Super interesting! I'm currently evaluating the new gemini models for hand-written text extraction and somehow Gemini 2.0 Flash has problems with actual words. German or English, doesn't matter. Haven't had the same problem with Flash Lite

Late last year I converted some of my projects to fastht.ml - and have been doing all new projects with it. My productivity and enjoyment has increased significantly. An absolute joy to use. Thousand thanks to @howard.fm et al for the fantastic tooling.

Here is why I think reasoning models (like DeepSeek R1) are a huge step forward. (and it's not necessarily their superior reasoning performance) blog.mariusvach.com/posts/i-love...

Llama 3.3 70B with speculative decoding on @groq.com is absolutely crazy. The answers come instant.

The effective use of LLMs IS a skill to be learned, just like using Google effectively is a skill too.

Is there a vim mode for microsoft word? vim motions really tend to infect everything you do on your computer.

chat.deepseek.com seems to have some performance issues right now :D The hype got real.

There is still room for an intuitive LLM app library like llama_index or langchain. Man, these two libraries are a mess and so bloated. The source code is unreadable. That's the problem when you try to do everything all at once.

Back from holidays to build cool shit and talk about it :)

That’s one one my favorite posts by @morganhousel.bsky.social and there is some striking similarities to the work of Hartmut Rosa on ‚Resonance‘. In the modern world we put too much emphasize on being efficient and correct. Super interesting stuff. collabfund.com/blog/intelli...

It’s amazing how hard it is to beat BM25 for retrieval, especially in realms with specialized language like medicine.

Super exiting stuff. I hope this will lead to smaller, more capable models!

I feel like there should be a ML/AI version of this: "Training!" xkcd.com/303/

editorialmanager.com is another business ready to be disrupted.

I can now run a GPT-4 class model on my laptop (The exact same laptop that could just about run a GPT-3 class model 20 months ago) The new Llama 3.3 70B is a striking example of the huge efficiency gains we've seen in the last two years simonwillison.net/2024/Dec/9/l...

This is just some nasty drumming www.youtube.com/watch?v=CQRX...

Resonates...

New Blogpost! How to use python decorators to define layouts in fastht.ml blog.mariusvach.com/posts/decora...

After a very stressful AoC 2024 day 6, maybe today we'll do something easier: building a random forest library or some web dev. That problem yesterday stressed me out :D

AoC animated #adventofcode #aoc

I love how fast you can spin up a jupyter lab instance using `uv`: `uv run --with jupyter jupyter lab` This `uv run` is so helpful.

I tried out the new SmolVLM model together with ColPali for a visual rag pipeline. It's really nice to run everything on 15gb of VRAM. Unfortunately, the QA abilities of SmolVLM are not as strong as I would have wished. Here is a notebook with the code: colab.research.google.com/drive/10ZnqA...

That was fast: The makers behind ColPali are already cooking up a version using the new SmolVLM model by @hf.co as the backbone. Exciting stuff! huggingface.co/vidore/colsm...

This little python class comes in very handy I you want to have javascript style function chaining like: "Pipe(sample_text).pipe(split_into_words).pipe(convert_to_lowercase).pipe(remove_punctuation).pipe(count_words).value" Python supports a much more functional programming style than I thought.

One big advantage of a new social network is that you can curate your feed new. I realized that I want to read more about AI/ML engineering/research and web development and less about indie hacking. The first one is much more authentic…

Very exciting release. I love these small LLMs/VLMs recently

Did I already say that I love alpinejs.dev ? Anyway, I love alpinejs.dev! h/t @calebporzio.bsky.social

Needed a simple carousel component for a client project. claude.ai needed one prompt to create a react carousel component that does exactly what I need. 5 years ago this would mean fiddling around with obscure packages. Crazy times.

Having 30 followers here is more satisfying than 500 followers on twitter of which 300 are bots and 20 people actually see your posts

All of this generative AI stuff is nice and all, but a lot of ML applications are still "discriminative" and it's a little sad that research for this moved into the background. Is there any good research on zero-shot/few-shot classification abilities of Llama, Claude etc. vs Bert?

Started reading refactoring.com after a recommendation by @jph.bsky.social Hopefully, it will help me avoid 🍝 code from now on.

Super exciting project.

I'm still amazed by how responsive the @bsky.app web app feels. It's. so. snappy.

That's really exciting. Missed the first version, but structured extraction is where small LLMs can really shine imo. And the demo playground is actually quite impressive. Worked really well with some data extraction from medical literature.

Had a real (positive) WTF moment with Claude in a long long time right now. It totally blew my mind. Might write about it later.

When the LLM is too well-behaved

I really hope this app persists. It feels so much more solid than X/Twitter