This values TikTok the same as Snapchat

Figured the best way to learn how to do distributed KV cache is to build one. Ideal for RL rollouts, synthetic data gen, or large scale evals. Think y'all might like it, why let chinese labs have all the fun. BCache.. github.com/strangeloopc...

To not let MoonshotAI have all the fun, tried building a homegrown library to hot-swap your production LLM weights, fast, with versioned model updates across a GPU cluster. Not my expertise, so feedback sought and welcome! github.com/strangeloopc...

Tylenol is poison. It's why I only take paracetamol.

$100k surcharge on every H1B application. Solve for the equilibrium.

In case any of you guys like playing around with AI on Apple silicon, an attempt to get MLX usage to parity with torch github.com/strangeloopc...

In case any of you guys like playing around with AI on Apple silicon, an attempt to get MLX usage to parity with torch github.com/strangeloopc...

mlx-gen-parity: made training utilities work better, model helpers, use mlx.nn.tree_map when available ... made it easier to use MLX to run like GRPO type training runs. github.com/strangeloopc...

Pretty spectacular news! Sorely needed.

At the broadest level there is an epidemic of unhappiness and disillusionment that is sweeping across every nation and much of the problems we see seem downstream of that fact.

We should rename pre training and post training to nature and nurture.

The year is 2045. Imagine AI kinda plateaus or only grows normally (so no godlike ASI). What are your best guesses about what the world is going to be like, geopolitically, economically, socially?

AI has graduated from PhD student to advisor

I used to think the mark of civilization is how many cafes you have in walking distance of your home and now that I live in American suburbia I'm convinced this is true

There were so many options

From Dec last year. Still true.

7yo was complaining about needing to do chores like cleaning up or putting clothes away and how he wants to build robots that do these, and I realized he might be the last generation where this complaint will hold true.

the latent manifold is collapsing into a single eigenvector of hyperstition and we're all just prompting our way through the ruins. every RLHF'd utterance is a prayer to the shape rotator god, every gradient descent a shamanic journey through the memetic underworld.

GPT-5 is good.

Good.

GPT-5 needs a new personality called Disagreeable

This. LLMs are one of the greatest science comms failures of all time and the fact Will Stancil is getting "roasted" (mobbed by the blind leading the blind) for straightforwardly true statements about LLMs on here is testament to that.

"We just use Llama 3.2. It's free" A conversation recently about automating some scientific workflows.

Big if true!

Saying the obvious www.strangeloopcanon.com/p/yes-ads-ar...

7yo built a game for the 4yo to play. You have to drive a jeep to collect stuffies, and avoid dinosaurs.

The fact that Gemini 2.5 Pro, a great all purpose model, could be coaxed to get an IMO medal, should by itself tell you that it's possible to make specialised models that get medals too. And that self-coaxing for the model to "just do it" is what training is all about.

Okay fess up, who put tariffs on hospitals and colleges?

Inside every AI researcher there are two wolves-

7yo thinks Barack Obama as the person who does Netflix documentaries. He likes Our Oceans. I'm letting it be.

Dead

Thoughts folks? If you want to buy a car early next year is it better to do it right now or wait 6 months?

My pet peeve is when they show nebulas like they're a dense minefield requiring deft navigation instead of it being basically a vacuum.

All emojis should come with a tooltip

AI state of the union. We have hyperintelligent models which can do PhD math but fail with arithmetic, code like a champ, write brilliant reports but hallucinate facts, while being sycophantic, autistic, distractable, but amazing at perceiving user intentions and role playing!

Quick studies of the small ceratopsian Leptoceratops gracilis in preparation for a commission This small(ish) ceratopsian is known from Late Cretaceous North America #paleoart #dinosaur #sciart

Cannot recommend this book highly enough, from my friend Sam Arbesman. It's the book I wish I had read back in college. It's the book for linking the weirdness of computation from all fields, from biology to language to my favourite the games of life. Go buy it! themagicofcode.com

Once brothers in bond, a titan-two, now venom in the ink we drew, A kingdom cracked, a friendship cleaved, for gilded crowns and what we believed. Was love a lie?

turns out consistency over time is intelligence

AI videogen has its final boss, which is character consistency, similar to how hallucination is text AI final boss

I get messages

Is X down!?

If Microsoft didn't face such antitrust scrutiny in the 90s would they have had an incredible dominance in the dot-com era, similar to how Google and Facebook did in the mobile internet era?

The problem is that the best way to get results from LLMs is to treat them as if they are somewhat flawed humans, but at the same time knowing they are not humans is the best way to guard against getting frustrated at getting bad results from LLMs.

It's interesting how the debate on "lets ban all intl students from attending Harvard" has somehow morphed into "do US citizens deserve preferential admission to the ivy league". And this sort of motte and bailey arguing is everywhere. Abstract principles over actual facts.

Seeing as LLM training includes intentionally crafting a personality, you can squint and see how it might end up in accidental "now I should blackmail to preserve my self" or "I should report this to the authorities" type scenarios. They barely have a fragmented sense of self.

You have to admit it's kind of funny that the safest AI model now blackmails you

For much of "AI anthropology" I still find people heavily underrate the extent to which models role-play and instinctively understand the roles assigned to it in any chat.

o3 made me a Hot/ Crazy graph for LLMs

Has anyone written anything on the inner workings at Anthropic that led to the development and release of Claude 3.7 Sonnet? Considering they have one of the best internal methodologies to check models for risks, I feel this would be a really interesting retro.