

Project Mariner, Google's web browser agent, was shutdown 2 days ago. People on the project have been moved, and the tech is being used elsewhere. Gemini Agent is coming out of Labs (US only) soon? labs.google.com/mariner/lan...

AI productivity is driving Coinbase’s 14%, or about 700 people, headcount cut and full shift to AI-native pods.

Gemini 3.2 Flash spotted in the Gemini app's model selector. "All-round help". More below.

Recent agentic models now handle judgement in complex long-run tasks. Ethan Mollick points out that you cannot complete the high-complexity work current AIs perform without some form of judgement.

Open models trail closed ones by roughly eight months in overall capability.

Gemini Agent coming soon, possibly at Google I/O. As shown, included options for schedules, tasks and skills. Previous coverage from last year (see thread below) suggested it could browse and interact with websites.

Google filed a trademark for ALPHAMATTER on 17th April 2026. The application from Google covers Ai software for simulating chemical and material science processes plus tools to run train and deploy models in chemistry physics and material design.

Anthropic is working on Orbit, a "proactive" assistant for Claude Cowork, similar to ChatGPT's Pulse. It delivers personalised insights and briefings pulled from Gmail, Slack, GitHub, Calendar, Drive, Figma and other connected apps without you needing to ask.

"We have no more manually written code anywhere at the company" - Anthropic

Jack Clark puts the odds of automated AI R&D at 60% by the end of 2028. After combing through hundreds of public data sources, he argues in Import AI 455 that frontier models may soon be able to autonomously train their own successors.

Google's new Omni reference leaked in Gemini's video generation interface. Omni sits close to Toucan the internal name for the current video generation tool powered by Veo and could outperform Veo 3.1 if released.

Frontier AI models still fail at basic physical tasks like planning CNC machining of a simple brass part.

Midjourney v8.1 image quality and sharpness just got better - particularly noticeable with SREFs, moodboards and HD images, though you'll see improvements globally. www.midjourney.com/imagine