sharing the best AI events in SF this week👇 🗓️May 12- 17

the number one thing that allows people to move fast with ai code is knowing that, at any given code change, random other shit isn't broken. whos building this?

I think I only tweet when I'm happy. Which is the wrong kind of tweeter

Performance art for the algorithm age.. oh god

"Before I generate the code, I need to make an assumption about how your data_access.messages layer works. I'll assume there's a method data_access.messages.create() that accepts a Pydantic model. I will proceed to verify this assumption" well done gemini. well done

Use the tool don't let the tool use you

guiding models CoT well with prompts makes a big difference. sometimes i wish i could just "ambiently" inject thoughts into its ongoing CoT process and it takes them in stride

An underrated aspect of llms is that more people than ever before are explicitly writing out what they want.

Some tech that makes me happy to be in tech - airplane landing/takeoff systems - llms - waymo - bicycles (yes)

An underrated UX of thinking models is that I actually like it when models spit out a ton of thinking tokens makes me feel like they are doing something useful

as code becomes easier to built its also easier to delete. I find myself doing a ton of rewrites just because its fast (and fun)

very simple q: gemini claims to support "realtime" video/audio streaming in the sense that it supports a continious video stream and can produce outputs without needing an explicit speaker turn. How does this work?? does the model just figure out when it needs to speak based on

this is really really cool

ai is natively well suited for clinical diagnoses

gonna keep tweeting about this until me or someone else does this but someone please build a voyager style rl agent that is good at building test suites. USE COMPOSABLE tests so that any new e2e test is easy to add and we can actually test every piece of code written by an ai