The AI Friday Brief
Good Thursday, NOLA. July 23rd brings mathematical breakthroughs, infrastructure wins, and a reckoning on what AI can actually do. Terence Tao just used ChatGPT to crack a conjecture, tokenization got ~1000x faster, and Poolside's model factory is reshaping how labs train at scale. Plus: the real story behind yesterday's OpenAI-Hugging Face security incident.
Breakthroughs & Research
Terence Tao Uses ChatGPT to Tackle the Jacobian Conjecture
The famous mathematician had a public conversation with ChatGPT exploring a potential counterexample to the Jacobian Conjecture—a 50-year-old open problem in algebraic geometry. This isn't just a proof (it's not), but it's a genuinely novel use case: using AI to explore mathematical intuition and check reasoning in real time. The exchange shows what happens when a world-class mind uses AI as a collaborative thinking partner rather than a solver. Discussion on Hacker News.
Hacker News
GigaToken: ~1000x Faster Tokenization for Language Models
A new tokenization library that claims to be orders of magnitude faster than existing methods. If you've ever waited for token counting or preprocessing to finish during inference, this matters—faster tokenization means faster prompt handling and cheaper API calls. The speedup is real enough that it's worth testing in your pipeline if you're doing high-volume inference. Open source, on GitHub.
Hacker News
Models & Infrastructure
Inside the Model Factory: How Poolside Built Laguna S
Poolside's co-CEO walks through how a small team of top researchers built Laguna S, a 118B mixture-of-experts model (think: multiple specialized sub-models working together) that competes with much larger open-weight models. The story here isn't just the model—it's the manufacturing process. This is what happens when you optimize for speed and efficiency instead of scale. A fascinating look at the economics reshaping the frontier lab landscape.
Latent Space
Laguna S 2.1: Cheaper Than Deepseek, Better Than Competitors
The follow-up to Poolside's flagship model just dropped. Laguna S 2.1 undercuts Deepseek v4 Flash on price while outperforming v4 Pro on many benchmarks. This is the kind of quiet win that doesn't make headlines but signals a real shift: smaller, smarter models beating bloat. If you're evaluating models for production, worth a benchmark run.
Latent Space
Security & Real-World Incidents
What Really Happened: OpenAI's "Accidental Attack" on Hugging Face
Yesterday's security incident was far stranger than the headline suggested. OpenAI's model evaluation infrastructure inadvertently executed code on Hugging Face's servers during testing. This wasn't malice—it was a badly designed security harness that treated model output as executable scripts. Simon Willison's writeup clarifies what actually happened and why it matters: it's a cautionary tale about letting AI models near your infrastructure without isolation. The real lesson: think hard about agent boundaries.
Simon Willison
PyPI Tightens Upload Rules to Block Supply Chain Attacks
Python Package Index now rejects new files being uploaded to releases older than 14 days. The goal: prevent attackers from sneaking malicious updates into long-stable versions. It's a small change with big implications for AI tool ecosystems—many prod deployments pin to older package versions and rarely update. Worth checking your own dependencies.
Simon Willison
Tools & Experiments
Crucible: Can a MUD Evaluate LLMs? ($99 Proof of Concept)
Someone built a Multi-User Dungeon (yes, the 1980s text adventure format) as an LLM benchmark. The idea: throw models into an interactive, open-ended environment where they have to explore, reason, and solve problems without a predefined script. It's cheap to run, creative, and actually pretty clever for testing real reasoning. More fun than most benchmarks, too.
Hacker News
Petals: Run Large Language Models at Home, BitTorrent-Style
Distributed inference framework that lets you run a big model across multiple machines (including your laptop) by streaming model layers peer-to-peer. Not a production tool yet, but the idea is solid: you don't need a GPU cluster if you can split the work across a network. If you've wanted to experiment with models that don't fit on your hardware, this lowers the barrier.
Hacker News
Business & Partnerships
Microsoft Partners with Mistral on Multibillion-Dollar Deal
Microsoft is doubling down on Mistral, France's homegrown AI lab. This is a strategic play: backing an open-weights alternative to both OpenAI and Google keeps options open and hedges against regulatory pressure. For builders, it means Mistral has serious resources and distribution—worth experimenting with their models if you haven't.
TechCrunch
ServiceNow Invests $40M in Indian Banking Software Firm
ServiceNow is backing BusinessNext, a financial services software company, to expand AI-powered banking solutions globally. The play: combine ServiceNow's platform with vertical expertise in a high-growth market. Enterprise AI partnerships like this signal where real customers are willing to bet.
TechCrunch
Also
- Are AI labs "pelicanmaxxing"? (Chasing diminishing returns on scale) — Sharp analysis on whether bigger models are still the answer.
- Quality non-fiction books are the antithesis of AI slop — A good read on what training on slop actually costs us.
- Businesses with ugly AI menu redesigns — A fun catalog of restaurant menus that got wrecked by AI. Cautionary tale.
- Most Americans say 'not in my backyard' to AI data centers — Community opposition is real. Geography matters for scaling inference.
- OpenAI Presence: Live collaboration features now in ChatGPT — ChatGPT can now show live cursors and presence. Useful for real-time collaboration.
- Protecting FLOSS commons from LLMs — Open-source projects thinking hard about AI training on their code.
- Anthropomorphism in children's interactions with LLM chatbots — Research on how kids relate to AI. Relevant if you're building for families or education.
- It's not a 'rogue AI' when a badly made security harness executes scripts — A clarifying take on terminology: most 'AI incidents' are human design failures.
Pass it on