Halting is sometimes preferable to thrashing around and running in circles.
I feel like if LLMs "knew" when they're out of their depth, they could be much more useful. The question is whether knowing when to stop can be meaningfully learned from examples with RL. From all we've seen the hallucination problem and this stopping problem all boil down to this problem that you could teach the model to say "I don't know" but if that's part of the training dataset it might just spit out "I don't know" to random questions, because it's a likely response in the realm of possible responses, instead of spitting out "I don't know" to not knowing.
SocratesAI is still unsolved, and LLMs are probably not the path to get knowing that you know nothing.
> if LLMs "knew" when they're out of their depth, they could be much more useful.
I used to think this, but no longer sure.
Large-scale tasks just grind to a halt with more modern LLMs because of this perception of impassable complexity.
And it's not that they need extensive planning, the LLM knows what needs to be done (it'll even tell you!), it's just more work than will fit within a "session" (arbitrary) and so it would rather refuse than get started.
So you're now looking at TODOs, and hierarchical plans, and all this unnecessary pre-work even when the task scales horizontally very well (if it just jumped into it).
I would consider that detecting your own limits when trying to solve a problem is preferable to having the illusion of thinking that your solution is working and correct.
Ah yes, the function that halts if the input problem would take too long to halt.
But yes, I assume you mean they abort their loop after a while, which they do.
This whole idea of a "reasoning benchmark" doesn't sit well with me. It seems still not well-defined to me.
Maybe it's just bias I have or my own lack of intelligence, but it seems to me that using language models for "reasoning" is still more or less a gimmick and convenience feature (to automate re-prompts, clarifications etc, as far as possible).
But reading this pop-sci article from summer 2022 seems like this definition problem hasn't changed very much since then.
Although it's about AI progress before ChatGPT and it doesn't even mention the GPT base models. Sure, some of the tasks mentioned in the article seem dated today.
But IMO, there is still no AI model that can be trusted to, for example, accurately summarize a Wikipedia article.
Not all humans can do that either, sure. But humans are better at knowing what they don't know, and deciding what other humans can be trusted. And of course, none of this is an arithmetic or calculation task.
I’ve had models “redirect the problem to someone who has a greater likelihood of not failing”. Gemini in particular will do this when it runs into trouble.
I don’t find all these claims that models are somehow worse than humans in such areas convincing. Yes, they’re worse in some respects. But when you’re talking about things related to failures and accuracy, they’re mostly superhuman.
For example, how many humans can write hundred of lines of code (in seconds mind you) and regularly not have any syntax errors or bugs?
> I’ve had models “redirect the problem to someone who has a greater likelihood of not failing”. Gemini in particular will do this when it runs into trouble.
I have too, and I sense that this is something that has been engineered in rather than coming up naturally. I like it very much and they should do it a lot more often. They're allergic to "I can't figure this out" but hearing "I can't figure this out" gives me the alert to help it over the hump.
> But when you’re talking about things related to failures and accuracy, they’re mostly superhuman.
Only if you consider speed to failure and inaccuracy. They're very much subhuman in output, but you can make them retry a lot in a short time, and refine what you're asking them each time to avoid the mistakes they're repeatedly making. But that's you doing the work.
> For example, how many humans can write hundred of lines of code (in seconds mind you) and regularly not have any syntax errors or bugs?
Ez, just use codegen.
Also the second part (not having bugs) is unlikely to be true for the LLM generated code, whereas traditional codegen will actually generate code with pretty much no bugs.
I'm highly skeptical about this paper just because the resulting images are in color. How the hell would the model even infer that from the input data?
It is an overfitted model thst use WiFi data as hints for generation:
"We consider a WiFi sensing system designed to monitor indoor environments by capturing human activity through wireless signals. The system consists of a WiFi access point, a WiFi terminal, and an RGB camera that is available only during the training phase. This setup enables the collection of paired channel state information (CSI) and image data, which are used to train an image generation model"
That's just a diffusion model (Stable Diffusion 1.5) with a custom encoder that uses CSI measurements as input. So apparently the answer is it's all hallucinated.
You can't even pick colour out of infra-red-illuminated night time photography. There's no way you can pick colour out of WiFi-illuminated photography.
There would be some correlation between the visual color of objects and the spectrum of an object in another EM frequency, many object's color share the same dye or pigment materials, but it seems pretty unlikely that it would be reliable at all with a spectrum of different objects and materials and dyes because there is no universal RGB dye or pigment set we rely upon. You can make the same red color many different ways but each material will have different spectral "colors" outside of the visual range. Even something simple like black plastics can be completely transparent in other spectrums like the PS3 was to infrared. Structural colors would probably be impossible to see discern however I don't think too many household objects have structural colors unless you got a stuffed bird or fish on the wall.
If it sees the shape of a fire extinguisher, the diffusion model will "know" it should be red. But that's not all that's going on here. Hair color etc seems impossible to guess, right? To be fair I haven't actually read the paper so maybe they explain this
This is largely guesswork but I think whats happening is this. The training set contains images of a small number of rooms taken from specific camera angles with only that individual standing in it, and associated wifi signal data. The model then learns to predict the posture of the individual given the wifi signal data, outputting the prediction as a colour image. Given that the background doesn't vary across images, the model learns to predict it consistently with accurate colors etc.
The interesting part of the whole setup is that the wifi signal seems to contain the information required to predict the posture of the individual to a reasonably high degree of accuracy, which is actually pretty cool.
What about the next fab? And the fab after that? Yes, for the short term the USA has a cutting edge fab on US soil, but what about 10 years after that? The US would lack the ability to build future cutting edge fabs.
Having TSMC build a fab on US soil doesn't prevent any parallel efforts. If anything having the workforce here would make it easier to build future fabs.
I never said it did. Having a fab on US soil is important for the short term, but you need cutting edge fabs on US soil in the future too and that doesn't happen unless you are doing a lot of research.
Okay, so Taiwan is blockaded or invaded by China. The USA has a 3nm TSMC fab that they can assume control over, and, yes, they have the labor of that fab, great, but what about 2 nm? 1 nm? Etc? Without TSMCs R&D does the US have a cutting edge fab in 10 years? 20 years? Beyond? There is literally no other company in the United States that could even hope to expand their capabilities to be considered on the cutting edge within the next 15 years.
The easiest way to get TSMC's R&D in the US is to have them build a fab here and have employees here. If China invades Taiwan, and TSMC has employees that want to flee to somewhere else, the US would be the most logical option. If they already have a fab here, an established workforce and infrastructure here, that's better than having to start from nothing.
Taiwan as a country doesn't want that. There is a reason they only fab their latest and greatest chips domestically in Taiwan. It is of Taiwan's national interest and of TSMCs interest to do everything they can to ensure that the US protects them from China.
Again, it is a net positive for the US to have TSMC manufacture chips on US soil, so I am not arguing against that point, I simply posit it is not enough from a US national security / technology leadership standpoint.
