sacred_numbers

3 jaar geledendiscuss

GPT-4 is not the same product. I know it seems like it due to the way they position 3.5 and 4 on the same page, but they are really quite separate things. When I signed up for ChatGPT plus I didn't even bother using 3.5 because I knew it would be inferior. I still have only used it a handful of times. GPT-4 is just so much farther ahead that using 3.5 is just a waste of time.

sacred_numbers

3 jaar geledendiscuss

Based on my research, GPT-3.5 is likely significantly smaller than 70B parameters, so it would make sense that it's cheaper to run. My guess is that OpenAI significantly overtrained GPT-3.5 to get as small a model as possible to optimize for inference. Also, Nvidia chips are way more efficient at inference than M1 Max. OpenAI also has the advantage of batching API calls which leads to better hardware utilization. I don't have definitive proof that they're not dumping, but economies of scale and optimization seem like better explanations to me.

sacred_numbers

3 jaar geledendiscuss

The reason cement is a major contributor to CO2 emissions is because of how much cement we produce. I don't know the lifetime or effectiveness of this catalyst, but typically you only need a tiny amount of catalyst to start a reaction and the catalyst material can be used over and over for a long time.

sacred_numbers

3 jaar geledendiscuss

Yes, with 4 bit quantization.

sacred_numbers

3 jaar geledendiscuss

We could not do that inadvertently. To block even 1% of light using Starlink sized satellites (~30 m^2 with solar panels deployed) would require tens of billions of satellites. We could do it on purpose with huge rotating solar reflectors, and it should honestly be considered as a real option, but we couldn't and wouldn't do so just by launching communication satellites.

sacred_numbers

3 jaar geledendiscuss

The quality difference is substantial. I don't care if it's wasteful to use something that has many uses for a supposedly narrow task (although I don't see translation as a particularly narrow task anymore than I see writing as a narrow task). I would gladly waste untold trillions of floating point operations for a 1% increase in translation quality. From my experiments, though, it's much higher than 1% increase in translation quality. And regardless of how wasteful the compute is, it's actually cheaper in terms of dollars. Using GPT-3.5 to translate Korean to English would cost about $11 per million words, based on the average characters per token of the small sample of text I gave it. DeepL (the best translation service I could find) costs $25 per million characters, or for my sample text, about $64 per million words. At $11 per million words I can have GPT-3.5 perform multiple translation passes and use it's own judgment to pick the best translation and STILL save money compared to DeepL.

sacred_numbers

3 jaar geledendiscuss

It's worse on English and a lot of other common languages (see Appendix C of the paper). It does better on less common languages like Latvian or Tajik, though.

sacred_numbers

3 jaar geledendiscuss

There will always be overhead, but that doesn't mean it will always be a huge amount of overhead. I believe the state of the art is 97% efficiency (https://www.osti.gov/biblio/1495980) which is better than a lot of wired chargers. Real world systems will be less efficient, and it may be too expensive, but a maglev system would be even more expensive.

sacred_numbers

3 jaar geledendiscuss

I did my own calculations based on plotting loss on benchmarks compared to models with known parameters and training data, as well as using a quote from Sam Altman that said that GPT-4 would not use very many more parameters than GPT-3. Based on this, I estimated that GPT-4 probably used about 250B parameters, and since I had an estimate for the total compute I was able to estimate that the training data was about 15T tokens. 250B parameters times 15T tokens times 6 (https://medium.com/@dzmitrybahdanau/the-flops-calculus-of-la...) means the compute was about 2.2510^25 FLOPs. I estimated that A100s cost about $1/hr and can process about 5.410^17 FLOPs at 50% efficiency per hour. Therefore, the compute cost would be (2.2510^25)/(5.410^17) or about $40 million.

Interestingly, my own calculations lined up pretty well with this calculation, although they approached the problem from a different direction (a leak by Morgan Stanley about how many GPUs OpenAI used to train GPT-4 as well as an estimate of how long it was trained): https://colab.research.google.com/drive/1O99z9b1I5O66bT78r9S...

Sam Altman has also stated that GPT-4 cost more than $100 million to train, and replication can cost 2-4x less compute. https://www.wired.com/story/openai-ceo-sam-altman-the-age-of...

If you know of an organization that can replicate GPT-4 for $400k to $4m I would love to know so that I can invest in them.

sacred_numbers

3 jaar geledendiscuss

I would bet money against that. Replicating GPT-4 pre-training with current hardware would cost about 40-50m in compute. Compute will continue to decrease in cost and algorithmic improvements may allow for more efficient training, but probably not 3 orders of magnitude in a few years. I think there will be plenty of open source models that will claim GPT-4 quality, and some of them will be close, but they will be models that used millions of dollars (probably from some corporate benefactor but possibly from crowdsourcing) in compute to train. You will probably be able to fine-tune and run inference on fairly cheap hardware, but you can't cheat scale. It's going to take a major innovation to move away from the expensive base model paradigm.

sacred_numbers

3 jaar geledendiscuss

They do update the model in the background, although I'm not sure how often or how much they update it. To avoid issues with this practice they offer gpt-4-0314 which says this in the documentation:

"Snapshot of gpt-4 from March 14th 2023. Unlike gpt-4, this model will not receive updates, and will only be supported for a three month period ending on June 14th 2023."

Unfortunately this experiment is using the frozen snapshot model gpt-4-0314 instead of the unfrozen gpt-4 or gpt-4-32k models, so any differences are literally 100% noise. This would be a somewhat interesting experiment if someone were to use an unfrozen model, though. I do appreciate the author for captioning the images with the exact model they used for generation so that this bug could be caught quickly.

[0]https://platform.openai.com/docs/models/gpt-4

sacred_numbers

3 jaar geledendiscuss

It's unlikely that OSS LLMs will ever be able to compete with corporate LLMs. I can only think of a few scenarios where this could work:

1. Someone develops a procedure for training models with distributed computing resources, including consumer CPUs and GPUs. Even this is not really guaranteed to work, since corporations will probably just buy up all the consumer GPUs, since they are cheaper on a FLOPS/$ metric.

2. One or more governments provide a lot of funding for OSS models, including being willing to pay competitive salaries for the best talent (potentially millions of dollars per year). This is unlikely for a lot of reasons. The only thing that could speed up the process enough to compete with private organizations is a major war that required AI to win. In that case, though, open source would be the least of their concern.

3. Scaling laws stop working and Moore's law catches up. In that case adding more compute won't really help and eventually even organizations with small budgets can afford to train a SOTA LLM. We can't know until we find the limit, but we haven't hit the limit of scaling laws so far, despite scaling up massively over the last few years, so I doubt we will find the limit any time soon.

4. A bunch of corporations that have no hope of reaching first place decide to combine their resources to beat OpenAI and thus prevent a monopoly. I'm not sure if there is precedent, but even if there is, that would still require a lot of coordination and resources for little direct monetary gain.

We'll see what happens, but I'm not really confident about any of these possibilities.

sacred_numbers

3 jaar geledendiscuss

I can think of a few ways: 1. The ChatGPT web search plugin becomes standard protocol for every prompt. If you ask a factual question ChatGPT will first look up an answer with a search engine, then use the results to craft an answer. This could also be implemented using a knowledge database created by scraping the web that is occasionally updated rather than actually performing a web search every time. This also allows OpenAI to shift the blame to whatever source it found rather than ChatGPT if it does find libelous content. 2. OpenAI will just add a disclaimer to every response (or at least every response that appears to be asking a factual question) saying that asking factual questions is unreliable. 3. If there is a blacklist of topics that require a disclaimer or cause the LLM to refuse to answer, it can generate an answer, check it against the blacklist using embeddings/semantic search and either re-generate an answer or generate a refusal before showing the answer to the user.

My best guess is that it will be a combination of 1 and 2. I have always maintained that LLMs are very unlikely to develop into AGI on their own, but are very likely to be a critical piece of an AGI. The most recent research into scaling laws (Chinchilla, Llama) finds significant improvements from scaling data size much farther than parameter size, so memorizing facts within the parameters will become less and less feasible. This is actually ideal, though, since you want your model parameters to encode language and reasoning patterns, not memorize facts. If it's memorizing facts you either need more data or better (i.e. deduplicated) data. I'm not an expert, though, and I'm too lazy for a research review, so please don't sue me for libel if my facts are out of date.

sacred_numbers

3 jaar geledendiscuss

If you bought an 8xA100 machine for $140k you would have to run it continuously for over 10,000 hours (about 14 months) to train the 7B model. By that time the value of the A100s you bought would have depreciated substantially; especially because cloud companies will be renting/selling A100s at a discount as they bring H100s online. It might still be worth it, but it's not a home run.

sacred_numbers

3 jaar geledendiscuss

When I checked yesterday I believe the signature said OpenAI CEO Sam Altman, so it was definitely a joke signature, not a case of two people having the same name.

sacred_numbers

3 jaar geledendiscuss

The Reflexion paper (https://arxiv.org/abs/2303.11366) that came out recently shows how this kind of mistake might be overcome. Asking the model to think about the answer after it's generated a first draft greatly improves accuracy. Also, prompt engineering such as copying the generated code, pasting it in a new chat and saying "There's a bug in this code, please find it" can go a long way. There is so much low hanging fruit in harnessing the power of these models that is just being ignored because some even lower hanging fruit (RLHF, system messages, context window size, plugins, etc) is being released seemingly every few days.

sacred_numbers

3 jaar geledendiscuss

Theoretically it should be way less energy intensive as well, since there won't be an animal expending energy to live for months before slaughter. Nor will there be a need to grow feathers, bones, or blood that end up as byproducts. Of course, this tech is still being developed, so it probably hasn't reached optimal efficiency, but it doesn't have to be that efficient to be better than standard animal agriculture.

sacred_numbers

3 jaar geledendiscuss

When you are speaking to a person, they have inner thoughts and outer actions/words. If a person sees a chess board they will either consciously or unconsciously evaluate all the legal moves available to them and then choose one. An LLM like ChatGPT does not distinguish between inner thoughts and outer actions/words. The words that it speaks when prompted are its inner thoughts. There is also no distinction between subconscious and conscious thoughts. Humans generate and discard a multitude of thoughts in the subconscious before any thoughts ever make it to the conscious layer. In addition, most humans do not immediately speak every conscious thought they have before evaluating it to see whether speaking it aloud is consistent with their goals.

There's already a lot of research on this, but I strongly believe that eventually the best AIs will consist of LLMs stuck in a while loop that generate a stream of consciousness which will be evaluated by other tools (perhaps other specialized LLMs) that evaluate the thoughts for factual correctness, logical consistency, goal coherence, and more. There may be multiple layers as well, to emulate subconscious, conscious, and external thoughts.

For now though, in order to prompt the machine into emulating a human chess player, we will need to act as the machine's subconscious.

sacred_numbers

3 jaar geledendiscuss

Alternatively:

1. Quickly reduce costs by increasing model and computation efficiency.

2. Massively reduce prices while still maintaining some gross margin.

3. Massively increase market size and take the vast majority of market share.

4. End up with a higher gross profit due to a much larger market size despite decreasing prices and gross margins.

5. Profit.

sacred_numbers

3 jaar geledendiscuss

It could be even smaller than a Chinchilla optimal model. The Chinchilla paper was about training the most capable models with the least training compute. If you are optimizing for capability and inference compute you can "over-train" by providing much more data per parameter than even Chinchilla, or you can train a larger model and then distill it to a smaller size. Increasing context size increases inference compute, but the increased capabilities of high context size might allow you to skimp on parameters and lead to a net decrease in compute. There's probably other strategies as well, but those are the ones I know of.

comments