OpenAI 5

As I read it, we have Bill Gates and Sam Altman out there in the world trying to round up something like 9 TRILLION dollars for the next version of OpenAI.....

I do not believe that is merely to produce an AI system that is sentient; I believe that OpenAI 4 whichis out there now is sufficiiiently sentient for any normal or reasonable purpose. I believe they are looking at recreating a major aspect of antediluvian reality i.e. a planetary super mind.

I believe that Q* should be treated as a 2024 version of something like the Manhatten project and that government has a role to play, including:

2, Determine if the OpenAI view of requirements is reasonable or if thre is some less gigantically expensive approach Inpuut from OpenAI's competitors would be required for that.

3. As per LaRouche/PrometheanAction suggestions, create a third national bank to finance such projects as was being done in the 19th century.

4.Insure that the projec t does not end in a manner siimilar to the 1956 Scyfi classic about the race of super beings that destroyed themselves in a quest for godlike powers.

I asked ChatGPT to format the video transcript and here it is (for speed readers):

For those of you who don’t remember, QAR is the secretive project inside of OpenAI that is essentially a new way for artificial intelligence to do logic, reasoning, and planning. All we’ve had are little drips and drabs of leaks out of OpenAI, and now we have more information in the form of a project code-named Strawberry. That’s what we’re going to be covering today.

Not only that, OpenAI has laid out the foundation for how they’re going to determine when we actually hit AGI and the different levels before that. Today we’re going to be covering this article from Reuters by Anna Tong and Katie Paul, titled “OpenAI Working on a New Reasoning Technology Under Code Name Strawberry.”

Strawberry is the new code name for QAR. ChatGPT maker OpenAI is working on a novel approach to its artificial intelligence models in a project code-named Strawberry, according to a person familiar with the matter and internal documentation reviewed by Reuters. Reuters could not ascertain the precise date of the document, but it details how OpenAI intends to use Strawberry to perform research. This is just another extension of the whole QAR saga, which is all about logic, reasoning, and most of all, planning. Today’s large language models just do not do this well; they are not able to actually plan ahead without the help of agentic frameworks around them.

The Strawberry models aim to enable the company’s AI not just to generate answers to queries but to plan ahead enough to navigate the internet autonomously and reliably perform what OpenAI calls deep research. One thing I’ve been thinking a lot about lately is something that Sam Altman said a couple of months ago about deep research in two ways: post-training and pre-training. Sam Altman mentioned that the total set of data available to train models is not really growing that quickly. So, you have two options: you can have other models create synthetic data to then train new models on, but I’m not sure how well that approach is going to work because it’s all going to be derivative data. The other option is to allow models to do a lot more with the existing data.

What does that actually mean? Well, when a model is trained, it trains on a set of data. But models that are pre-trained in this way basically look at the data once, and that is not how the human brain works. If you think about it, if you’re trying to learn something sufficiently complex, you read about it, take your time to think about it, reference other information sources while building up your own knowledge about it. It takes weeks, months, and sometimes even years to learn about a subject in depth. That is just not something that large language models do today. But what if they could? What if they could actually take that same set of data and think about it over and over again in different ways to come up with potentially the best set of knowledge?

Post-training is kind of similar. When you prompt a model, it gives you the answer to your prompt, but it doesn’t really think about it; it just gives you the first answer that it can come up with. What happens if it has time to think and is actually able to iterate? We achieve that through a number of different prompting techniques today, like chain of thought and even mixture of agents in a sense. But large language models by themselves can’t do that. That is what QAR, now code-named Strawberry, is aiming to do apparently.

Here, OpenAI actually commented on this story. An OpenAI company spokesperson said in a statement: “We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a practice in the industry with a shared belief that these systems will improve in reasoning over time.”

I’ve actually heard a few things about what QAR could actually be, including everything from planning to deep research, but also being really good specifically at math. Let’s keep reading. A different source briefed on the matter said OpenAI has tested AI internally that scored over 90% on a math data set, a benchmark of championship math problems. However, they couldn’t determine if this was the Strawberry project.

On Tuesday, at an internal all-hands meeting, OpenAI showed a demo of a research project that it claimed had new human-like reasoning skills, according to Bloomberg. This article is also tied to the release of information about how OpenAI plans to gauge where they are in the AGI timeline, which I’ll go over in a moment.

Here’s a really important paragraph because it is very true for how models work today: large language models can already summarize dense text and compose elegant prose far more quickly than any human. The technology often falls short on common sense problems whose solutions seem intuitive to people, like recognizing logical fallacies and playing tic-tac-toe. We’ve seen these models get a lot better at these logic and reasoning problems. If you’ve watched any of my videos, I run them through a series of logic and reasoning tests. Over the last six months, and especially the last four months, we’ve seen tremendous progression in the large language model’s ability to solve some of these problems, but they still fall pretty short of getting it right consistently.

AI researchers interviewed by Reuters generally agree that reasoning in the context of AI involves the formation of a model that enables AI to plan ahead, reflect on how the physical world functions, and work through challenging multi-step problems reliably. A few things here: AI that can plan ahead is not how LLMs work today. You are required to either prompt them to do this, but even that goes only so far. However, if you’re able to give them memory and wrap it in an agentic framework, then you’re starting to achieve the ability to plan, especially when you combine multiple models together.

Reflecting on how the physical world functions, Yann LeCun has talked a lot about large language models being world models, world simulators, and that the Transformers architecture alone is not enough to achieve that. He believes that other people think that scaling laws and just using Transformers models, the same thing we have today, is enough to be able to have world simulations. He gives a number of examples of how large language models are not able to really simulate the physical world today because he basically says language alone is not enough to model the world. An example he gives is, say you’re starting at the North Pole, and someone tells you to walk 200 meters in one direction, then turn left, walk 200 meters, and then asks you if you will ever cross over your first point. When you’re thinking about that in your head, you’re not actually using any language whatsoever; you’re using spatial reasoning and world modeling in your head. This is super easy for humans to figure out, but it’s actually quite difficult for large language models to figure out. That’s his point.

Next, it says they want this new Strawberry model to be able to work through challenging multi-step problems reliably, something again that multi-step planning and multi-step problems LLMs generally struggle with. Improving reasoning in AI models is seen as the key to unlocking the ability for the models to do everything from making major scientific discoveries to planning and building new software applications. If we do have this greatly enhanced planning and reasoning from AI, we might actually be at that point of AGI. If you watched my video about Leopold Aschenbrenner’s situational awareness paper, you know that AGI is just one step before superintelligence, and this is probably one of the ingredients to hitting AGI.

Here’s a quote from Sam Altman earlier this year: “The most important areas of progress will be around reasoning ability.” Right here, they actually mention Yann LeCun, for instance, one of the pioneers of modern AI. Yann LeCun, who works at Meta, has frequently said that LLMs are not capable of human-like reasoning, which is what I just mentioned.

OpenAI has teased a number of new technologies, such as GPT-4’s voice, and they keep delaying it or having no firm date of when they’re going to release it. In fact, GPT-4T, which was really just a much faster version of GPT-4, is the only recently released innovation that we’ve had from OpenAI in quite a while. So, we’re kind of due for a big release from OpenAI, which is exciting, but we haven’t had one yet. In the article, they say in recent months, the company has privately been signaling to developers and other outside parties that it is on the cusp of releasing technology with significantly more advanced reasoning capabilities, according to four people who have heard the company’s pitches.

A little bit more on Strawberry now: it includes a specialized way of what is known as post-training. OpenAI generative AI models are adapting the base model to hone their performance in specific ways after they have already been trained on reams of generalized data, one of the sources says. Now, that’s interesting because if it required training a new model from scratch, that would be an incredibly big monetary and time investment by OpenAI. If this technology is able to take an existing model and do post-training on it, that’s going to decrease the time to release and decrease the cost of training it tremendously.

Strawberry has similarities to a method developed at Stanford in 2022 called Self-Taught Reasoner, or STAR. We’ve covered that in a previous video. STAR enables AI models to bootstrap themselves into higher intelligence levels via iteratively creating their own training data. In theory, it could be used to get language models to transcend human-level intelligence.

More on Strawberry: it is aiming to perform long-horizon tasks, referring to complex tasks that require a model to plan ahead and perform a series of actions over an extended period of time. This is something Sam Altman has talked about, both pre- and post-training. Rather than the model simply taking your prompt and spitting out the first thing it thinks of as soon as it can, it will actually take seconds, minutes, hours, even days potentially to think through it, come up with the best possible answer, and then give it to you. That might actually be a setting that you can give to the model, saying, “Hey, take as much time as you need,” or “I need this within 10 minutes,” and the longer you give it, the better the result will be. Currently, there is not any really known technology that can do that, and that would be an enormous unlock if it could be done like that.

Imagine that in a multi-agent system: you might have a few agents working together in real-time, and then they task a long-thinking agent to go off and take its time while it’s doing other stuff. Then that long-thinking agent comes back, reports what it’s found to the other agents, and then they put it all together and deliver you whatever that deliverable was. OpenAI specifically wants its models to use these capabilities to conduct research by browsing the web autonomously with the assistance of QUA, a computer-using agent that can take actions based on its findings, according to the document and one of the sources.

That’s it. Great reporting by Anna Tong and Katie Paul. Thank you for putting this great article together. Now, I want to show you one last thing. This is an article by The Verge about how OpenAI will actually determine how powerful its AI models are. You’ll be a little bit surprised to hear that its current iteration of ChatGPT is only level one of five. They really think that the current version of ChatGPT is really dumb, and that’s exactly what Sam Altman has said in the past. He has said the current version of ChatGPT is really bad, and in fact, it will be the worst it ever is right now. As impressive as it has been over the last year to see ChatGPT and its iterations in performance, we are still just at the very beginning.

OpenAI has created an internal scale to track the progress its large language models are making towards artificial general intelligence. So, ChatGPT today is level one, and OpenAI claims it is nearing level two. Level two is defined as a system that can solve basic problems at the level of a person with a Ph.D. Level three refers to AI agents capable of taking actions on a user’s behalf. In my opinion, this level three is going to change humanity. When you can have agents that are working on your behalf and actually accomplishing useful tasks 24 hours a day, not simply answering world knowledge questions, that is a tremendous productivity increase for humanity. When Apple made their Apple Intelligence announcements, they were just scratching the surface of this, where you can get Siri to do things on your phone, which is super interesting. I have a lot of thoughts about what the architecture of this needs to be, but it certainly needs to be local. That is because, for an AI agent to actually accomplish things on my behalf, I want it to have access to all of my private information, and thus I don’t want it leaving my phone. I definitely don’t want to send it to OpenAI. So, all of this needs to be done locally on the phone or on any edge device.

Here’s the thing, and I have a video that I’m creating all about this: I strongly believe in a future where we’re going to have very narrowly defined, very vertical models that are small, hyper-efficient, low-cost, and can run on edge devices. More on that in a coming video.

Next, level four involves AI that can create new innovations, such as scientific discoveries, math discoveries, and much more. If you remember the chart of Leopold Aschenbrenner’s situational awareness paper, that would be when we reach AGI. At that point, when we reach AGI and AGI can go do research about itself and self-improve, that is when we’re going to have the superintelligence explosion. Level five, the final step to achieving AGI, is AI that can perform the work of entire organizations of people. This ranking feels very similar to how they ranked autonomous driving vehicles: level zero, level one, level two, and so on, where at level five you basically don’t even need to pay attention, the car is just driving you.

That’s it. What do you think of QAR? Do you think Strawberry is the next iteration of QAR? Do you think it’s coming soon? Is that GPT-5? Let me know in the comments. If you enjoyed this video, please consider giving a like and subscribe. I’ll see you in the next one.

Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.