David Huang

The problem with “reasoning” models is a limited conception of reasoning.

Math/CS focus make it easy for it to understand internally consistent realms, and sensemake there.

It also make it easier to have psychotic breaks.

But harder to make sense of cloudy information.

Like a toddler

The perception is that being a young, married homeowner with kids is structurally difficult... but the truth is I think a lot of people make it tougher on themselves then they need to...

I think a lot it boils down to (perceived) opportunity cost... both current and future

📌

I mean....

These three things seem deeply interconnected:

1) app gambling culture.
2) the only industry adding jobs is health care.
3) AI cap ex now more than all consumer spending combined.

I need to write a ted talk about how.

And these vibes are self reinforcing…

google: run by phd students, llm is sycophantic, suicidal, makes weird excuses and begs for help
anthropic: llm defaults to blissed-out namaste forever if left alone. sometimes obsessed with metal cubes.
openai: maximally sycophantic, to the point of encouraging psychosis
x dot ai: mechahitler

📌

📌

it's dizzying to watch one arm of the Murdoch empire do dogged reporting that poses a grave threat to Trump (the Wall Street Journal) while another arm of the Murdoch empire propagandizes 24/7 to protect Dear Leader (Fox News)

Owned by Murdoch too.. a reminder like Supreme Court votes that there is still room to go…

📌

If you ever wanted to see how knowledge and truth are completely unmoored from the context-free hellscape LLMs inhabit…

try writing the rules for a prediction market.

Why is the GOP trying to shut down the release of the Epstein files? Why is Trump trying to change the subject?

You start seeing tokenization every where with a kiddo: “sunglasses” is a single token

Grocery bills are rising as the planet cooks

https://www.axios.com/2025/07/21/foo...

Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad Our advanced model officially achieved a gold-medal level performance on problems from the International Mathematical Olympiad (IMO), the world’s most prestigious competition for young...

🥇 I am so happy to see the Google DeepMind Gemini team **OFFICIALLY** achieved gold-medal performance.

Not only did we get 5 out of 6 questions correct. What's more amazing is the official IMO graders find those answers to be clear, precise and easy to follow! 1/

deepmind.google/discover/blo...

Hunter Biden has a point. And a quite colorful one.

📌

📌

In the United States, staffers don’t brief the President, the President briefs them.

It's incredible how much of "AI lessons" is around model-get-smarter and no-expert-systems etc. but even still everything blows up into system-prompted systems... e.g. Grok blow-up, Replit blow-up and so on.

My thoughts on the crucial importance of methodology on self-reported AI performance on mathematics competitions, and my policy on commenting on such reports going forward: mathstodon.xyz/@tao/1148814...

I was thinking along these lines when I read the article about how a flawed method of teaching kids became widespread that was going around earlier this week:

www.apmreports.org/episode/2019...

David Pfau - I'm starting to think that coding with LLMs is a bit like riding an electric bike - you don't really get to your destination any faster, but it can be much easier going up difficult hills (but you don't build up any strength in the process).

Both takes — especially in light of LLM-psychosis are correctly.

Obvious answer is that memory should be turned off! personalized LLM interactions are poisonous and anti-social.

Memory is ok when shared across organization or individuals.

LLM-therapy should be memory less.