Hold on, isn't that just standard practice for post-training LLMs for agentic use? Give task, generate n rollouts, grade rollouts (either at termination or after each tool call)? Or is the difference that the rollouts are generated ahead of time and then graded? (Of course, then it's not really on-policy.)
We started off naively as just another intelligent document processing SaaS. Years of experience have shown us where the real pain points and value lies. Join my AI team and be horrified by how many critical business processes rely on email. Then help me execute on our ambitious roadmap!
If you find yourself nodding along with the following statements, you may be a good fit:
- You enjoy getting your work into the hands of users
- You’re okay debugging a stray tensor in the morning and designing services in the afternoon
- You care about the field of AI and interact with its community
- You can ‘sense’ when a loss curve is doomed
- You’ve fine-tuned LLMs, even just for fun
- You’ve been thinking a lot about agents lately, for some reason
- You don’t mind the odd bit of customer support (it’s where we get our best insights!)
All credentials welcome, including recent graduates. We’ll shape the role around you. What matters is creativity and visible engagement with the field.
Who may apply: EU citizens or holders of a Belgian work permit. (Everything else is more admin than we have time for.)
Hah, I use the same probe when I’m unsure which provider openrouter is routing me to!
Fwiw, deepseek v4 will happily discuss it. Only chinese providers will stop it in its tracks and give a canned answer. Streamed responses sometimes start with what the model was actually generating before it got cut off.
It’s top bad, really. Sometimes the Chinese providers are the model labs themselves, like deepseek. I’d like my money to go directly to deepseek, since they did all the work. But data protection concerns aside, how do I trust a system that denies objective reality? (Kind of like how Grok will tell me that wikipedia is ‘woke’.)
The serving endpoint can censor. At TrustedRouter we ran the same GLM-4.7 weights on both hosts: Cerebras answered all 60 FreedomBench questions; one Z.ai endpoint went blank on 27.
That’s what a dictator seeking power through lies wants you to think — that the outcome of the 2020 election and Russia’s invasion of Ukraine are matters of opinion.
Try this simple test: Visit a mental hospital. Speak to some people with delusions. Usually,if you let people at the nearest bus pickup spot bum cigarettes they'll talk to you all day.
Keep talking to them until you meet two or more who think they are the ONE true god, or maybe God's only son. Or Napoléon.
If you can't bring yourself to accept that the both are who they believe they are, then there is an objective reality and you knew it all along.
Kind of a coincidence to see this post here, just after I’ve gotten done playing with Kakoune and poking around the docs.
After using Helix fulltime for two years, I’m considering a switch. Helix is fantastic OOTB and I have considerable muscle memory in it, but the pace of development is glacial. There are I think three active maintainers, and they’re completely swamped by PRs, many of which have sat unreviewed since they were openend. It’s enough to wonder why anyone would even consider contributing. That would be understandable if it weren’t for the fact that, no, Helix isn’t ‘finished’. Newer LSP features aren’t supported; DAP for python has never worked; I also miss code folding, bread crumbs, and tab name disambiguation. The steel scheme plugin branch is functional and several neat plugins are beginning to emerge. However, I switched back to main after noticing that the plugin branch adds about 1s to startup time. (I’m constantly opening and closing helix.)
Anyway, Kakoune has an attractive philosophy. There’s ‘editor for life’ potential. A really minimal core that I can shape to my needs in the moment. The scripting language looks even gnarlier than vimscript but no matter – we have coding agents for that now.
But Kakoune’s minimalism also brings downsides that make me wonder if it’s more effort than it’s worth. For example, that minimalism rules out a dependency on treesitter. But you can hardly expect to shell out for treesitter highlighting and expect good performance. With Helix, nvim, and Zed all consolidating on treesitter, I don’t see why I’d want to spend time on an already solved problem.
All in all, neovim is feeling like a better option, and as another commenter says, configuring it has become remarkably easy thanks to coding agents. And apparently native multiple cursors are on the roadmap.
https://sr.ht/~hadronized/kak-tree-sitter/ exists. I have not used it personally but the author has been pretty active and we had multiple discussions around performance, AFAIU it does not rely that much on the shell as most communication with Kakoune can take place directly over fifos.
Similarly https://github.com/kakoune-lsp/kakoune-lsp exists and has been a pretty good showcase of how complex plugins can be implemented in other languages than shell (rust here) with the extension model.
I am of course pretty biased, but I am regularly surprised by how simple/quickly things can be done with Kakoune's shell based model, and I its a part of the design I am pretty happy about in retrospect.
> I switched back to main after noticing that the plugin branch adds about 1s to startup time
Very curious what’s happening there. With empty helix.scm/init.scm, there is essentially zero measurable impact on startup for me.
Bytecode caching isn’t yet enabled (it is in progress), so I can see that kind of delay if I’m giving it tens of thousands of lines of code to evaluate on startup, but it took a lot of effort to get there.
Nono has been my daily driver since the start of the year. It's not a perfect sandbox -- that's for sure. For example, the default network rules let you escape via a global TMUX server. But it is extremely practical. It gives me enough guarantees to feel confident about running in YOLO mode. So far nothing has gone awry.
I rebuild herdr from master every few days to gauge the direction it’s moving in.
What herdr gets right over zellij:
- Default keybinds don’t conflict with helix. Even the leader key mode in zellij keeps alt-bindings around. Herdr keeps things nicely tucked away.
- There’s beta support for the kitty image protocol. Three years on, we’re still waiting in zellij.
- Notifications get passed correctly through to the terminal.
As for herdr’s killer feature, agent status monitoring, there is apparently a zellij plugin.
Big win for zellij: the docs and code feel human. By contrast, herdr’s docs have all the tells of a vibe coded project. Not to say the maintainer hasn’t put care into it! But I’m getting burned out on LLM docs.
I started to use tmux-claude-session-manager (https://github.com/craftzdog/tmux-claude-session-manager) in tmux, to have a pop-up to see all active sessions. I don't have many in parallel, so usually I don't even need, but if I'm lost or forgot one, it's really helpful.
reply