Lukas Kidzinski

Building generative AI since 2015. Co-founder of Saliency acquired by Clario.

[Lukas Kidzinski @kidzik]

Jev is (likely) a diffusion model. It seems to match other diffusion approaches in speed and performance (@InceptionLabs, @googlegemma, @adityagrover_)

I built Jiffy to test whether an open diffusion model could reproduce similar decision performance. It got surprisingly close without task-specific retraining.

On 231 public JevBench tasks:
• Jiffy: 84.4%
• Jev’s published results: 86.6%
• Jiffy latency: 259 ms median on an H100 (so likely can be substantially improved just with better hardware).

After days of distillation experiments, my strongest accuracy-latency result was essentially a thin wrapper around frozen DiffusionGemma.

Open source, including Doom and Tetris demos: https://t.co/nRVFugdxeg

[Lukas Kidzinski @kidzik]

@TexDoggie I don't know looks quite legit to me.

[Lukas Kidzinski @kidzik]

@legen_eth

[Lukas Kidzinski @kidzik]

@danielnewmanUV

[Lukas Kidzinski @kidzik]

@FinansowyUmysl

[Lukas Kidzinski @kidzik]

Much has been written about $NVDA tracking the dot-com bubble. Less attention has been paid to the Australian sheep population.

After normalizing the series, compressing 150 years of agricultural history, and aligning the peaks, a compelling historical parallel emerges.

Both experienced substantial growth. Both had setbacks. Both eventually reached the point I aligned with the other peak.

I believe this deserves the same analytical weight.

[Lukas Kidzinski @kidzik]

@coinbureau Let's not forget Australian sheep!

[Tomek Czajka @TCzajka]

Alice earns $50K a year. Bob earns $1M a year. Next year, both earn $10K less. Whose tax bill should drop by more?

[Piotr Migdal @pmigdal]

I created a voxel map of central Warsaw.

Explore as you wish, change seasons and weather, or even possess a pedestrian.

Made with Opus 5.5, heavily based on open data @openstreetmap, LiDAR @GUGiKgovpl, and the tree database @warszawa.

Live: https://t.co/axuA22CkFB

[Lukas Kidzinski @kidzik]

@sundarpichai @ArtificialAnlys

[Lukas Kidzinski @kidzik]

@mtpennycook Tell me you don't understand logic without telling me you don't understand logic.

[Lukas Kidzinski @kidzik]

@ylecun @elonmusk Why stack perceptrons when a single linear model is so much simpler?

[Lukas Kidzinski @kidzik]

I did my PhD in mathematics.

Today, work similar to what took me three years can be done in a week with a few prompts.

I'm not particularly bothered by that.

What I'm more anxious about is how this changes who gets to contribute. Mathematics felt meritocratic: with a paper, a pencil, and access to a library, you could take a serious shot at an important open problem.

As research becomes something we can purchase as computation, access to capital will increasingly determine who can compete. How do we even know whether an open problem requires a new idea or simply a larger token budget?

Many other disciplines will likely follow. More money buys more computational capacity. That capacity produces discoveries. Those discoveries can generate more money, which buys still more capacity. One could say it's always been like that, but as we replace humans with compute, this loop will accelerate, and inequality will widen.

[Lukas Kidzinski @kidzik]

@GaryMarcus The entire discussion misses one important issue. A year ago, all you needed to do math was access to a pen, paper, and a library. Now, to "push the frontier," you need $10M in tokens, and AI labs will gladly keep presenting it that way.

[François Chollet @fchollet]

What if the jagged frontier is mainly math + code (which you can push arbitrarily far with RLVR), and everything else starts to plateau because it is still bottlenecked by human generated data?

Model performance in non-verifiable areas has kept improving steadily, albeit much slower than for math and code. But is that steady improvement a side effect of a higher G (itself driven by RLVR), or only a function of the amount of new human data getting injected into training (which is still continually happening on a massive scale)?

A lot of things depend on the answer to this question

[Jungkyu Park @jpatrickpark]

Neoadjuvant therapy means months of treatment before surgery. Knowing a patient's chance of pathological complete response (pCR) before treatment could help tailor that treatment.

Today we're releasing two preprints on NEO, our AI model that predicts a patient's chance of pCR from a routine biopsy slide. 🧵

[Piotr Migdal @pmigdal]

Navier-Stokes Equation, explained with colors
https://t.co/TcTMh5vYQL

[Guanyang Wang @GuanyangW]

KLS conjecture is proven by AI 👀

[Lukas Kidzinski @kidzik]

@thsottiaux Isn’t it a little counter productive? Users now come and hope for a reset, praying you don’t release anything and get disappointed when you do.

[Lukas Kidzinski @kidzik]

@OngDevLab To be fair it does feel like ten years already.

[Christian Szegedy @ChrSzegedy]

Reading all those recent beautiful essays by Terry Tao, Kevin Buzzard and @wtgowers, I could not resist, but write up my deeply personal perspective on how I perceive and see things:

https://t.co/fvWb4b7jGU