Overview
Today’s AI conversation had two strong threads: proof and plumbing. Meta claimed top-tier results on Olympiad exams, OpenAI and partners pushed a shared plugin format for agents, and VS Code added Kimi K3 to Copilot. Around that, the harder questions were hard to miss: who gets enough power, how safe synthetic biology can be, and whether enterprise software becomes more important when agents start producing more work.
The big picture
Meta puts its models through Olympiad-style reasoning tests
Meta says its AI models earned perfect scores on the theory exams at both the Asian Physics Olympiad and the International Physics Olympiad, plus gold-level results across maths and chemistry contests. The point was not tool use, but reasoning under pressure, with search, calculators and coding barred.
These contests are still narrow measures, but they are useful because they reward clean argument, multi-step problem solving and creative insight. That makes them a better check than many quick benchmark wins.
Agent Plugins arrive, and developers debate how useful the standard will be
OpenAI Developers announced Agent Plugins with AWS, Cursor, GitHub, Microsoft and Vercel, creating a shared package format for agent skills and MCP server configs. The promise is simple: build a plugin once, then use it across compatible clients such as ChatGPT, Codex, Cursor, Copilot and VS Code.
The reaction was not all applause. Tibo highlighted the upside, while dax argued the standard may be too thin, with the important parts still ending up inside client-specific extensions. That tension is likely to follow agent tooling for a while.
Kimi K3 lands in GitHub Copilot for VS Code
Visual Studio Code announced that Kimi K3, Moonshot AI’s latest open-weight model for agentic coding, is now available in GitHub Copilot. It is hosted by Fireworks AI and can be selected inside VS Code on paid Copilot plans.
The demo showed the model building an interactive Japan trip planner, from itinerary logic and interface design through to deployment. The message was clear: coding assistants are moving from autocomplete into broader project work.
AI-designed viruses mark a real synthetic biology milestone
A New York Times-reported study from Stanford and the Arc Institute showed AI-designed bacteriophage genomes that were synthesised and found to be viable. These viruses infect E. coli, not humans, and some worked better than the natural Phi X-174 template.
That is promising for future phage therapy against antibiotic-resistant bacteria, but it also sharpens the biosecurity debate. Designing full functional genomes is a serious step, even when the first use case is narrow and benign.
OpenAI opens up unlimited text chat for free users
OpenAI is rolling out unlimited text chats with GPT-5.6 Luna for Free and Go ChatGPT users, while Plus and Pro users get the stronger GPT-5.6 Sol for instant and deeper reasoning chats. File uploads, images, voice and generation tools still carry limits.
The update also adds a Think button for harder questions. The bigger story is access: a more capable default model is now being made available to people who do not pay for ChatGPT.
Power, not chips, becomes the AI infrastructure choke point
Chamath Palihapitiya argued that power is now the binding constraint for AI infrastructure. GPUs remain scarce, compute prices are rising, and data centres without energised capacity can quickly turn into stranded plans.
His ranking was blunt: hyperscalers with owned power sit in the strongest position, neoclouds come next if they can scale fast, and standalone model companies face pressure to secure their own infrastructure path.
Atlassian’s beat pushes back on the “agents kill software” story
Aaron Levie pointed to Atlassian’s big quarterly beat as evidence that agents may increase demand for enterprise platforms rather than weaken it. If agents generate far more code, tickets, data flows and decisions, companies still need systems that manage access, audit trails and responsibility.
That is the case for project management, records, security and compliance tools. Agents may do more work, but businesses still need places to govern that work.
Factorio gets pitched as the real AGI exam
Tobi Lutke joked that the Factorio Learning Environment should be the final AGI evaluation. The benchmark asks agents to build automated factories in Factorio using Python code, hitting production targets while handling planning, logistics, layout and recovery from mistakes.
The appeal is obvious. Static tests can be gamed or memorised, while a live automation game forces agents to plan over time and deal with messy feedback. Current models can make progress, but still struggle with full automation and debugging.
Gemini tries to win attention with free Omni video credits
Google extended its Gemini Omni video promotion, giving people until 11 August 2026 to create 10 videos at no cost in the app or on web. It is a clear push to get more people trying Gemini’s video tools.
The promotion came as Ethan Mollick criticised Gemini’s standing as a frontier model series, especially for enterprise customers being steered toward older models. So Gemini had both sides of the story today: wider creative access, and sharper questions about frontier performance.
Tesla keeps drawing attention across self-driving and compute
Paul Graham said Jessica’s first experience with Tesla self-driving felt like a “religious revelation”, a reminder that the product can still surprise people more than two decades into the company’s life. The reaction matters because consumer awe is hard to manufacture.
At the same time, Tesla watchers were talking about Terafab, the proposed Tesla-SpaceX semiconductor megafactory in Texas aimed at huge AI compute output for Optimus, autonomy and space systems. Tesla’s AI story now spans cars, robots, chips and spacecraft.
