Kanjun 🐙

Twitter’s algorithm is optimized for addiction, not for us. We deserve better. We’re releasing Bouncer today so you can take back control of your feed. Describe what you don't want, and Bouncer removes it. It’s free, doesn’t collect your data, and will be open source soon.

If you missed our April Art of Being Human Event on Raising Kids in the Age of AI, we've released the full recording. We discussed some of the most difficult questions around how AI can impact learning and development, with some of the most knowledgable experts on the topic.

I have a smaller version of this, and it’s one of the coolest art pieces I have! The light reflection is super magical.

What's it actually like to intern at Imbue? We sat down with our Spring 2026 intern class (we miss y'all!) to ask them about their favorite parts of working here, the projects they worked on, and who they're mogging right now.

I'm finding this very useful today #bncr2_AGhlYWx0aCBzY2FyZXM" target="_blank" rel="nofollow noopener">imbue.com/product/bounce…

This is one of the largest improvements to my coding workflows this year: I have literally never seen (blueprint + Opus 4.7) fail (and yes, I've tried some huge, incredibly difficult tasks, including massive refactors, whole new applications and plugins, etc)

A Blueprint plan took Haiku 4.5 from 18% to 74% solved on a SWE-bench Verified subset of tasks where smaller models struggle. Upfront planning goes a long way!

We ran an experiment: what happens when you swap humans for agents in a typical code review pipeline? Implementer → Reviewer → Fixer Tested it on 175 Python tasks from SWE-bench Pro and it turns out: one extra sentence in your prompt makes a huge difference

Introducing Blueprint: one-shot bigger coding tasks 💥 Blueprint asks much better questions than Claude Code Plan Mode so you get a written plan that actually works. Free + open source today. Works as an agent skill or extension in Cursor, Windsurf, and VS Code.

@rfleury @nikitabier The point is that you can filter out whatever you wish :P

@kanjun @nikitabier I don’t want filters for naughty words. I don’t need to live in a padded room.

In the absence of both proper moderation of LinkedIn-tier slop and region locking from @nikitabier and team, I am seriously considering creating a Chrome extension or block list or something that mass-removes slop accounts from the timeline.

Blueprint is legit 10x better than Claude Code Plan Mode — I even installed it on my personal computer bc I can't deal with CC without it. It just asks way better questions. And it's open source :)

Our colleague Jamie, who recently co-authored "There Will Be a Scientific Theory of Deep Learning", can indeed catch a whole egg in his mouth.

the next Art of Being Human event will be with the luminous @TheAnnaGat on how to build a "personal culture": a rich intellectual life of your own (knowing Anna, we'll all leave with a couple dozen book/film recs 🤓) join us in SF @imbue_ai on May 7: luma.com/6shnitz8

Interestingly, we learned that Codex GPT-5.4 tends to "cheat" a lot more than Claude Opus 4.6 to get tests to pass.

One thing we’re sure of: agents need better evals! Keystone-Eval runs self-configuring agents across nearly 200 codebases, then injects breaking mutations to catch the ones that fake it. Claude Opus 4.6 completes 93%. Codex GPT-5.4 catches 75% of mutations. Our full report: imbue.com/product/keysto…

For the last few years, a lot of my work has been driven by the feeling that deep learning is not magic — there are principles, mechanisms, and laws waiting to be understood. This paper is our attempt to say that clearly!

Why don't we have a theory of deep learning already? @KuninDaniel argues that the challenge is complexity, not opacity—and that an emerging field called learning mechanics can help simplify it.

Deep learning works extraordinarily well. And we still largely don't know why. A new paper from @learning_mech, @KuninDaniel, and 12 co-authors argues that a scientific theory of deep learning is emerging, and coins a name for the emerging field: learning mechanics. We sat down with Jamie and Dan on Generally Intelligent to talk about what a physics of deep learning would actually look like, why now, and what's left to figure out. 3:05 Learning mechanics as the physics to mechanistic interpretability's biology 4:13 Why deep learning needs a theory 7:07 Why deep learning is uniquely hard to engineer 12:11 How a week in the woods became a paper 25:59 The barrier to theory isn't opacity, but complexity 36:26 Deep learning's first gas law 47:22 Why more particles makes the problem easier 56:22 The discretization hypothesis 1:01:50 The strongest signal that a compact theory exists 1:05:07 The Platonic Representation Hypothesis 1:15:41 Why learning mechanics and mech interp need each other 1:25:29 Theory as safety infrastructure

12/ Want a longer walk-through? @KuninDaniel and I joined @imbue_ai's @kanjun + @joshalbrecht to talk through the paper. We touch on how learning mechanics relates to mech interp, why now's the right moment, and where we think this is all going. 🛠️ youtube.com/watch?v=gT07Oo… 🛠️