Or it matches their search needs?

it is interesting. i believe they may be hitting a wall as well

A paper from a week ago found that models trained on multiple data modes perform an order of magnitude better than text-only models of the same or even larger size.

Genuinely curious, what does it mean in this context to perform better? (hopefully that doesn't come across as snarky as text sometimes does).

Some of these large models are able to do zero shot learning and perform tasks they weren't explicitly trained on since the training objective is very general.

Being able to perform more advanced types of zero shot learning tasks would be comparable and further the accuracy on those tasks can be evaluated

Any chance this could improve coding LLMs like Co-pilot? Or would that sort of thing be limited to source code feeds (not that Github has a shortage).

The next big step for coding LLMs will be context window increases, leaked docs have OpenAI pricing for up to 16K I believe, 4x the current maximum. Now you're talking "write a class" instead of this line and maybe sometimes a method

It can already reliably write a todo web browser application.

With 16k and some other techniques, I’m guessing it could write a custom CMS database backed web application.

Not 16k, 32k. 8x the current window.

Nice

What is 16k referring to here

I’ve begun to grok it as “the amount of ram I have to play in before I have to start sharding work”

more literally and correctly, it’s the maximum number of tokens in the input and output, combined, where a token is 4/3 of a word

So we’re shifting from 5K words maximum to 40K (per sibling comment, who pointed out 32K context leaked as well)

A minor correction: a token is 3/4 of a word. ie. it’s slightly smaller than a word, not larger.

I figure it’s a stand-in for embodiment.

Are you referring to PALM-E? It didn't have any positive transfer for NLP tasks, in fact the unfrozen model performed slightly worse after the finetune. That being said, PALM-E wasn't really a multimodal model from the start, it's still basically just a text model with a visual one glued on top. Whether a truly multimodal model will be better at reasoning and data efficiency is still an open question though.

Can you share a link to this paper?

There is substantial evidence that cross training on multiple tasks improves performance.

ChatGPT struggles with intuition about the world, if it has image/video knowledge than it can potentially reason about what happens when you flip a plate upside down.

I wonder if there’s any reason this wouldn’t also apply in reverse: could a corpus of textual training data from which the role of a hand can be derived (including the fact that it naturally has five fingers) help DALL-E reliably render hands correctly?

Almost certainly. There is a vast corpus of text which references hands and what they are used for. Likewise there is a vast corpus of images which include hands. A sufficiently advanced statistical model can infer that any image with 6 fingers is a fabrication provided that it has learned how to classify fingers and has enough linguistic knowledge. Contrast this situation with a primarily image based model which will struggle to differentiate what situations detailed hands appear in compared with a patterned rug.

There is an interesting “takeoff” point where models are able to generate enough “interesting” data to train themselves. Asking chatGPT for ml training data on various tasks makes me think we are close to that inflection point.

Can’t wait for AI to tell people with atypical bodies that they don’t exist.

Definitely a bias, chatGPT will likely believe that only individuals with figures similar to models are real.

The hands on the clock pointed to 15 past 2...

Dalle/GPT: that's not right, there should be 5 fingers on this clock.

I think it would say “My name is Inigo Montoya. You Killed my Father. Prepare to die.”

One of the biggest improvements to text models like Image/Parti which are much better than Stable Diffusion is that they typically either train a much larger text model from scratch, or plug in a very good pre-trained text model like T5. This instantly upgrades their images.

It turns out that the image models already understand the low-level details of images well, and it's high-level reasoning/composition that is screwing their samples up - which is due to the poor text embedding they receive as a summary. (You can't generate a good image of 'a blue ball left of a red ball' if your text embedding is confused what 'left' is, no matter how high quality you are able to generate 'blue ball' or 'red ball' individually.) So, I would not be surprised if hands are improved by simply using a much better text model. (And why not plug in GPT-3 or PaLM for even better results...)

That said, hands are also like text-inside-images in being small, highly variable, and intrinsically difficult. And we know that the larger models solve text rendering simply by scaling up an OOM or 2 past SD. (You can see this from the text samples in the papers, and also more recently from Stability's Deep Floyd samples they've been teasing on Twitter.) So either way, scaling will probably solve hands soon.

It doesn’t “reason.” It predicts.

In the limit of 100% certainty, is prediction any different from reasoning?

Cross-training doesn't always help, I think the architecture is still being worked out. I can't remember the name of it but Google had some multimodal model they released a paper for, maybe six months or a year ago, and it did slightly worse than SOTA ("negative transfer" is the term they used for doing worse from additional modalities), and I recall the paper talking about how reaching near-parity with single mode models was an improvement.

One of their new models, PaLM-E, does show some positive transfer, so it's definitely a potential path for increased capabilities.

i think it's likely that they hired a diversified AI team with expertise in multiple areas and they're all working on their own areas. parallel efforts.

There's no more text to train on.

Also, amazingly, transfer learning works better than anyone could have hoped! As an analogy, imagine that you could show that football players learned musical instruments 50% quicker than non-athletes. That would be amazing!

But LLMs can write music and training them on music makes them better writers. It's almost like intelligence & skills are completely general!

That's just not true. There's no more super cheap data, but there's still probably about 5x the data we are currently feeding these models just waiting to be scanned and OCRd, and most of it is higher quality (if less topical).

The fact that AI requires so much data does make me think that we are somehow on the wrong track. Evolution didn't scan every atom on earth in order to produce a brain. FSD is a good example of this. It is ridiculous to me that we need to show the car videos of every possible situation in order for it to know how to drive.

It’s not the wrong track, it’s just very early and specialized. As you get more multi-modal models with better ability to learn/remember on the fly, they will require less and less data to train.

I'm pretty sure we will look back and laugh about how this was done in 50 years. No disrespect to anyone working in the area now of course, but I think better, more natural approaches will emerge.

> The fact that AI requires so much data does make me think that we are somehow on the wrong track.

I know the entire AI industry on the wrong track, but there's nothing that can be done about it. Sometimes you just have to wait for the crash and burn, and then things can change. It's like being at the ocean and trying to stop a wave from crashing after it has already formed, you just have to let it resolve itself of its own volition. As Einstein said, stupidity is infinite.

Not just that. Language acquisition. Teacher never stands in front of class and says "here's what the rules of grammar aren't. You today learn want do not how." There is a paucity of negative examples in the "training". Same goes for so many other tasks we learn quicker than can be explained with a Hebbian punishment/reward model. Most people can play a board game somewhat competently after just being told the rules. Some are even good at it. You don't need to crash your car dozens of time to figure out how to drive. All over we see something fundamental missing, and the industry is trying to hammer the screw into the wall while pretending that's the best way to do it.