Kacper

Founding engineer at @attentioninc

prev Citadel, Optiver, IBM Research,
CS @ ETH Zurich, CS&Math @ Sorbonne

[Kirill @kegashin]

helping GPT6 Astra become agi

[Kacper @kozieblowski]

hey slopbot, here is an empty pr with a set of tests asserting a bug, please fix so the tests pass
slop: hello, i changed... slop slop slop, hides the answer in a long text, slop slop.
me: tldr, what did you do?
slop: i restricted the harness so that the tests pass, < insert longwinded explanation of golden files >
me: slopbot did you just skip the tests?
slop: well actually no, i just restricted the tests so the failing cases dont run on failing examples. its now green
me: slopbot, did you fix the issue stated in the pr description
slop: "Direct answer: there was no product bug behind that failure, so there was nothing of that kind to fix."
me: <paste the original issue explanation, same one as in the PR description>
slop: "Let me verify that fix against the actual code rather than take the description on trust."
4 mins and 5L of water later
slop: blah blah "The bug class stays open" blah blah

this happened today, fresh chat btw, so no context poisoning, opus 5

[Valentin Ignatev @valigo]

The future of solved software is looking bright

[ol’ stocky ⛳️ @oldstocky]

Treat your body like a data center

Consume as much water and literature as possible

Upset local city council with your existence

[Julian Windeck @julianwindeck]

we cut cpu usage of https://t.co/V8dwJ9yqqZ by half! 🏎️🌊

[Mike Wang @zmwang]

The amount of food I can eat to “not let it go to waste” feels endless and infinite.

[Paul Bertrand @pbertrand_dev]

Spawn a worker agent that does the work
Then a reviewer agent that reviews the work
Then a worker that fixes the work
Then a reviewer that reviews the reviewed work
Then open a PR
Then review the PR with a reviewer agent that reviews PRs

Token inflation is insane right now

[Lucas Beyer (bl16) @giffmana]

The real reason everybody is worried for Gemini after Jeff (pic related) has left:

Jeff has a keyboard with this little volume wheel thingy (pic2 related) on it. During meetings, he always looks at the pretraining curves, especially signal propagation ones (pic3 paper related). If something trends in the wrong direction, he can type a regex and then use the wheel to change the learning rate of *the specific params matching the regex* smoothly and continuously by turning it slowly (pic4 related).

Effectively, Jeff has been the per-block adaptive learning-rate scheduler for Gemini models. Without this, the loss would always explode as they scale up and train for longer. Now that he's gone, who can take over that job? Sergey? Not Jeff enough. Andy? Can't take the pressure. Nobody! That's the real struggle for Gemini 4 pretraining which is ongoing now, so time is ticking: can they find a new learning-rate scheduler worthy enough, quick enough, before the first spike shows up??

I'm in a GC with the wandb CEO who regularly brags about looking at customer dashboards, so I'll keep y'all updated on any news i hear.

[Kacper @kozieblowski]

@0baylor these were built out of cardboard, planks and stucco. never were meant to last. they fell apart on their own before being demolished. heck palace of fine arts had to be demolished and rebuilt from stronger materials at some point

[Patrick Collison @patrickc]

I only recently learned that this pedestrian bridge sits where San Francisco's first Hall of Justice did. Before/after below. I submit this in my evidence packet to anyone who maintains that nothing has gone wrong.

[Kacper @kozieblowski]

To add to this, AI can already find useful formalizations if prompted but the question is more about autonomous creativity and discrimination.

AI currently follows directions but it needs to define the direction itself. And choosing the right research direction is non trivial.

[Kacper @kozieblowski]

The question now becomes, when will AI become capable of finding the right formalizations/abstractions and be able to judge if they are useful/general enough.

If it then is capable of making statements and conjectures about those entities and if also is able to judge where to steer the formalization and research, then we effectively solved math.

[Mert Ünsal @mertunsal2020]

insane times we live in

I recall spending a few hours on this problem couple years back after it was presented in my Mathematics of Data Science class at ETH... it felt really hard and now LLMs zero-shot it

[Marvin von Hagen @marvinvonhagen]

Rewind was ahead of its time!

Since agents took off, video data of work getting done has become incredibly helpful:
(a.) Personal context
(b.) Enterprise-level insights.

I'm both personally a Coast Local DAU, and @Interaction has equipped every Poke Human with @AttentionInc cloud software.

Two real examples from my last months with Coast Local:

1. Coast is the de facto-memory system for my Claude Code & Devin CLI.
Local agents naturally (without being asked to!) start querying the Coast CLI to check what I have done, and what they have done before

2. I watched a presentation during a Zoom call, and then needed it, but couldn't easily ask for it.
Claude Code one-shotted going through the Coast recordings, and cropped each screenshot / video frame, before finally combining them all into a finished PDF. I got the whole presentation without asking ever!

Both examples show very novel use cases that simply weren't possible before. But Coast Local also has feature parity with Rewind, while being free instead of $30/mo!
So you get the system of record for every form you ever fill, every website refresh that loses drafted text, a queryable index of everything you've ever looked at, and all that.
Try it out for yourself!

Pic of @Interaction, @AttentionInc, and Palo Alto friends this weekend below 🌊🌴

[Aurther Nadeem @AurtherNadeem]

figma can show you what it looks like. only code can show you what it feels like. i have spent many days on making animations and interactions feel "just right" - super excited to get this in your hands

[Silas Alberti @silasalberti]

"context is the bottleneck for agents" - I’ve heard this thesis many times. It‘s surprisingly hard to execute on it. @aidangch & @julianwindeck have assembled a stellar team and built a product that’s delightful to use & delivers. Lots of potential for both personal use & enterprise!

[Anika @AnikaSomaia]

we optimize for capability because capability is legible. but what you actually want from a model is illegible – to the model, and probably to you. it never fits in a single request and it's incredibly personal. it's in what you're doing, what you've abandoned, what you keep choosing

in other words, the context that actually matters can't be authored. it has to be accumulated

i couldn't be more excited that @aidangch @ssanticalderon and team are working on this!! congrats 🥳

[Ryan Zhu @ryanzhuuuu]

this is a thing that i always hope someone can build. we have left so much information on our computers, it’s so valuable for intelligence

[Ayush S @ayushswrites]

Congrats on the launch! Let’s goo 🔥