Q* Hypothesis: Enhancing Reasoning, Rewards, and Synthetic Data
interconnects.ai
interconnects.ai
Q* or any X* for that matter is extremely common for referring to the optimal function under certain assumptions. (usually cost / reward structure).
An example I heard was that one of the programmers working on the original Unreal engine saw a demo of John Carmack's constructive solid geometry (CSG) editor. He incorrectly surmised that this was a real-time editor, so he hurriedly made one for the Unreal game engine to "keep up" with Quake. In reality, the Quake editor wasn't nearly as responsive as he assumed, and in fact he had to significantly advance the state of the art to "keep up"!
If this was an extremely competitive market that'd be more plausible. But they enjoy some pretty serious dominance and are struggling to handle the growth they already have with GPT.
If Q* is real, you likely wouldn't need to hype up something that has the potential to solve math / logic problems without having seen the problem/solution before hand. Something that novel would be hugely valuable and generate demand naturally.
What is that though? I've seen a lot of tools created for it. Custom AI Characters. Things that let you have an LLM read a DB etc. But I haven't much in regards to customer facing things.
How about ChatGPT? It’s a game changer. It has allowed me to learn Rust extremely quickly since I can just ask it direct questions about my code. And I don’t worry about hallucinations since the compiler is always there to “fact check”.
I’m pretty bearish on OpenAI wrappers. Low effort, zero moat. But that’s largely irrelevant to the value of OpenAI products themselves.
I was asking today about polars, and it hallucinated all answers..
How is this helpful?
Where it is helpful is that it's way more convenient than asking a human, doing a web search, or looking it up in a book. Even if a followup question is required. In this particular case there is a lot of transferability between knowing programming languages, too, so one can filter most implausible answers even if they don't know the programming language they are asking about.
https://www.zdnet.com/article/microsoft-has-over-a-million-p...
For the most part usages of GenAI have been sharing output on social media. It is mind-blowingly fascinating, but the utility of it is far far behind.
Remember Altman saying that they shouldn't release GPT-2 because of it being too dangerous? It's the same thing with this Q* thing.
And to set a precedent that models should be released cautiously, and he was right about that too, and it is to our detriment that we don't take that more seriously.
Why?
edit: You know what, let's take a concrete issue that could happen today. You've made a generative image network. Five weeks after releasing it on Huggingface, you discover to your chagrin that the dataset that you used to train it contains an astonishing amount of child pornography, something like 1%. Your spot checks didn't find this because it's all in a subfolder that you forgot to check. Who knew it wasn't a good idea to download datasets from 4chan? As a result, this network is now extremely good at generating images of children in sexual situations, and because of mode collapse, it's creating fake images of real children, something which all but the most libertarian consider morally abhorrent. At any rate, you consider this morally abhorrent, and you'd love to work with the police to prevent any further misuse. Unfortunately, your network has been downloaded at least ten thousand times and it has already been fine-tuned to be even better at child porn by the nice folks at <insert dubious discord here>. Now you have an appointment with a senator in three days, and you have to explain to her why you thought it was a good idea to publish this network for open download, even though you could have made way more money by keeping it closed. Good luck?
Now of course you can argue that in this case all the material was already out there. But that doesn't change the fact that you were the one who did the training run, and released the network, and you're the reason why perceptual hashes now won't find collisions on the generated pictures anymore. If there was a limited amount of generated images in circulation, you could just take the API down, apologize profusely, donate 10k to RAINN or whatever and restart your project under a new name. But as it is, that option is no longer available. The point is, we don't know what a network is doing, and so we don't know what it's going to do in the wild. We cannot prove the absence of capability, so we should hedge our bets.
Company’s have competing interests and personalities. That’s normal. But there is no indication that GPT was held back for marketing.
Q* is already a thing and it's the Bellman equation describing the optimal action-value function.
Because having similar acronyms or notations used for multiple contexts that end up collapsing with cross-pollination of ideas is far too frequent these days. I once made a dictionary of terms used in A/B testing / Feature Flags / DevOps / Statistics / Econometrics, and most keywords had multiple, incompatible acceptions depending on the exact context, all somewhat relevant to A/B testing. Every reader came out of it so defeated, like language itself was broken…
1. https://spinningup.openai.com/en/latest/algorithms/ddpg.html
Language is about communication of information between parties. One instance of an LLM doing one-shot inference is not leveraging much of this. Only first-order semantics can really be explored. There is a limit to what can be communicated in a context of any size if you only get one shot at it. Change over time is a critical part of our reality.
Imagine if your agent could determine that it has been thinking about something for too long and adapt strategy automatically. Increase to higher param model, adapt the context, etc.
Perhaps we aren't seeking total AGI/ASI either (aka inventing new physics). From a business standpoint, it seems like we mostly have what we need now. The next ~3 months are going to be a hurricane in our shop.
A significant part of intelligence comes from existence in meatspace and the ability to manipulate and observe that meatspace. A two year old learns much faster with much less data than any LLM.
Beyond modeling the world, text is also a great way to model human thought and reason. People like to explain their thought process in writing. LLMs already pick up on and mimic chain of thought well.
Contained within large datasets is crystallized thought, and efficient descriptions of reality that have proven useful for processing modalities beyond text. To me that seems like a great foundation for AGI.
It's only one part, predicting text is relatively straightforward because it doesn't require predicting complex sequences like 'a S23mz s.zawsds'. Based on statistical analysis, there is a limited number of word combinations that humans use. With hundreds of billions of parameters, significant compression is possible. Mathematics is different as it requires actual reasoning, an area where LLMs often struggle significantly because they lack the capability for genuine reasoning.
These senses can be described via text, but I’m highly skeptical that the learning outcomes will be the same.
This is wrong. There’s nothing magical about human perception. You see the world because a 2D image is projected onto your retina.
GPT-4 was trained on text and generalized the ability to output 2D images. There’s absolutely nothing to suggest text can’t generalize further to new modalities. GPT4 is forced to serialize images as SVGs to output them (a crazy emergent ability btw), but that demonstrates an inherent spatial reasoning capability baked into the model.
GPT4V was created with a transfer learning step where image embeddings are passed as input in place of text. That’s further evidence of models ability to generalize to new modalities.
Everything you need to do multimodal input and output is already trained in, GPT-4V I’m sure is just the start.
And it shows. It has a poor grasp of reality. It does a poor job with complex tasks. It cannot be trusted with specialized tasks typically done by expert humans. It is certainly an amazing technical achievement that does a decent job with simple tasks requiring cursory knowledge, but that’s all it is at this time.
>There’s absolutely nothing to suggest text can’t generalized further to new modalities
Inversion of burden of proof.
Nope. OpenAI has already demonstrated the ability to generalize GPT4 to a new modality. Your claim that text models can only generalize to images and not other modalities is utterly unconvincing. Explain to me why vision is so much different than say audio?
>> And it shows. It has a poor grasp of reality. It does a poor job with complex tasks.
GPT4 is a proof of concept more than anything. I’m excited to see how much reliability improves over time. It’s grasp of reality isn’t prefect, but at least it understands how burden of proof works.
Hilarious walk-back. “Text can generalize anything” —-> “It’s just a demo, bro” in the same post.
Lmao
I don’t understand why some people have a such hard time envisioning the potential of new technologies without a polished end product in their hands. Imagine if AI researchers had the same attitude.
Technology can be both real and unpolished at the same time. Those two things are not contradictory.
https://openai.com/research/improving-mathematical-reasoning...
I am a techno-optimist and I believe this is possible and all I need is a lot of money. I think $80B would be more than sufficient. I will be awaiting a reply from other techno-optimists like Marc Andreesen and those who are techno-optimist adjacent like millionaires and billionaires that read HN comments.