Jeremy Howard

Seems likely that at some point AI will make it much easier to hack critical infrastructure, create bioweapons, etc etc. Many argue that if so, a hardcore nonproliferation strategy is our only option.

Rising Tide launch week post 3/3 is on why I disagree 🧵

helentoner.substack.com/p/nonprolife...

Explaining Human Preferences via Metrics for Structured 3D Reconstruction

@jacklangerman.bsky.social Denys Rozumnyi, Yuzhong Huang, @ducha-aiki.bsky.social

tl;dr: we asked 3D modelers to rank wireframe reconstructions & compared it to ranking by metrics. Observations🧵
1/
arxiv.org/abs/2503.08208

regex.search - Carlo Lepelaars Using regex.search for simple string parsing. `regex` is a superset of `re` with additional functionality and more thorough Unicode support.

Struggling with Python `re`? regex is a superset of re with additional functionality and more thorough Unicode support. It is very powerful and still kind of an open secret. When I saw
@howard.fm explain this I just had to write about it!

Today I Learned:
carlo.ai/til/regex-se...

Currently reading a study about influenza aerosol transmission, published in 2020 and shows more RNA copies (log scale) in fine aerosol fraction.

In 2020…

and we still haven’t sorted out air safety in hospitals, and we are still whinging about flu being such a big problem (which it is).

This is fascinating, so LLMs.txt is essentially a hybrid of a sitemap and schemas, accepted by a fair few AI chatbots. You can reference all manor of information about you and your brand in there - to improve AI interpretations. towardsdatascience.com/llms-txt-414...

This may be a professional bias of someone who was originally trained to study philosophy, but when my friends tell me a thing, I generally try listen to them respectfully and carefully.

FastHTML Modern web applications in pure Python

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.

Yeah it basically killed it. Lots of folks have told me rather apologetically that they couldn’t deal with all the hate here so gave up on it.

I have to admit that except for the occasional cross-posting, not sure I’ll keep very active on here. There was a big opportunity to migrate the core AI/ML community that was unfortunately squandered with all this hf dataset backlash.

Thanks to @howard.fm for the llms.txt initiative! For context: (Analogy) -> llms.txt is like robots.txt; helps LLMs to provide information by using a website at inference time.

2024: A year of extraordinary progress and advancement in AI As we move into 2025, we’re looking back at the astonishing progress in AI in 2024.

Demis Hassabis, James Manyika, and I wrote up an overview of the AI research work & advances across Google in 2024 (Gemini, NotebookLM, robotics, ML for science, & advances in responsible AI+more). 🎊

Given it a read or paste it into NotebookLM to listen, if you prefer!

blog.google/technology/a...

no no no you misunderstood. I said “fuck YOUR feelings”. MY feelings are very important and must be handled gently, like a tiny baby hummingbird

I would like a small non-lethal but painful electrical shock to come through the keyboard of anyone who uses the phrase “do better.”

Welcome to my filter bubble! ;)

Yeah my feed is nearly exclusively AI research content, which I like.

Sky Follower Bridge - Chrome Web Store Easily transfer your following users and list members from X to Bluesky.

Getting myself set up here. I found the Sky Follower Bridge Chrome plugin pretty helpful (thanks @kawamataryo.bsky.social!)

chromewebstore.google.com/detail/sky-f...

🔥 Host your apps for FREE and simplify your infrastructure! 🚀

Tired of complex cloud hosting and the recurring bills that come with it? Host locally on hardware you already own and access anywhere with a browser. It's simpler than you think!

Learn how in my latest post. 🔗👇

General consensus in the replies and quotes of this seems to be that the entire concept of "AI skills" is a joke - how hard is typing text into a chatbot, really?

I will continue to argue that it's genuinely difficult, and that the challenge in using these tools is widely underestimated

A new paper dropped from DeepMind: Deliberation in Latent Space via Differentiable Cache Augmentation

The trouble is, it's not very readable. I tried making a thread here, but it got far too long, so it's a blog now:

timkellogg.me/blog/2024/12...

Thanks for checking it out Simon :)

Wrote up my notes on ModernBERT, the brand new modern alternative to 2018-era BERT released by @benjaminwarner.dev and @howard.fm and team simonwillison.net/2024/Dec/24/...

Carlo Lepelaars - Personal website Hi, nice to meet you! I'm Carlo Lepelaars and I enjoy exploring Artificial Intelligence (AI), Mathematics, Quantum Computing, Investing and Music.

Are you a Python programmer interested in Quantum Computing? This blog post provides a crossover. Using
@howard.fm's "fastcore" library for quantum.

I explain what fastcore is, how its Transform functionality works and dive into the quantum use case. (1/2)

carlo.ai/posts/fastco...

This week we released ModernBERT, the first encoder to reach SOTA on most common benchmarks across language understanding, retrieval, and code, while running twice as fast as DeBERTaV3 on short context and three times faster than NomicBERT & GTE on long context.

Both!

What's our head of T&S doing? @whats-aaron-doing.bsky.social + Follow Aaron liked: BEST GIRLS TIKTOK • Changed Handle •SHE'S GORGEOUS » @tiktokgirls.bsky.social • 14h THER PROFILES go.sendmemessage.bio - #nfsw #adultcontent #fetish #model #onlyfans #petite #xxx #contentcreator #censored #18+ #pov #sexy #tiktokhot #tiktok #tiktok18+

Good morning. Let me check if I’ve got this right.
- Juni created a bot that shows what Aaron (head of trust and safety) likes
- His likes are public information
- Aaron likes a porn post
- Trust and safety ban the bot and creator in 16 minutes
- Creator appeals and ban is upheld

To turn a base model like ModernBERT into a retrieval model you need to fine tune it. We've provided scripts in our repo.

If there were high quality small efficient Enc-Dec models where the decoder on its own worked well, they would indeed be useful to compare to.

That's agreeing, actually -- rip out the decoder, as he says, and now you have an encoder-only model. Like ModernBERT.

And if that's what you wanted, why not just train that directly?

We don't all have infinite resources like Google, so cost matters.

Thank you! :)