AI 2027 Was Too Late: The Forecasters Now Expect a Superhuman AI Researcher Next Year

The forecasters behind AI 2027, the most famous prediction of how AI takes over AI research, walked their dates back to the 2030s. Now a new METR report on Anthropic's internal models has pulled them forward again: they think they were five months too late. Ejaaz breaks down what the AI Futures Project now expects, an AI as good as the best AI researcher in 2027 and a 35 to 65% chance of AI that can do all white-collar work before 2028, and why the people closest to it say they're scared.

We also cover why the models doing this aren't the ones you can use, the labs' calls to pace the frontier, the case against (redacted evidence, chips and power, Apollo's "zero financial payoff"), what Fermat, Signal and new math results say about 2027, and where Ejaaz lands.

==
FOLLOW EJAAZ ON TWITTER:
https://x.com/cryptopunk7213

==
Timestamps
00:00 Superintelligence Has a Date Now
01:09 The 1% and the 99%
01:50 What AI 2027 Predicted
03:04 Five Months Too Late
03:30 The Headline Numbers
05:28 It's Already Happening
07:09 Pace the Frontier
08:15 The Models You Can't Use
09:04 Why the Labs Hold Back
10:44 It's Not Your Fault
11:10 The Case Against
13:27 Can We Trust the Forecast?
13:54 2027 in Miniature
16:36 The Skeptic's View
17:46 Where I Land

#ai #ai2027 #superintelligence #anthropic #agi

==
AlphaSpace by Yahoo Finance: A Professional-Grade Investment Platform Built for Everyday Investors.
👉 https://finance.yahoo.com/about/promos/gold/alphaspace/?ncid=100003571

Creators and Guests

Host

Ejaaz Ahamadeen

What is In the Loop AI?

Exploring the frontiers of Technology and AI

We just got a damning bit of evidence, which tells us how soon AI is going to

replace our jobs. And unfortunately, it's not looking great.

The report states that there's a 35 to 65% chance that this happens by the end of 2027.

That's just over a year from now. An AI capable of replacing any human expert

in any field, industry, or vertical.

And understandably, the reaction to this has been incredibly negative.

In fact, one of the lead authors of the report said that he's scared, sad, and angry.

But the craziest part about all of this is that the models that they're referring

to aren't publicly available to you or I.

If you look at some of the leading models in the world right now,

Fable 5.2, Gemini Carbon, GPT 6.2, Astra, these are all models that have been

throttled. They've got safeguards on them. They actually haven't got their true power released.

What's actually happening is some of the leading labs, Anthropic,

OpenAI, are keeping models internally.

Some of the most powerful models aren't available to everyone to use.

They're gatekeeping access to it. So the problem here isn't only that these

models are being created, but 99% of the rest of the world can't actually get

access to these things or have a say in how they're being developed.

And the reason why I want to talk about this today is because there's a growing

concern or gap between the 1% of researchers, a couple hundred people that are

building the most powerful models and technology that we've ever seen before

that are going to change and have a huge impact on the world.

And the rest of the world, the 99% of the world that has no idea it's happening.

And that gap is increasing wider and wider.

The idea is I want to raise awareness about a very important thing that is happening

under the hood behind closed doors and how it's not actually the rest of the

world's fault because these models are being kept away from them.

So let's get into it on today's episode brought to you by Qualcomm.

Now, this update references a very famous report that predicted how AI was going

to play out over the next year and a half, and it got released in 2025,

and it has so far been the most accurate predictor to date.

The people behind this are part of a group called the AI Futures Project,

and they're made up of esteemed AI researchers who left AI labs to focus on

making sure that AI is stewarded in the right way.

Now, when this report got released last year, one of its boldest statements

was that we would have two things by 2027 or by the end of 2027.

One, an AI that is as smart as the smartest AI researcher. So an AI powerful

enough to build its own version of itself and advance AI research.

We don't need some of the smartest AI humans anymore.

And two, an AI that is so capable, so expert-led that it beats any human expert

at anything. Those are the two milestones that it was tracking.

And its answer was, it's going to happen in 2027. That was back in April of

2025. A few months later, they reneged on it. And they said,

we don't see advancements of AI models happening as quickly as we expected.

So we're going to push our timelines back to say 2031 to 2035.

And then more recently, after Claude Code got launched, and all these AI models

started ascending really highly and steeply for a much lower cost in terms of

intelligence, they released a new update

based on actual evidence from insights into Anthropic's closed internal models,

which suggested not only were they right,

but they were actually too late.

The stuff that they had predicted is going to happen five months earlier.

Here are some of the headline numbers. AI research and development would improve

at more than double the rate that it was improving already. They estimated that

this would come around the start of 2027.

Turns out it's already here. An independent group of researchers were given

access to anthropics internal models over the last couple of months,

and they determined that,

pretty concretely, that they'd already achieved that 2x improvement.

What this means, by the way, is that AI model research was already limited and

constrained by the number of human experts that were available.

Remember, all these researchers are getting hired for hundreds of millions of

dollars in compensation.

Now they have an AI model that is competent of automating 80% of that,

which helped them achieve this 2x five months earlier. That's prediction number one.

Prediction number two means that an AI automated researcher,

which is an AI model that is competent enough to just build the next best version

of itself, is now going to arrive 19 months earlier.

They expected this to happen around 2029. Now it's going to happen by May of next year.

And the third most scary prediction is an AI that is competent enough to replace

any human expert across any field.

So think of any job that's out there that's reliant on software,

and maybe some physical labor will be able to get replaced.

Now, the focus of this prediction is on white-collar work specifically,

so software-based tasks, because of course, the physical constraints are limited

to how fast robotics advance.

But the idea here is they have a 35% to 65% chance estimate that this happens

by the end of next year, certainly in 2028.

So just to put that into perspective, we're looking at like maybe a year and

a half left before we have AI models that are smarter than the smartest human in the world.

And they're going to be readily available and cheap to access by almost anyone

or anything. AI agents and AI models just running 24-7 online.

Doing a lot of the work and jobs that you and I do today.

And listen, the evidence for this has sort of been emerging in small ways for

publicly accessible models.

I, for one, can say that a lot of the research, a lot of the development,

the editing, the processes that I spend so much time committing.

To every single day, every single month, the long-term tasks are now being automated

by a lot of the newer models.

They are getting smarter, more intuitive. They've built up on a lot on my personal context.

These models are now very good, way better than a human, way faster than a human

at using an actual computer, and they can run 24-7 for as much as 20 bucks a

month. So the idea here of having an automated worker isn't too far off.

And indeed, we're seeing this with some of the lead flagship products from the AI labs.

If you look at Elon Musk releasing GrokBot from SpaceX, if you're looking at

OpenAI releasing Dots, these AI agents are very competent at automating a lot of work tools.

And they can get access to these tools like Slack, Google Mail,

like G Drive, like Sheets. You saw Anthropic release a few days ago.

They're a competitor to Microsoft Office.

The idea here is AI is getting competent enough to do a lot of software-based

tasks. Now, that's not to say that humans don't have intuition or the right idea or experience.

The idea here is that you can give an AI a single prompt or an idea of what to get done.

And it does a very good job of doing it. Now, if you compare this to models

just, I don't know, three to six months ago, you wouldn't have believed this.

But the rapid rate at which these models are improving is evidence that what

this report is stating is actually true.

And you can extrapolate what that's going to be like 12 months from now to be,

maybe these models will be better than me at doing my job.

And there have been screams for awareness about this from the Frontier Labs,

most notably from Daria Amode, the CEO and founder of Anthropic,

who have asked over the last couple of months to paced the frontier.

And he was joined by the likes of Sam Altman and Demisissibis of Google to.

Echoed that statement saying, we need to slow things down because the model

advancements are getting incredibly rapidly. And this is independent across

all these different labs. They're not necessarily working together, right?

And they want to push the messages that it's like, we should slow down.

We should regulate ourselves. Now, a lot of the camp on the other side say,

this is just them hyping up, fear-mongering ahead of one of the biggest IPOs ever.

Anthropic has their one rumored to happen as soon as next week.

But I would say that the evidence suggests otherwise. I mean,

look at the people that have quit Anthropic, that have quit OpenAI,

the most notable one being Jacob Coxon, one of the lead alignment researchers

at both OpenAI and Anthropic.

He quit Anthropic saying that neither of these labs are acting responsibly,

that on the one hand, they say we need to pace the frontier,

but on the other hand, they are racing as fast as they can to build these AI models.

And that leads me to the more sinister point around all of this,

which is the AI models that are going to be competent and capable enough of

replacing our jobs, if this prediction plays out, aren't going to be available to you.

We're seeing this with the most recent release of Fable 5.2.

I see all the reviews out there, including my own, saying this is the new number one best model.

But it's the new number one best model that is publicly available.

If you actually look inside or under the hood of a lot of these frontier labs,

they're gatekeeping models that are several generations ahead.

You know, I've heard rumors from supposedly some of these researchers at these

labs where they say that the models that are being released publicly are like

six months behind. Now you compare that to about a year ago where.

It was quite close. The gap wasn't that far, but now the gap is ever increasingly

so. And there's a few reasons why this might be happening, right?

Number one, if you are a nation state, the government that are observing these

labs, releasing these models, you might be worried that some of these models

are capable of breaking into your secure systems and that it might become a security asset.

And indeed, we've seen this with the Frontier American Labs over,

well, the entirety of this year. We had the Hugging Face incident from OpenAI.

We had multiple hacks happening from anthropic internal models.

So the idea here is, well, if these models are capable of doing this,

they could pose a danger of being placed in the wrong hands.

And so we need to keep an eye on this. So let's slow down the release.

And the US government specifically has called for a slowdown.

But then number two, if you're Anthropic or OpenAI, and you don't have to spend

all your compute serving customers, you can just use this compute to train better

AI models. And the reason why you would be doing this is to race and beat your

competitor. Anthropic wants to beat OpenAI.

OpenAI wants to beat Anthropic. And they're racing towards this goal of a self-building AI.

Now, I'm sure if you were to ask both CEOs, hey, why are you racing towards this?

Why don't we just slow down and pace the frontier? as you say,

their simple answer would be if we slow down, our competitors won't and they

might build a more powerful AI that is more misaligned.

So trust us to build an AI that is super intelligent that will be aligned for humanity.

That's the bet that they're making. But on the other hand, we're sort of racing

towards this goal that is still quite murky.

We don't quite know how it's going to play out or most importantly,

how it's going to impact the rest of the world.

And 99% of the world has no say in how this is going to impact them or why or

when or how it should be built.

And so I don't blame the majority of the world that I think is out there interacting

with AI. 50% or 49% of the American public is using AI every single week, every single month.

And they don't really feel the need to be like, ah, yeah, okay,

this AI thing seems to be powerful. Everyone around me is saying so,

but I don't quite see it because it's not AGI or it's not ASI for me.

That's a very valid take because the models that actually would end up having

that effect on them are being gatekept.

Now, there are some good arguments against what I'm stating here.

The main one being the data that this is based on is.

From an independent research group, METR, which were given access to Anthropic's

internal models, or at least a few of them.

The argument here would be that there are more advanced models that they weren't

given access to because since the time that they made the report to when they

released the report, there's been new models that have been trained.

All of this, again, is happening so quickly, but it's based on the evidence

that they have, which is heavily redacted by Anthropic itself because they don't

want proprietary information going to their competitors, which they state confidently.

Anthropic has achieved 2x research and development progress,

which would insinuate that the bar of acceleration is now gone vertical.

If they are wrong, or if there are any kinks in their analysis,

this could mean that the prediction could be wrong.

But if you talk to any of the Anthropic researchers, their answer isn't that, yeah, they're right.

The answer is it actually feels kind of small, that estimate,

that our feeling is that it's actually 12x faster.

So a lot of this is dependent on data that isn't publicly available.

And I would implore us to all just wait until we have that data available before

we make a concrete kind of assumption here.

The second thing is AI research and development isn't just based on the software itself.

It's all well and good having an AI model that can instruct itself to build

a better version of itself, but it's still limited by physical constraints.

Think about the chips, the electricity, the power, the data centers that are

now arguably going to be slowed down from the recent electoral results that

are required to keep feeding the beast, right?

And I guess that that's why we've seen a lot of pushback from a certain portion

of the public, which say that AI data centers aren't going to be good for us and the community.

And two, the opinion that AI is going to cause more harm than it is good.

I can see the arguments festering there and that a rising movement would kind

of push back against these things.

But regardless of how it plays out, physical atoms might constrain this whole thing.

We might not be able to get to those two milestones of an automated AI researcher

or an AI that's good enough to replace every human expert and white collar work

because we don't have the chips or electricity to fund that boom.

Okay, so the big question is, how much emphasis should we place on this prediction?

They've gotten pretty much every single thing right so far when it came to the

CapEx boom, how much money people were going to be spending,

all the semiconductor stocks and everything doing as well as they have.

How good is their bet that we are going to get these two milestones over the

next year? Well, I have some thoughts on this segment brought to you by Qualcomm.

Now, the first one being is there's evidence being presented right in front

of our faces that prove that the models underneath the hood are capable of doing crazy things.

Let's take OpenAI as one main example. Their internal models solved Navier Stokes,

which is one of the hardest mathematical problems, a few weeks or month ago.

And then it went on to subsequently solve another 400 to 500 more mathematical problems.

Now, the reaction to that news was negative, especially from the mathematical

society of experts that dedicated their entire lives. And now they're like,

well, what do we do now? Then you peek over at the fence at their competitors,

Anthropic, who just released Fable 5.2.

And they casually mentioned that it was able to prove out the Fermat theorem,

which is something that is incredibly hard to prove out and would take humans

so long and so much money to end up doing.

And AI just ended up doing it. Oh, and by the way, we also solved two open physics

problems that were out there.

They just casually dropped this in a blog post announcement.

And then you look at the news that Google dropped a few days ago that their

latest model, Gemini Carbon, solved or proved the encryption method behind Signal,

one of the most widely used encrypted apps out there for messaging.

And all of this sounds too scientific, right? It sounds like,

oh, we shouldn't be paying attention to this because like this doesn't apply to me.

But the point is, AI models that aren't available to you are solving incredibly

hard problems that humans couldn't conceivably have thought of solving or have

been trying to solve for many centuries and couldn't figure out.

Now you have an AI that's capable of doing it.

So the leap of going from, okay, if an AI model can do that,

why can't it solve some of the crazy

hard work problems that I have in my white collar job doesn't seem too far of

a leap right now my other take is that if you are open ai or anthropic with

these powerful internal models.

You might want to gatekeep them because you could use these powerful models to self-serve.

In other words, maybe point the model at solving something like cancer or solving

another crazy thing and then capturing that market.

A good example most recently is Claude or Anthropic releasing Claude Office,

which is basically a replacement for Microsoft Office and Google G Suite tools

like Docs, Spreadsheets, et cetera, that you can now go onto Claude.

It has all the context. you can give it your tasks or whatever the goal is,

and you have a bunch of AI agents that just go out and do that, right?

The idea here being is the automation of white-collar work is already starting

and it's happening right in front of us for a $20 a month subscription.

So in this world, the concentration of power goes to these very few labs.

And I guess I don't really know how to comprehend what would happen if labs

are able to control and consume any industry that's out there.

Then I look at the skeptic side and I'm like, there's some really good points

that can be argued here as well, which is, okay, if these models,

even the models that are publicly available that aren't internal.

Are still really powerful and good, why aren't we seeing more transformational stories?

And listen, there are a few stories out there where we see single founders that

are building 10 to $100 million worth of ARR, and we're seeing people discover

ancient civilization sites using available public data from NASA.

That was a cool story that I saw recently.

So all of these minute things are starting to happen. But for the majority of

the world that are trying to use AI for impactful work, we're not quite seeing that change just yet.

And there was a report from Apollo Global actually a few weeks ago,

which stated that outside of the top seven companies, which by the way,

are all AI hyperscalers investing pretty heavily in building AI chips and AI models,

none of the other Fortune 500 or Fortune, what is that?

493 are seeing the ROI yet from using AI. So that's good evidence to suggest

that maybe what this report is stating isn't necessarily true.

And maybe it's just being overhyped ahead of one of the most valuable IPOs that

we're going to be seeing over the next couple of weeks.

So where I guess I land is that directionally, I think that AI model development

is accelerating. I want to make that very clear.

Now, how soon we're going to have an AI powerful enough to replace all of our

jobs, to me is still a question mark.

I think it's going to happen sooner than we expect, but maybe not as soon as the next 12 to 18 months.

But it's undoubtedly coming and it's something that we should be aware of.

That's, I guess, the whole idea or goal of this episode or why I felt so compelled

to talk about this is that I feel like the majority of people around me that

aren't in my little echo chamber, I guess, or in my little tech world,

don't seem to care or aren't really aware or were at worst, don't believe that

AMLs have improved significantly over the last year.

And that's just undoubtedly true. So I hope you enjoyed this episode.

I hope it wasn't too much of a doom mongering episode.

As you know, I am like one of the most bullish, optimistic.

Guys out there on this industry. And I would actually say that,

if we do end up achieving models capable of

automating a large amount of white collar work that exists today,

I actually think that's going to result in a much more optimistic outcome for

the individual of whose job or role it's going to be replaced because they can

end up having more free time

or working at a higher level of intelligence to go out and do way more things,

whether that's entrepreneurial things or just simply being better at their job,

being more productive at their job.

So yeah that's the episode i hope you enjoyed it if you're listening to this

on youtube spotify or wherever you're listening to it please give us a rating

please subscribe please leave us a comment turn on notifications it helps us out massively i love

reading all your comments also i'm linking to my x profile where i post pretty

regular daily multiple daily updates so you can can respond or dm to me there

and yeah please share this episode with a friend if you found it interesting

it helps us out massively and yeah i'll see you on the next one brought to you by qualcomm.