Agrin Hilmkil

[Agrin Hilmkil @agrinh]

Hello world! We're out of stealth.

[Simon Kohl @saakohl]

@Latent_Labs comes out of stealth today with $50M funding. Our goal? To push the frontiers of generative biology, giving partners instant access to tools capable of accelerating drug design.

Every biotech or pharma company searching for the best therapeutic molecules understands the role AI can play - but not all are in a position to develop their own advanced models. That’s where @Latent_Labs comes in.

[Simon Kohl @saakohl]

Introducing Latent-X2 — AI-generated antibodies with drug-like developability and low immunogenicity in human panels, zero-shot.

Technical report: https://t.co/9uU6GQCdM7
Blog: https://t.co/TFyCzlZ4qN
Apply for access: [email protected]

[Simon Mathis @SimMat20]

Today at 11am I’ll present some of the exciting work in AI x Bio we’ve been up to at @Latent_Labs at the @RenPhilanthropy booth (Hall A/B, booth 1343 @NeurIPSConf - straight through the entrance at the back left). Come say hi!

[Simon Kohl @saakohl]

Introducing Latent-X — our all-atom frontier AI model for protein binder design.

State-of-the-art lab performance, widely accessible via the Latent Labs Platform.

Free tier: https://t.co/NamdznPWjL
Blog: https://t.co/2UkYDEe8a9
Technical report: https://t.co/0m2s3y7vwN

[Tycho van der Ouderaa @tychovdo]

🔵New paper!🔵 Our latest work on Pyramid Vector Quantization for LLMs achieves state-of-the-art post-training quantization with a Pareto-optimal trade-off between performance, bits per weight, and bits per activation. A thread. 👇 1/15

[Marc Rigter @MarcRigter]

Video models like Sora and Gen 3 can generate realistic videos, but can they produce useful synthetic data for planning/RL?

Our work (AVID) explores how pretrained image-to-video models can be adapted to accurate action-conditioned world models.

1/n

[Chao Ma @ChaoMaLaurant]

Congrats to the Causica team @MSFTResearchCam . We got two papers accepted in #ICML 24, contributing to our efforts on integrating #Causality with modern #FoundationModels. 1/3

TL;DR: For the potential outcome framework, self-attention = causal inference via optimal balancing [1]; For the SCM framework, SCM learning = fixed-point problem given TO = a specific causally-consistent attention mechanism [2].

Both discoveries contributes towards enabling transformer-like architecture to directly solve causal reasoning tasks, even when presented with an unseen dataset in a foundational setting. @pwnic @JiaqiZhangVic @ScetbonM @agrinh
[1] Towards Causal Foundation Model: on Duality between Causal Inference and Attention (https://t.co/oXzvfSSZ8e)
[2] FiP: a Fixed-Point Approach for Causal Generative Modeling (https://t.co/FwyawE2a92)

[Agrin Hilmkil @agrinh]

@Finnair I booked a flight with you from Stockholm to London, which was cancelled yesterday. It happens to be operated by @British_Airways. I've spent hours taking to each party and your service desks keep referring to each other. Which one of you are actually capable of helping?

[Agrin Hilmkil @agrinh]

@Tim_Dettmers I love the notation itself and it feels like it could be commonly understood if it became more popular. One argument against its current implementation would be how it makes static checks hard. But then again, we're not exactly blessed with great static checks for tensors 😅

[Agrin Hilmkil @agrinh]

@JonnyRothwell @foodora_se @McDonalds @UberEats What's more, the app lies about what's actually going on. According to the app, the food delivery only took 10 minutes. But the restaurant leaves a receipt with the pickup time, so we know it was just more than 1h.

[Agrin Hilmkil @agrinh]

@JonnyRothwell @foodora_se @McDonalds @UberEats Found this because I made the mistake of ordering through Foodora again (restaurant not available on Uber), though closer to 2h. Never again.

[Agrin Hilmkil @agrinh]

@benedictevans How should one read this figure? Seems like one is supposed to draw some conclusion regarding how the countries compare. But it looks like each plot is shifted and potentially scaled independently, making a direct comparison very hard.

[Agrin Hilmkil @agrinh]

@SebastianBujwid I hear it's really good at telling hotdogs apart from things that are not hotdogs 🙃

[Agrin Hilmkil @agrinh]

@chriswolfvision That's super helpful. Reminds me a bit about unit checking and annotating in physics.

[Agrin Hilmkil @agrinh]

@SebastianBujwid I guess this was a weird suggested query instead though

[Agrin Hilmkil @agrinh]

@SebastianBujwid Pretty surprising how bad it gets, this was also fun: (right side in english: cat, <is this english?>, cat, second world war)

[Agrin Hilmkil @agrinh]

@srchvrs Unfortunately not, trying to keep things as consistent as possible.

[Agrin Hilmkil @agrinh]

People have been asking so I figured I should share. Here are some very early real-world performance numbers for transformer fine-tuning on the A100s we're getting access to through @AIsweden. 2 epochs of Yelp 5 with ALBERT (base):

A100: 2h 7m
K80: 13h 47m

1/2

[Agrin Hilmkil @agrinh]

@Tim_Dettmers @srchvrs Definitely worth it. As opposed to CPU instances, GPU instances on GCP go down for maintenance. So for long running jobs it's worth building resumable jobs. At that point running on pre-emptible instances has been a no brainer for me.

(https://t.co/MjEvh1pC3g)