Qwen 3.6 27B performs similarly to sonnet 4.5 (note I said 4.5, not 4.6) when it comes to coding. It runs amazingly well on my PC with a 7900xtx.
It's worse at general tasks, but in the precise domain of coding I actually prefer to use it over my claude subscription because it has 0 latency (and no privacy concerns whatsoever).
I believe the direction of UI I was exploring there has more than what graphs currently have, although I didn't have the time to build it out and I saw that the site has been offline for a while.
reMarkable is doing a decent job, their first generation device launched in 2017. Still getting updates. It is discontinued for sale, but there is no reason to believe reMarkable will stop updating their other devices if they're _still_ updating a device they don't even sell anymore.
On top of that, their aftermarket and open source situation is pretty good.
They're not ideal e-readers though, but if you're in the market for a good e-ink device with long-term support and that works well with calibre? Might be worth a look.
I was researching how to predict hallucinations using the literature (fastowski et al, 2025) (cecere et al, 2025) and the general-ish situation is that there are ways to introspect model certainty levels by probing it from the outside to get the same certainty metric that you _would_ have gotten if the model was trained as a bayesian model, ie, it knows what it knows and it knows what it doesn't know.
This significantly improves claim-level false-positive rates (which is measured with the AUARC metric, ie, abstention rates; ie have the model shut up when it is actually uncertain).
This would be great to include as a metric in benchmarks because right now the benchmark just says "it solves x% of benchmarks", whereas the real question real-world developers care about is "it solves x% of benchmarks *reliably*" AND "It creates false positives on y% of the time".
So the answer to your question, we don't know. It might be a cherry picked result, it might be fewer hallucinations (better metacognition) it might be capability to solve more difficult problems (better intelligence).
I wrote a program that has programmable brushes about ten years ago, it's a bit different from moss in that it has a physics simulation underneath rather than a sort of shader, but I've always thought this kind of approach has a lot of potential.
It feels _amazing_ to draw a bird in a single stroke!
If I understand the author correctly, he chose the hyperbolic model specifically because the story of "the singularity" _requires_ a function that hits infinity.
He's looking for a model that works for the story in the media and runs with it.
Your criticism seems to be criticizing the story, not the author's attempt to take it "seriously"
I would love to have a nix flake to install it easily with nix! Given it's built in bash that should be basically no issue whatsoever.
Thanks a lot for this! I was interested in beads but found the author's approach to software development quite erratic and honestly a bit unprofessional. Yes, LLMs are great, but no they shouldn't be the lead developer.
Beads is an incredibly difficult-to-follow mess for something that is at its core a pretty simple idea. You distilled it to its core, I will absolutely be checking this out :)
A new kind of science is one of my favorite books, I read the entirety of the book during a dreadful vacation when I was 19 or 20 on an iPod touch.
It goes much beyond just cellular automata, the thousand pages or so all seem to drive down the same few points:
- "I, Stephen Wolfram, am an unprecedented genius" (not my favorite part of the book)
- Simple rules lead to complexity when iterated upon
- The invention of field of computation is as big and important of an invention as the field of mathematics
The last one is less explicit, but it's what I took away from it. Computation is of course part of mathematics, but it is a kind of "live" mathematics. Executable mathematics.
Super cool book and absolutely worth reading if you're into this kind of thing.
> S-expressions are indisputably harder to learn to read.
Has this been studied? This is a very strong claim to make without any references.
What if you take two groups of software developers, one which has 5-10 years of experience in a popular language of choice, let's say C, and then take a group of people who write LISP professionally (maybe clojure? Common lisp? Academics who work with scheme/racket?) and then have scientists who know how to evaluate cognitive effort measure the difference in reading difficulty.
Isn't the space you're talking about the input images that are close to the textual prompt?
These models are trained on image+text pairs. So if you prompt something like "an apple" you get a conceptual average of all images containing apples. Depending on your dataset, it's likely going to be a photograph of an apple in the center.
https://laura.fm