prideout

They were speaking from an implementation perspective.

Ah, I see it as a "reasoning level" in codex after typing /model

Confused about how to access Ultra; I don't see it in on their plans page.

Kolmogorov Complexity looks interesting. It seems to formalize Occam’s Razor and the notion that intelligence = compression.

This is a fascinating mathematical framework, but the post title might be a bit of an overreach. I often wonder if "a theory of deep learning" could exist that could be stated succinctly and that could predict (1) scaling laws and (2) the surprising reliability of gradient descent.

Note that I said "predict" not "describe". It feels like we're still in the era of Kepler, not Newton.

Yes! My first encounter with spreadsheets was SuperCalc for CP/M, and I suppose you would call it a "TUI" nowadays.

What does the "Session Picker" section refer to? Claude Code does not have a session picker, as far as I can tell.

Neat. Would be nice if I could easily drop the replayer into a slack thread. Maybe a video file would be better for that than an HTML file?

This is not written by a former Google employee, it's written by a former Google user. Confusing title to see on hackernews.

Why is Minecraft mentioned several times in the post?

This post is well written, useful, and original. I'm surprised by the negative comments.

Neat but I would prefer simply using a dylib for the part of my code that I want to be reloadable.

I have never seen a Koreeda film but he sounds compelling -- which movie would you recommend for a first-timer?

Did you need to write an app for the phone?

There's less ageism in embedded systems because older folks are more likely to be experienced with a lean & mean tech stack (for example vanilla C programming).

Currently we simply copy a slice of the heap's ArrayBuffer from WASM to JS. In the past we exposed the heap slice directly but it was technically "unsafe" (perhaps because the heap can grow), and doing a copy did not hurt performance in any measurable way.

The fact that triangles have proliferated is not due to half-assery. Hardware can rasterize them very quickly, and a triangle can have only one normal vector. Quads can be non-planar. It's true that quads are nice for humans and artists though!

As an aside, Catmull-Clark subdivision has been around since 1978, which, as a first step, breaks an arbitrary polyhedron into a mesh of quadrilaterals.

nm, found a nice open source one:

https://github.com/downforacross/downforacross.com

Neat! I've been getting into crossword puzzles recently, does anyone know of nice places to play them, other than newspaper sites?

I lost interest fairly quickly because the entire article seems to rely on a certain definition of "intelligent" that is not made clear in the beginning.

This article has a superb diagram of the DeepSeek training pipeline.

What model does this use?

Am I the only one who is irked by by ads on Read the Docs pages?

Very nice analysis, wish I could upvote your comment more than once.

Thanks, that helps! I still don't quite understand the mechanics of this, since backprop makes adjustments to steer the LLM towards a specific token sequence, not towards a score produced by a reward function.

Reinforcement learning seems to be key. I understand how traditional fine tuning works for LLMs (i.e. RLHL), but not RL.

It seems one popular method is PPO, but I don't understand at all how to implement that. e.g. is backpropagation still used to adjust weights and biases? Would love to read more from something less opaque than an academic paper.

If you ask it to "Pick 100 numbers from 0-10", it will invoke a Python snippet and give you a much better distribution.

Today I realized "MIT Technology Review" is another compelling print publication.

Wow Asterisk looks exactly like what I'm interested in!

The "Comment" magazine looks nice, nowadays I don't see many appealing printed magazines. I wonder if there are print publications out there on technical subjects that interest me.