3 ms·

Extremely hot LLM take: You will often get better results with few-shot prompting (with good examples) on a modern LLM than with a finetuned LLM. Finetuning wa

Extremely hot LLM take: You will often get better results with few-shot prompting (with good examples) on a modern LLM than with a finetuned LLM. Finetuning was the best option for weaker LLMs with lower context windows (e.g. the original GPT-3): both problems have been solved nowadays. The cost economics are much better with few-shot prompting to modern LLMs too: input tokens are super cheap (especially with the recently-released Claude Haiku), so giving a lot of examples per call will still end up cheaper than finetuning. Meanwhile, a finetuned ChatGPT costs 4-6x of normal ChatGPT usage.

viksit 3y ago
can you give a few pointers on articles or examples of this?
- minimaxir 3y ago
  A low-tech example to create a good blog post title for submission to Hacker News would be a system prompt like: You are an expert copywriter. Write five distinct blog post titles optimized for high clickthrough for Hacker News for the article the user provides. Your response must follow the style of these titles: - The ü/ü Conundrum - Why isn't preprint review being adopted? - Majority of web apps could just run on a single server - Weather Planning for Eclipse Day - PSChess – A chess engine in PostScript Then provide the blog post as the user message input. I just ran one of my blog posts (https://news.ycombinator.com/item?id=39495476 https://news.ycombinator.com/item?id=39495476) with the workflow through Claude Haiku and got this: Here are five distinct blog post titles optimized for high clickthrough on Hacker News for the article provided: 1. Tipping ChatGPT: Does Offering Monetary Incentives Improve AI Text Generation? 2. Quantifying the Impact of Incentives on Large Language Model Performance 3. Carrot or Stick? Exploring the Effects of Positive and Negative Prompts on ChatGPT 4. Gamifying AI: Using "Generation Golf" to Test ChatGPT's Ability to Follow Length Constraints 5. The Curious Case of ChatGPT's Motivations: Can an AI Be Incentivized Like Humans? Not bad titles, although more verbose than the 5 input examples I gave. I only gave 5 for simplicity: my main point is that you can give it a lot more than five and/or be more aggressive with constraints, like the blog post linked incidentially.
  - viksit 3y ago
    interesting thank you. intuitively, prompting like this to get an answer seems basically like the first part of a fine tuning process (more exemplars). what is your thought here behind why reinforcing good output via a loss optimization is worse than the one shot example? does the model start to over fit at some point towards some local minima? and this is avoided in this scenario?
    - minimaxir 3y ago
      Prompt engineering in general is necessary because LLMs optimize for the average output, and average output is not good. So LLMs need a slight nudge.
- netdur 3y ago
  Use Gemini 1.5 Pro, which has 1.5 million tokens. Prompt it with a logical question and observe it struggling to answer. Then, upload a book on logical thinking in PDF format and ask the same question again. Notice how it can now answer the question effectively.
thorum 3y ago
That might be true for finetuning ChatGPT 3.5, but if you can finetune a small model (7B or less) to perform on par with GPT-4, while being faster and private, that’s a different story.
- smallnamespace 3y ago
  You definitely can't in the general case (for example, your 7B model is never going to be able to help much with coding, fine tuning or no). It can make sense if you have a particularly simple use case.
  - qeternity 3y ago
    By definition you wouldn’t fine tune a 7B model to be generally as good at GPT4. You would just be trying to overfit some small amount of functionality in a narrow domain.
    - smallnamespace 3y ago
      Yes but from the context of this discussion, we’re trying to figure out the “sweet spot” model size where it’s worth attempting fine tuning. My guess is it’s only worthwhile for matching simple tasks with small models, and any sufficiently complicated task it’s better to do few/zero shot instead.
zapperdulchen 3y ago
Seems like the bitter lesson is still right: http://www.incompleteideas.net/IncIdeas/BitterLesson.html http://www.incompleteideas.net/IncIdeas/BitterLesson.html
- Solvency 3y ago
  Whoa this guy says "computation" and not grammatically bastardized techbrospeak "compute" like some neckbeard equivalent of a caveman! For that alone I commend him.
  - jorvi 3y ago
    Compute is.. I don’t know the exact English grammatical term but it’s like water. Computation is not. “I have 1000 flops of compute” - works. “I have 1000 flops of computation” - doesn’t work. “That compute failed” - doesn’t work. “That computation failed” - works. They’re different.
    - thewakalix 3y ago
      https://en.wikipedia.org/wiki/Mass_noun https://en.wikipedia.org/wiki/Mass_noun
    - Solvency 3y ago
      Literally not true. Compute is a verb. Computation is the right word in all of those cases. Or computational <noun>.
    - xanderlewis 3y ago
      As far as I know, it’s usually called an uncountable noun. …but ‘computation’ is also uncountable, and your second sentence seems to be perfectly fine to me. Your examples do not constitute an argument. You haven’t articulated the (purported) difference between the two words; you’ve just decided arbitrarily that some sentences don’t work, and not elaborated or explained at all. I can make up words too, and provide example sentences: “karrotz are delicious” works. “carrots are delicious” doesn’t. “inside the karrotz” doesn’t work. “inside the carrots” does. I don’t actually think there is any difference. The above comment about ‘brospeak’ was snarky but I do think it’s more of a cultural phenomenon than a semantic one — unless someone is willing to kindly explain the difference rather than just rolling their eyes! What exactly is wrong with the sentence ‘this would require huge amounts of computation’? Saying ‘compute’ seems more to be a synonym of ‘computation’ that’s caught on recently than a useful gap-filling addition to the language. Again: reasoned arguments please. Or just ‘we think it sounds cool so we use it’ — that’s fine, too. EDIT: pondering briefly, perhaps one could argue the difference is something like ‘you can own compute, but you can’t own computation.’ ‘Compute’ is the capacity to carry out computation. …although ‘compute’ seems to be used to refer to the ‘abstract’ computation being done as well as the computational resources, so I don’t know. I’m stretching it. To be honest I’m not sure it’s a useful (or even real) distinction. I think it’s a matter of fashion, and that’s fine and normal.
marviel 3y ago
Several MSFT AI/ML friends actively dissuaded me and my team from fine-tuning. They said that it's pretty clear in all their internal tests that it "lobotomizes" the general reasoning capabilities of the model, unless you're really careful. "All work and no play makes GPT a very dull AI"
- FrustratedMonky 3y ago
  "bitter lesson that building in how we think we think does not work in the long run" Guess. Stop trying to shape the NN. And let it learn on its own.
  - TMWNN 3y ago
    >And let it learn on its own. The best single work of fiction ever created about LLMs' capabilities (and, perhaps, dangers) is Colossus by Jones. Although I think the film is even better than the book, only the latter mentions how, despite being created specifically for US national defense, Colossus is also fed unrelated data including Shakespeare's sonnets, because its creators do not know if it could be important.
    - FrustratedMonky 3y ago
      I think the movie Colossus still holds up today. Saw it last year, It was pretty scary.
- Ambix 3y ago
  Yes, that's what I've seen from a lot of my experiments with fine-tuning. One should be really careful to not "lobotomize" already capable model and achieve better results at the end. It's trickier than seems from multiple of tutorials. But I believe that most of the data stored in foundation models are just useless for some particular domain. So it's better to forget something, getting really useful info instead.
jna_sh 3y ago
“Modern” is an extremely funny delineation given the small temporal window of this whole thing

viksit 3y ago

can you give a few pointers on articles or examples of this?

minimaxir 3y ago
A low-tech example to create a good blog post title for submission to Hacker News would be a system prompt like: You are an expert copywriter. Write five distinct blog post titles optimized for high clickthrough for Hacker News for the article the user provides. Your response must follow the style of these titles: - The ü/ü Conundrum - Why isn't preprint review being adopted? - Majority of web apps could just run on a single server - Weather Planning for Eclipse Day - PSChess – A chess engine in PostScript Then provide the blog post as the user message input. I just ran one of my blog posts (https://news.ycombinator.com/item?id=39495476 https://news.ycombinator.com/item?id=39495476) with the workflow through Claude Haiku and got this: Here are five distinct blog post titles optimized for high clickthrough on Hacker News for the article provided: 1. Tipping ChatGPT: Does Offering Monetary Incentives Improve AI Text Generation? 2. Quantifying the Impact of Incentives on Large Language Model Performance 3. Carrot or Stick? Exploring the Effects of Positive and Negative Prompts on ChatGPT 4. Gamifying AI: Using "Generation Golf" to Test ChatGPT's Ability to Follow Length Constraints 5. The Curious Case of ChatGPT's Motivations: Can an AI Be Incentivized Like Humans? Not bad titles, although more verbose than the 5 input examples I gave. I only gave 5 for simplicity: my main point is that you can give it a lot more than five and/or be more aggressive with constraints, like the blog post linked incidentially.
- viksit 3y ago
  interesting thank you. intuitively, prompting like this to get an answer seems basically like the first part of a fine tuning process (more exemplars). what is your thought here behind why reinforcing good output via a loss optimization is worse than the one shot example? does the model start to over fit at some point towards some local minima? and this is avoided in this scenario?
  - minimaxir 3y ago
    Prompt engineering in general is necessary because LLMs optimize for the average output, and average output is not good. So LLMs need a slight nudge.
netdur 3y ago
Use Gemini 1.5 Pro, which has 1.5 million tokens. Prompt it with a logical question and observe it struggling to answer. Then, upload a book on logical thinking in PDF format and ask the same question again. Notice how it can now answer the question effectively.

thorum 3y ago

That might be true for finetuning ChatGPT 3.5, but if you can finetune a small model (7B or less) to perform on par with GPT-4, while being faster and private, that’s a different story.

smallnamespace 3y ago
You definitely can't in the general case (for example, your 7B model is never going to be able to help much with coding, fine tuning or no). It can make sense if you have a particularly simple use case.
- qeternity 3y ago
  By definition you wouldn’t fine tune a 7B model to be generally as good at GPT4. You would just be trying to overfit some small amount of functionality in a narrow domain.
  - smallnamespace 3y ago
    Yes but from the context of this discussion, we’re trying to figure out the “sweet spot” model size where it’s worth attempting fine tuning. My guess is it’s only worthwhile for matching simple tasks with small models, and any sufficiently complicated task it’s better to do few/zero shot instead.

zapperdulchen 3y ago

Seems like the bitter lesson is still right: http://www.incompleteideas.net/IncIdeas/BitterLesson.html http://www.incompleteideas.net/IncIdeas/BitterLesson.html

Solvency 3y ago
Whoa this guy says "computation" and not grammatically bastardized techbrospeak "compute" like some neckbeard equivalent of a caveman! For that alone I commend him.
- jorvi 3y ago
  Compute is.. I don’t know the exact English grammatical term but it’s like water. Computation is not. “I have 1000 flops of compute” - works. “I have 1000 flops of computation” - doesn’t work. “That compute failed” - doesn’t work. “That computation failed” - works. They’re different.
  - thewakalix 3y ago
    https://en.wikipedia.org/wiki/Mass_noun https://en.wikipedia.org/wiki/Mass_noun
  - Solvency 3y ago
    Literally not true. Compute is a verb. Computation is the right word in all of those cases. Or computational <noun>.
  - xanderlewis 3y ago
    As far as I know, it’s usually called an uncountable noun. …but ‘computation’ is also uncountable, and your second sentence seems to be perfectly fine to me. Your examples do not constitute an argument. You haven’t articulated the (purported) difference between the two words; you’ve just decided arbitrarily that some sentences don’t work, and not elaborated or explained at all. I can make up words too, and provide example sentences: “karrotz are delicious” works. “carrots are delicious” doesn’t. “inside the karrotz” doesn’t work. “inside the carrots” does. I don’t actually think there is any difference. The above comment about ‘brospeak’ was snarky but I do think it’s more of a cultural phenomenon than a semantic one — unless someone is willing to kindly explain the difference rather than just rolling their eyes! What exactly is wrong with the sentence ‘this would require huge amounts of computation’? Saying ‘compute’ seems more to be a synonym of ‘computation’ that’s caught on recently than a useful gap-filling addition to the language. Again: reasoned arguments please. Or just ‘we think it sounds cool so we use it’ — that’s fine, too. EDIT: pondering briefly, perhaps one could argue the difference is something like ‘you can own compute, but you can’t own computation.’ ‘Compute’ is the capacity to carry out computation. …although ‘compute’ seems to be used to refer to the ‘abstract’ computation being done as well as the computational resources, so I don’t know. I’m stretching it. To be honest I’m not sure it’s a useful (or even real) distinction. I think it’s a matter of fashion, and that’s fine and normal.

marviel 3y ago

Several MSFT AI/ML friends actively dissuaded me and my team from fine-tuning. They said that it's pretty clear in all their internal tests that it "lobotomizes" the general reasoning capabilities of the model, unless you're really careful. "All work and no play makes GPT a very dull AI"

FrustratedMonky 3y ago
"bitter lesson that building in how we think we think does not work in the long run" Guess. Stop trying to shape the NN. And let it learn on its own.
- TMWNN 3y ago
  >And let it learn on its own. The best single work of fiction ever created about LLMs' capabilities (and, perhaps, dangers) is Colossus by Jones. Although I think the film is even better than the book, only the latter mentions how, despite being created specifically for US national defense, Colossus is also fed unrelated data including Shakespeare's sonnets, because its creators do not know if it could be important.
  - FrustratedMonky 3y ago
    I think the movie Colossus still holds up today. Saw it last year, It was pretty scary.
Ambix 3y ago
Yes, that's what I've seen from a lot of my experiments with fine-tuning. One should be really careful to not "lobotomize" already capable model and achieve better results at the end. It's trickier than seems from multiple of tutorials. But I believe that most of the data stored in foundation models are just useless for some particular domain. So it's better to forget something, getting really useful info instead.

jna_sh 3y ago

“Modern” is an extremely funny delineation given the small temporal window of this whole thing