The Weekly Waypoint, Issue #18
For two years, the promise of AI has been exactly that: a promise. Impressive demos, contested benchmarks, lots of "could" and "might" and "in the future." This week, the future showed up.
An OpenAI model disproved an 80-year-old math conjecture that humans couldn't crack. Anthropic's Claude found over 10,000 zero-day vulnerabilities in critical software that the world's best security teams missed. Alibaba's Qwen3.7 ran for 35 hours straight, autonomously optimizing code for its own custom chip. And Google didn't just update search -- it replaced the search box with AI agents.
This wasn't another week of incremental progress. This was AI proving itself across math, security, autonomy, and infrastructure in the same seven days. The question is no longer whether AI can do the work. It's whether we're ready for what happens when it does.
Let's get into it.
OpenAI Disproves an 80-Year-Old Erdos Conjecture
An OpenAI model just did something no mathematician could in eight decades: it disproved a central conjecture in discrete geometry attributed to Paul Erdos. The model produced a counterexample that human mathematicians verified as correct. Scientific American called the result "amazing." The mathematicians who checked it were most surprised not by the proof itself, but by how the AI arrived at it. This is AI doing real mathematical discovery -- not pattern matching, not approximation, but a genuine contribution to human knowledge.
Claude Mythos Finds 10,000+ Zero-Days
Anthropic's Project Glasswing, running with 50 partners including Apple, Google, Microsoft, and Amazon, used the Claude Mythos Preview model to find more than 10,000 high- and critical-severity vulnerabilities in widely used software. Zero-days that the original developers didn't know existed. Anthropic explicitly warned that Mythos finds bugs faster than developers can patch them -- which is why the model isn't being released publicly. The most powerful security AI in existence is too dangerous to put in anyone's hands without strict access controls.
Google Kills the Search Box
At I/O 2026, Google didn't update search. It replaced it. The traditional search box is now an AI-powered interface with "information agents" that monitor the web continuously and answer complex multi-step questions. Gemini 3.5 Flash powers the whole thing. Sundar Pichai called it the "agentic Gemini era." The ten blue links are officially over. This is the biggest change to the most visited website on earth, and it happened with a keynote, not a beta.
Qwen3.7-Max: 35 Hours of Autonomous Work
Alibaba's new Qwen3.7-Max model ran continuously for 35 hours without human intervention, optimizing code for its own custom chip. It wasn't following a script. It was planning, executing, testing, and iterating on its own. This is the longest confirmed autonomous agent run from a major AI lab, and it's a glimpse at something most people aren't ready for: AI that doesn't just assist you, but works independently for days at a time.
Karpathy Joins Anthropic
Andrej Karpathy, OpenAI co-founder and former Tesla AI director, joined Anthropic to lead pre-training efforts. In practice, that means using Claude to help train the next Claude. The person who helped build GPT is now helping build Claude's next generation. The talent war in AI has always been fierce -- but this one is personal. OpenAI and Anthropic are no longer just competing on benchmarks. They're competing for the people who build the benchmarks.
Altman Offers $2M in Tokens for Equity
Sam Altman announced at Y Combinator that OpenAI will give every YC startup in the current batch $2 million in API tokens in exchange for equity. He called these "tokenmaxxing startups" -- companies built on the premise that cheap, abundant AI compute is the new seed capital. It's a land grab disguised as a founder-friendly deal. OpenAI is buying distribution and loyalty at the earliest stage, before startups even know what model they'll standardize on.
EXIM Bank Backs AI Exports with Billions
The US Export-Import Bank launched the ExportAI Initiative, channeling over $100 billion in unused lending capacity toward American-built AI exports. The Commerce Department will help structure full-stack AI packages -- chips, models, data centers -- for foreign buyers. The US government is now explicitly treating AI as a strategic export, subsidizing it the way it once subsidized agriculture and aerospace. If you're building AI infrastructure, this is a new source of government-backed financing.
OpenAI Gets Serious About AI Detection
OpenAI announced it's adopting C2PA content credentials and Google's SynthID watermarking for all images generated by ChatGPT. They're also building a public verification tool. This is the most serious commitment any major AI lab has made to content provenance. The timing isn't coincidental -- as AI-generated content floods the internet, the companies building the models are starting to feel the political and social pressure to clean up their own mess.
The takeaway: This was the week AI stopped being a promise and started being a proof. It disproved conjectures. It found vulnerabilities. It ran for days without supervision. It replaced the world's most visited website. The debate about whether AI can deliver is over. The debate about what happens next is just beginning.
Deep Dive This Week
The Proving Week: How AI Earned Its Keep Across Math, Security, Autonomy, and Infrastructure
The full breakdown: what each development means on its own, how they connect, and how to position yourself for a world where AI doesn't just assist -- it proves. Pro members get the full deep dive.
Not Pro yet? Upgrade for $6/month
See you next Monday. -James
This week, the machines didn't ask for permission. They just proved it.