bernat sampera

Context for AI agents is infrastructure, not docs.

@ObtainerOf U vagueposter

Do I even have to say it?

@Gargotejant Fer fora a un funcionari??? Us heu tornat bojos!! Què serà el següent, reduïr les pensions als nostres avis??

Tots ells? O els que ho facin malament els podrem fer fora per a agafar-ne de millors? Trobo que la funció dels bons salaris han de servir la funció per a atreure talent en camps que n'escassegi.

@Jordifii @randMot7 @3CatInfo Jo en volia 5 més, a nosaltres ens falten peus

@randMot7 @3CatInfo Doncs per mi no era la prioritat. Jo volia 4 mestres més per centre. Ens falten mans.

El preacord entre Educació i sindicats preveu un augment salarial en la part autonòmica de 388,03 € a primària i de 389,50 a secundària. Això, sumat a les pujades estatals previstes, fa que l'augment total sigui d'uns 600 € en 4 anys 3cat.cat/3catinfo/sindi…

@theo So basically they reinvented 4.6 from first principle. We should really stop giving a fuck about new model minor versions.

@XMihura muy bueno este post sobre eso rlancemartin.github.io/2025/07/30/bit…

@XMihura la bitter lesson en su maximissima expressión

una cosa que estoy viendo es que algunos frikis de la IA/data science se están quedando atrás porque creen que los sistemas son menos potentes de lo que ya son que si bases de datos vectoriales, que si embeddings que si... no necesitas nada de eso coge toda tu información, divídela en archivos de texto con nombres informativos y metadatos para que sea fácilmente indexable por grep/ls etc... y deja que el agente se encargue del resto no necesitas nada más

@Pps831 @me_bruno_dev @dhh there's indeed other explanations for it

@me_bruno_dev @dhh Because they are funded by china/russia. No other explanation why would germany destroy its nuclear while getting fully loaded on russian natural gas

And climate hysteria continues to haunt the European psyche. Without confession, without forgiveness, the secular circus of degrowth keeps clowning the old world. Pantomiming hell with apocalypse theatrics.

@nerder_ Looks kinda goofy though

I will probably never understand the rationale of: "It was better when it was worse" Are humans really so resistant to change?

This validates my thesis samperalabs.com/notes/empathy-… x.com/geohotarchive/…

The Eternal Sloptember geohot.github.io//blog/jekyll/u…

this is so fucking cool

@zench4n ok bro, u can tell the AI to stop replying to this

@bsampera97 @bsampera97 Unscoped context windows are essentially wide-open lateral movement paths for prompt injection. Treating LLM context as raw data rather than structured, typed objects turns every RAG retrieval into a potential remote code execution vector for the model

LLM orchestration frameworks like Dify are expanding the attack surface for API-driven vulnerabilities. Recent findings show that even authenticated users can exploit logic flaws to bypass authorization or perform path traversal via improper input sanitization.

crazy stuff, built a claudogram for it x.com/curtis_yarvin/…

@curtis_yarvin Found the conversation very interesting and build this claudogram to understand it better

Owning the libs. Your competitor already super good at this (long read): claude.ai/share/9e3c8ef1… Please hurry up with the continuous learning so I can redpill your whole AI in one giant bolus

@LeonPalaiodimos this is the real one

Το αν αρέσει ή όχι ο ουρανοξύστης στο Ελληνικό είναι θέμα προς συζήτηση. Αυτή η εικόνα όμως είναι ψεύτικη. Είναι προϊόν τεχνικής νοημοσύνης. Οποιος έκανε στοιχειώδη μαθήματα γεωμετρίας στο γυμνάσιο πρέπει να μπορεί να το καταλάβει.

@grouchydiogenes It has potential

Απίστευτο αυτό που έκαναν στην Αθήνα μας.

A trolley is about to hit 5 people laying on the track You can redirect the car, but the other track has not yet reached regulatory approval or completed its 1 year environmental testing period, so operating a train car on it is a violation of transit regulations What do you do?

@sarahwooders Static prompts cannot capture this because each agent needs diverge and keep changing. You end up needing context modules the agent can restructure itself.

Letta Code agents have ownership over their own context (this include system-prompt level memory blocks, skills, and external files) -- and it's pretty interesting how different agents will end up structuring context very differently over time.

@tryadaline Tool selection is a context window problem. When everything lives in one blob prompt, the model loses the signal for which tool matches which subtask.

Your agent called the wrong tool. To resolve the issue, you rewrote the system prompt, reran the test, and watched it call the wrong tool again. The system prompt was never the problem. The model selects its tool before writing a single word, and that decision is based entirely on the tool’s description. Most production tool definitions are one sentence long. One sentence is not enough to avoid selection errors in production. Here are the four failure modes that cause most of these problems: • Ambiguous overlap: Two tools with similar descriptions cause the model to pick inconsistently between them. • Missing constraints: Without a rule for when not to use a tool, any matching request becomes a trigger. • Misleading parameter names: Parameter names carry their own selection signal, separate from the description text. • Unnecessary calls: Agents invoke tools on queries they could answer on their own. A wrong call leads to a wrong answer about 62% of the time. The full breakdown is in the first reply.