Carles Navarro

Chemisty & Physics | Machine Learning at @acellera

[Carles Navarro @11krls]

There is only one basic human right, the right to do as you damn well please. And with it comes the only basic human duty, the duty to take the consequences.

[Ariel @redtachyon]

When will people realize that GRPO is still basically reinforce?

[Carles Navarro @11krls]

writing is the most important skill of the next 10 years

[Carles Navarro @11krls]

See a lot of hype on the new app/feature/product from a big AI company -> try it because of FOMO -> lose my time -> go back to VSCode + cli agents + Obsidian

[Carles Navarro @11krls]

This is the way, thought about this last year when training reasoning models.

And this implementation is pretty elegant 👌

Latent reasoning is the future

[Chris Hayduk @ChrisHayduk]

Paper link: https://t.co/AtbxIYDVip

[Carles Navarro @11krls]

@hwchase17 I think sub* agents will kinda go away, the original blocking/recursive ones. We will go towards async ones, which is basicall multi* agents

[Carles Navarro @11krls]

guys all important stuff (for us) of claude code was mostly public, system prompts and hooks were all leaked. All the rest is just an envelop to put an agent into production and make it work on our computers / permissions / observability.

[Carles Navarro @11krls]

@barckcode todas estas cosas eran ya públicas https://t.co/uybmFxuQII

[Carles Navarro @11krls]

vibe-leaked!

[Chaofan Shou @Fried_rice]

Claude code source code has been leaked via a map file in their npm registry!

Code: https://t.co/jBiMoOzt8G

[Jo Kristian Bergum @jobergum]

Exactly this and why we are building https://t.co/UtnhzBwqvK

[Carles Navarro @11krls]

@teortaxesTex We just don't think about it

[Carles Navarro @11krls]

Irónicamente, la universidad sí se ha adaptado a la ley de la oferta y la demanda: a la infinita demanda de madres, padres y chavales a quienes se les ha hecho creer que la única salida digna es estudiar una carrera.

La mayoría de grados no deberían existir o deberían ser especializaciones.

[Carles Navarro @11krls]

There’s one thing people who say ‘MCP or CLI’ and people who say ‘Transformers or diffusion’ have in common: they have no idea.

[Carles Navarro @11krls]

@davidgomes They love backwards compatibility too

[Matt Pocock @mattpocockuk]

Good tip for avoiding cognitive debt in codebases where AI has run wild:

Design the interface, delegate the implementation

[rahul @rahulgs]

seems obvious but:

things that are changing rapidly:
1. context windows
2. intelligence / ability to reason within context
3. performance on any given benchmark
4. cost per token

things that are not changing much:
1. humans
2. human behavior, preferences, affinities
3. tools, integrations, infrastructure
4. single core cpu performance

therefore,

ngmi:
1. "i found this method to cut 15% context"
2. "our method improves retrieval performance 10% by using hybrid search"
3. "our finetuned model is cheaper than opus at this benchmark"
4. "our harness does this better because we invented this multi agent system"
5. "we're building a memory system"
6. "context graphs"
7. "we trained an in house specialized rl model to improve task performance in X benchmark at Y% cost reduction"

wagmi:
1. product/ui
3. customer acquisition
4. integrations
5. fast linting, ci, skills, feedback for agents
6. background agent infra to parallelize more work
7. speed up your agent verification loops
8. training your users, connecting to their systems and working with their data, meeting them where they are

[Carles Navarro @11krls]

Memory should be only per chat. So I can have a long chat without losing context, if I start a new chat is because I want fresh context. When context engineering gets good enough I may never leave the conversation, but for now it just messes up both experiences (not enough good for infinite conversation nor for global context)

[Carles Navarro @11krls]

Ehh that's really good. Been saying this for a while

[Anthony Kroeger @kr0der]

just found out Claude Code has a new (unreleased?) feature called "Auto-dream" under /memory

according to reddit, this basically runs a subagent periodically to consolidate Claude's memory files for better long-term storage

this is pretty crazy because that's basically how humans store long-term memories if you think about it - by sleeping

[Kpaxs @Kpaxs]

This is a man who has been haunted since childhood and built a billion dollar company as a side effect of trying to make the haunting stop.

[Carles Navarro @11krls]

With agents getting better and working for longer, feels no longer productive to wait several minutes for each task to finish. I resigned, I don't want to know what code they produce anymore, I fully embrace multi-agentic coding.

This codebase is modular enough so I can coordinate with a simple agent_inbox/ files communication protocol.

[Carles Navarro @11krls]

Cursor's job is not to compete with model providers IMO, creating an agentOS was (may still be) a better strategy. Easy to say it a posteriori though