Paweł Kotowski

AI dev

[Paweł Kotowski @p_kot1]

@jiangkoumo_

[Paweł Kotowski @p_kot1]

@lukerramsden https://t.co/wlkYw0hWY4

[Paweł Kotowski @p_kot1]

@Justinncase10 @lukerramsden https://t.co/wlkYw0iuNC :)

[Paweł Kotowski @p_kot1]

We built Mos with coding agents at the core: lots of background agents, a ton of integrations (messengers, wearables), and deep domain knowledge from Dr. Oliver Zolman, Bryan Johnson's longevity collaborator. And it works.

We developed and validated it with a custom user simulator plugged straight into our AI harness. The whole thing is tuned for one goal: making it as easy as possible to get healthier.

Stack, for the curious: RLM, Pydantic Monty, GPT-Live, Astra, Opus, Gemini (yes, that too).

Want to see what happens when you throw the best of today's AI at helping you feel better? → https://t.co/eqocMmznt8

[Pati Brzozowska @heypatihere]

We’ve never had more health data.

And somehow, we’ve never been more confused about what to do with it.

So today, we’re launching @moshealthco.

Mos comes to you, not the other way around.

Try it for free - link in the comments!

[Pati Brzozowska @heypatihere]

We’ve never had more health data.

And somehow, we’ve never been more confused about what to do with it.

So today, we’re launching @moshealthco.

Mos comes to you, not the other way around.

Try it for free - link in the comments!

[Pawel Sobkowiak @popotam]

12 years building Booksy. Now launching Mos Health, your AI health partner.
AI powers the product and helps our small team build it without slop code.
The trick? Less yolo and more rigor and harness engineering.
Try it free -> https://t.co/Ehm74p0r0i
Tell me what you think ;)

[Paweł Kotowski @p_kot1]

@tobi

[Paweł Kotowski @p_kot1]

@steipete @zeeg :D

[Paweł Kotowski @p_kot1]

@michuk Jeszcze może być dobra:)

[Paweł Kotowski @p_kot1]

@theo You should not scroll down along with the stream to allow reading it right away. Otherwise streaming indeed does not make sense at all.

[Paweł Kotowski @p_kot1]

@hwchase17 Wait so until now it just lost data on compaction? And we need bringing RLM into discussion for such obvious things?:D

[Paweł Kotowski @p_kot1]

@RhysSullivan I find it odd that people are surprised or even don't believe it.

[Paweł Kotowski @p_kot1]

@KBezbailis Yep e.g. https://t.co/S6LcYh7rfD

[Paweł Kotowski @p_kot1]

@dzhng I am building dynamic graphs / timelines based on traces from simulated users to understand my ai backend e2e.

It contains also review layers, e.g. if models use tools given to them, prompts consistency, etc.

Can call coding clis from there (less exploration)

[Peter Steinberger 🦞 @steipete]

code mode is now in the modern harnesses, so none of that really matters anymore.

cli, mcp, tools, it’s all just javascript the agent writes.

[Paweł Kotowski @p_kot1]

@Shashikant86 @ClementDelangue @pydantic @samuelcolvin If only it supported pandas/numpy..;)

[Paweł Kotowski @p_kot1]

@banteg At https://t.co/owrNG34RQy we are using this approach in background agents that constantly figure out what's best for your health

[Paweł Kotowski @p_kot1]

@jmbollenbacher @_can1357 For my use case it started working well with Sonnet 3.6, Opus 3 would be fine though I think if it was cheaper and faster.

[Paweł Kotowski @p_kot1]

@_can1357 @jmbollenbacher I've had stateful codegen in my harness since feb 2024 with subagents since sep 2024, thought everyone would do it in a month, 2 years later people argue who came up with it, may as well join the crowd:D