iiinova — where your digital self keeps evolving

Lucas Beyer (bl16) @giffmana.ai · 26/01/2025

I noticed that I'm not using bsky much anymore. Not sure why, vibes. Anyways, someone noticing that DeepSeek refuses to answer *anything* about Xi Jinping, even the question whether he exists at all, triggered me writing a short snippet on safety fine-tuning: lb.eyer.be/s/safety-sft...

6888

Lucas Beyer (bl16) @giffmana.ai · 02/01/2025

First candidate for banger of the year appeared, only 2 days in:

0280

Lucas Beyer (bl16) @giffmana.ai · 19/12/2024

Not quite because this one is not stacked, so I give it better chance to scale:

010

Lucas Beyer (bl16) @giffmana.ai · 18/12/2024

A post by @cloneofsimo on Twitter made me write up some lore about residuals, ResNets, and Transformers. And I couldn't resist sliding in the usual cautionary tale about small/mid-scale != large-scale. Blogpost: lb.eyer.be/s/residuals....

3819

Lucas Beyer (bl16) @giffmana.ai · 10/12/2024

Good morning Vancouver! Things are different here: this guy is alone, chonky, and not scared at all, I was more scared of him towards the end lol. Also look at that … industrialization

3310

Lucas Beyer (bl16) @giffmana.ai · 09/12/2024

Good morning! On my way to NeurIPS, slightly sad to leave this beautiful place and my family for the week, but also excited to meet many new and old friends at NeurIPS!

3650

Lucas Beyer (bl16) @giffmana.ai · 06/12/2024

Yes. ViT is not inspired by nonlocal paper at all, and I may even have been the only person (or close to that) in the project who knew nonlocal paper. I’m not saying this as a hater, I liked nonlocal:

270

Lucas Beyer (bl16) @giffmana.ai · 03/12/2024

It's relatively long, so just a few highlights: 1. It works really well on GPUs out of the box. 2. I made ready-to-go docker templates for runpod. 3. Fine-tuning is cheap!! A sweep of 72 runs cost me less than $1 4. PaliGemma and SigLIP nail the "line intersections" task with 128 examples:

0130

Lucas Beyer (bl16) @giffmana.ai · 03/12/2024

Our big_vision codebase is really good! And it's *the* reference for ViT, SigLIP, PaliGemma, JetFormer, ... including fine-tuning them. However, it's criminally undocumented. I tried using it outside Google to fine-tune PaliGemma and SigLIP on GPUs, and wrote a tutorial: lb.eyer.be/a/bv_tuto.html

311519

Lucas Beyer (bl16) @giffmana.ai · 02/12/2024

Haven't actually read it yet, but man, this image is beautiful and perfect, well done:

150

Lucas Beyer (bl16) @giffmana.ai · 30/11/2024

Yeah but don’t think this was somehow a US exposure thing. Funnily relevant to this discussion, „Bundeswehr university“ literally means defense/military university.

110

Lucas Beyer (bl16) @giffmana.ai · 29/11/2024

Some recent discussions made me write up a short read on how I think about doing computer vision research when there's clear potential for abuse. Alternative title: why I decided to stop working on tracking. Curious about other's thoughts on this. lb.eyer.be/s/cv-ethics....

1917320

Lucas Beyer (bl16) @giffmana.ai · 28/11/2024

I must be on some kind of list or something lol

170

Lucas Beyer (bl16) @giffmana.ai · 28/11/2024

segment cow not on the beach, but as you may know I like cows in "out of distribution" places and poses. huggingface.co/spaces/big-v...

2190

Lucas Beyer (bl16) @giffmana.ai · 27/11/2024

Here's Gemini latest, which is wrong in both variants, and also yaps way too much :(

030

Lucas Beyer (bl16) @giffmana.ai · 27/11/2024

Here's a fun real-life prompt where 30min of Googling didn't really help me. The small/fast/mini chatbots all failed miserably The large ones work great (except Gemini). Also: 1. I really like C3.6 giving only answer and asking if explain 2. Wild that Chat understood the calls and added comments!

1151

Lucas Beyer (bl16) @giffmana.ai · 27/11/2024

Reminder that robots.txt is a gentleman’s agreement, not a legal document. The good actors all respect it, but there’s many bad actors out there on the internet.

3411

Lucas Beyer (bl16) @giffmana.ai · 26/11/2024

A sunny winter morning near the lake in Zürich is probably one of the best mood enhancers available world-wide. I don't think I'll ever get tired of this.

2900

Lucas Beyer (bl16) @giffmana.ai · 25/11/2024

Also switched my main server from nginx to Caddy. I'm quite liking it so far: https works out of the box and for free, and the whole configuration got shorter and more readable.

2100

Lucas Beyer (bl16) @giffmana.ai · 24/11/2024

doubt it, must have been something else:

100

Lucas Beyer (bl16) @giffmana.ai · 24/11/2024

Yeah pretty sure i learned muting from either you or Plumard (sorry, you two occupy the same spot in my mind). Though i use it slightly differently lol:

130

Lucas Beyer (bl16) @giffmana.ai · 23/11/2024

PSA: force-pushing to GitHub does NOT actually delete/overwrite anything. Everything you squash and force-push over is still there:

2301

Lucas Beyer (bl16) @giffmana.ai · 23/11/2024

And description of the model is in the code, I guess. And they do say it's a DiT.

020

Lucas Beyer (bl16) @giffmana.ai · 23/11/2024

A real-time (or very fast) open-source txt2video model dropped: LTXV. HF: huggingface.co/Lightricks/L... Gradio: huggingface.co/spaces/Light... Github: github.com/Lightricks/L... Look at that prompt example though. Need to be a proper writer to get that quality.

68910

Lucas Beyer (bl16) @giffmana.ai · 23/11/2024

soooo, this marks the 10th anniversary of using digitalocean for my personal server. It's been running smoothly on 512MB RAM all that time. Now I'm going to splurge on a 1GB RAM, just so it feels like an upgrade! (Same price.)

5420

Lucas Beyer (bl16) @giffmana.ai · 21/11/2024

The third example in the sft mix is a mean joke about Poland in German, using the stereotype of Polish people being thieves. I wasn’t even trying to find something bad 🙃

3220

Lucas Beyer (bl16) @giffmana.ai · 21/11/2024

I switched from htop to the much prettier `btop`. Love it, but few caveats: 1. ubuntu pkg too old, need to build from source (easy) 2. for me, it crashes because of some assert with 8 GPUs, works with 6. 3. Researching the above, found a comment where author recommends using nvtop in that case :(

6850

Lucas Beyer (bl16) @giffmana.ai · 20/11/2024

www.astralcodexten.com/p/how-did-yo...

89818

Lucas Beyer (bl16) @giffmana.ai · 19/11/2024

French startup H Company released their first model/product today: Runner H, a web agent that's a 3B VLM. The startup was co-founded by Charles Kantor and 4 DeepMinders: Laurent Sifre, Karl Tuyls, Daan Wierstra and Julien Perolat in May. The latter 3 left in August. www.hcompany.ai/blog/a-resea...

1284

Lucas Beyer (bl16) @giffmana.ai · 19/11/2024

Hello, Python! Do you know the reason?

10331

Lucas Beyer (bl16) @giffmana.ai · 19/11/2024

Still spamming refresh on that page, we’re getting very close!

120

Lucas Beyer (bl16) @giffmana.ai · 19/11/2024

This is how i worked on conference deadlines during PhD.

16890

Lucas Beyer (bl16) @giffmana.ai · 18/11/2024

They really missed an opportunity for "Willkommen in La Forêt" here. That being said, nice to see Europe going strong!

2401