Max Idahl

AI Research @ ellamind

[Max Idahl @maxidahl]

Great to see a tech report with some details on data. Even better to see propella being put to good use

[Poolside @poolsideai]

Today we’re publishing the technical report behind Laguna M.1 and Laguna XS.2.

This report opens up more of what went into them: Model Factory, pre-training data, distributed training, post-training, agent RL, quantization, and evaluation.

https://t.co/RWk2F9IrAI

[Max Idahl @maxidahl]

three /goal per day

[Frank Hutter @FrankRHutter]

Huge news: @prior_labs has signed a definitive agreement to be acquired by @SAP.

€1B+ invested over four years to build a globally-leading frontier AI lab for structured data — in Europe, in the open.

Independent entity. Same team, same mission, same open models. A massive boost to what we can do. The mission just got accelerated.

Founders’ statement: https://t.co/7ZEi7q8a8l

(Deal subject to regulatory approval; terms not disclosed.)

[Max Idahl @maxidahl]

@damian_b Last time I booked Starlink on a flight they just sent out an email that it won't work 5 mins before takeoff...

[Max Idahl @maxidahl]

a tiny bug snug between my mac's display glass and the panel. my inner monk is gonna have a very rough time

[Max Idahl @maxidahl]

managed to print two posters last minute today in rio. was less of a hassle than expected

[Max Idahl @maxidahl]

Going up the stairs to Christo redentor the day before @iclr_conf, everyone talks about GPU clusters

[Max Idahl @maxidahl]

Just landed in Rio. A few days early to clear the head before @iclr_conf . Ping me if you want to chat about agents, evals, data, or to get out for some hiking/sightseeing

[Max Idahl @maxidahl]

deep crawling 150 domains doubles the number of publicly available German HQ tokens?

[Max Idahl @maxidahl]

Quality-guided crawling. The easter hack project is turning out too good.

[Max Idahl @maxidahl]

Find me here today.

[PyTorch @PyTorch]

GPU MODE (@GPU_MODE) & PyTorch Foundation are organizing an ML systems hackathon in Paris on April 9, immediately following PyTorch Conference Europe 2026.

Researchers and engineers will compete across two tracks:

-Distributed training (LLM speedrun) and inference optimization (leaderboard)
- Access to a B300 cluster from Verda and H200 instances from Sesterce
- Cloud credits as prizes, including 48-hour access to a GB300 NVL72 rack
- Talks from PyTorch (Helion), vLLM, Prime Intellect, and more
- Food and refreshments

Doors open at 9:30, with the closing ceremony at 20:00. Attendees can join with a pre-formed team or match on-site. Location details are shared upon registration.

Spots are limited. Register: https://t.co/1BYgfGWJIr

#OpenSourceAI #PyTorchCon #PyTorch #vLLM #Helion

[Max Idahl @maxidahl]

This cluster is great. Loads 1T Kimi-K2 from disk to GPU in 108 secs. 512 GPU job, no crash ~5 days in.

[Lisan al Gaib @scaling01]

Telekom and NVIDIA building a $1.1B datacenter in Munich with 10k GPUs including DGX B200 and RTX PRO Servers

[Max Idahl @maxidahl]

@llm_wizard What I need is checkpoints, e.g. at 1T,2T,4T tokens trained

[Max Idahl @maxidahl]

Joe Nemotron also has a pretty good compute multiplier it seems.

[Luca Soldaini 🎀 @soldni]

This plot undersells how much of a compute multiplier Olmo Hybrid is: 2x compute multiplier on many downstream tasks (and solid LC performance!!!)

[Max Idahl @maxidahl]

-> https://t.co/lsPpaGPj0J

[Max Idahl @maxidahl]

propella got an oral at the DATA-FM workshop at ICLR 2026. See you there!

[Max Idahl @maxidahl]

If only they would have avoided the drop at stage change

[Max Idahl @maxidahl]

Looking at Olmo 3 midtraining. Within 100B tokens, roughly the same gains on downstream evals as within the last 5T tokens.
Time to make midtraining alltraining. First 5T token open midtraining dataset when?

[Alexander Doria @Dorialexander]

Actually we have started to use Propella internally to curate common corpus subcollections and run comparisons with other pretraining dataset. Really filling a missing piece of training/synthetic infra in Europe. https://t.co/VnKk5MoIvt