News for hackers. Stuff that matters.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by loudmax on Thursday March 16, 2023 @04:10PM (#35187738)
It's frankly amazing how much information is summarized into those weights. You do need to take whatever it says with a very healthy pinch of salt. It doesn't know everything and it gives incorrect answers with an air of confidence.

Here's an example:

  > who preceded emperor Commodus?
  The first emperor of Rome was Nerva (98-100 AD). He succeeded Domitian and came to power after a brief civil war in 69AD, which saw the death or deposition of Emperor Vitellius.
On the one hand, that's just factually wrong. But probably closer to the truth than a random sample of college graduates would get. This is truly impressive.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by pr337h4m on Thursday March 16, 2023 @04:21PM (#35187890)
The thesis behind the Hutter Prize was spot-on, except for focusing on lossless compression instead of lossy compression.

http://prize.hutter1.net/

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by londons_explore on Thursday March 16, 2023 @04:29PM (#35188032)
There is a mathematical mapping between lossy and lossless compression. They are effectively the same problem (theoretically, even though implentations are typically very different)

LLM's would make excellent (but very slow) lossy or lossless compressors.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by muyyatin2 on Thursday March 16, 2023 @05:09PM (#35188680)
Do you have some links/references for someone wanting to learn more about this?
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by optimalsolver on Thursday March 16, 2023 @06:51PM (#35190027)
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by telotortium on Thursday March 16, 2023 @07:05PM (#35190193)
Here's a simple explanation: https://news.ycombinator.com/item?id=34726719. Basically, with a sufficiently good lossy compressor, you can use entropy coding to encode the difference between the decompression output and the original file.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by rm999 on Thursday March 16, 2023 @06:25PM (#35189704)
That brings me back! I just dug up an email I sent Matt Mahoney in August 2006. I still stand by my stance (with some caveats) - we're starting to see emergent intelligence coming out of these networks, and it really did come from brute forcing parameter capacity (and data, and innovation on ML algorithms) - smaller networks didn't make for smarter models. My guess is the algorithms that would win the contest likely cannot process and compress new text nearly as well as the GPT style algorithms because that is not what they were optimized for.

> I read your page "http://www.cs.fit.edu/~mmahoney/compression/rationale.html" and found it very interesting. As an AI grad student, I never really thought of data compression as equivalent to AI.

>After reading your page, I have a problem with the wikipedia contest. The contest is trying to achieve something "intelligent," but penalizes large decompressors. If the decompressor can be thought of as the "brain" that we are trying to make intelligent, this is penalizing a large brain. The human brain, the pinnacle of intelligence as we know it, evolved to contain a huge amount of memory. The real-world never penalized better human brains during the brain's evolution. You have a larger brain? Great, you're smarter, and will probably fare better!

>That is my philosophical argument. My practical argument is using the size of the decompressor does not scale with the contest. If the contest was on 10 GB, the size of the decompressor would suddenly matter 10x less (assuming compression size is linear to the input size). Isn't that somewhat arbitrary? I entirely understand why the size of the decompressor cannot be ignored, but in its current state I think the limitation is drastically hurting the chances of anything moderately "intelligent." Adding the size of the decompressor into the rules of the contest encourages hacks instead of something more cognitively intelligent-like.

>Perhaps a more fair rule would be to not penalize decompressors up to a certain point. For example, anything less than 200 MB is free. I think this would be a more fair compromise, because IMO truly intelligent artifical agents, whenever they arise, will have to use large amount of memory.

>I would appreciate to hear what you think of this analysis.

The reply I got:

> Good compressors will use lots of memory, but this is not the same as requiring a large decompressor executable. The program will typically allocate much more memory than its size. There are practical limits on memory like 1-2 GB but that is another matter. However we must include the decompressor size or else it is easy to write a decompressor to keep a copy of the data and compress to 0. This is easy even with a 200 MB limit, even for enwik9 by keeping a copy internally compressed with ppmd or such.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by modeless on Thursday March 16, 2023 @07:15PM (#35190297)
Lossless is actually the appropriate test. These models output probabilities and can easily be turned into very effective lossless compressors with the use of arithmetic encoding. For a very interesting discussion see: https://www.youtube.com/watch?v=dO4TPJkeaaU

The real problem with the Hutter prize is that the data and compute are too small. It seems obvious now that the regime where compression turns into AGI is far beyond the scale of the current contest (1 GB data, 50 CPU hours). If you simply bumped those limits up five or six orders of magnitude then the Hutter prize could be relevant today. But you'd also have to increase the size of the prize a few orders of magnitude because the cost of competing would be more than the current prize money.

Of course a prize isn't really a necessary incentive to develop AGI, since getting an AGI is about the best incentive you could possibly imagine.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by optimalsolver on Thursday March 16, 2023 @07:49PM (#35190633)
>Lossless is actually the appropriate test

Could you elaborate on why that is. It seems very non-obvious to me.

In the realm of vision, after a certain point you would just be trying to compress noise that has nothing to do with the content of image/video.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by modeless on Thursday March 16, 2023 @07:57PM (#35190703)
That's OK. The purpose of the test is to compare approaches to compression. Noise in the dataset is a handicap but it affects all approaches, so it doesn't invalidate the test as a way to compare approaches. And true white noise is very rare, almost all real noise has characteristics that are at least somewhat compressible. And hopefully whoever constructs the dataset tries not to include tons of noise in it.

Lossless compression is a great objective because it's impossible to cheat. When you do lossy compression you have to define a quality metric and as soon as you do that the game becomes cheating the quality metric rather than actually compressing the data in a useful way.

I highly recommend watching the video I linked. Arithmetic encoding reduces the task of lossless compression to assigning probabilities for the next token, which is the exact task these models are trained to do.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by inimino on Friday March 17, 2023 @07:15AM (#35195504)
Answered in detail in the FAQ:

http://prize.hutter1.net/hfaq.htm#lossless

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by radarsat1 on Thursday March 16, 2023 @04:43PM (#35188257)
I think a really interesting area of further research into LLMs is going to be how to further decouple "intelligence" from "knowledge". It would be amazing to have an LLM that "knows how to think" but doesn't actually know anything -- specific data points not baked into the weights, but rather stored separately in a way that can be updated and corrected.

Perhaps a step forward is the "toolformer" router, making the LLM recognize when it is referencing a fact and have it look that up (use a "tool") instead of trying to infer it. I realize that even semantically this is kind of a fine line, hard facts vs inferences, so there's some philosophy to ponder here. What can be 'stored' and what can be 'calculated'. What kind of knowledge is innate to intelligent reasoning and what kind acts more like a lookup table.

But, exactly because we can't easily define these two categories, I have a feeling that if LLMs are trained to actually learn by themselves how to distinguish them in a "fuzzy" way, we might start getting somewhere.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by Y_Y on Thursday March 16, 2023 @05:06PM (#35188634)
If you want something that "knows how to think" then I don't think you'll find it in an LLM. You can't divorce knowledge from this kind of model, it guesses what word comes next based on its training on a corpus of knowledge.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by georgelyon on Thursday March 16, 2023 @05:19PM (#35188817)
No, but you could build something that guesses the next step in a list of steps to access knowledge from an external source.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by gamegoblin on Thursday March 16, 2023 @08:50PM (#35191217)
One could imagine training a much smaller LLM on synthetic data that didn't contain much or any real knowledge, but just examples of in-context reasoning.

E.g. the training data might look like "a fooble is to a tooble as a fabble is to a" with the answer "tabble".

So you feed it tons of these kind of nonsense training data that forces it to only learn the in-context reasoning part of language, and none of the world knowledge.

That said, it is unclear to me how much value such a model would be.

You could imagine a middle ground where it does have basic knowledge, such as dictionary definitions and things derivable from those.

E.g. "A tree has leaves. Leaves are green. Therefore a tree has a part that is green." type stuff.

So you could give it some amount of world-grounding and common sense knowledge, but nothing involving history, proper nouns, etc.

I imagine you could make such a model much smaller than these giant LLMs.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by chii on Thursday March 16, 2023 @09:25PM (#35191549)
The reason, i suspect, that this method won't work, is because the training dataset has knowledge embedded in it, regardless of whether you think it only contains in-context logic.

The evidence is that there are these "glitched" tokens : https://www.youtube.com/watch?v=WO2X3oZEJOA

if you watched to the end of that video, it explores a very plausible reason why such glitched tokens exist - it is because the training dataset has a bunch of "nonsense" (a large dump of the subreddit that just counts a number, as well as various other "data").

So my hypothesis is that if you attempted the in-context logic training dataset, it would produce large number of glitched tokens (different ones, but equally as "illogical" to the user).

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by gamegoblin on Friday March 17, 2023 @12:47AM (#35193259)
Tokenization is purely an implementation detail. If OpenAI had cared, they could have deleted those obviously glitched tokens from their tokenizer. They just didn't inspect it carefully and/or care.

GPT4 does not suffer from the same glitched tokens as GPT3, presumably because it uses a different tokenizer.

Furthermore, there are LLMs that operate on single bytes instead of multi-character tokens, totally obviating that problem.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by pyinstallwoes on Saturday March 18, 2023 @03:40PM (#35212390)
You can separate that from which someone can recite from that which someone can perform in practice or do. To remember facts and present them in language is different than being able to craft a bow.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by zamnos on Thursday March 16, 2023 @05:52PM (#35189270)
Bing Chat, powered by GPT-4, looks stuff up on the Internet. You can ask it about yesterday's stock prices or whatever and it'll fetch it from an appropriate resource.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by Tostino on Thursday March 16, 2023 @06:52PM (#35190049)
The noise on the search index seems like it introduces quite a few issues with accuracy. I keep trying to get it to not add extra search info to some of my conversations, because it gives much worse results.

Blog-spam really sucks.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by telotortium on Thursday March 16, 2023 @07:00PM (#35190145)
Bing Chat is doing something like that, although it was far from perfect. Langchain can also do this to some extent. Then Steve Hsu claims to have been able to get a model working very well that can answer questions from college textbooks (https://twitter.com/hsu_steve/status/1636337165775908864). So it seems like someone will figure it out in a few years.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by up2isomorphism on Thursday March 16, 2023 @05:21PM (#35188854)
“Closed to truth” is actually the worst property of LLM models. Because it hard for human to detect if it is wrong or not.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by Fomite on Thursday March 16, 2023 @05:39PM (#35189085)
Turns humans from writers into fact checkers, which it turns out is a harder skill.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by up2isomorphism on Thursday March 16, 2023 @05:44PM (#35189147)
It is actually much harder skill. And what worse is that any successful check makes the future checks harder. But there might be no significant improvement on the quality, just the difficulty of detecting the fallacy.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by zamnos on Thursday March 16, 2023 @06:03PM (#35189413)
What sort of facts are you thinking of? There are plenty of surface level facts that another model or a human should easily be able to verify as true or not. Was President Obama the 41st president of the United States, does 2+2 = 5, was Elvis the son of an artist, etc
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by Fomite on Friday March 17, 2023 @12:28AM (#35193127)
In my sort of informal "How would I like to use this?" testing, this ranges from easily verifiable facts (i.e. "Did you get where I got my degree right?") to structural misunderstandings of important concepts.

The former would be easy enough to fix - though it undermines the utility and if not caught could carry ramifications for people, but the latter genuinely concerns me.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by chii on Thursday March 16, 2023 @09:27PM (#35191576)
But if you create an adversarial ai to do fact checking - aka, a reward function where if a fact is found on the internet that contradicts the given GPT response, then would it work?
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by Fomite on Thursday March 16, 2023 @05:38PM (#35189075)
This is one of the things that LLMs seem to be struggling with. They're very good at tone, and feeling right without actually being write.

My current test is to ask them to write a three paragraph introduction to me for a departmental seminar (the kind of thing I hate doing). Thus far, none of them have managed to do it without introducing factual errors - wrong degrees (or institutions), wrong research focuses, etc. But they're all credible sounding in their wrongness.

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by bleuchase on Thursday March 16, 2023 @06:49PM (#35190010)
> They're very good at tone, and feeling right without actually being write.

The text they generate is probably free of errors like this one as well

Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by Fomite on Friday March 17, 2023 @12:31AM (#35193143)
Probably.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by bleuchase on Friday March 17, 2023 @04:11AM (#35194494)
I’m sure someone is working on adding support for human-like errors in LLM outputs.
Re:Show HN: Alpaca.cpp – Run an Instruction-Tuned Chat-Style... by crucialfelix on Friday March 17, 2023 @06:19AM (#35195181)
Write a crappy intro for yourself with all the facts included.

Ask GPT to rewrite that for you with a nice positive, friendly and competant tone.

It's really good at that.