azakai

6 ms·

31.
by azakai 4mo ago
What about the cognitive capacity of understanding?
32.
by azakai 4mo ago
If you want examples of this, see the recent book "The AI Con" https://www.goodreads.com/en/book/show/217432753-the-ai-con which describes LLMs as "souped-up autocomplete", complex statist
33.
by azakai 5mo ago
There was a lot new in calculus, but it also didn't come out of nowhere. That Newton and Leibniz came up with similar ideas in parallel, independently, around the same time (what are the odds?), supports that. https://en.wik
34.
by azakai 5mo ago
Thanks! I missed that part before.
35.
by azakai 5mo ago
I had the same question. I think that could be answered by using the predicted activation, but I don't see that in the paper. That is, rather than just translate activation to text, then text to activation, that final activation coul
36.
by azakai 6mo ago
The hardware can also add nondeterminism. GPUs reorder operations, leading to different results. Vendors might also be running A/B testing or who knows what, even when you ask for a temperature of 0. But, if you run a fixed model with
37.
by azakai 6mo ago
A carb counting app might use API calls to these frontier models and then do some kind of analysis. It could see if different models agree or not, or multiple calls, and with how much variance. So it would be more accurate to test the apps
38.
by azakai 6mo ago
fwiw, asking the model directly, "who is the ruler of England at present?" returns "Queen Victoria is the reigning sovereign of England."
39.
by azakai 6mo ago
Another way to put it: if training a model cost 72,000 tons of carbon, and it then gets used by 100 million people (typical of major models), the cost per person is 0.00072 tons. Per the article, the average human uses over 5 tons per year
40.
by azakai 6mo ago
It is academically interesting what pure neural networks can do, of course. But when someone goes to Claude and tries to do something, they don't care if it solves the problem using a neural network or a call out to Python. So long as
41.
by azakai 6mo ago
An "elaborate harness" that can break down a problem into sub-tasks, write Python scripts for the ones it can't solve itself, and then combine the results, seems able to solve a wide range of cognitive tasks? At least in theo
42.
by azakai 6mo ago
You are trying it on a production model. The paper is using models with tool calls disabled.
43.
by azakai 6mo ago
It has "outsourced" it to another component, sure, but does that matter? What the user sees is the total behavior of the entire system, not whether the system has internal divisions and separations.
44.
by azakai 6mo ago
I do think this is a tool issue. Here is what the article says: > For the multiplication task, note that agents that make external calls to a calculator tool may have ZEH = ∞. While ZEH = ∞ does have meaning, in this paper we primarily e
45.
by azakai 7mo ago
There are many examples of current limitations, but do you see a reason to think they are fundamental limitations? (I'm not saying they aren't, I'm curious what the evidence is for that.)
46.
by azakai 7mo ago
You're generally right - rewrites let you improve the code - but they do have an actual reason the new language was better: avoiding copies on the boundary. They say they measured that cost, and it was most of the runtime in the old ve
47.
by azakai 7mo ago
O(N²) -> O(N) was 3.3x faster, but before that, eliminating the boundary (replacing wasm with JS) led to speedups of 2.2x, 4.6x, 3.0x (see one table back). It looks like neither is the "real win". both the language and the algo
48.
by azakai 7mo ago
> Figma is one site. There are also a handful of other sites that use wasm. But most of the web does not use wasm. Most of the web also doesn't use the Video element, but it isn't 'a dud' either. Video and wasm are cr
49.
by azakai 7mo ago
Sorry, I'll ask my question in a better way: what applications written in wasm that exist today would benefit from this? Now, maybe there aren't many because of performance - maybe they haven't used wasm because it was too sl
50.
by azakai 7mo ago
> DOM performance is a big deal and bottleneck for a lot of applications What are examples of such applications? Honest question - I'm curious to learn more about issues such applications have in production. > But we really shoul
51.
by azakai 7mo ago
> LLMs do not encode nor encrypt their training data. The fact they can recite training data is a defect not a default. About this specific point, it is unclear how much of a defect memorization actually is - there are also reasons to se
52.
by azakai 7mo ago
> There is no solution without tradeoffs here, but the only reason JS-glue-code is winning out is because the complexity is moved from browsers to each language or framework that wants to work with wasm Correct, but this is has been one
53.
by azakai 7mo ago
> How would a compiler toolchain ship a debugger for webassembly? You can integrate external debuggers, like Uno documents here: https://platform.uno/docs/articles/debugging-wasm.html I assume that uses some br
54.
by azakai 7mo ago
> Wasm should be able to just talk to the browser directly. Web APIs are designed for JavaScript, though, which makes this hard. For example, APIs that receive or return JS Typed Arrays, or objects with flags, etc. - wasm can't oper
55.
by azakai 7mo ago
1. This is a kind of fuzzer. In general it's just great to have many different fuzzers that work in different ways, to get more coverage. 2. I wouldn't say LLMs are "better" than other fuzzers. Someone would need to meas
56.
by azakai 8mo ago
At the very least, computers are still getting faster. Models will get faster and cheaper to run over time, allowing them more time to "think", and we know that helps. Might be slow progress, but it seems inevitable. I do agree th
57.
by azakai 8mo ago
I see now, thanks. Yes, in poker you need more of a mental model of the other side.
58.
by azakai 8mo ago
> AlphaGo or AlphaZero didn’t need to model human cognition. It needed to see the current state and calculate the optimal path better than any human could. I don't think this is right: To calculate the optimal path, you do need to m
59.
by azakai 8mo ago
Plenty of reasons to be skeptical, but also we know that LLMs can find security vulnerabilities since at least 2024: https://projectzero.google/2024/10/from-naptime-to-big-sleep... Some followup findings reported
60.
by azakai 9mo ago
If this interested you, here is another detailed and precise article by a historian, on the same topic: https://acoup.blog/2024/10/25/new-acquisitions-1933-and-the-...

More ›