Petr Baudis

> be @RossumAi > take all the AI advances we are hyped about here > gloriously plug them in a ✨B2B SaaS✨ > run AI agents on many Mpages/week > automate a common business process (transactional paperwork) for 100s enterprises > fix a menial clerical job that people hated to do

In NYC next week. Anything I shouldn't miss?

i think this meme is hilarious. my take on all this: the point of introspection is to end up thinking less, not more, to be more in the flow, more productive, to dissolve into being itself. if your introspection is making you think more i recommend getting another one

@scaling01 that graph screams "benchmaxxed" though :(

that looks pretty fucking good

This is really cool. It got me thinking more deeply about personalized RL: what’s the real point of personalizing a model in a world where base models can become obsolete so quickly? The reality in AI is that new models ship every few weeks, each better than the last. And the pace is only accelerating, as we see on the Hugging Face Hub. We are not far away from better base models dropping daily. There’s a research gap in RL here that almost no one is working on. Most LLM personalization research assumes a fixed base model, but very few ask what happens to that personalization when you swap the base model. Think about going from Llama 3 to Llama 4. All the tuned preferences, reward signals, and LoRAs are suddenly tied to yesterday’s model. As a user or a team, you don’t want to reteach every new model your preferences. But you also don’t want to be stuck on an older one just because it knows you. We could call this "RL model transferability": how can an RL trace, a reward signal, or a preference representation trained on model N be distilled, stored, and automatically reapplied to model N+1 without too much user involvement? We solved that in SFT where a training dataset can be stored and reused to train a future model. We also tackled a version of that in RLHF phases somehow but it remain unclear more generally when using RL deployed in the real world. There are some related threads (RLTR for transferable reasoning traces, P-RLHF and PREMIUM for model-agnostic user representations, HCP for portable preference protocols) but the full loop seems under-studied to me. Some of these questions are about off-policy but other are about capabilities versus personalization: which of the old customizations/fixes does the new model already handle out of the box, and which ones are actually user/team-specific to ever be solved by default? That you would store in a skill for now but that RL allow to extend beyond the written guidance level. I have surely missed some work so please post any good work you’ve seen on this topic in the comments.

@VictorTaelin It is very easy to add web search too! You might enjoy github.com/pasky/pi-ampli… as a nice starter kit of extra primitives (or, as @badlogicgames says, as a source for pi to steal the stuff you like from). (Mix it with github.com/pasky/pi-side-… as a "subagents" solution, but ymmv.)

To elaborate - I never liked the "Linux way": > ship a half assed software, and let the user modify it Instead, I always bought the Apple way: > pay us and we'll give you the best possible defaults This worked for me, because I wanted to spend my time writing compilers, NOT fixing driver issues. So, when people told me "Pi comes with just the bare essentials and you can add what you want", that definitely did NOT paint a good picture. But it is different. The time to modify is minimal. "Pi, extend yourself so that I can spawn sub agents in a specific way that works Bend2's prelude" One or two prompts later, and it is done. It modifies itself for what I need and I suddenly have a new tool to help me get things done. It just works. You can't do that with Claude Code and Codex. That said, I'm still not sure that'll always work. How the hell do I make my Pi browse the web, for example? Seems like the author doesn't want it, it is definitely important, and there's no easy / satisfactory way to add 🤔

Ok so I thought that was a dumb gimmick but now I'm completely sold on how pi is a self-modifiable software. It literally knows how to modify itself very cleanly and that's extremely useful in practice I'm not using Codex / Claude Code anymore Bend2 should definitely be like this! I mean, constructed in a way that AI's can easily navigate it and know how to modify it to add any feature the user wants. Perhaps we're past the era of open source software and into the era of forkable software, where the most hackable project wins?

thanks, google <3

@scaling01 benchmark mimo-v2 models instead :)

@xpasky gonna run lisanbench for GPT-5.4, Mistral Small 4 and M2.7 this weekend

Looks like an absolute banger on paper, but zero hype on my timeline. (I know, @scaling01 currently consumed by Dune 3, but even so.) Is it a good model, or is it benchmaxxed?

the -omni model (healer-alpha) is insane, as surprisingly good as disappointing m2.7 was at the first impression it's a good model, sir - i think we have a new default for @MuaddibLLM serious mode

Stealth Model Reveal: Hunter and Healer Alpha are @XiaomiMiMo MiMo-V2-Pro and MiMo-V2-Omni Both models are live now on OpenRouter, and free to use in @OpenClaw via the OpenRouter provider for the next week!

Ooh, you can go help save Gemini by fixing their sad state of posttrain!

@_fallpeak @scaling01 It would be to possibly replace GLM-5 for me if the completion API wasn't so janky

@xpasky @scaling01 I like it, feels a lot like M2.5 but with a bit less of that autistic tendency to take you extremely literally and get to work without asking for clarification. Definitely not frontier-level smart, and it degrades above ~80k context, but it's fast and takes direction well.

@lisa44Yes @francoisfleuret Yeah, I'd count 4090 as "on the beefy side" :) Beautiful tho

@xpasky @francoisfleuret RTX 4090 gets me Qwen3.5-27B Q4 @ ~40 t/s Qwen3.5-35B-A3B Q4 @ ~100 t/s

1. What are the best open source coding / general purpose models? 2. What hardware to run them comfortably? 3. How do they compare to the flagships?

@scaling01 doesn't seem to be a good model, at least unusable in its current form x.com/xpasky/status/…

@scaling01 what the heck is this response, it's the first time i see randomly interleaved thinking/text blocks (and i called *lots* of models)

@MiniMax_AI @OpenRouter It's been a long time since I had a (near-)frontier model stop on me like that w/o calling any tools when it said it's going to do something. It's hard to believe it's posttrained so poorly - looks more like the deployment got botched in some way?

@MiniMax_AI Hey guys, looks like a nice release, but this kind of interleaving thinking/text blocks in the middle of a sentence on @OpenRouter seems pretty unreasonable (and is breaking API clients - what even *is* the correct behavior, should text blocks be pasted together w/o separators?)

Introducing MiniMax-M2.7, our first model which deeply participated in its own evolution, with an 88% win-rate vs M2.5 - Production-Ready SWE: With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%), M2.7 reduced intervention-to-recovery time for online incidents to 3-min on certain occasions. - Advanced Agentic Abilities: Trained for Agent Teams and tool search tool, with 97% skill adherence across 40+ complex skills. M2.7 is on par with Sonnet 4.6 in OpenClaw. - Professional Workspace: SOTA in professional knowledge, supports multi-turn, high-fidelity Office file editing. MiniMax Agent: agent.minimax.io API: platform.minimax.io Token Plan: platform.minimax.io/subscribe/toke…

sure, web search using ollama, your local model execution tool * sure, the web search is all wrapping ollama cloud API, oh, we didn't mention that? well, isn't it kinda obvious for an ollama skill mentioned right next to "local models", silly!