I’ve read the paper and the skeptical comments here, to wit: it’s just an actor/critic pipeline by another name.
I’ll bite and say this is actually interesting — and the paper title is misleading.
What they’ve done here is hooked up a text-only LLM to multimodal critics, given it (mostly) an image diffusion generation task, and asked it to improve its prompting of the multimodal generation by getting a set of scores back.
This definitely works, based on their outputs. Which is to say, LLMs can, zero shot, with outside tool feedback, iteratively improve their prompting using only that tooling feedback.
Why is this interesting? Well, this did not work in the GPT-3 era; it seems to do so now. I see this as an interesting line to be added in the ‘model capabilities’ box as our models get larger and more sophisticated — the LLMs can perform some sort of internally guided search against a black box generator and use a black box scorer to improve at inference time.
That’s pretty cool. It’s also generalizable, and I think is worth keeping in mind on the stack of possible approaches for, say agentic coding, that you can use a critic to not just ‘improve’ generated output, but most likely do some guided search through output space.
I really wish we would find a different term for this.
Doing something always takes at least one attempt, i.e. "one shotting". "Zero shotting" is an oxymoron, which makes it a term that only creates more confusion rather than succinctly conveying something.
"One shot" is simply about the action itself, but it says nothing about how much preparation was done beforehand. "Zero shot" additionally implies without training or preparation.
TCGs have a related "zero turn win" concept, where the opponent goes first and you win without getting a turn due to the set of cards you randomly drew and being able to activate them on the opponent's turn.
I think of a shot as an example, not a try: “One shot” is “One example”. Zero shot is “Zero examples”. I don’t love it, but I don’t hate it, got a better word for it?
> I think is worth keeping in mind on the stack of possible approaches for, say agentic coding, that you can use a critic to not just ‘improve’ generated output, but most likely do some guided search through output space.
The one issue I keep finding with those approaches is that there’s already good tools for the problem, but we keep searching for wasteful approaches because “natural languages” for something humans are not going to interact without a good deal of training.
I do understand the hope of getting LLMs do the bulk of the work, and then after audit, we fix the errors. But both audit and fixing will require the same mental energy as writing the code in the first place. And possibly more time.
Specialist tools are always more expansive and offer more controls than general public tools. Most approaches with agentic coding is offering general interfaces instead of specialized interfaces, but redirecting you to a bespoke and badly designed specialized interface whenever you want to do anything useful.
I hear that. Counterpoint - if you all you have is a Philips-head screwdriver, all you have is a Philips-head screwdriver. On the other hand if all you have is a six axis CnC mill, well, then you have a lot.
I think of this less as audit misses, and more as developing a permanently useful tool. For open model weights, humanity will not (unless we’re talking real zombie apocalypse scenarios) lose these weights. They are an incredible global asset, so making them more generally useful and figuring out how to use them is super helpful.
While it is hard to argue with the wisdom of crystallizing intellectual capital into our tools, I do wonder if these models might be as likely to diminish as to develop the person using them, in which case we trade an implement's iterative improvement for ours, in a way
Monks in the Middle Ages: “The Printing Press will destroy people’s ability to memorize.”
This was accurate. But mostly humans gained from books. I think we will develop the social technology to use these tools over time; giving some things up and gaining others.
If we don’t, the Amish can just take over and be like “Stupid English, using the devil’s weights.” :)
Maybe they are useful. But I think there’s more usefulness in specialized databases and optimized approaches than betting everything on big llms models. Kinda like deriving linting rules and combining it with a rule engines to catch errors. Efficient and useful instead of continuously running a big llm model.
Are they using the same diffusion models as the GPT-3 area? Meaning is it the LLM that has improved or is it the diffusion model? I know it's probably a foolish take but I am really skeptical of the "larger models will solve all our problems" line of thinking.
They don’t compare in the paper. I will say I experimented extensively with GPT-3 era LLMs on improving ouput by trying to guide early diffusion models with critical responses. It was a) not successful, and b) pretty clear to me that GPT-3 didn’t “get” what it was supposed to be doing, or didn’t have enough context to keep all this in mind, or couldn’t process it properly, or some such thing.
This paper has ablations, although I didn’t read that section, so you could see where they say the effectiveness comes from. I bet you thought that it’s emergent from a bunch of different places.
FWIW, I don’t think LLMS will solve all our problems, so I too am skeptical of that claim. I’m not skeptical of the slightly weaker “larger models have emergent capabilities and we are probably not done finding them as we scale up”.
> FWIW, I don’t think LLMS will solve all our problems, so I too am skeptical of that claim. I’m not skeptical of the slightly weaker “larger models have emergent capabilities and we are probably not done finding them as we scale up”.
100% agree. I'd classify the time now as identifying the limits of what they can functionally do though, an it's a lot!
They only did that for image generation. The more interesting part is that an LLM can approach or find the correct caption for an image, video or audio during test time with no training using only the score as a guide. It's essentially working blind almost like the game Marco Polo where the scorer is saying "warmer" or "colder" while the LLM is finding its way towards the goal. This is an example of emergent capabilities since there are no examples of this in the training data.
Actually, it's the name of the paper. And while the team also developed and released a system to elicit the behavior by doing what you described, it's entirely possible that the researchers thought the title to be the most important finding in their work.
My photoresistor nightlight can "see" that it is dark and it "knows" to turn on the light - not only does it not have training, it does not have any code!
And if you think that is amazing, my bi-metallic strip thermostat "feels" the temperature and then modifies the environment because it "knows" if it's hot to turn on the A/C, and if it's cold to turn on the heat - no training or code!
All of this AI stuff is just unbelievably incredible - what a brave new world (of word games)!
The nightlight and thermostat's response to stimulus is nowhere near analyzing a picture of a clock tower and responding with "Image of a city's tallest, historic landmark with a sepia filter." To me, recognizing the umbrella in the spoon is one of the most impressive items they list.
You don't have to train the LLM soecifically for the tasks and even the auxiliary tools aren't trained on the tasks they are used as scorers for (because they aren't doing the task,just evaluating how well the LlM is), so there is no task-specific training.
Through the story and experience of a blind man, they end up getting into the question of what does it mean to see
The podcast is pretty straightforward, but it does end up showing that defining “seeing” is a philosophical question, rather than a simple obvious answer
In actor/critic the actor and critic are normally learned, i.e., their weights are adjusted during the process. The paper is correct that their method is zero-shot, but it doesn't mention that their method is essentially equivalent to a few rounds of training but then discarding the training update.
Anyone who works with deep architectures and momentum-based optimizers knows that the first few updates alone provide large improvements in loss. In this paper the breakthrough is that computing these first few updates at test time enables one to describe the algorithm as "without training" and therefore attract hype.
But they aren't updating the model weights. They're iteratively updating the prompt. It's automating the process that humans use with generative models.
I think there is potentially a powerful method here. Specifically, the optimal context for a given task can be saved and a meta-learner can be trained to map the task to the context. This would allow fine tuning a model for some specific task without retaining the LLM. For example, generating an SEM image with of some material with a specified porosity and grain size.
I just remember Zuck's comments about AI and how the idea of it dooming our species is a bit silly, etc
This is the wrong approach to take. At minimum you have to say things like "well yes we're always on the lookout for this kind of thing". With him? Not a care in the world
Exactly how little training is "without any"? I'm assuming that companies haven't been spending billions trying to train LLMs to better understand things when they can do it without any training.
Is the LLM essentially playing "Wordle" with an external system that rates the quality of its output, gradually climbing the score ladder until it produces good results?
The paper certainly contradicts my expectation from the title. I.e. it does not present an LLM that can generate images without any access to images before.
The base layer is just electronic circuitry. As long there is electricity it will do stuff (like a radio producing noise). GPU, CPU, is mostly software embedded in hardware.
I’ll bite and say this is actually interesting — and the paper title is misleading.
What they’ve done here is hooked up a text-only LLM to multimodal critics, given it (mostly) an image diffusion generation task, and asked it to improve its prompting of the multimodal generation by getting a set of scores back.
This definitely works, based on their outputs. Which is to say, LLMs can, zero shot, with outside tool feedback, iteratively improve their prompting using only that tooling feedback.
Why is this interesting? Well, this did not work in the GPT-3 era; it seems to do so now. I see this as an interesting line to be added in the ‘model capabilities’ box as our models get larger and more sophisticated — the LLMs can perform some sort of internally guided search against a black box generator and use a black box scorer to improve at inference time.
That’s pretty cool. It’s also generalizable, and I think is worth keeping in mind on the stack of possible approaches for, say agentic coding, that you can use a critic to not just ‘improve’ generated output, but most likely do some guided search through output space.
I really wish we would find a different term for this.
Doing something always takes at least one attempt, i.e. "one shotting". "Zero shotting" is an oxymoron, which makes it a term that only creates more confusion rather than succinctly conveying something.
TCGs have a related "zero turn win" concept, where the opponent goes first and you win without getting a turn due to the set of cards you randomly drew and being able to activate them on the opponent's turn.
The one issue I keep finding with those approaches is that there’s already good tools for the problem, but we keep searching for wasteful approaches because “natural languages” for something humans are not going to interact without a good deal of training.
I do understand the hope of getting LLMs do the bulk of the work, and then after audit, we fix the errors. But both audit and fixing will require the same mental energy as writing the code in the first place. And possibly more time.
Specialist tools are always more expansive and offer more controls than general public tools. Most approaches with agentic coding is offering general interfaces instead of specialized interfaces, but redirecting you to a bespoke and badly designed specialized interface whenever you want to do anything useful.
I think of this less as audit misses, and more as developing a permanently useful tool. For open model weights, humanity will not (unless we’re talking real zombie apocalypse scenarios) lose these weights. They are an incredible global asset, so making them more generally useful and figuring out how to use them is super helpful.
This was accurate. But mostly humans gained from books. I think we will develop the social technology to use these tools over time; giving some things up and gaining others.
If we don’t, the Amish can just take over and be like “Stupid English, using the devil’s weights.” :)
This paper has ablations, although I didn’t read that section, so you could see where they say the effectiveness comes from. I bet you thought that it’s emergent from a bunch of different places.
FWIW, I don’t think LLMS will solve all our problems, so I too am skeptical of that claim. I’m not skeptical of the slightly weaker “larger models have emergent capabilities and we are probably not done finding them as we scale up”.
100% agree. I'd classify the time now as identifying the limits of what they can functionally do though, an it's a lot!
Interesting, but title is definitely clickbait.
And if you think that is amazing, my bi-metallic strip thermostat "feels" the temperature and then modifies the environment because it "knows" if it's hot to turn on the A/C, and if it's cold to turn on the heat - no training or code!
All of this AI stuff is just unbelievably incredible - what a brave new world (of word games)!
Through the story and experience of a blind man, they end up getting into the question of what does it mean to see
The podcast is pretty straightforward, but it does end up showing that defining “seeing” is a philosophical question, rather than a simple obvious answer
Anyone who works with deep architectures and momentum-based optimizers knows that the first few updates alone provide large improvements in loss. In this paper the breakthrough is that computing these first few updates at test time enables one to describe the algorithm as "without training" and therefore attract hype.
But they aren't updating the model weights. They're iteratively updating the prompt. It's automating the process that humans use with generative models.
Agreed that it's conceptually equivalent though.
This is the wrong approach to take. At minimum you have to say things like "well yes we're always on the lookout for this kind of thing". With him? Not a care in the world
Anyway, this looks like a case of human trying to understand article without reading it.
I’m guessing the iterative approach burns a lot of tokens though, though that may not matter too much with 8B Llama as the LLM.