Aleh Shydlouski

[Aleh Shydlouski @_electronick]

@mitsuhiko Experimenting with defining outter loop as a code (written by model) expressing retries and relationships between sub-agents. Think of it as Programmatic agents orchestration before starting the loop. Maybe some ideas will be useful: https://t.co/ya4TsunYpM

[Aleh Shydlouski @_electronick]

@jasonzhou1993 To make outter loop more stable and better react on agents outcomes you can predefine outter loop as a code (with retry policies, and relations between agents). Think of it as programmatic sub-agents orchestration. Maybe some ideas here will be useful: https://t.co/ya4TsunYpM

[Aleh Shydlouski @_electronick]

@djfarrelly Hey @djfarrelly
I'm working on a similar concept, and made a skill that translates goal into code-based orchestrator, creates missing agents and tracks the state in append-only log, maybe will be interesting:

[Aleh Shydlouski @_electronick]

https://t.co/3CmwGrlEm9

[Aleh Shydlouski @_electronick]

https://t.co/3CmwGrlEm9

[Aleh Shydlouski @_electronick]

@samueljmcd Did you consider to have outer loop defined as code plan? Curently working on this approach, maybe can be helpful: https://t.co/ya4TsunYpM

[Aleh Shydlouski @_electronick]

https://t.co/MSnDRVO77H

[Aleh Shydlouski @_electronick]

@RLanceMartin Im currently working on the library that uses Assembly emulation on python for tool orchestration and achieve what "Programmatic tool calling" does in a simplier way. Assembly getting things done to handle control flows in tool calling: https://t.co/5ZRNlCJWwM

[Aleh Shydlouski @_electronick]

@samuelcolvin Assembly quite easy to emulate in cotrolled environment, and its covering conditional stmts and loops to handle complex tool execution logic

[Aleh Shydlouski @_electronick]

@samuelcolvin Im working on the library to emulate Assembly instead of the python to orchestrate tool calls. LLM pretty good in generating Assembly for tools control flows. And curently working on PydanticAI integration for it

https://t.co/5ZRNlCKumk

[Aleh Shydlouski @_electronick]

@mitsuhiko Hi @mitsuhiko Im working on library for LLM tools orchestration with Assembly-like code that you can emulate in a strict environment (e.g. with python). Showing pretty good results for some cases, maybe will be usefull for your case as well https://t.co/5ZRNlCJWwM

[Aleh Shydlouski @_electronick]

🟦🟦🟦
🟨🟨🟨

[Aleh Shydlouski @_electronick]

@PotHix @goncalossilva Discovered Qubes OS a few days ago. Didn't tried it yet, but looks awesome. It runs your application in isolated "VMs" called Qubes, you can run Qube which will be completely destroyed when you close it, very useful when you do a lot of experimentation work.

[Bartosz Ciechanowski @BCiechanowski]

Internal Combustion Engine – https://t.co/NLDoCFmqsd

[Aaron Patterson @tenderlove]

Vaccines wouldn't have side effects if they had written them in Haskell

[Aleh Shydlouski @_electronick]

@jankratochvilcz It's hard to test this. This guy from stackoverflow tested performance for 100 users only. SQL kind of solutions has queries caching, and it will behave like Redis in this case. (Invalidating MySQL cache manually usually not working, BTW).

[Roman Imankulov @rdotpy]

Next Monday at 18:30 GMT I will be talking about #Python for beginners.

- The event is virtual.
- Also, free.
- Also, in Lisbon, but open for everyone.
- Also, with me.

Come on, how can we make it even better? RSVP!

https://t.co/qVMQQpX8CO

[Roman Imankulov @rdotpy]

Your RT counts as a contribution to the #Python community.

https://t.co/Ti0910g2ua

[Aleh Shydlouski @_electronick]

@tenderlove Also paper about GPUnet from 2014:
https://t.co/vYSIydAdYr

[Aleh Shydlouski @_electronick]

@tenderlove Crazy thoughts: 3090 has 10K cuda cores. Image 10K super light threads on GPU, which just collecting data from sockets in parallel. CPU instead of iterating all 10K opened sockets will just request GPU.
There are some steps forward in this direction:
https://t.co/0p26RwLGED

[Amir Salihefendić @amix3k]

We are also live on Product Hunt. Please upvote 🙏 https://t.co/55q0MBKt1i