quinten

building @provnai @provncloud verifiable infra for the agentic era 🇧🇪

[quinten @quint3ns]

@matei_zaharia Strong signal that the agent stack is moving above individual harnesses.

The interesting next question is what kind of control belongs at the meta-harness layer versus what needs to be independently verifiable at the execution boundary.

[quinten @quint3ns]

This is why I started building ProvnAI.

The next security boundary for AI is not only the prompt.

It is the moment an agent is about to do something real:

touch a file,
run a command,
call an API,
query a database,
trigger internal automation.

That boundary needs enforcement.

[provnai @provnai]

MCP security is not just prompt safety.

It is an execution-boundary problem.

The dangerous moment is often not what the model says.

It is what the agent is about to do.

[quinten @quint3ns]

Proof without possession.

[quinten @quint3ns]

@Railway Juicy

[quinten @quint3ns]

Built the 'ULTIMATE PERFORMANCE LAUNCHER' for @antigravity
→ 64GB RAM
→ RTX 3060
→ Max V8 heap
→ Every GPU flag known to man
Result: Extension hosts still dying and the profiler having a panic attack.
10/10 would over-engineer again.

[quinten @quint3ns]

@sama Democratization: $200/mo wall.
Empowerment: lawyers gaslighting “Open.”
Universal Prosperity: your portfolio.
Resilience: ghosting safety letters.
Adaptability: adapting “non-profit” in real-time until it bends like a fun-house mirror.

[quinten @quint3ns]

@theonejvo That em-dash is a load-bearing structural hyphen for the entire federal government.

[quinten @quint3ns]

@heychaarah @puasdfjasdf @marktenenholtz Measuring 'quality' by chat vibes is a toy-tier error. I’m using a Gemini harness to hunt P1 server crashes on malformed tokens in live infra. If you think orchestration is just a prompt, you're looking at the paint while I'm building the engine.

[quinten @quint3ns]

@heychaarah @puasdfjasdf @marktenenholtz Not really.

You’re comparing a language/runtime question to a capability-vs-harness question.

The point is that UX here is heavily shaped by orchestration quality, not just the underlying model. Those are different variables.

[quinten @quint3ns]

@puasdfjasdf @heychaarah @marktenenholtz If you need someone to hand-feed you a 'reproducible task' to realize the value, you're missing the point. SOTA isn’t a magic prompt; it’s what you build when you stop treating AI like a magic 8-ball and start building actual state management. 🤡

[quinten @quint3ns]

@puasdfjasdf @heychaarah @marktenenholtz Tell me you are lazy without telling me you are lazy. example problem/prompt that showcases "giga sota" capabilities? 🤡

[quinten @quint3ns]

@heychaarah @puasdfjasdf @marktenenholtz Different variable.
Harness quality != model capability.

Harness is state/tooling/validation/retries/decomposition.
Capability is what the model can reason through once the setup isn't sabotaging it.

Bad harnesses hide strong models. Good harnesses can prop up weaker ones.

[quinten @quint3ns]

@steipete Leave us some codex please 😅

[quinten @quint3ns]

@puasdfjasdf @marktenenholtz @heychaarah SOTA isn't a prompt. If you don't know how to orchestrate state, validation, and tool calls, every model is going to look bad to you.

[quinten @quint3ns]

@puasdfjasdf @marktenenholtz @heychaarah This only means you haven't figured out how to govern the system properly.

[quinten @quint3ns]

@icanvardar Yeah, makes sense… Google should totally quit AI — right after inventing TPUs, TensorFlow, and half the research the industry runs on

[quinten @quint3ns]

@iruletheworldmo Govern the agent.

[quinten @quint3ns]

You want a perfectly precise system using an imprecise AI

[quinten @quint3ns]

@WillGeek4Food @provnai You can poison what the agent sees, but that still shouldn’t give you a free path to sensitive execution. The real lesson is that agent safety can’t live only in prompting or only in policy, it needs a hard boundary at runtime.