Borja Odriozola Schick

Context Architect. Building the theory underneath persistent AI systems.

[Borja Odriozola Schick @borcho23]

https://t.co/bra8QlMJpB

[Borja Odriozola Schick @borcho23]

@badlogicgames one quick question, is there a reason why you chose TypeScript for Pi instead of Go for instance?

I was just curious because Im considering as a project trying to have Pi run on Go but wanted to know if you considered it or not.

Thanks in advance!

[Borja Odriozola Schick @borcho23]

Im team TokenEfficiency.

We are going to Tokenmogg Tokenmaxxers

[Borja Odriozola Schick @borcho23]

@ThePrimeagen would you be interested in a mathematical proof that agents cannot supervise another agents without a human in the loop?

https://t.co/iTfgEPQwSx

big fan by the way

[Borja Odriozola Schick @borcho23]

@dair_ai This is very interesting. Would you say it intersects with persistence ?

https://t.co/qOWQ9nQwyF

[Borja Odriozola Schick @borcho23]

@Alibaba_Qwen It would be really nice to have the data for MRCR v2 8needle at 1M context instead of just at 128k because in that regard Opus 4.6 dominates the rest like by 20 points which is crazy to me.

[Borja Odriozola Schick @borcho23]

@BrendanFalk Id rather show the harness I built myself but dont know if that qualifies

[Borja Odriozola Schick @borcho23]

I asked Claude Code and Codex the same questions. Same repo. No coaching. No editing. Both confirmed: → Irrelevant context is never free → Task switching has a cost nobody tracks → Output quality degrades gradually and invisibly → The system gives you a fuel gauge, not a performance gauge
You can try it yourself if you want!! Transcripts and the questions available here:

https://t.co/3p7oYH3PQE

[Borja Odriozola Schick @borcho23]

@mem0ai I rarely see any mention to gate the context window of users to trigger a compaction at high fidelity. If we think of LLM's as compressors, we should optimize the other variables to be able to compress signal when the compressor is at high fidelity.

[Borja Odriozola Schick @borcho23]

@benedictk__ I do.

[Borja Odriozola Schick @borcho23]

So funny to read this when so many people have been doing this for months

[Matthew Berman @MatthewBerman]

Imagine writing an article on X and changing an entire industry instantly. @trq212 has powerful ideas.

[Borja Odriozola Schick @borcho23]

@mem0ai This is called context provenance and trajectory preservation. Attention dows not decay equally across different dimensions. One thing is retrieval, other thing is continuity.

[Borja Odriozola Schick @borcho23]

@badlogicgames I know you must be super busy, but I would love to have your take on this paper about persistence in LLM's within the boundaries of the context window.

https://t.co/qOWQ9nQwyF

[Borja Odriozola Schick @borcho23]

@badlogicgames Extremely annoying. I have to add “read top to bottom every word”

[Borja Odriozola Schick @borcho23]

@badlogicgames Im still in love with Opus 4.6 for long context and general exploration/ planning.

I do think it is the best for that particular purpose.

[Borja Odriozola Schick @borcho23]

@sama Can you please optimize a model to get close to Opus 4.6 on long context? That would be awesome to have a better model optimized for persistence and planning that reduces downstream drift to subagents via handoffs.

[Borja Odriozola Schick @borcho23]

@badlogicgames @mitsuhiko Can we get no details on this? Not even hints?

[Borja Odriozola Schick @borcho23]

@omarsar0 https://t.co/dgDtcfmXY9

[Borja Odriozola Schick @borcho23]

@rohit4verse @Pseudo_Sid26 I beg to differ:

There is a concrete theory and implementation for persistent memory

https://t.co/dgDtcfmXY9