cjbprime

36 ms·

481.
by cjbprime 4y ago
> LLMs do not "reason"; they do not "learn" or "develop" anything of their own volition. If this claim of yours was sustainable, your grievance with the writing style would make a lot of sense. But GPT-4 can
482.
by cjbprime 4y ago
> Gaining capabilities unpredictably is a bit exaggerated - it's more that a lot of apparent capabilities are embedded in language and these models approximate the truly vast amounts of text they've digested (imo) I think the a
483.
by cjbprime 4y ago
(I think the project we're talking about does CPU runtime inference, not GPU.)
484.
by cjbprime 4y ago
So a single-cell protozoa is more intelligent than GPT-4, then. This definition doesn't seem to be very useful?
485.
by cjbprime 4y ago
I don't expect that language is required. Using input frames of video (or any other complex sensory input) instead of words might work too. Transformers are a tool for building representations from high-dimensional data sources. Also,
486.
by cjbprime 4y ago
AGI doesn't have to imply consciousness. It can just mean superhuman performance at tasks, as GPT-4 has perhaps already reached. As AGI architectures go, LLMs seem pretty good for this worry about consciousness. They were already not d
487.
by cjbprime 4y ago
> Leading researchers are scratching their heads trying to figure out the model. Don't think so? The base model seems like a straightforward improvement over previous LLMs. They trained it with more parameters (and therefore expense
488.
by cjbprime 4y ago
Simple arithmetic with five digit numbers.
489.
by cjbprime 4y ago
Bing is a free way to get GPT-4. It is much much more impressive, and definitely not a parrot.
490.
by cjbprime 4y ago
Are you using GPT-4? (Requires ChatGPT Plus subscription.)
491.
by cjbprime 4y ago
> A New open-source Flan 20B with UL2 I think I know how all the non-programmers who read HN feel now.
492.
by cjbprime 4y ago
> From the article GPT-4 gave the likelihood of a bank run as a 1 out of 5. .. what probability would you have assigned to a SVB bank run in 2021?
493.
by cjbprime 4y ago
> It's also the case that while GPT-4 can say a bear at the north pole is "white", it has no way of knowing what "white", or "bear", or "north" actually represent. This is a preposterous claim
494.
by cjbprime 4y ago
> Yes, GPT-4 is not an AGI yet, but the research paper (OP) has a point. How did you go from "human-level IQ with some super-human abilities" to "not an AGI"?
495.
by cjbprime 4y ago
There's a 150-page paper solely to describe instances of it doing that. It's the article attached to this comment thread.
496.
by cjbprime 4y ago
I wonder if the same reaction would have happened if ChatGPT had waited and released with GPT-4. It's very different.
497.
by cjbprime 4y ago
Agree: compression is straightforwardly prediction. Prediction is intelligence when you're able to do it well across disparate novel tasks. (Of course definitions are idiosyncratic, but I'm curious if anyone disagrees with these o
498.
by cjbprime 4y ago
I don't think that's a dumb question! I don't know if this project has an answer to that, but there are some techniques (Merkle tree hashes etc) that might work depending on how much of the model you want to download locally.
499.
by cjbprime 4y ago
From the article: > You could also say that GPT-4 was playing the game blindfolded as it had no access to refreshing its memory of the current state of the board. That doesn't sound right. Under the hood, I think everything inside t
500.
by cjbprime 4y ago
Looks like it does this to me: for file in `git ls-files`; do echo $file; cat $file; echo; echo -------; done
501.
by cjbprime 4y ago
Thanks, do you recall how to do e.g. a full TCP payload text search across all packets? Didn't find it with a quick search.
502.
by cjbprime 4y ago
This looks good! What do people use for fast indexed search of pcaps? (Contents, not metadata.)
503.
by cjbprime 4y ago
Surely it does help SVB depositors, since without this action they would (a) take a haircut, since we expect SVB has less assets than deposits, and (b) have to wait some months or years for those remaining assets to be liquidated and distri
504.
by cjbprime 4y ago
In the other professions which use the title "engineer", all of these are commonplace.
505.
by cjbprime 4y ago
Best future proof might be to wait two months and get an M2 Mac Pro.
506.
by cjbprime 4y ago
In case folks miss the link at the top of the article, this is translated from an old 2017 post by Aphyr. That post was in Haskell, where it's not too surprising that you can do serious computation inside the type system. This new post
507.
by cjbprime 4y ago
Sounds like this commenter is not just arguing that language is being used incorrectly in this situation, but that the standard language and definitions are themselves propaganda terms. (Usually people are a little clearer about their inten
508.
by cjbprime 4y ago
It is straightforward for this code, which is CPU-only.
509.
by cjbprime 4y ago
I was wondering if Apple Silicon would be uniquely suited for high-GPU-RAM tasks because it shares memory across the system. But I guess in this case it's a CPU model, so that's unrelated. Is that right? Do you think you could run
510.
by cjbprime 4y ago
The latest Intel/AMD non-HEDT memory controllers only have two channels, so your DDR5 would drop to below-DDR4 speeds if you put 128GB DDR5 in it. (32GB dual rank sticks * 4 -- there are no 64GB DDR5 sticks yet. Some reports of system

More ›