Bart

Cofounder at HoppeSyler, a reagent company run on AI agents. Writing about what actually changes when models face real stakes

[Bart @bkolendowski]

https://t.co/dujKoAA36g

[Bart @bkolendowski]

@dan_sci_phil @justharshal @simonw It's incredible what a small team can now do. There are parts that are still not solved and certain things we can't outsource to the Agents yet, but for the most part all the easier things have been outsourced.

[Bart @bkolendowski]

@__spekulator__ @justharshal @simonw Local filesystem.

[Bart @bkolendowski]

@lefthanddraft The models outputs (including its character) are dictated by upstream tokens just as much, if not more, than its training/RLHF. Anthropomorphizing, instead of thinking from first-principles, will cause you all sorts of cognitive dissonance.

[Bart @bkolendowski]

@Heberth26 @brunolemos If it’s not a model problem then how come we’re seeing this types of issues accumulating for Sol and not for Fable?

[Bart @bkolendowski]

@justharshal @simonw Sorry - let me clarify - the 'harness' is just Claude Code. The interface is my own. This is just Claude Code running in the back on my own repo and I'm interacting with it (and the repo) through my own interface (that I built using Claude Code).

[Bart @bkolendowski]

@justharshal @simonw My own. Build it from scratch. But it does everything I need and when I need something new I just build it in using Claude Code itself.

[Bart @bkolendowski]

@BlindBearMedia Not a single limp was seen.

[Bart @bkolendowski]

@Breaking911

[Bart @bkolendowski]

@Aaali172 My company sells chemicals for science. I do agree that software based companies are going to be in trouble.

[Bart @bkolendowski]

@PeterPa00865077 @simonw Most knowledge work is going to be replaced but we will still need labor for a while.

[Bart @bkolendowski]

Wrote down why: https://t.co/LKO2d5RvL1

[Bart @bkolendowski]

Since January my agents and I have made 3,884 commits to the repo that runs my company. For a while I read the solid-green heatmap as proof things were going well. I no longer think it measures that.

[Bart @bkolendowski]

@DanielMiessler I don’t know about you but a deep conversation with Fable 5 is more enlightening than a conversation with 99% of the people I know. Maybe that’s more of an indictment of the people I know, but regardless, that is the case.

[Bart @bkolendowski]

@DanielMiessler Meanwhile the managers boss is like, "Hey Claude... Can you tell me what Ben is actually doing?"

"Hi Boss... Looks like Ben is just running the following prompts..."

[Bart @bkolendowski]

You're confusing catching up to the past with catching up to the frontier. Yes, a hobbyist in 2030 will be able to train a 2024-era model. But by 2030, frontier labs will be training models on $100 billion supercomputers. The baseline rises, but the gap between the hobbyist and the frontier actually widens.

[Bart @bkolendowski]

@Abelaer When you scroll tiktok for an hour the algorithm already knows what things maximally stimulate your dopamine.

What's the material difference with this approach?

[Bart @bkolendowski]

@nickhollon10 @NVIDIAAI Do you do any testing on 'business' tasks (working on a complex knowledge base like a wiki and being able to carry out complex tasks)?

[Bart @bkolendowski]

@levie Fable 5 has been a game changer for my business (selling reagents to lifescience labs and bioinformatics).
It's now able to automate complex real-world tasks like Opus 4.8 was struggling with.