Daniel Ukaegbu

PhD Student in Computational Toxicology @JohnsHopkins @JohnsHopkinsSPH @jhucaat Toxicology and AI
Views are mine

[Daniel Ukaegbu @brah_coco]

@hayoholla @Mrbankstips LMAO ! Consider what exactly?
He’s still lying and being deceitful!! That is not why they lost

[slight fever @lifeofibrahim23]

I didn't sleep and I won't regret it cause I've been in a masterpiece with an insane playlist
Bass X Machina 🔥

[Daniel Ukaegbu @brah_coco]

@amindavison @AskMichaelTaiwo 😂😂😂😂

[Daniel Ukaegbu @brah_coco]

@marshalwahlexyz Fucking Insane !!!

[Daniel Ukaegbu @brah_coco]

@NewsWire_US yessssss!!!!!🥳🥳🥳🥳

[Daniel Ukaegbu @brah_coco]

@ThePoliticalHQ 🥳🥳🥳🥳🥳🥳🥳

[Shuiwang Ji @ShuiwangJi]

With the help of AI, I prepared a 400+ page tutorial slide deck on foundations and frontiers of generative models. The slides are at https://t.co/MwlRefxQTg and there is also a way to submit your feedback for which we really appreciate.

[DiscussingFilm @DiscussingFilm]

New teaser for 'BLUE EYE SAMURAI' Season 2.

Releasing January 19 on Netflix.

[Nicholas Ho @eigenNick]

Contextualized gene representations work surprisingly well for predicting multigene perturbation effects.

I’m excited to introduce EMBER, a pretrained gene-centric foundation model designed specifically to learn context-specific gene representations. A joint work with @jmuiuc and @ericxing.

With innovations in single-cell pretraining and dataset curation, EMBER gene representations can be combined with a lightweight probe for single gene knockouts, genetic interactions (double knockouts), and even drug perturbations.

Both a 🧵 and a 🎥, got in front of a camera to share some highlights! (1/n)

[Aaron Mueller @amuuueller]

Interested in putting interpretability on more rigorous foundations? Check out our #EMNLP2026 tutorial slides!

I learned a ton from my fantastic collaborators @rpatrik96, @_shruti_joshi_, @dhanya_sridhar. Recommended for newcomers and experts alike!

📚 https://t.co/qK9cJdOId0

[Probability and Statistics @probnstat]

Principal Component Analysis (PCA) is a fundamental unsupervised learning algorithm for reducing the dimensionality of data while preserving as much variation as possible. Given centered observations x₁,…,xₙ ∈ ℝᵖ with sample covariance matrix S, the first principal component is the direction v₁ solving

v₁ = argmax_{‖v‖=1} vᵀSv.

Thus, v₁ is the eigenvector of S corresponding to its largest eigenvalue. Subsequent principal components are obtained from orthogonal directions associated with the remaining eigenvalues. Equivalently, PCA can be computed through the singular value decomposition of the data matrix.

PCA has extensive applications in statistical machine learning, including dimensionality reduction, visualization, feature extraction, noise reduction and exploratory data analysis. In high-dimensional statistics, PCA is particularly useful for identifying latent low-dimensional structure hidden within noisy observations. It also forms the foundation of methods such as probabilistic PCA and factor analysis. In modern ML, PCA can reduce computational complexity before classification or regression and provide compact representations for downstream models. Its close relationship with covariance estimation, spectral theory and least-squares reconstruction makes PCA a natural bridge between classical multivariate statistics and machine learning.

[General Reasoning @GenReasoning]

🐝 How to Train a Swarm

In our first blog post, we explain how latency pressure can incentivise parallel work, the communication primitives agents need to co-ordinate, and outstanding challenges in training large agent collectives.

https://t.co/8HmsNqt4wQ

[Dmitry Krotov @DimaKrotov]

Our new paper in Nature Communications reveals an interesting connection between generative diffusion models and Associative Memory.

It identifies 3 phases of generated samples:
👉 Memorization: reproducing training examples verbatim.
👉 Spurious states: discovering new patterns between stored memories, while maintaining a well-defined basin of attraction around those samples.
👉 Generalization: generating genuinely novel samples with flat energy landscape around them.

The surprising twist? What was traditionally considered a flaw in Associative Memory — spurious states — may actually be the first sign of generative capabilities.

It was fun to work on this project with @baophamhq, @gaboraya, Matteo Negri, @mj_zaki, and @LucaAmb, and I am happy to see it published!

Paper: https://t.co/kkXTePfIQ0

[Shuming Hu @ShumingHu]

One of the best blogs I read recently! Both the presentation and interative widgets to illustrate retrieval from associative memory are delightful.

Random thoughts while reading:
1 Gated-MLP like SwiGLU has become the default. Have anyone tried Gated^2-MLP? or Gated^N-MLP?
2 Are there any study into network behavior when it transitions from higher to lower MDL solution during training.

[antisense. @razoralign]

EMBER: Contextualized gene representations enable compositional prediction across cellular perturbations https://t.co/IqKgo7I3RQ

[AI Quanting @AIQuanting]

Looping a transformer means running the same weights over the input more than once. The model gets deeper without getting bigger, and a new study works out when that actually pays.

The sweep covers dense decoder-only models trained from scratch, up to 1B parameters, 155B tokens and 32 loops. Which part you loop turns out to matter. Looping the whole transformer gives the best quality for a fixed parameter count. Looping only a middle block of layers gives a better return when training compute is the binding constraint. More loops is not automatically better either. In the 70M and 100M runs that go out to 32 loops, the gain from each extra loop shrinks, and it can reverse or go unstable.

Forcing every loop to use identical weights is part of that limit. The authors add small loop-specific low-rank adapters so each pass can learn a slightly different transformation. The gain is marginal at two loops and clear at four and eight. At 300M these adapters add 1.1 to 4.3 percent to the non-embedding parameter count, against 51.4 percent for simply doubling the depth.

Looping also slows generation down, because each loop has to wait for the one before it. Parallel Loop Transformers pass information between loops across neighbouring token positions so the stages can overlap. Under the published version, the paper reports that quality can degrade as more loops are added. Keeping the current token's own embedding, and normalising it separately from the carried state, brings the gains back.

My read: the loop count and the data multiplier in this paper are the same number. Two-loop models trained at 80 tokens per parameter reach a given bits-per-byte score using fewer training FLOPs than plain models trained at 160. Two passes over one set of weights bought more than twice the data did.

https://t.co/Aq1ymsAr93

[AVB @neural_avb]

Been studying mechanistic interpretability seriously for a month. Here is a web book that contains all the resources, experiment results, and interactive widgets to learn mech-interp at any level. Written by Claude, curated by me. Free, no signups. WIP.

https://t.co/g8L9CKDb2C

[Ryan Kidd @ryan_kidd44]

If you read any blog post on AI safety from this year, make it be this one
https://t.co/VYWF1iKnim

[alphaXiv @askalphaxiv]

Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve it.

The Platonic Representation Hypothesis suggests that AI models trained on different modalities, like images and text, eventually learn similar representations of the world. But this paper shows that how models are trained matters more than simply scaling them up.

So they mathematically show that, under certain conditions, two JEPAs trained completely separately can learn essentially the same underlying representation of the world. One learns from different images of the same scene, while the other learns from different descriptions of that scene, without ever being trained on image-text pairs.

Their approach achieves the same image-text alignment with 42x fewer training samples than next-token prediction. A 109M-parameter text JEPA even outperforms LLMs up to 72B parameters on cross-modal alignment benchmarks.

https://t.co/3j1enq9qF7

[Prof. Nikolai Slavov @slavov_n]

UMAP was introduced to biology data with the argument that it 'preserves the global structure of the data' better than t-SNE.

The claim was falsified in this matters arising article.

1/