What is OpenAI’s super-secret Project Q*? | About That

If I asked you to solve this maze how would you do it like when the path Forks how do you decide as a human which path to take is it all just kind of you know like luck and guessing or is there some skill involved or even intelligence hold

That thought we’re going to come back to this in a second because first I need to be real with you it is very hard to know right now what qar actually is we know from reuter’s reporting that according to their sources it may be some kind of powerful artificial intelligence

Discovery at open aai the company behind chat GPT and that there are fears it is so powerful it could threaten Humanity that sounds really dramatic but this discovery was apparently alarming enough that at some point after a group of open AI researchers took their concern to the board like oh my God are

You all aware of what this company is working on the CEO Sam Multon was fired well we have the latest in The Shakespearean drama gripping artificial intelligence Sam Alman is back as CEO of open AI the previously unreported letter and AI breakthrough were key developments ahead of alman’s ouer that

Whole Saga we just did a whole other explainer on but what is this AI Discovery and what makes it so threatening open ai’s very existence is the Quest for what’s called artificial general intelligence the day will come when the digital brains that live inside our computers will become as good and

Even better than our own biological brains we call such an AI and AGI artificial general intelligence that’s like AI but way more advanced you see large language models like chat GPT have been trained to absorb patterns woven into the the structure of language and

I’ll explain what I mean so if I ask a question to chat gbt like how do I scramble an egg and it gives me you know like this whole thing about how to scramble an egg it’s using its raw computational power to discover patterns make inferences and then generate a

Response based on Words and patterns that had seen before like all the times it’s ever seen the word scramble and egg together those help lead it down a specific spefic Road about breakfast foods human brains do that too but we also think on an intuitive level because

We understand Concepts not just data and importantly if chat GPT had been trained exclusively on bad recipes it might struggle to creatively come up with a better one so uh we could also Imagine a robot a robot is acting in the world so to train these kinds of systems that

Take uh decisions and actions and then get you know some feedback back and you know the world changes and then they act again just like us so that’s yosua Benjo he’s been dubbed One of the Godfathers of AI who laid the foundation for the developments we are seeing in AI if

Anyone understands how this works it’s him reinforcement learning has been designed to train those systems based on their experience so that they can achieve goals for example just like you train your cat or your dog and if we come back to our maze how would artificial intelligence solve this well

Maybe you know it would start at the beginning and at every Fork it would just kind of like pick a path and so let’s say it goes this way and if it hits a dead end it would backtrack pick a different path and it would keep doing this until eventually hopefully it finds

The exit or maybe we can do better right we can give it some rules of thumb to follow like generally if it’s going down a path you’re looking for paths that seem to take you towards the exit so at every decision point it says okay well

The exit is down so let’s go that way and the exit’s still down if we’re going to keep moving in that direction until ultimately we get to the exit you don’t waste time right going in wrong directions those tidbits of wisdom are called heuristics and you know we can

Teach it all kinds of things like maze exits tend to be on the edges of the Maze and not say you know in the middle or in some random spot on the map but what if the AI could formulate its own heuristics on the Fly What If instead of

Bumping into every dead end it could actually think like a human and predict the optimal solution without having to pursue all the suboptimal ones how do we allow these Learning Systems to think in ways similar to how we think how we reason um how we take different pieces of information

That come to our mind and are relevant and combine them in order to come up with an answer to a question earlier this year open AI published Research into a new kind of problem solving a way of training AI by rewarding not just whether it came up with the right answer

But by rewarding the way it thinks by giving the AI model feedback for each individual Chain of Thought rather than just its final output researchers wound up with more reliable answers and an AI that could actually solve math problems that’s hard because if you remember the

Way AI normally works you know the way it thinks is by linking features of things it has studied in the past like when chat GPT knew how to scramble an egg variations on that question it has seen before help inform how to assemble words in a way that answers the question

Today but math is so much harder because instead of a finite number of words with a finite number of Arrangements we’re literally talking about an infinite problem set like if I write a random math problem on the board so you know 8 + 362 minus 147 *

89 / 45 * 67 + 111 minus you know 3 million 60,000 whatever whatever the question I want to ask I can almost guarantee that this exact problem has never been written before in the history of humankind there’s just way too many ways to arrange all of these numbers nobody

Has ever asked this so if some hypothetical AI model learned exclusively by looking for examples it would get stuck fast to independently and reliably solve math you actually have have to be able to reason these systems are not as good as us still on certain number of fronts um in

Particular things having to do with reasoning mathematics planning chat GPT for example is kind of terrible at this if I ask it 14,709 + 45 + 879 + 78 + 12 let’s see the answer it gives me 15,622 that’s the wrong answer it’s not that it’s actually supposed to be

15,723 these are like higher level cognitive functions um and and that’s a gap that currently exists uh between current AI systems and human level intelligence even even a child doing you know their math homework might be doing better in some cases but open AI is reportedly closer to building AI

That can do this kind of grade school math not just because it’s practiced on billions of problem sets but because it kind of understands math it is unknown how broadly these results will generalize beyond the domain of math and we consider it important for future work to explore the impact of process

Supervision in other domain so if we bring this all back to qar what is it is it this is it capable of solving math problems is it capable of carrying out novel scientific research on its own a lot of uh work we do in science involve

Reasoning like this it would be good to solve that problem but also it’s a bit concerning because if we can bridge that Gap to human intelligence we might have ai systems that are uh smarter than us or at least as smart as us and I don’t think Society is ready for That the sort of technological IAL change happening now that is going to so change the constraints of the way we live and the sort of economy and social structures and what’s possible um I think this is like going to be the greatest Leap Forward that we’ve had yet

Far yet so far and the greatest Leap Forward of any of the of the big technological revolutions we’ve had so far we have no way of knowing how close to the concept of artificial general intelligence qar comes but it’s fun to speculate on just how close to approaching or approx proxim ating human

Thinking open AI may be we’re working on something that will will change everything will change the way that we work uh the way that we interact with each other and the way that we think four times now in the history of open ey the the most recent

Time was just in the last couple of weeks I’ve gotten to be in the room um when we sort of like push the front the sort of the veil of ignorance back and the frontier of Discovery forward and getting to do that is like the professional honor of a lifetime I think

There’s still this contingent of people that are deeply afraid of the existential risk of AI within the company um and that potentially might do something drastic because if you put yourself in the shoes of people who genuinely believe that this could be catastrophic for Humanity like you will

Go to Great Lengths to try and prevent that from happening so there there could still be like fireworks so buckle up

A breakthrough in artificial intelligence at OpenAI preceded former CEO Sam Altman’s firing and was part of a list of the board’s grievances, according to sources cited in a Reuters report. Andrew Chang explains what we know about the AI technology referred to as Q*, and also breaks down the gap between current AI technology and “human” intelligence.

»»» Subscribe to CBC News to watch more videos:

Connect with CBC News Online:

For breaking news, video, audio and in-depth coverage:
Find CBC News on Facebook:
Follow CBC News on Twitter:
For breaking news on Twitter:
Follow CBC News on Instagram:
Subscribe to CBC News on Snapchat:

Download the CBC News app for iOS:
Download the CBC News app for Android:

»»»»»»»»»»»»»»»»»»
For more than 80 years, CBC News has been the source Canadians turn to, to keep them informed about their communities, their country and their world. Through regional and national programming on multiple platforms, including CBC Television, CBC News Network, CBC Radio, CBCNews.ca, mobile and on-demand, CBC News and its internationally recognized team of award-winning journalists deliver the breaking stories, the issues, the analyses and the personalities that matter to Canadians.

Reference