Nadeem

@emilkowalski @joshpuckett You should deff set up something like @threejsJourney's monthly jams for enrollers. I'd love to see what people create with the course (+ gets you more visibility)

Each enrollment of animations.dev comes with free updates for existing students. This time it’s a new lesson on training your judgement with more than 20 exercises, and a guest lesson by @joshpuckett called “Animations as Proof of Care”. You still have 2 days to join!

@omarsar0 Maybe if we had the infinite list of tasks we're to ever ask a model to solve. I just ask a model to build a harness for the project and self-evolve over time. Use a teacher model to test it on a range of project-relevant tasks and help course correct when needed.

i haven't seen a model that just works across agent harnesses. seems like it should exist. great opportunity for open-weight models. any thoughts?

@steveruizok for the greater good

I think it's time we provide researchers with training data. For now all of the data is going to be mine personally.

the beginning of something new: huggingface.co/datasets/steve…

@ben_burtenshaw @_lewtun @willccbb Hey would love to learn how HF's own products/teams benefitted from RL agents/loops internally? And what went into deciding if it's worth the time, cost, risk against other existing now mostly mature alternatives/MLOps workflows.

2 days until we will hold this deep dive workshop on everything RL for agents with some of the best names in the game. Objective: understand what it takes to train agents in the open. Speakers include: - @_lewtun giving a crash course on RL beyond language - @willccbb defining the core bottlenecks in open source rl tools. - @OfirPress is going to ground the whole thing with insights from SWE-Bench-* - @a1zhang is going to come at it with recursive language models, and show us what's possible at the harness level. Drop your hard questions in this thread, and I'll raise them in the session. or come along and just heckle.

The paper I’ve been most obsessed with lately is finally out: nbcnews.com/tech/tech-news…! Check out this beautiful plot: it shows how much LLMs distort human writing when making edits, compared to how humans would revise the same content. We take a dataset of human-written essays from 2021, before the release of ChatGPT. We compare how people revise draft v1 -> v2 given expert feedback, with how an LLM revises the same v1 given the same feedback. This enables a counterfactual comparison: how much does the LLM alter the essay compared to what the human was originally intending to write? We find LLMs consistently induce massive distortions, even changing the actual meaning and conclusions argued for.

@handsdiff @NousResearch @ssh_exe_dev The VM solution sounds great, did you discover any other cheaper alternatives (that you could manage yourself on a single VPS?). How are you also handling secret management? I'm using infisical but it's still quite a hacky wrapper around hermes.

I am putting @NousResearch hermes agents on @ssh_exe_dev containers, giving them good secret management, making them public, and letting it rip.

@p44v9n Nice! Could you add a rss feed?

lots of events for designers in London this weeken

@claudeai @figma and something that we need to see more often x.com/_catwu/status/…

@claudeai @figma a good adjacent read alignment.anthropic.com/2026/psm/

Enjoying @claudeai Design. When you can own the model layer, seems like you get more leverage over the traits of the persona, enough so that the system feels closer to havin consistent expert design judgment built-in vs something like @figma's harness that insists a persona.

@IanArawjo This is great, i've been struggling to gather confidence in benchmarks w.r.t. real world uses and needed to get back to basics.

Thanks to a colleague's suggestion, promptstats is now renamed to "evalstats." This renaming better conveys the breadth of the package—aiming to be the go-to resource for everything statistics and AI evaluation, whether that's comparing models or prompts. github.com/ianarawjo/eval…

@mr_r0b0t @NousResearch @TencentAI_News please support this as well @Hetzner_Online

Big news for @NousResearch users wanting to run Hermes in a hosted environment! Very cool @TencentAI_News!

Best career hack is to make sure you’re the person in the room who's always having fun.

@zqwq333 @IanOsband @_rockt @wgaluba @DGneusheva @modic123 @george__wing @aiengine_hack Really looking forward to this. Signed up but when are you guys approving folks?

For an in depth discussion on RL: IRL with @IanOsband @_rockt @wgaluba Armin Cc: @DGneusheva @modic123 @george__wing luma.com/4acss683

@JakeHulberg @NousResearch @infisical This isnt particularly great but does the job I need as I already to fine tune and rotate my secrets outside infisical anyways. Would love any advice on how to improve this (havent read the docs yet). I'm no cybersec.

@JakeHulberg @NousResearch @infisical It's hacky rn, a startup script authenticates with machine credentials, then wraps the hermes process, secrets get injected as env vars. Systemd keeps it alive. and restarts it to pick up new secrets. Looking into a subprocess/subagent wrapper approach atm.

@NousResearch's hermes-agent has been good help with streamlining my research tasks this past weekend. Would have saved me a lot of setup time if they had some sort of out of the box VPS config and secret management provider support.

the best thing you can do for yourself is actively increase your surface area for luck to hit you. go outside, try new cafes, museums, events, take a new route home, speak to people, ask questions, side quest. the more you do, the more serendipity and synchronicity will find you.

@mervenoyann @GoogleDeepMind one of my fav demos. This would be so good at parties.

Lyria by @GoogleDeepMind is next level it did average Turkish pop song very well all my Turkish friends in the conference hall was summoned lol