Florian Brand

@ADarmouni Yes, they are one tool in the tool belt

@xeophon Personal belief is that they’re practical Not a panacea, not a wincon ; didn’t like the hype around personally But definitely thinking they’re always nice to have and pretty good to provide to a model when availabl, useful and well defined

hating mcps is such a midtwit opinion

@HamelHusain Love the post!!

New Blog Post: Do Automated Evals Work? There has been a rise of tools that look through your traces with AI and identifies issues. We tested these tools with real production data to see how good they are. Where they shine - They often spot issues human miss - Integrate into your workflow: viewing traces, creating LLM judges etc. Where they fall short - They miss problems that require domain expertise and taste - Don't have great mechanisms to learn from human feedback - You can get similar results from using your coding agent So you should use them? Yes, BUT do so iteratively with you in the loop. We describe how in the post: parlance-labs.com/blog/posts/aut… It's also a good idea to try using your coding agent with you in the loop, which we discuss in the post. This was written with @doesdatmaksense , who led the research and collated the results.

@eliebakouch @Laz4rz 3T params or bust

linear attention 6:1 ratio, DSA for global attention, 70 layers, 1M context, 1.5T parameters, ~30T training tokens, use latentMoE (or a variant of it), kimi residual attention, no engram, natively multimodal (only text out tho) (i have 0 information)

@GrantSlatton 😍😍😍😍😍 happy birthday!!!!!

he's 10 years old today happy birthday buddy

🚨 BREAKING: Kimi K3 is launching tomorrow with a discount on top-ups, according to a briefly live article on their docs site

@eternal_twil 100+ tools is not a problem of the protocol, people nowadays install 1000 skills and think its better. lazy loading is a solved problem and security issues are as well

@xeophon I mean restrained use yes, but people are often deranged, shoving in 100+ tools, bloating context; not gonna mention security issues

@eternal_twil nah man

@ben_burtenshaw @paradite_ @PrimeIntellect @willccbb You are getting it

@xeophon @paradite_ @PrimeIntellect @willccbb just pivot these attempts into a verifier-implementation-env and add it into the mix.

@ben_burtenshaw @PrimeIntellect you should be able to zero-shot this thing and tailor it to your specific needs / setup in a few hours with claude code?

@ben_burtenshaw @paradite_ @PrimeIntellect @willccbb And: Can you one-shot such a library? Yes, we’ve done so. Thrice. By three people. It worked but it kinda sucked and had some rough edges. So that’s why we took two steps back and made the leash shorter. Result: code is better + more confidence in letting others build on it

@ben_burtenshaw @paradite_ @PrimeIntellect @willccbb Rather work together with you and add the missing pieces 🤗🤝🦋

@GergelyOrosz That’s why we wrote the docs for our library refactor by hand 🫡 removed like 60% of useless stuff, especially since concepts are meant to be read by humans while agents fill in the code gaps

The amount of AI writing on OpenAI's docs makes me sick. Filler sentences for nothing. Mannerisms and phrases humans would not write. Has a human even read this? And all of this will change how other docs are written, and how we all talk - for the worse IMO. 🤮

@willccbb @usr_bin_roygbiv Yeah I should do a write up over this journey TL;DR is: Yes, you can one shot this with good specs but it’ll suck

@usr_bin_roygbiv thinking about how florian vibecoded a version with gpt-5.5 over a couple weeks and then i was like "nah we can do better" and i vibecoded a version over a couple weeks with gpt-5.5 and then mika was like "nah we can do better" and then basically wrote the final v1 by hand

I gotta be honest prime intellect and the new verifiers looks completely legit and would have saved my life two months ago but i'm pretty sure sol could one shot this now. I don't doubt they're cooking something even more insane right now though.

@xlr8harder @willccbb Do the reverse of this

@willccbb @xeophon how are you going to train an ai model to feel existential pain when it reads benchmarks tho

2027 is gonna be so sick

@willccbb Hell yeah

sorry @xeophon we're automating u next year

New addition as of this morning: Verifiers v1 now supports GEPA, allowing you to get very fast and robust performance gains before you do any RL.

@1thousandfaces_ did you block @basedjensen

wait twitter is good again and I’m not seeing shit about katseye anymore?

we’ve tested prl both with professors and also in production runs, so come grab the update

@1thousandfaces_ yearning…

when you are 21 someone will tell you not to move to San Francisco after college. it’s very important that you do not listen to them

@code_star Hell yeah

@xeophon That’s a good point. Looks like twitter beers are back on the menu. You want some?

I used to immediately spend all my X money on drinks with friends (because I don’t make much and that’s fun) But now they offer 6% apy so I have to be financially responsible with my “dumb jokes” money