Hi tech & AI enthusiasts, this week's highlights:
GPT-5.2 Made a Real Physics Discovery (First Time AI Has Done This)
Gemini 3 Deep Think Solved 18 Unsolved Research Problems
Claude Sonnet 4.6: Near-Flagship Performance, Now Free
Plus: OpenClaw Joins OpenAI, Manus Hits Telegram, Grok 4.2 Beta
The Lab: How to Use Gemini Deep Think for Business Research
AI Image of the Week: Contemporary art print of a sleek black horse
Let’s get into it.
▲ AI SCIENCE
Midjourney
OpenAI's GPT-5.2 derived a new mathematical formula in theoretical physics that researchers had spent years trying to crack. This is the first time an AI has independently discovered something genuinely new in science. The AI didn’t just helping with calculations, but spotted a pattern scientists had missed entirely.
What you need to know
It spotted what decades of research couldn't: Physicists had been calculating specific particle interactions by hand up to a certain complexity level but couldn't find a general formula. GPT-5.2 identified the pattern, proposed a formula, and spent 12 hours proving it was correct.
Verified by Harvard, Cambridge and Princeton: The AI's formula was independently confirmed by researchers at multiple universities and published as a peer-reviewed scientific paper.
Already applied to gravity: The same method is now being used to analyse gravitons, the particles behind how gravity works. More results expected in future publications.
Key Ideas
▲ Why it matters: AI has been useful for searching and summarising existing science for a while now. But this is different. It found something genuinely new that humans hadn't worked out. If AI can start producing peer-reviewed discoveries, research timelines in medicine, materials, and engineering could shorten dramatically. This felt like a different kind of milestone this week.
▲ AI RESEARCH
Google upgraded Gemini 3 Deep Think. An AI built for complex problem-solving in science and engineering. It solved 18 previously unsolved academic problems and disproved a mathematical concept that had stood since 2015.
What you need to know
Caught an error that peer reviewers missed: A Rutgers mathematician used Deep Think to review a physics paper. It spotted a logical flaw that reviewers had not picked up. Researchers at Duke also used it to advance semiconductor manufacturing.
Scored 48.4% on the hardest AI test: Humanity's Last Exam is designed to be nearly impossible for AI. It includes questions that stump world experts. Gemini 3 Deep Think hit 48.4%, the highest score recorded.
Now available via API: Previously limited to Google AI Ultra subscribers in the app. Developers and enterprise teams can now access it through the Gemini API for the first time.
Key Idea
▲ Why it matters: Deep Think is being positioned less as a chatbot and more as a research co-worker. One that handles the hard analytical work whilst you focus on what to investigate. For teams doing data analysis, technical research, or complex modelling, this is moving from interesting demo to genuinely useful tool.
▲ AI MODELS
Early Sonnet 4.6 users are seeing human-level capability
Anthropic released Claude Sonnet 4.6, making it the new default for all users including free accounts. It performs close to Opus 4.6 (their most powerful model) on most tasks, costs significantly less to access via API, and now the free tier gets it too.
What you need to know
Free users now get near-flagship AI: In early testing, users preferred Sonnet 4.6 over the previous Opus 4.5 flagship 59% of the time. That level of performance is now available at no cost for standard use.
Reads your entire codebase or contract in one go: The context window expanded to 1 million tokens. That means it can hold a full codebase, dozens of reports, or a lengthy legal document in memory whilst working through it.
Better at navigating software: Anthropic says it can use software "the way a person would". Clicking, typing, filling forms across multiple browser tabs. Making it more reliable than previous versions for automation tasks.
Key Idea
▲ Why it matters: The AI performance-to-price gap is collapsing fast. What paid users had access to a few months ago is now free. Businesses running Claude via the API can now get near-Opus results at Sonnet pricing. If you haven't revisited what tier you're on, now is a good time.
▲ EVEN MORE NEWS
Other big ideas from the past week.
Peter Steinberger joining OpenAI to work on bringing agents to everyone
▲ OpenClaw's Creator Joins OpenAI: OpenClaw is an open-source AI agent with 196,000 GitHub stars that went viral in January. OpenAI just hired its creator Peter Steinberger to lead personal agent development. The project stays open source.
▲ Manus AI Agents Now in Telegram: Meta-owned AI startup Manus launched its agent on Telegram. It can search for flats, book hotels, build websites, and analyse documents all inside a chat. (did someone say OpenClaw?)
▲ Grok 4.2 Public Beta Live: xAI released the public beta of Grok 4.2 today. Musk says it learns rapidly with weekly updates baked in. In early testing it outperformed GPT and Gemini on a stock-trading simulation, turning $10,000 into $12,193 over 14 days. Grok 5 is targeting Q1 2026.
The Lab
▲ AI EDUCATION
Gemini 3 Deep Think just opened up via API this week and it works differently to standard AI. Instead of answering immediately, it spends time reasoning through a problem, checks its own logic, and flags where it's uncertain. This week it solved 18 unsolved research problems and caught an error a human peer reviewer missed.
What you'll learn:
How to access Deep Think via the Gemini app or Google AI Studio
How to structure research prompts that get specific, useful answers
How to read the reasoning panel, where the real value is
Use cases for competitive analysis, document review, and strategic planning
🔬 Interested in learning about AI for business? Join The Lab waitlist.
▲ AI IMAGE SHOWCASE
Create this type of image with Midjourney
--sref 1784241165
Graphic contemporary art print of a sleek black horse eating a stacks of US dollars and facing to the viewer, strong red backdrop, halftone texture, subtle grunge scratches, symmetrical layout, bold negative space, Asian poster design aesthetic, screen printed look --sref 1784241165Image Usage Suggestion: Chinese New Year, finance and investment content, startup fundraising campaigns, articles about AI costs and ROI, economic commentary, "burning through budget" narratives, fintech marketing, or any bold visual that needs to say money without being boring about it.
💌 Reply to this email with your AI image generation or suggestion of what you’d like to see and I'll feature it in a future newsletter.
▲ SHARE YOUR THOUGHTS
Help shape the content you see here by giving feedback
Have specific feedback or want to get in touch? Reply to this email and I’ll get back to you.
Know someone who’d love this newsletter? Forward it to a friend and have them sign up here.
Thanks for reading, until next time.
Stay curious,
Matt Lok, Editor
Brought to you by Metalabs. Digital marketing consulting specialising in AI.