Thanks a lot for your lucid explanations of some of the latest developments in the area of large language models. The manner in which you have linked together ideas from diverse papers and code sources is exquisite.
Great explanation. I like the simplicity of just predicting the next word. Sounds promising for generalization and ChatGPT is already producing amazing results. With exponential progress of AI this could get really interesting within just a few years.
I think the fact that predict-the-next-word works so well has been surprising to everyone in the field. Imagine someone standing up at one of the big NLP conferences in 2015 and saying "Forget all of these specialised models you're building, and just concentrate on predicting the next word" - they would have been laughed out of the room. (Shout out to @samwitteveen1806 for this hypothetical) As for progress - one of the interesting aspects of ChatGPT is that the vast majority of 'smarts' in the model already exist (hidden beneath the surface) due to the pretraining of the huge LLMs : So it may be that a lot of the current rough edges might be smoothed off with relatively little nudges training-wise. We'll see : Exciting times ahead!
Thanks so much, Martin, for this fascinating talk! I'm glad I was there at your in-person event -- the red umbrella is the coolest swag ever! This might be a whimsical idea, but as machines start to 'understand' reasoning, 'direct' their own rationales, learn to take actions in an environment, and, soon for sure, observe the real world within their embodied presence (and continually update themselves from such observations), I can't help but think that we are edging ever closer to the day when some version of Asimov's 3 laws of robotics might actually become practically useful -- perhaps with lots of step-by-step prompting built into the model? Just a year ago, I would have thought this idea to be still very much out of reach.
And Asimov was always my go-to author for cybernetic SciFi... And I'm puzzled that the Three Laws aren't even part of the discussion - perhaps because a lot of the plot lines revolved around how brittle/twistable they were. OTOH, it seems reasonable that you'd want to bake something 'ethical' into the network (using, for instance, RLHF or other data) - but that also having some external legalistic watchdog on overall machine behaviours would still make sense. I also think the umbrellas were a nice gesture from Google : The logo design (Machine Learning Singapore + frameworks) was left up to us, and turned out nicely, IMHO.
Hey Martin, this video is exquisite, thank you so much for making it. Could you explain a bit why they're still going with a successor to GPT3? I thought the RETRO movement of last year would mean we'd see only retrieval style large language models from now on.
Taking the liberty of writing some keywords to hopefully draw more eyeballs to this video. Keywords: Large language models. RLHF (reinforcement learning with human feedback). Using a smaller model to train a larger model. Prompt engineering. Prompting a model to programme in Python. Action transformer -- a robot acting in a virtual environment, instructed to pick things up and put them around in a circle. LangChain which integrates ChatGPT, acting like a human, doing reasoning to look for resources it needs to write a programme. Language model cascades -- new action capabilities using large language models. Prompt design. Computers finding prompts for themselves. Robot programming itself. Automated Prompt Engineering (APE). Self-improvement. Can we get the machine to teach itself? Chain-of-thought (CoT) reasoning, distillation. Machine generating rationales. Distilling from a large model to a smaller model, while retaining the benefits. Cautionary note from Yann LeCun: AI trained on words alone will never approximate human understanding.
Thanks for the work you put into this : Do you think it would help the video to add this to the description? Or convert it into stuff for the keywords field? I'm open to suggestions!
@@MartinAndrews-mdda I'm no TH-cam expert, but I'm certain the description would be used for search results. If anything, a human glancing through it might catch something of interest and decide to watch it, when they might otherwise put off for later and then forget about it altogether. Feel free to copy and paste!
Thanks for linking a Wiki entry! I'm assuming that you mean to point people who are not so interested in the broader LLM themes to another source of information about ChatGPT. Yes, the ChatGPT demo is a game-changer in terms of what everyone has been able to play with for free, but it'll probably turn into a branding thing going forwards (like the Dall-E model morphed over time). One issue (that perhaps came out more strongly during the in-person version of this talk) is that these LLMs actually "know" much more about the world than single roll-outs of text would indicate. It seems to me that other publicly released LLMs (eg: from Meta, and Stability AI-backed sources) are actually already really capable internally, it's just that they haven't had the
![[1hr Talk] Intro to Large Language Models](http://i.ytimg.com/vi/zjkBMFhNj_g/mqdefault.jpg)
Thanks a lot for your lucid explanations of some of the latest developments in the area of large language models. The manner in which you have linked together ideas from diverse papers and code sources is exquisite.
Great explanation. I like the simplicity of just predicting the next word. Sounds promising for generalization and ChatGPT is already producing amazing results. With exponential progress of AI this could get really interesting within just a few years.
I think the fact that predict-the-next-word works so well has been surprising to everyone in the field. Imagine someone standing up at one of the big NLP conferences in 2015 and saying "Forget all of these specialised models you're building, and just concentrate on predicting the next word" - they would have been laughed out of the room. (Shout out to @samwitteveen1806 for this hypothetical)
As for progress - one of the interesting aspects of ChatGPT is that the vast majority of 'smarts' in the model already exist (hidden beneath the surface) due to the pretraining of the huge LLMs : So it may be that a lot of the current rough edges might be smoothed off with relatively little nudges training-wise. We'll see : Exciting times ahead!
Thanks so much, Martin, for this fascinating talk! I'm glad I was there at your in-person event -- the red umbrella is the coolest swag ever!
This might be a whimsical idea, but as machines start to 'understand' reasoning, 'direct' their own rationales, learn to take actions in an environment, and, soon for sure, observe the real world within their embodied presence (and continually update themselves from such observations), I can't help but think that we are edging ever closer to the day when some version of Asimov's 3 laws of robotics might actually become practically useful -- perhaps with lots of step-by-step prompting built into the model? Just a year ago, I would have thought this idea to be still very much out of reach.
And Asimov was always my go-to author for cybernetic SciFi... And I'm puzzled that the Three Laws aren't even part of the discussion - perhaps because a lot of the plot lines revolved around how brittle/twistable they were. OTOH, it seems reasonable that you'd want to bake something 'ethical' into the network (using, for instance, RLHF or other data) - but that also having some external legalistic watchdog on overall machine behaviours would still make sense.
I also think the umbrellas were a nice gesture from Google : The logo design (Machine Learning Singapore + frameworks) was left up to us, and turned out nicely, IMHO.
Hey Martin, this video is exquisite, thank you so much for making it. Could you explain a bit why they're still going with a successor to GPT3? I thought the RETRO movement of last year would mean we'd see only retrieval style large language models from now on.
Taking the liberty of writing some keywords to hopefully draw more eyeballs to this video.
Keywords:
Large language models.
RLHF (reinforcement learning with human feedback).
Using a smaller model to train a larger model.
Prompt engineering.
Prompting a model to programme in Python.
Action transformer -- a robot acting in a virtual environment, instructed to pick things up and put them around in a circle.
LangChain which integrates ChatGPT, acting like a human, doing reasoning to look for resources it needs to write a programme.
Language model cascades -- new action capabilities using large language models.
Prompt design. Computers finding prompts for themselves. Robot programming itself.
Automated Prompt Engineering (APE).
Self-improvement. Can we get the machine to teach itself?
Chain-of-thought (CoT) reasoning, distillation.
Machine generating rationales.
Distilling from a large model to a smaller model, while retaining the benefits.
Cautionary note from Yann LeCun: AI trained on words alone will never approximate human understanding.
Thanks for the work you put into this : Do you think it would help the video to add this to the description? Or convert it into stuff for the keywords field? I'm open to suggestions!
@@MartinAndrews-mdda I'm no TH-cam expert, but I'm certain the description would be used for search results. If anything, a human glancing through it might catch something of interest and decide to watch it, when they might otherwise put off for later and then forget about it altogether. Feel free to copy and paste!
en.m.wikipedia.org/wiki/ChatGPT
Thanks for linking a Wiki entry! I'm assuming that you mean to point people who are not so interested in the broader LLM themes to another source of information about ChatGPT. Yes, the ChatGPT demo is a game-changer in terms of what everyone has been able to play with for free, but it'll probably turn into a branding thing going forwards (like the Dall-E model morphed over time).
One issue (that perhaps came out more strongly during the in-person version of this talk) is that these LLMs actually "know" much more about the world than single roll-outs of text would indicate. It seems to me that other publicly released LLMs (eg: from Meta, and Stability AI-backed sources) are actually already really capable internally, it's just that they haven't had the