have been extremely impressed with Gemini 2.5 Flash for writing of late. increasingly my go to.

RT @vitrupo: Google DeepMind's Logan Kilpatrick says AGI will be a product experience. Not a model. His bet: whoever nails memory + contex….

Keynote Summaries (6/3 and 6/4): . Useful General Intelligence (Danielle Perszyk, Cognitive Scientist, PhD, Amazon) . Designing AI-Intensive Applications (@swyx) . Spark to System: Building the Open Agentic Web (Asha Sharma, Microsoft) . State of Startups and AI 2025

@xanderatallah @sgrove h/t to @matijagrcic for the notes format.

Keynote Summaries (6/5, continued): . fun stories from building OpenRouter and where all this is going (@xanderatallah). The New Code (@sgrove)

Keynote Summaries (6/5, continued): . The infrastructure for the singularity (@jessemhan). State of AI Engineering 2025 (@barrnanas). Trends Across the AI Frontier (@_micah_h and @grmcameron). Evals Closing Keynote (@ankrgyl)

Keynote Summaries (6/5): . A year of Gemini progress + what comes next (@OfficialLoganK). Thinking Deeper in Gemini (@jack_w_rae). Why should anyone care about Evals? (Manu Goyal, Braintrust). Containing Agent Chaos (@solomonstre)

Keynote Summaries (6/4, continued): . 2025 in LLMs so far (@simonw). Windsurf everywhere, doing everything, all at once (@kevinhou22). Define AI Engineer (@gdb, interviewed by @swyx)

"Code is 10-20% of your value. The other 80-90% is structured communication." (@sgrove). In the AI era, specs become source code. Implementation becomes compilation. The most valuable programmers will be the best communicators.

"The prompt is a bug, not a feature." (@saranormous). We're building UIs that force users to become prompt engineers. The best AI products feel like mind-reading. They collect context automatically. They present solutions thoughtfully.

In AI, B2C is winning today. B2B will dominate tomorrow. Consumer apps can ship fast and break things. Enterprise needs security, reliability, compliance. 2026 is when the enterprise wave hits.

Best practices are emerging, not established. Prompting techniques. Eval frameworks. Design patterns. Multi-agent architectures. We're writing the playbook as we play the game.

Plot twist: Most practitioners are new to LLMs. (h/t @barrnanas). The "AI engineer" role is still being invented in real-time by people figuring it out. There are no experts. Only early adopters.

Developers want configurability, not vendor lock-in. (h/t @xanderatallah). They've seen this movie before. Cloud providers became overlords. API companies became gatekeepers. This time they're choosing flexibility.

AI hallucinations aren't bugs. They're features. "Controlled hallucinations are necessary for flexibility, similar to human perception." (Danielle Perszyk, Cognitive Scientist, Amazon). The challenge isn't eliminating them. It's controlling them.

Multi-agent systems are where the real breakthroughs happen. Solving branching and multi-threading isn't just technical: it's the key to AI that actually works. Single agents hit walls. Swarms break through them.

Voice is the next iPhone moment. We're still thinking in text, but the world is moving to conversation. The interface is disappearing.

Forget large teams and specialized roles. The new winning formula: Small, full-stack generalist squads. 5 people who can ship end-to-end vs 50 people who need handoffs. Coordination cost is the killer. "Fewer people means more context per head." (@EricSimons @boltdotnew).

"The future isn't one giant model. It's a 'menagerie' of specialized models." @gdb. Distillation works. Specialized can beat general. Small + specific > large + generic. The monolith era is ending.

The most successful teams obsess over evals. Not features. Evals. "If you invest in good evals, you're building a laboratory that lets you do 90% of product iteration before going to prod." (Manu Goyal, Braintrust). Discipline beats talent when talent lacks discipline. But.