Boris Golden

This is wild. 143 million people thought they were catching Pokémon. They were actually building one of the largest real-world visual datasets in AI history. Niantic just disclosed that photos and AR scans collected through Pokémon Go have produced a dataset of over 30 billion real-world images. The company is now using that data to power visual navigation AI for delivery robots. Players didn't just walk around with their phones. They scanned landmarks, storefronts, parks, and sidewalks from every angle, at every time of day, in lighting and weather conditions that staged photography would never capture. They documented the physical world at a scale no mapping company with a fleet of vehicles could have replicated on the same timeline or budget. Niantic collected this systematically, data point by data point, across eight years, while users thought the only thing at stake was catching a rare Charizard. The most valuable AI training datasets in the world aren't being assembled in data centers. They're being built by people who have no idea they're building them.

I got the chance to write a guest post on @OpenAIDevs Blog about 15 lessons learned building ChatGPT Apps. developers.openai.com/blog/15-lesson… If you're building Apps, I think it's worth a read!

Today we’re launching KORA, the first public benchmark for AI child safety. x.com/korabench/stat…

Stéphane Mallat (médaille d'or CNRS): l'IA n'est pas un perroquet stochastique, elle généralise ; "calculer des probabilités conditionnelles, c'est découvrir la structure du problème", et c'est cela l'intelligence, bien davantage que la logique... même en maths, et même pour nous

programming always sucked. it was a requisite pain for ~everyone who wanted to manipulate computers into doing useful things and im glad it’s over. it’s amazing how quickly I’ve moved on and don’t miss even slightly. im resentful that computers didn’t always work this way

Un article paru ce jeudi 12 juin dépeint un tableau très éloigné de la réalité concernant Sorare. La trajectoire économique de notre entreprise est fidèle au plan que nous exécutons depuis plusieurs années. Après avoir investi des sommes importantes pour conquérir le marché entre les années 2021 et 2023, Sorare est désormais proche de la rentabilité. Ces résultats récompensent nos efforts importants menés ces dernières années pour (i) faire évoluer le modèle d'affaires de Sorare, en générant $50m de CA dès la première année (2024), (ii) réduire méthodiquement nos coûts (notamment de partenariats, divisés par 5), (iii) mettre à disposition de nos utilisateurs ce qui reste à ce jour le plus vaste portefeuille de partenaires sportifs du gaming. L'ensemble des ces éléments nous permettra de finir l'année 2025 avec un “burn” en forte baisse, une trésorerie plus importante qu'initialement projetée, et des projets forts de croissance pour accélérer notre développement durable et rentable. Nous construisons, étape par étape et avec détermination, un groupe indépendant et innovant avec les moyens de ses ambitions : être le leader mondial du divertissement sportif.

Some key takeaways: 1. 90% of Claude’s code is now written by Claude Code, and this has completely transformed how they build products. The bottlenecks have shifted from engineering (writing code) to decision-making (what to build) and merge queues (getting code into production). This is happening faster than anyone expected—Mike thinks most companies will reach this point within a year. 2. Claude Opus 4 has crossed a critical threshold where it has become a genuine thought partner for strategy. Mike now uses Claude as his go-to product strategy partner, and for the first time, it provides novel angles he hadn’t considered. This shift from “helpful but obvious” to “genuinely creative” happened just in the past month. 3. Anthropic isn’t trying to beat ChatGPT at consumer mindshare—instead, it’s doubling down on differentiation and focus. Anthropic is leaning into their strengths: developers love them, builders use them to create things, and they excel at agentic behavior and coding. Mike’s advice: “Embrace who you are and what you could be rather than who others are.” 4. Product teams working directly with AI researchers drive 10x more value than those just building UX on top of models. Mike has shifted almost all product resources to embed with research teams, working on post-training and fine-tuning rather than just using models off the shelf. If you’re building something anyone could build with public APIs, you’re missing the opportunity. 5. MCP (Model Context Protocol) might be the most important thing Anthropic has shipped. It’s already the fastest-growing standard in tech history, with Microsoft integrating it into Windows. The vision: everything becomes an MCP endpoint, making the entire digital world scriptable and composable by AI agents. 6. The skills to teach kids in an AI world: curiosity, scientific thinking, and maintaining independent thought. Mike’s daughter perfectly captured it: “You can ask Claude, but I know I’m right.” Don’t outsource all cognition to AI—maintain the ability to think independently and verify claims. 7. When building AI products, work at the edge of model capabilities and be willing to break things. The best companies using Anthropic’s APIs are those that pushed the limits with earlier models, hit walls, and then were ready when new capabilities emerged. Cursor and Lovable both took off when Claude 3.5 came out because they’d been testing the boundaries. 8. For AI startups worried about getting crushed by big companies, focus on three moats: deep domain expertise (like Harvey in legal), differentiated go-to-market with specific customer knowledge, and completely new interaction paradigms that incumbents can’t easily copy. Plus, don’t underestimate the power of true startup urgency. 9. The future of product metrics in AI isn’t engagement—it’s actual value delivered. When Claude helps Mike prototype something in 25 minutes that would have taken six hours, that’s the metric that matters. Traditional engagement metrics can be misleading, when one good conversation could be 2 messages or 200. 10. Mike’s previous startup, Artifact, failed despite being loved because mobile web is broken, news doesn’t spread virally, and remote work made pivots nearly impossible. The biggest lesson: know when to call it. They had “10 units of input for 1 unit of output”—the energy just wasn’t there. Sometimes shutting down is the right call to free everyone up for more impactful work.

What AI agents need to be trusted, robust & relevant and thus adopted widely: 1/ Company-specific context (knowledge, culture, goals), 2/ Accountability (transparency & auditability), 3/ Coordination (w/ agents+humans), 4/ Dedicated tools (to work well). buff.ly/4bnYKAe

"A good idea means a bird’s eye view of the idea maze, understanding all the permutations of the idea & the branching of the decision tree, gaming things out to the end of each scenario. [F]ew can think through all the branches [beyond the maze entrance]." buff.ly/4jUl7kF

Very inspiring! When not taking for granted the expert opinions saying it's impossible, working from first principles without knowledge from the space, and being very lean, make an unexpected disruption happen! Still so many things left to invent & build! buff.ly/4hXxhas

Great post on how what a company needs to move forward changes dramatically before & after product-market fit! buff.ly/4fMRhuQ

Today OpenAI announced o3, its next-gen reasoning model. We've worked with OpenAI to test it on ARC-AGI, and we believe it represents a significant breakthrough in getting AI to adapt to novel tasks. It scores 75.7% on the semi-private eval in low-compute mode (for $20 per task in compute ) and 87.5% in high-compute mode (thousands of $ per task). It's very expensive, but it's not just brute -- these capabilities are new territory and they demand serious scientific attention.

2021 is back, folks!

@2lr 🫡🦾

RISK REWARD AT SERIES A VS SEED From a LP update: “we saw the Series A go from the least attractive, risk adjusted round in 2020-21 to the most attractive in 2023.” For the most part, I think this is right. Investing in a pre-product company at $15m post money valuation versus investing in a $1-2m ARR company at $24m pre / 30m post, means you are paying a 60% premium (24 / 15) in exchanging for massive derisking. Different from when A was done at $42m pre / $50M post, so you were paying 3x premium (42 / 15) for the same derisking.

🚨 IMPORTANT : le très sérieux journal Washington Post, vient de publier un article très fouillé sur l'appui de la Russie au Rassemblement National de Marine Le Pen depuis 2022. Pire : Poutine a mobilisé des armées de trolls pour diviser la société française. Je vous explique 🧵

@dioscuri Interesting thread and discussion in the replies

Quick lesson in the dangers of data contamination. Years ago, I came up with an acronym for remembering the periods of the Paleozoic era — “Catastrophic Overthrow Started Different Colder Period”. I was curious if ChatGPT could guess what it stood for. 1/4

@debarghya_das If you outsource everything including taking care of your kids, is this really “raising your kids” anymore? 😇Irrespective of the country btw

Raising a child in the US seems hellish compared to India. US—Wake up early, cook, drive kid to work, both parents work, do groceries, chores, cleaning. Lonely. India—Maid cleans, driver drops you, groceries in 15mins, cook cooks, parents babysit. Community. (requires $$) 1/4

"It’s become fashionable to say that everything can be empirically answered by data. The ethos, especially in product, is unless you are doing everything as an experiment, then you are doing it wrong. And I disagree with that." buff.ly/47Ua9ox