Parade. Thursday. Manhattan.

Internal AI adoption at mid-stage cos. Have seen this play out dozens of times. > phase 1: grant broad access to tools, data, and tokens. big spread between the p50 and p99 employee emerges. > phase 2: juice adoption through incentives, leaderboards, show-n-tells, and a culture of tokenmaxxing > phase 3: finance panics after company spends the entire annual forecasted token spend in ~1 month > phase 4: intelligence routing, including going to older versions of closed models as well as going open source. realize that you don't need a ferrari to drive to costco when you probably want a minivan for the task. reduce cost without sacrificing quality and latency.

OG agent went hard in the paint last night

The absolute wildest betting line swings I have ever seen in live sports.

an AI reading a 64k-token document (roughly a short novel) is holding 9 GB of memory just to track what it's read. that number grows linearly. at 128k tokens, 18 GB. that cost is why AI agents struggle with long-running tasks. the memory just gets too expensive to hold. the Baseten research team published a fix for this. instead of the current approach (deciding which tokens to delete as context grows) their system compresses the entire memory into a synthesized, smaller version. 9 GB becomes 180 MB at the same context length. the accuracy hit is small enough to be practical. on long-context RULER benchmarks, it beats the strongest existing method by 8-22 accuracy points in 16 of 18 tested conditions. the near-term implication is straightforward: longer context at lower infrastructure cost. the less obvious one is what this unlocks for agent products -- anything that requires holding context across a long work session runs into this memory wall today. this is what clearing it looks like at the infrastructure level. excited to see what gets built with it. credit: Charles, Alex, Harry, Mudith, Max, Baseten

That millions of people tune in to the updates of a random German tourist here for the World Cup is a testament to how refreshing it is to see your own country through fresh eyes. Very easy to take it for granted day to day.

AI actors are getting scary good.. spent 2 day making this short film.. if you still think actors are safe, ihave nothing to say.. this is so over check my prompts and workflow on buzzy now:

I redesigned Trump's bar chart according to "The Visual Display of Quantitative Information" by Edward Tufte, with the only egregious violation being the colors of the bars, which I kept and simplified.

NYC tech week in the books. Chill Baseten x Verci cafe vibes all day. Perfect weather. Knicks Game 2 tonight. Life is good

outside of being a brilliantly entertaining piece of media, it's a fascinating look into the psyche of some very influential people. must watch. also, bryan johnson going off his blueprint diet and snacking is sending me.

Transportation in NYC includes trains, buses, ferries, pedestrian space, and bikes. In San Antonio, transportation includes a left lane, middle lane, and right lane.   Let's Go Knicks!

big cos investing in AI give their employees codex/claude and call it a day. no real investment in fundamentally changing workflows or restructuring product/gtm. huge miss.

“If the game has gotten harder, the question is not how to work more hours. The question is which moments, chapters, and books are worth writing at all. Every choice to work on one thing is a choice not to work on something else. The opportunity cost of our attention has never been higher.” Not just for work but any expenditure of time for any reason

2 million businesses registered in delaware. all paying franchise tax through this website today. myself included. you get what you get when the state has a monopoly on incorporation. (same developer as the india visa site, clearly) btw, squatting on companies is the next level up from squatting on domains. same output (no actual company) but a different level of commitment to the fantasy.

On quality: My first day at @baseten, the very first person I met was @philipkiely on the AI Education team. He told me he's working on a textbook. An absurd thing to say on the surface. I asked for an early pdf version that I started to read on the plane ride home. It was reviewed for accuracy, comprehensiveness, and clarity by dozens of people across multiple disciplines and hand-edited. The book became a useful reference for me and deeply accessible to someone new to the space and something I still return to often. A few months later, the book was complete. Printed by an old school shop in Europe, the cover, charts, and diagrams meticulously crafted pixel-by-pixel with @lucas_dehaas's hands. Sourcing the printer and getting the proofs just right added many weeks to the launch beyond the original deadline. Then, it was time for the launch. The team wanted to get the books in the hands of ML researchers and the broader tech community. Instead of shipping it via USPS, a gift box was custom printed and hand-delivered. @ad0rnai, who was mere days into her first week at Baseten, sourced a custom metal bookmark, tea from Fortnum & Mason, and created an art deco experience that culminated in an event at The American Bookbinders Museum (planned end-to-end down to the last custom gobo detail for lighting the sidewalks by Ashley). None of this needed to be done. But it was an ambitious idea executed with an extremely high quality bar under a reasonable timeframe and budget. Execution wins across the board. And a no-brainer for leadership to sign off on. People assume the market prizes speed and therefore marketers have to make a tradeoff between speed and quality. It often degenerates to slop to keep up with the Joneses. I believe this is a false dichotomy. I think the market actually prizes value received and human effort. And it's worth taking the big swings. You don't have to make a tradeoff, you just have to be intentional with your choices. But it's important to keep the quality bar extremely high, both when the world is watching your effort and, more importantly, when nobody is watching what you're doing. By the way, none of the above-mentioned people had ever done a project like this. They just kept standards high and figured it out together, thinking how to deliver maximal value to the reader. And the results spoke for themselves. 📷: @katedeyneka

Accessible data is indeed the limiter on pre-training AI models yet at the same time there exists troves of it both online and offline that just needs to be captured. Clever. Though, as someone who worked at a cleaning company for three years (Homejoy) and has personally cleaned dozens of homes, this is gonna be a gnarly endeavor with a lot of fun stories.

love him or hate him, this is exactly the type of stuff low-code/no-code/agentic software enables. is it complicated to make? no. and that's the point. shipping tiny, discretely-useful product where there's a real user demand for it. meeting users where they are, quickly. great to see local gov adopt.

People ask me all the time how to pick a winner in the AI application layer. This is how you pick a winner. Seek the companies post-training their own models.

Been craving cereal lately