Language Models are Few-Shot Learners

prompt and 2) asking them to think through and 'reason' through their suggestions and the products have always been superior to what most humans have produced. The 'LLMs are probabilistic predictors' comments though keep appearing on threads and I'm not quite sure I understand them -- yes, LLMs don't have 'human context' i.e. data needed to understand human beings since they have not directly been fed in human experiences, but for the most part -- LLMs are not simple 'statistical predictors' as everyone brands them to be. You can see a thorough write-up I did of what GPT is / was here if you're interested: https://photonlines.substack.com/p/intuitive-and-visual-guid...

there is no obvious way to go from approximately human-level AGI to the kinds of sci-fi super-super-human AGIs that some AI catastrophists imagine. It's worth reading more about the topic, it's less that we'll have some human comparable AI and then be stuck with it - more so that things will continue to scale. Stopping at human level might be a harder task (or even getting something that's human like at all). > This is all not to mention that we have no way right now of tackling the problem of teaching the vast amounts of human common sense knowledge that is likely baked into our genes to an AI, and it's hard to tell how much that will impact true AGI. This is a good point and basically the 'goal alignment problem' or 'friendly AI' problem. It's the main reason for the risk since you're more likely to get a powerful AGI without these 'common sense' human intuitions of things. I think your mistake is thinking the goal alignment is a prerequisite for AGI - the risk comes from the truth being that it's not. Also humans aren't entirely goal aligned either, but that's a different issue. I understand the skepticism, I was skeptical too - but if you read more about it (not pop-sci, but the books from the people working on the stuff) it's more solid than you probably think and your positions on it won't hold up.

that they have to expend so much more energy to do their magic (and yes, to me they are a bit magical), which makes some wonder if what these models do is a form of refined brute-force search, rather than ideating. Well, of course, all they are doing is searching and curve-fitting. To me, the magical thing is that they have shown us, more or less undeniably (Penrose notwithstanding), that that is all we do. Questions that have been asked for thousands of years have now been answered: there's nothing special about the human brain, except for the ability to form, consolidate, consult, and revise long-term memories. 1: E.g., https://arxiv.org/abs/2005.14165 from 2020

human, feeling the need to change the output and would likely consult a dictionary but I don't think that's a fair comparison here as LLMs don't have these same biases. King - crown really seems like it should be something like "man" or "person". The crown is the manifestation of the ruling power. We still use phrases like "heavy is the head that wears the crown" in reference to general leaders, not just monarchs. king - princess I honestly don't know what to expect. Man is technically gender neutral so I'll take this one. king - queen I would expect similar outputs to the previous one. Don't quite agree here. queen - king I get why is removing royalty but given the previous (two) results I think is showing a weird gender bias. Remember that queen is something like (woman + crown) and king is akin to (man + crown). So subtracting should be woman - man. The others I agree with. These were actually done because I was quite surprised at the results and was thinking about the aforementioned gender bias. > But keep in mind that this doesn't do embedding math like OP! I think you are misunderstanding the architecture of these models. The embedding sub-network is the translation of text to numeric tokens. You'll find mention of the embedding sub-networks in both the GPT3[3] and GPT4 papers. Though they are given lower importance than other works. While much smaller than the main network, don't forget that embedding networks are still quite large. For the smaller models they constitute a significant part of the total parameter count[4] After the embedding sub-network is your main transformer network. The purpose of this network is to perform embedding math! It is just that the goal is to do significantly more complicated math. Remember, these are learnable mappings (see Optimal Transport). We're just breaking it down into their two main intermediate mappings. But the embeddings still end up being a bottleneck. It is your literal gateway from words to numbers. [0] https://en.wikipedia.org/wiki/Mass_noun [1] https://www.merriam-webster.com/dictionary/data [2] https://www.sciotoanalysis.com/news/2023/1/18/this-data-or-t... [3] https://arxiv.org/abs/2005.14165 [4] https://arxiv.org/abs/2303.08774 [4] https://www.lesswrong.com/posts/3duR8CrvcHywrnhLo/how-does-g...

speech, NYT content, copyrighted content and more getting used to train models. https://foundation.mozilla.org/en/research/library/generativ... (This article is to establish the issues with CC as a source of LLM training) https://facctconference.org/static/papers24/facct24-148.pdf (this details those issues.) Firms, such as the NYT are now stopping common crawl from archiving their pages. https://www.wired.com/story/the-fight-against-ai-comes-to-a-... ----- TLDR: 'NYT' and other high quality content has largely been ingested by models. Reddit and other sources play a large part in training current models. While I appreciate your being blunt, this also means not being sharp and incisive. Perhaps precision would be required here to clarify your point. Finally - >Factual accuracy is not binary, it is a matter of degrees. Obviously training on content that is more factually correct will result in more factually correct next tokens What. Come on, I think you wouldnt agree with your own statement after reading it once more -Factual correctness is not a matter of degrees. Furthermore, facts dont automatically create facts. Calcuation, processing, testing and verification create more facts. Just putting facts together creates content.

non-Mistral open-source LLMs like Yi and Qwen that should do well too). I'm not sure what BloombergGPT has to do with LLMs vs non-LLMs; BloombergGPT is an LLM [2], and it defeating other LLMs on financial benchmarks doesn't prove much about large language models other than "LLMs can be trained to be better at specific tasks." 1: https://arxiv.org/abs/2005.14165 2: https://www.bloomberg.com/company/press/bloomberggpt-50-bill...

applied to a fine-tuned language model to further align model behavior with human preferences and instruction following. We collect data that represents empirically sampled human preferences, whereby human annotators select which of two model outputs they prefer. This human feedback is subsequently used to train a reward model, which learns patterns in the preferences of the human annotators and can then automate preference decisions. The rest of 3.2 should be enlightening as well (in service to understanding how RLHF is actually quite different from the training algorithms used in pretraining or finetuning). > The data is input to the learning algorithm which generates the model. > To make ChatGPT less "racist" they provided new training data and rewarded it for the correct answers. For frontier language models, this is strictly incorrect. Multiple rounds of different training algorithms are used, each with its own objectives. RLHF is more complex than just 'put better data into same old pretraining algorithm.' > The process or algorithm of RHLF was not modified specially to handle "racism". These are general learning algorithms. They simply added new training data. Nobody is claiming RLHF was modified, but that the 'logic behind the weighting methods' are modified. Pretraining, then finetuning on specialized dataset does not produce a sufficiently aligned (non-racist, gender-neutral, etc) model, no matter how much extra data you throw at it: the answer to this is the introduction of a different algorithm, RLHF, which more accurately fits the model to desired alignment than unsupervised autoregressive loss minimization does. The sequence of events here is: 1. unsupervised pretraining on mixed dataset eval result: "it's just continuing the prompt instead of answering questions or being useful" researcher response: "let's put more focused data in to make it respond better" 2. instruction tuning on task-specific dataset eval result: "it's answering questions now but it has inherited biases from the pretraining corpus (it's too racist, sexist, or otherwise not aligned with desired values)" researcher response: "we've hit the limit on what you can reasonably accomplish with this training algorithm, let's try something else" (see GPT3, section "5. Limitations") 3. rlhf on human preference dataset eval result: "well, it's less racist now" > You are confusing yourself and still not quite understanding the basic concept. Once again, as per my last reply: "If you only look at the very high level superficial details, they could be said to be similar. That is all. The details differ in nontrivial ways." The basic ontology that you're trying to argue exists, is counterfactual. It boils down to: "put data into training algorithm, get model out. if it's too racist, put aligned data in and get non-racist model out." It's not hard to understand, you're just wrong. SOTA models are not produced this way. Full stop. The details of training phases are nontrivially different both algorithmically, and in the datasets used. At this point we've exhausted all reasonable discourse; it's clear that you're not engaging in good faith and I remain convinced that your position falls apart under the most trivial application of rigorous analysis. If you have any actual source to back up your claims, I'd love to see it. Beyond that, good day.

instructgpt/rlhf) https://arxiv.org/pdf/1707.06347.pdf). - 2018 lead author on GPT1 https://www.mikecaptain.com/resources/pdf/GPT-1.pdf - 2019 lead author on GPT2 https://insightcivic.s3.us-east-1.amazonaws.com/language-mod... blog https://openai.com/research/better-language-models - 2019 coauthor on RLHF https://arxiv.org/pdf/1909.08593.pdf) - 2020 coauthor on gpt3 https://arxiv.org/abs/2005.14165 - 2020 coauthor on scaling laws https://arxiv.org/pdf/2001.08361.pdf%E4%B8%AD%E5%BE%97%E5%88... - 2021 coauthor on DallE http://proceedings.mlr.press/v139/ramesh21a/ramesh21a.pdf - 2021 coauthor on Codex https://arxiv.org/pdf/2107.03374.pdf?trk=public_post_comment... - 2023 lead author on Whisper https://proceedings.mlr.press/v202/radford23a/radford23a.pdf name a more successful 7 year CS career post undergrad... Update: FYI openai just announced a "developer day" in Nov - somehow not blessed by the HN gods https://news.ycombinator.com/item?id=37408234

is doing something else than simply memorising a table of sums, which is an altogether much grander claim). OpenAI themselves seemed to be dubious enough about their own claim that the Arithmetic section of their paper was only included in the preprint (on Arxiv) and not in the published paper (in the 34th NeurIPS). Yet, the claim in the preprint was still enough for people to forcefully argue that GPT-3 can do arithmetic, that it can learn the rules of arithmetic, and other impossible things before breakfast. This is a discussion that goes back at least 3 years (judging from my comments where I point out that it's nonsense). It seems that the arithmetic ability of large language models is now a well accepted truth in the minds of the general public, who will casually use it to do, say, their maths coursework etc. So blame OpenAI who made the big claims.