What an impressive waste of resources. It’s portrayed as THE most important race and yet what has been delivered so far?
Slightly better TTS or OCR, photography manipulation that is commercially unusable because sources can’t be traced, summarization that can introduce hallucinations, … sure all of that is interesting in terms of academic research, with potentially some use cases… but it’s not as if it didn’t exist before at nearly the same quality for a fraction of the resources.
It’s a competitions where “winners” actually don’t win much, quite a ridiculous situation to be in.
Image gen did not exist in any way shape or form before. Now we’re getting video gen like a few years later.
Let’s not forget we started by playing the game of Go better. My prediction as a hobby Go programmer (the game, not language) in 2015 would be that better than human AIs would be there by 2020 and they got there by 2016.
Before the AlphaGo match with Lee Sedol people predicted the AI would just put up a decent fight since a previous version played questionably against a weaker player. It blew one of the best players ever out of the water, losing only one game of the series.
Future matches even against the world #1 with the better models showed it to be invincible against humans
You’re making the same mistake. You’re looking at the current capabilities and predicting a human speed of improvement. AI is improving faster.
Image gen did not exist in any way shape or form before.
Typical trope while promoting a “new” technology. A classic example is 1972’s AARON https://en.wikipedia.org/wiki/AARON which, despite not being based on LLM (so not CLIP) nor even ML is still creating novel images. So… image generation has been existing since at least the 70s, more than half a century ago. I’m not saying it’s equivalent to the implementation since DALLE (it is not) but to somehow ignore the history of a research field is not doing it justice. I have also been modding https://old.reddit.com/r/computationalcrea/ since 9 years, so that’s before OpenAI was even founded, just to give some historical context. Also 2015 means 6 years before CLIP. Again, not to say this is the equivalent, solely that generative AI has a long history and thus setting back dates to grand moments like AlphaGo or DeepBlue (and on this topic I can recommend Rematch from Arte) … are very much arbitrary and in no way help to predict what’s yet to come, both in terms of what’s achievable but even the pace.
Anyway, I don’t know what you actually tried but here is a short list of the 58 (as of today) models I tried https://fabien.benetou.fr/Content/SelfHostingArtificialIntelligence and that’s excluding the popular ones, e.g. ChatGPT, Mistal LeChat, DALLE, etc which I also tried.
I might be making “the same mistake” but, as I hope you can see, I do keep on trying what I believe is the state of the art of a pretty much weekly basis.
Creating abstract art by moving pixels around is not anywhere close to what we mean by image generation. At no point did this other software generate something from a prompt
I’d normally accept the challenge if you didn’t add that. You did though and it, namely a system (arguably intelligent) made an image, several images in fact. The fact that we dislike or like the aesthetics of it or that the way it was done (without prompt) is different than how it currently is remains irrelevant according to your own criteria, which is none. Anyway my point with AARON isn’t about this piece of work specifically, rather that there is prior work, and this one is JUST an example. Consequently the starting point is wrong.
Anyway… even if you did question this, I argued for more, showing that I did try numerous (more than 50) models, including very current ones. It even makes me curious if you, who is arguing for the capabilities and their progress, if you tried more models than I did and if so where can I read about it and what you learned about such attempts.
So LLMs can trace their origin back to the 2017 paper “Attention is all you need”, they with diffusion models have enabled prompt based image generation at an impressive quality.
However, looking at just image generation you have GANs as far back as 2014 with style GANs (ones that you could more easily influence) dating back to 2018. While diffusion models also date back to 2015, I don’t see any mention of use in images until early 2020’s.
Thats also ignoring that all of these technologies go back further to lstms and CNNs, which go back further into other NLP/CV technologies. So there has been a lot of progress here, but progress isn’t also always linear.
This sort of stuff has been said about pretty much every technological breakthrough in history. Language models on their own do indeed have lots of limitations, however there is a lot of potential in coupling them with other types of expert systems. We simply don’t know what all the potential applications are for this tech. However, the iron rule throughout history has been that people dismissing new technological developments have typically been proven wrong.
Language models on their own do indeed have lots of limitations, however there is a lot of potential in coupling them with other types of expert systems.
Still, while keeping this in mind we also must remain mindful of what each system can actually do, not conflate with what we WANT it do yet it can not do yet, and might never will.
Sure we have to be realistic about capabilities of different systems. Thing is that we don’t know what the actual limitations are yet. In the past few years we’ve seen huge progress in terms of making language models mode efficient, and more capable.
My expectation is that language models, and the whole GPT algorithm, will end up being a building block in more sophisticated systems. We’re already seeing research shift from simply making models bigger to having models do reasoning about the output. I suspect that we’ll start seeing people rediscovering a lot of symbolic logic research that was done back in the 80s.
The overall point here is that we don’t know what the limits of this tech are, and the only way to find out is to continue researching it, and trying new things. So, it’s clearly not a waste of resources to pursue this. What makes this the most important race isn’t what it’s delivered so far, but what it has potential to deliver.
If we can make AI systems that are capable of doing reasoning tasks in a sufficiently useful fashion that would be a game changer because it would allow automating tasks that fundamentally could not be automated before. It’s also worth noting that reasoning isn’t a binary thing where it’s either correct or wrong. Humans are notorious for making logical errors, and most can’t do formal logic to save their lives. Yet, most humans can reason about tasks they need to complete in their daily lives sufficiently well to function. We should be applying the same standard to AI systems. The system just needs to be able to function well enough to accomplish tasks within the domain it’s being used in.
What an impressive waste of resources. It’s portrayed as THE most important race and yet what has been delivered so far?
Slightly better TTS or OCR, photography manipulation that is commercially unusable because sources can’t be traced, summarization that can introduce hallucinations, … sure all of that is interesting in terms of academic research, with potentially some use cases… but it’s not as if it didn’t exist before at nearly the same quality for a fraction of the resources.
It’s a competitions where “winners” actually don’t win much, quite a ridiculous situation to be in.
Image gen did not exist in any way shape or form before. Now we’re getting video gen like a few years later.
Let’s not forget we started by playing the game of Go better. My prediction as a hobby Go programmer (the game, not language) in 2015 would be that better than human AIs would be there by 2020 and they got there by 2016.
Before the AlphaGo match with Lee Sedol people predicted the AI would just put up a decent fight since a previous version played questionably against a weaker player. It blew one of the best players ever out of the water, losing only one game of the series.
Future matches even against the world #1 with the better models showed it to be invincible against humans
You’re making the same mistake. You’re looking at the current capabilities and predicting a human speed of improvement. AI is improving faster.
Typical trope while promoting a “new” technology. A classic example is 1972’s AARON https://en.wikipedia.org/wiki/AARON which, despite not being based on LLM (so not CLIP) nor even ML is still creating novel images. So… image generation has been existing since at least the 70s, more than half a century ago. I’m not saying it’s equivalent to the implementation since DALLE (it is not) but to somehow ignore the history of a research field is not doing it justice. I have also been modding https://old.reddit.com/r/computationalcrea/ since 9 years, so that’s before OpenAI was even founded, just to give some historical context. Also 2015 means 6 years before CLIP. Again, not to say this is the equivalent, solely that generative AI has a long history and thus setting back dates to grand moments like AlphaGo or DeepBlue (and on this topic I can recommend Rematch from Arte) … are very much arbitrary and in no way help to predict what’s yet to come, both in terms of what’s achievable but even the pace.
Anyway, I don’t know what you actually tried but here is a short list of the 58 (as of today) models I tried https://fabien.benetou.fr/Content/SelfHostingArtificialIntelligence and that’s excluding the popular ones, e.g. ChatGPT, Mistal LeChat, DALLE, etc which I also tried.
I might be making “the same mistake” but, as I hope you can see, I do keep on trying what I believe is the state of the art of a pretty much weekly basis.
Creating abstract art by moving pixels around is not anywhere close to what we mean by image generation. At no point did this other software generate something from a prompt
I’d normally accept the challenge if you didn’t add that. You did though and it, namely a system (arguably intelligent) made an image, several images in fact. The fact that we dislike or like the aesthetics of it or that the way it was done (without prompt) is different than how it currently is remains irrelevant according to your own criteria, which is none. Anyway my point with AARON isn’t about this piece of work specifically, rather that there is prior work, and this one is JUST an example. Consequently the starting point is wrong.
Anyway… even if you did question this, I argued for more, showing that I did try numerous (more than 50) models, including very current ones. It even makes me curious if you, who is arguing for the capabilities and their progress, if you tried more models than I did and if so where can I read about it and what you learned about such attempts.
It’s irrelevant because it wasn’t a precursor technique. The precursor was machine learning research, not other image generation technology
So LLMs can trace their origin back to the 2017 paper “Attention is all you need”, they with diffusion models have enabled prompt based image generation at an impressive quality.
However, looking at just image generation you have GANs as far back as 2014 with style GANs (ones that you could more easily influence) dating back to 2018. While diffusion models also date back to 2015, I don’t see any mention of use in images until early 2020’s.
Thats also ignoring that all of these technologies go back further to lstms and CNNs, which go back further into other NLP/CV technologies. So there has been a lot of progress here, but progress isn’t also always linear.
You can see with image generation progress was extremely quick
This sort of stuff has been said about pretty much every technological breakthrough in history. Language models on their own do indeed have lots of limitations, however there is a lot of potential in coupling them with other types of expert systems. We simply don’t know what all the potential applications are for this tech. However, the iron rule throughout history has been that people dismissing new technological developments have typically been proven wrong.
Absolutely, I even have a dedicated section “Trying to insure combinatoriality/compositionality” in my notes on the topic https://fabien.benetou.fr/Content/SelfHostingArtificialIntelligence
Still, while keeping this in mind we also must remain mindful of what each system can actually do, not conflate with what we WANT it do yet it can not do yet, and might never will.
Sure we have to be realistic about capabilities of different systems. Thing is that we don’t know what the actual limitations are yet. In the past few years we’ve seen huge progress in terms of making language models mode efficient, and more capable.
My expectation is that language models, and the whole GPT algorithm, will end up being a building block in more sophisticated systems. We’re already seeing research shift from simply making models bigger to having models do reasoning about the output. I suspect that we’ll start seeing people rediscovering a lot of symbolic logic research that was done back in the 80s.
The overall point here is that we don’t know what the limits of this tech are, and the only way to find out is to continue researching it, and trying new things. So, it’s clearly not a waste of resources to pursue this. What makes this the most important race isn’t what it’s delivered so far, but what it has potential to deliver.
If we can make AI systems that are capable of doing reasoning tasks in a sufficiently useful fashion that would be a game changer because it would allow automating tasks that fundamentally could not be automated before. It’s also worth noting that reasoning isn’t a binary thing where it’s either correct or wrong. Humans are notorious for making logical errors, and most can’t do formal logic to save their lives. Yet, most humans can reason about tasks they need to complete in their daily lives sufficiently well to function. We should be applying the same standard to AI systems. The system just needs to be able to function well enough to accomplish tasks within the domain it’s being used in.
LLM’s are great at generating boilerplate code or sending you in the right direction.