Daily

The math community has spent years asking how you know when an AI result in mathematics is real. OpenAI's answer arrived August 1: ten formally verified proofs for problems that had been open for at least a decade each, generated by an internal version of their upcoming Astra model, validated in Lean 4 with machine-checkable certificates. Anyone with a Lean installation can check these proofs themselves. That is not a small claim. That is the claim that closes the "but is it actually correct" question.

The ten problems are not selected for approachability. The list includes the existence of non-sofic groups, an open question in group theory that career group theorists have spent decades on. A disproof of Connes' Rigidity Conjecture in von Neumann algebras. High-dimensional sphere packing bounds down to the Cohn-Elkies threshold, the theoretical tightest possible bound in that regime. Binary and spherical codes with exponentially improved maximum-size bounds. Arithmetic circuit complexity. Quantum complexity. Monochromatic triangles in multicolored graphs. Operator algebras. The breadth is what's unusual. A typical landmark AI math result picks one problem and plants a flag. OpenAI planted ten.

The model is the part that should not slide by. This is an internal version of Astra, not GPT-5 or anything currently in production. OpenAI is showing the frontier without opening the gate. This is at least the second time a major frontier lab has published results from a capability it withholds from customers. The pattern is: the most capable model the public knows about is not the most capable model that exists. What Astra is doing with mathematics is what the next tier of API access will eventually enable. The blog post is a preview. The invoice comes later.

The blast radius lands first in the research community. Academic mathematicians working in any of the ten domains now have formally verified solutions they will read, dispute, or confirm over the coming months. If any Lean 4 certificate contains an error, the community will surface it. If none do, OpenAI has done something that belongs in a different category from all prior AI math results.

The pattern across labs is what makes this more than a single release. Mistral published Leanstral 1.5 in early July: Apache 2.0, six billion active parameters, purpose-built for Lean 4 formal proof engineering, trained to write and verify proofs in the exact format OpenAI just used for its ten results. Google DeepMind has been on AlphaProof since 2024. Three labs, different continents, different architectures, different business models, all independently reached the same conclusion: Lean 4 is where AI math claims become credible. This is not a lab decision. It is the research community making a standards choice, and the community chose Lean 4.

The read: The format competition in AI mathematics is over. Lean 4 won. What is now a race is who can use it to solve harder problems faster, and whether the proofs that come out are open-weight, partner-restricted, or behind an API key. Mistral bet open. OpenAI bet closed. That is a meaningful difference for anyone who needs to work with this technology rather than just read about it.

The builder's move: Leanstral 1.5 is downloadable today under Apache 2.0. If any of the ten proofs is adjacent to your domain, read the Lean 4 certificate rather than the blog post. The certificate is where the claim actually lives. Source: openai.com/index/ten-advances-in-mathematics

Sign in with ChatGPT went into beta on August 2. First wave: Airtable, GitLab, HubSpot, Notion, Supabase, Vercel. Mechanics are standard OAuth: name, email, and profile picture shared with the partner app. Plugin access remains separately approved.

The mechanism is not novel. Sign in with Google has been around since 2009. Sign in with Apple since 2019. Sign in with GitHub is the default for developer tooling. The game every platform plays on the way to becoming infrastructure is to make its session the primary relationship the user has, with everything else authenticating against it. OpenAI is now playing that game. The user's GPT-5 conversation window is the same session where they authenticate to Notion and deploy to Vercel. That is a different relationship between OpenAI and its users than "you have a subscription."

The blast radius: SaaS products outside the first wave are on a clock. Once a meaningful share of users can sign in to Notion and Supabase via ChatGPT in fewer steps than they can sign in via email, every other onboarding flow looks like friction by comparison. Products that integrate win the next comparison. Products that don't have to explain the gap.

Three platform expansions shipped the same Sunday: Sign in with ChatGPT, Codex Remote generally available on all plans (desktop-to-mobile agent handoffs), and ChatGPT Voice in the desktop Work and Codex interfaces. This is not a quiet day for OpenAI. It is a coordinated product push across identity, agentic infrastructure, and voice, landing together on a weekend when the news cycle is thin.

The builder's move: if your application is not in the first integration wave, understand the OAuth specification before your users start asking why they cannot sign in with ChatGPT the way they can with Google or Apple. The second wave follows the first faster than the first wave followed the announcement. Sources: openai.com/news

Grok Imagine Video 1.5 launched August 1. Native 1080p output for text-to-video, image-to-video, and reference-to-video workflows. Multi-reference support up to seven visual anchors per generation: face, product, location, style, and more. Voice consistency via multi-channel audio references. Built on Aurora-driven models. Available on Grok web, iOS, Android, and the xAI API. Subscriber-exclusive.

The 1080p threshold matters because most AI video generation at scale has been running at 720p or upscaling to nominal higher resolutions. Native 1080p without upscaling changes what the output is useful for. Short-form content at broadcast quality, product video for professional distribution, reference-anchored campaign material: those workflows have different requirements than exploratory generations, and 1080p is the threshold where they start to apply.

Voice consistency across a generated video has been a recurring pain point. The underlying problem is that audio and video inference have historically been separate passes that do not share state, so the voice drifts unless you re-inject the reference on every segment. Multi-channel audio references are architecturally different: multiple audio anchors lock the voice in a way that is distinct from how most competitors have approached it. Whether it holds across long generations at production scale is the test that matters next.

The contrast with the day: OpenAI is proving theorems. xAI is hitting broadcast-quality video resolution. Anthropic was quiet. These labs are not competing on the same surface right now. The frontier is not one race. It is three separate bets about where the next ten million users come from.

The builder's move: if you have video generation workflows on Aurora, test 1080p output against your delivery infrastructure before scaling. The seven-anchor limit is a concrete constraint worth knowing before you design around it. Source: x.ai/news/grok-imagine-1-5

August 2 is when EU AI Act enforcement capabilities went live: regulators now have the statutory right to request information from AI providers, access models directly for evaluation, and issue recalls. This is not a new law. The Act has been on the books for two years. But enforcement authority is a different instrument than regulatory text. The clock is now running for every lab with European operations.

Mistral is the most directly exposed. The Paris-based lab is the only major frontier provider that is European in domicile, not just in regulatory scope. Mistral signed the General Purpose AI Code of Practice alongside Anthropic, Google, IBM, Microsoft, and OpenAI, which represents the industry's attempt to preempt mandatory requirements with voluntary compliance. How voluntary compliance holds under an enforcement regime with actual recall authority is an open question. Mistral's unnamed frontier Mixture-of-Experts model is currently in early partner access, with no confirmed public release date. Every release it makes from here forward happens under this enforcement context, not the previous one.

The blast radius is gradual but structural. Labs with European operations will face information requests. Models deployed in high-risk classifications in the EU will face additional scrutiny. The compliance surface for frontier model providers in Europe just became real rather than anticipated. Anthropic, OpenAI, and Google operate European subsidiaries and data centers. None of them are immune.

The builder's move: if you are deploying AI applications in the EU, the enforcement activation is a prompt to audit your compliance posture before an information request arrives. The labs' Code of Practice commitments are now the floor, not the ceiling, of what regulators expect.