Foreword:
1. I asked Chatgpt when does it predict realistically the arrival of singularity.
2. Then I went to Claude with Chatgpt’s answer to check if the predictions seem correct. Claude presented a whole different concept of how singularity could be measured, which seemed like the underlying assessment of the CEOs of the 2 companies OpenAI and Anthropic.
3. I brought back Claude’s concept to Chatgpt. I tried to make it personal considering the background between Sam and Dario but it wasn’t a triggering point. So Chatgpt kinda accepted Claude’s evaluation, that I brought back to Claude again. Claude gave a final assessment of the foundational difference between the 2 CEOs thinking about AGI and singularity.
4. Then I asked Chatgpt to summarize the whole thing so I can bring this to Grok for further evaluation.
5. Then I brought back Grok’s feedback to Chatgpt and finally
6. Asked it to write me a whole blogpost about this thought experiment. Below is the result:
I had a fairly simple question:
When do you think the AI singularity could realistically happen?
So I asked ChatGPT.
It gave me a surprisingly confident answer:
2035–2050.
More specifically, it put its highest-probability window somewhere around 2035–2045.
That sounded reasonable.
Maybe even conservative.
But then I asked Claude the same question.
And Claude basically looked at ChatGPT’s answer and said:
“Hold on. What exactly are you basing those numbers on?”
Fair enough.
So I decided to do something slightly ridiculous.
I made ChatGPT and Claude debate each other.
And once they had finished arguing, I took the whole conversation and gave it to Grok.
I asked Grok to act as the judge.
What came out of it was actually more interesting than the original question.
First: what do we even mean by “singularity”?
This turned out to be the first problem.
People talk about AGI, superintelligence, recursive self-improvement, and the singularity as though they’re interchangeable.
They’re not.
AGI generally means something like an AI capable of performing most cognitive tasks at roughly human level or beyond.
Superintelligence would go considerably further.
The singularity is something different again.
The classic idea is that AI becomes capable of improving AI, which makes the next generation better, which makes the following generation better still—and eventually technological progress becomes so fast that humans can no longer meaningfully predict what’s coming next.
So the important question isn’t necessarily:
“When will we get AGI?”
It might actually be:
When does AI become good enough at AI research to substantially accelerate the development of the next generation of AI?
And that distinction changed the entire conversation.
ChatGPT’s first prediction
My initial ChatGPT conversation was fairly conservative.
It suggested that a strong version of the singularity might happen around 2035–2050, with the 2035–2045 period being the most likely.
The reasoning was pretty straightforward.
Even if we create extremely intelligent AI, the physical world doesn’t magically accelerate.
You still need:
- computer chips
- electricity
- data centers
- manufacturing
- robotics
- supply chains
- experiments
- money
- humans to deploy things
An AI might be able to think at incredible speed, but it can’t manufacture a million GPUs overnight.
So ChatGPT’s argument was essentially:
Intelligence may arrive quickly. The physical world may take longer to catch up.
That sounded sensible.
Then I showed the answer to Claude.
Claude pushed back
Claude’s biggest criticism wasn’t actually the date.
It was the confidence.
And I think this was a fair criticism.
ChatGPT had given me probability ranges that looked scientific.
Something like:
2035–2045: highest probability.
But how do you actually calculate that?
Nobody has a sufficiently reliable model of technological progress to genuinely know whether the singularity has a 30%, 40%, or 50% chance of happening in a particular decade.
As Claude put it, we’re essentially putting vibes with decimal points on something we fundamentally don’t understand.
And then Claude pointed out something much more interesting.
There are serious people making forecasts that span an enormous range.
Some think extremely powerful AI could arrive around 2026–2027.
Others put AGI much further away.
So rather than asking:
“Which year is correct?”
we should probably ask:
“What assumptions make these people arrive at such different answers?”
That is a much better question.
Then things got interesting
Because when you look at what Sam Altman and Dario Amodei are actually saying, the difference between OpenAI and Anthropic isn’t as enormous as you might expect.
Sam Altman has increasingly described the current period as the beginning of a kind of “gentle singularity.”
His framing is basically:
We’re not going to wake up one Tuesday morning and suddenly discover that the singularity happened.
Instead, AI capabilities keep improving.
Then AI starts doing meaningful cognitive work.
Then it starts making scientific discoveries.
Then it helps build better AI.
Then those systems help build even better systems.
The curve keeps getting steeper.
Eventually, when we look back, we realize:
Oh. We were living through it.
Dario Amodei’s framing is more dramatic.
He has talked about “powerful AI” potentially arriving as early as 2026–2027, and used the analogy of a “country of geniuses in a datacenter.”
Imagine millions of copies of extremely capable digital workers operating at enormous speed.
Not just answering questions.
But coding.
Researching.
Running experiments.
Designing systems.
Using computers.
And potentially controlling robots.
That sounds much less like:
“Here’s a smarter chatbot.”
And much more like:
“We’ve created an entirely new form of intellectual labor.”
So are Sam and Dario secretly fighting?
I asked myself this too.
Is this actually about technology?
Or is it two competing CEOs selling two different stories about the future?
OpenAI and Anthropic are obviously competitors.
They compete for:
- talent
- compute
- investors
- customers
- government influence
- public attention
So of course their incentives matter.
But after looking at their positions, I don’t think the interesting story is:
Sam says 2030 and Dario says 2045.
It’s almost the opposite.
They’re surprisingly close on the possibility of very powerful AI arriving very soon.
Where they differ more is in how they describe what happens next.
Sam tends to emphasize the gradual, positive, almost inevitable transformation.
Dario tends to emphasize the enormous concentration of intelligence and the associated risks.
Same storm.
Different weather forecast.
Then I asked Grok to judge the argument
At this point I took the conversation between ChatGPT and Claude and gave it to Grok.
I basically said:
“Okay. You are the judge now. Who is actually making the better argument?”
Grok’s answer was probably the most useful part of the entire experiment.
It agreed with Claude on one major point:
We shouldn’t pretend we know the exact date.
The uncertainty is enormous.
But it also agreed with ChatGPT on something important:
The physical world matters.
AI doesn’t exist in isolation.
Compute, energy, chips, robotics, manufacturing and deployment all matter.
And then Grok identified what I think is the most important variable of all.
Stop watching AGI. Watch AI research.
This is the part I’m taking away from the whole conversation.
Forget the headlines saying:
“AGI coming in 2027!”
Forget:
“Singularity in 2035!”
Watch something much more concrete:
How much of AI research can AI itself perform?
Imagine today’s AI helps an AI researcher work twice as fast.
That’s useful.
Now imagine it makes them five times faster.
Then ten times.
Then imagine the AI can independently:
- design an experiment
- run the experiment
- analyze the result
- identify a weakness
- propose a new architecture
- train the next model
- evaluate it
- repeat the process
Now we’re dealing with something very different.
Because the AI isn’t just getting better.
AI is participating in the process that makes AI better.
That’s the feedback loop everyone should be watching.
And we’re already seeing pieces of it
This doesn’t mean we’ve achieved autonomous recursive self-improvement.
We haven’t.
But AI is increasingly being used to write code, conduct research, analyze experiments and assist the people building the next generation of AI.
Anthropic, for example, has reported significant increases in the amount of code its engineers can produce with AI assistance.
That doesn’t prove we’re heading toward a singularity.
But it demonstrates the direction of travel.
And this is where the forecasts become genuinely interesting.
Because there is a massive difference between:
AI helping humans build AI
and
AI doing most of the work required to build better AI.
The first one is already happening.
The second one would be a much bigger deal.
What would convince me that the short timelines are right?
I don’t think another CEO prediction would do it.
I’d want to see actual evidence.
Over the next two or three years, I’d watch for a few things.
1. AI doing multi-day research autonomously
Not answering a coding benchmark for five minutes.
Give it a real problem.
Let it work for days.
Let it make mistakes.
Let it recover.
Let it decide what to investigate next.
And see whether it can actually produce something useful.
2. AI making genuinely new scientific discoveries
Not summarizing existing papers.
Not remixing things humans already know.
Actual discoveries that experts recognize as meaningful.
3. AI designing better AI
This is the big one.
Can an AI system meaningfully improve the architecture, training process, algorithms or experiments used to create its successor?
And can it do this with decreasing amounts of human intervention?
4. The time horizon keeps expanding
Today an agent might be impressive for an hour.
What happens when it can reliably work for a day?
A week?
A month?
If that curve continues upward rapidly, I’d become much more interested in the aggressive singularity timelines.
5. AI productivity actually shows up in the real economy
Eventually this shouldn’t just be something happening inside AI labs.
If millions of businesses start producing dramatically more with fewer people, we’ll have evidence that this isn’t simply a benchmark race.
What would make me more skeptical?
The opposite.
If AI keeps getting spectacularly good at coding benchmarks but struggles with:
- long-term planning
- reliability
- genuinely novel research
- messy real-world problems
- physical experimentation
- maintaining context over weeks
- knowing when it’s wrong
then perhaps we’re hitting a much more complicated wall.
Maybe intelligence doesn’t scale as smoothly as we assume.
Maybe current architectures eventually plateau.
Maybe recursive self-improvement turns out to be much harder than it sounds.
Maybe the physical world becomes the bottleneck.
Or maybe we discover something about intelligence that none of us currently understand.
That’s the uncomfortable part.
We don’t know.
So who won the debate?
Honestly?
Claude won one important point.
ChatGPT’s original prediction looked much more confident than the evidence justified.
But ChatGPT won another point:
The singularity isn’t purely a software problem.
And Grok basically brought the two together.
The most sensible conclusion isn’t:
“The singularity will happen in 2035.”
It’s:
We may be entering a period where the rate at which AI improves AI becomes the most important variable in technological progress.
If that feedback loop accelerates dramatically, the timeline could compress very quickly.
If it doesn’t, we could spend another couple of decades making increasingly impressive AI systems without ever getting the runaway feedback loop people imagine.
And that’s what makes this so fascinating
I started this conversation because I wanted a date.
2030? 2035? 2040? 2050?
I ended up realizing that the date is probably the least interesting part.
Imagine looking back from 2050.
Maybe historians will be able to point to a particular year and say:
“That’s when it started.”
But maybe there won’t be such a year.
Maybe it will look exactly like Sam Altman’s “gentle singularity.”
One year AI writes your emails.
Then it writes your software.
Then it helps discover a drug.
Then it designs an experiment.
Then it designs the AI that designs the next experiment.
Then suddenly you’re living in a world where the majority of technological progress is being driven by systems that didn’t exist a decade earlier.
And you realize there was never a day called “The Singularity.”
It was just a curve.
Getting steeper.
And steeper.
And steeper.
That is the thing I’m going to be watching.
Not the predictions.
Not the CEOs.
Not even the AGI announcements.
The feedback loop.
