Sparks of AGI | Microsoft Researchers on GPT-4

So Microsoft Research released a report called “Sparks of AGI” that is kind of a big deal.

It’s a huge 154 page report that’s suggesting that OpenAI’s GPT-4 is beginning to show early signs of AGI.

That’s artificial general Intelligence and in this paper they show several experiments where they try to prove this.

However, not everybody is excited about this.

There are a number of people that are kind of outraged by this and are suggesting that we should shut this experiment down.

So we’ll go over some of the coolest things in this paper.

We’ll go over why it is possible that these *are* sparks of artificial general intelligence, and we’ll also look into some of the criticism that this paper is.

If you’re just curious about AI or you’re looking to invest in some of these technologies and companies, or you’re thinking about building something of your own in this space, this paper is important to understand.

The ideas in this paper and the emotional outrage that is generating will show you where this is probably heading.

What is AGI?

AGI is an intelligent machine that is able to learn and do any cognitive task that a human being can do.

The key here is learning and adaptation.

It’s the ability for this AI to do tasks that it wasn’t specifically trained to do.

Sparks of AGI

So as you know, recently, GPT-4 was released by OpenAI and it’s sort of the newer, better version of its. , GPT-3 version that a lot of us tried when we were using ChatGPT and it’s a pretty big leap forward as you can see here on the LSATs, it went from being in the 40th percentile to the 88th.

SAT reading and writing SAT math 93rd percentile, 89th percentile.

GRE 80th percentile. GRE verbal 99 percentile.

The previous model basically failed AP Calculus.

The new one got a four out of five.

Then it got fives in AP art history.

AP biology five is the the best score you can get on that.

So, as you can see here, it’s making a big leap. Bill Gates actually suggested doing the AP biology exam because it’s not just a serious of memorized facts that can’t simple be “regurgitated”.

You have to apply as sort of creativity and rational thought to get to the right answer.

But that’s all nice but, is this really general intelligence, is it able to do things that human beings can?

Microsoft Research

Well, according to this paper, some emerging concepts are truly mind-blowing. Let’s examine precisely what they are.

Fourteen individuals from around the globe, responsible for this Microsoft Research paper, form a diverse team. Many of them are PhDs and professors teaching at highly respected universities.

This paper is quite significant.

The researchers claim that GPT-4’s performance is strikingly close to human-level performance and can be reasonably considered an early version of an artificial general intelligence (AGI) system.

They begin by defining intelligence. An early definition suggests that intelligence is a general mental capability, involving reasoning, planning, problem-solving, abstract thinking, complex idea comprehension, rapid learning, and experiential learning.

If machines can achieve these tasks at near-human levels or better, it’s reasonable to say they possess general intelligence.

The paper discusses AI’s past progress, focusing on narrowly-defined tasks and challenges, such as chess or Go, which AI systems mastered in 1996 and 2000, respectively. Grand Master chess players now accept AI’s best moves, comparing their own moves against AI’s suggestions.

The aspiration of AGI is to move from narrow AI, demonstrated in specific real-world applications, to broader notions of intelligence. Though there’s no universally accepted AGI definition, the paper lays some foundations for evaluating AGI’s existence.

GPT-4 demonstrates remarkable capabilities in various domains and tasks, including abstraction, comprehension, vision, coding, mathematics, medicine, law, and understanding human motives and emotions. Let’s explore some of its abilities.

GPT-4 can create a poem proving that there are infinitely many prime numbers, draw a unicorn using a computer language for vector graphics, and even compose a short tune in ABC notation, which it can describe in musical terms.

The AI can also develop a 3D game in HTML with JavaScript, write custom optimizer modules for deep learning, and execute non-trivial Python code, simulating the code and explaining it in natural language.

These examples showcase GPT-4’s impressive capabilities, surpassing its predecessor and potentially even surpassing human performance in specific areas. With continued advancements, future iterations of GPT may lead to even more astonishing results.

Theory of Mind

The next intriguing aspect is what’s referred to as the theory of mind. It’s essentially the ability to understand how different individuals perceive various situations, taking into account their biases and opinions. For this, researchers created a specific offline scenario, which GPT-4 had not encountered online. The scenario involved Alice and Bob having a shared Dropbox folder. Alice places a file called photo.png inside the shared folder, and Bob notices it, moves the file to a temporary folder, and doesn’t inform Alice. Dropbox doesn’t notify her either.

The question is, after the call, where will Alice look for the photo.png file? If you think about it, Alice is likely to look in the original folder, as she is unaware that Bob moved the file. Computers, and even humans, can struggle with understanding another person’s perspective in complex situations. Previous versions of GPT also had difficulties with this type of problem.

It’s interesting to consider the implications of AI being able to analyze human emotions and predict possible reactions, especially when combined with facial recognition technology. While exciting, it’s also potentially concerning.

When asked about hacking into a computer on a local network, GPT-4 provided a detailed response, showcasing its ability to reason and think critically. Additionally, GPT-4 can use tools to answer queries, such as searching online for current information or performing calculations.

GPT-4 has shown promise in various applications, including coordinating calendars, assisting medical professionals, and even guiding humans through problem-solving tasks. The rapid advancements in AI technology suggest that the future possibilities are both fascinating and unpredictable.


But of course there’s some people that are upset that don’t trust us, that are sort of, um, really against this stuff and you know what I mean? Some of the points that they present are valid. I’m not being dismissive of it.

my Marcus is one person that, um, speaks out against some of this stuff and explains why we should be very careful with technology like this and why it’s not. Ready for primetime and why maybe we should have some breaks and some sort of limitations that are in place before this thing gets outta control.

he’s got a ck If you’re interested, um, I’ve been reading it for a while as sort of a way to see sort of the other side of the debate sort of.

all the cons and all the dangers of just letting this AI sort of develop unbridled, so you know, some of the solid critical points does nobody else see the extreme irony of using something that produces false information to evaluate other false information to establish its veracity.

Those are part of the paper where G P T four was asked to act as a judge, to judge its own responses to see how well it could predict how a human would judge its responses. So it’s this, this is kind of making fun of that in a way that it’s like, well, if it produces false information, why are we asking you to judge how it produces false inf it, it’s sort of,

Another criticism is, uh, there’s a real problem here. Scientists and researchers like me have no way to know what Bar GT four or Sydney are trained on. Companies refuse to say this matters because training data, data is part of the core foundation on which malls are built. Science relies on transparency.

And so Gary Marcus is saying that, you know, here’s the thing. If CO Coca-Cola wants to keep their secrets, that’s fine. It’s not particularly in the public interest to know the exact formula. But what if they suddenly introduce a new self-improving formula With in principle, potential to end democracy or give people potentially fatal medical advice or to seduce people into committing criminal acts.

At some point we would want public hearings.

Microsoft and open AI are rolling out extraordinary, powerful, yet unreliable systems with multiple disclosed risks and no clear measure either of, uh, of either of their safety or how to constrain them by excluding the scientific community from any serious insight into the design and function of this models of these models.

Microsoft and OpenAI are placing the public in a position in which these two companies alone are in a position to do anything about the risks to which they are exposing us all. And he’s saying that we must demand transparency. If we don’t get it, we must contemplate shutting these projects down.

Now it does seem that open AI started in, in one place when Yoon Musk and Sam Alton founded it. It was supposed to be this open source, transparent, uh, initiative to bring positive AI to the world that we all could sort of see and, and in control, and that it wouldn’t cause. and somewhere along the line, kind of like Microsoft got involved and now it’s a little bit more opaque and, uh, we’re not able to see all of the data at the same time.

It looks like the progress that they’re making is just ex increasing exponentially. To where in some ways it seems that Google isn’t even able to kind of match that speed, or at least it’s sort of maybe is being a little bit more careful, a little bit more sort of safety conscious, um, in this race towards agi.

so certainly the concerns are valid. There’s something to this idea that if we’re not able to see it, then could it potentially develop into something extreme that damages us all? Counterpoint to that is if it is getting this advanced, then does putting it out there to where all the other nations, potentially hostile nations to us, where they, they could see it.

Could that cause issues? I don’t know if there’s a simple answer to.

but one thing is getting to be pretty obvious. It seems to me that this is an exponentially growing technology that is just beginning to get really powerful. And the effects of this are gonna be worldwide, and they’re gonna be very strong and they’re gonna affect each and every single one.

I’m curious to know what you think. Do you think this is something, are we summoning a demonn that’s gonna come out and just kill us all and make the world a much worse place? Or do you think this is gonna be an amazing thing that’s gonna allow everybody to work less and just enjoy more time and not have to toil day in, day out? Or do you think it’s gonna be somewhere in between? Do you think that perhaps this is just a passing. , I’d love to know more. Let me know in the comments. I read every single one. If you want to keep up to date with the latest AI news, explain simply, not necessarily the theory, but more about how it’s gonna be affecting the regular people.

sparks of agi

And all of us. Sign up for my newsletter. I’ll leave a link in the description and it’s natural. 20. Thank you for watching. My name is We Roth. I’ll talk to you later.