LLMs become much less useful the moment you dismantle bureaucracy.

The modern internet penalises anyone who thinks for more than 3 seconds before forming a strong opinion. You’ll be happy to know that I only thought for 2 seconds before typing this.

The idea that scaling up LLMs on human data will produce superhuman performance is magical thinking. The highest attainable performance in any given domain is simply the best human performance. Yes, maybe some insights span across domains, but I don’t think there’s actually evidence for that.

The people who laugh about hacky “vibe coding” with LLMs are the same people who think that grok 3 is better than a doctor. Absolutely nuts.

“Below is the updated code” followed by absolutely no code is such an o3-mini thing to do that at this point I don’t really understand why this model exists.

After much time spent looking at reasoning traces from DeepSeek R1 for medical cases, I have to conclude that there isn’t a strong correlation between good reasoning and a good answer.

These R1 distilled models are absolutely amazing on a single turn, but truly horrible on multi-turn conversations.

Meta’s new response to the Rohingya genocide.

"Mildly elevated rheumatoid factor has a very low positive predictive value that is completely overwhelmed in magnitude by the negative predictive value of not having any signs or symptoms of rheumatoid arthritis."

My 2 year old has 3 adjectives for the size of things. In increasing order: small, mummy, big.

👀👀👀👀 huggingface.co/blog/modernb...

Releasing Jupyter Agents - LLMs running data analysis directly in a notebook! The agent can load data, execute code, plot results and following your guidance and ideas! A very natural way to collaborate with an LLM over data and it's just scratching the surface of what's possible soon!

What percentage of “rhupus” is just misdiagnosed Sjögren?

o1 is equal parts brilliant and boring. Very, very boring.

Llamafile is a cheat code.

"Zuckerberg's eyes brimmed with tears, and his heart felt full. He truly loved Big Brother!"

Sora’s idea of a hand exam. This 60 year-old rheumatologist is giving me very 2nd year medical student vibes with this bizarre technique. No synovitis was detected this day. #rheumsky

So much delving

Four-year follow-up of patients with hand osteoarthritis found that the course of pain was variable. Changes in pain were associated with BMI, employment status, mental wellbeing, illness perceptions, and coping styles In Arthritis Care & Research doi.org/10.1002/acr....

Vlado Chernozemski, who assassinated King Alexander I of Yugoslavia, belonged in his youth to a club that was devoted to a game of regicide. In the game, known as "chess," contestants subtly maneuver into a position that allows them to topple the opponent's monarch.

For five years, I have been covering what is unquestionably the biggest criminal justice scandal in American history. Today it ended when the most corrupt cop in history took his life. But it's not over. A thread.

Are we the baddies?

I love the fact that Wikipedia’s awful contributor guidelines have lead them to work on algorithms that force contributors to adhere to the guidelines, rather than just making better guidelines www.mediawiki.org/wiki/Edit_ch...

Finally have COVID. 5 year streak comes to an end 😢

I think anthropic and pretty much anyone can beat o1 with better training data. I give o1 a complex problem, o1 thinks for 4 seconds, outputs a bad answer. Seriously, who trained this thing?

We need some normies here, too many word smiths, phrasing police, you can’t use that word posters, ie annoying left people

Reports of Australian cricket’s death are greatly exaggerated*

This is actually good to show skeptics of structured generation @dottxtai.bsky.social. Demonstrates very clearly that you can stop the model from fighting the constraints by telling it how it’s constrained, even if those constraints might seem absurd (like only using words from the King James Bible)

There's often a lot of confusion about what structured generation actually is. Brief thread. Structured generation, also called structured output or constrained decoding, is always correctly formatted. The model literally cannot fail to output the format you specify.

i wrote a custom llm sampler for llama-3.1-8b so it could only say words that are in the bible github.com/vgel/biblica...

The best feed (and it’s not even close) is this one. The signal:noise ratio is phenomenal bsky.app/profile/did:...

Liking a product without liking the person/company that makes it is extremely normal. I am dubious of Meta, but I like the Llama models. I don’t like Elon, but I like jokes about his weird, obsessive fans.

#ACR24 people would anyone be interested in a Chrome extension that adds the top 5 most similar abstracts at the end of the page.

When we prompt a VLM with an image and text, we don’t know to what extent the response depends on the image vs text. These are clear illustrations of this concept. This is important to consider when looking at benchmarks for these models – quite often the answer comes simply from the text.

My 2-year-old and I are debating: brown bear or wombat?

clearsky.app “blocked by” is so much fun. Very pleased to see that I’m blocked by @trashmuppet.bsky.social. Only good muppets on my timeline, thank you very much.

I think people would like AI more if our most aesthetic poster (roon) didn’t sound like every precocious 16 year-old who has just read Nietzsche and vaguely heard of Heidegger.

The first rule of using the bluesky API: don’t talk about using the bluesky API.

Rheumatology-adapted version of Llama 3.1 8B, running on my MacBook at 20 tok/s, correctly answering a question that no open weights models other than Llama 405B can answer. No instruction tuning involved, lots of synthetic data.

This conforms with my anti-paracetamol/acetaminophen bias.

Repeated acetaminophen use among older adults in the UK was associated with significant incidence of renal, cardiovascular, and GI side effects Results suggest recommendations for repeated use for long-term conditions such as osteoarthritis should be reconsidered In AC&R acr.tw/41lP46n #Medsky

Does anyone actually use this and, if so, why? cloud.google.com/blog/topics/...

Optimist: The cup is half full. Pessimist: The cup is half empty. Rheumatologist: Do you find it difficult to use a cup in the morning?

Optimist: The cup is half full. Pessimist: The cup is half empty. AI:

AI generated gaits are very odd.

X is insufficiently academic and this place is too academic. Where’s the place for people who just want to build things, but are not the kind of people who post “you can just build things”?

Trying to make a starter pack of people and organisations on here who produce content and tools for practical AI use. All suggestions welcome. go.bsky.app/Vr73Ggw (also, how do I remove myself from a starter pack? 😂)

How many GPUs to fry this thing?