After listening to Ilya Sutskever's one-hour interview, I realized for the first time: #AI 's second half, the rules have completely changed.

Honestly, after listening, I felt a bit of scalp tingling.
As someone who has been following AI trends for a long time and also believed in the Scaling Law,
Ilya's words almost overturned all my intuition about AI evolution over the past few years.
While the whole world is still frantically competing for computing power, stockpiling GPUs, and scaling parameters,
this father of ChatGPT gave an unusually calm judgment:
The era of "great efforts produce miracles" is coming to an end.
The second half of AI is no longer a resource war but a paradigm war.
I have summarized the five most important and most disruptive viewpoints from his interview.
1️⃣ The Age of Scaling is really coming to an end
In recent years, the formula for AI progress has been extremely simple and brutal:
More data + more compute = stronger models
This logic has been valid for so long that the entire industry has defaulted to it will always be.
But #Ilya clearly pointed out:
The marginal returns of Scaling Law are rapidly diminishing.
The reason is not mystical—
High-quality pre-training data is being "consumed."
Everyone is still crazily adding coal to the furnace,
but the flames no longer grow bigger.
What does this mean?
👉 Relying solely on resource stacking for "brute force aesthetics" is no longer feasible.
👉 The industry is forced to return to the most difficult and fundamental stage: finding new paradigms.
2️⃣ "High scores, low ability": existing models are falling into a dangerous trap
This is a problem overlooked by many but extremely deadly.
Current models can:
Win gold medals in programming competitions
Dominate benchmark scores beyond humans
But in the real world:
Fix one bug, introduce two new bugs
Slightly change to an unseen scenario, and they start to mess up
Ilya pointed out sharply:
This is not intelligence, but overfitting.
Models are more like "test-takers" who memorize the entire question bank,
rather than systems with genuine generalization ability.
This also means:
👉 The current reinforcement learning path may only optimize "exam ability"
👉 and does not lead to true intelligence.
3️⃣ The most explosive insight: Emotions are essentially the most efficient value functions
This was the most enlightening point in the entire interview.
We have always regarded "emotion" as the opposite of rational intelligence,
and even thought it was a human flaw.
But Ilya's view is completely the opposite:
Emotions are the most efficient algorithms evolved by biology.
What do they do?
No need for massive samples
No need to recall a whole lifetime
Provide decision feedback in milliseconds
Fear, excitement, disgust, boredom—
These are highly compressed value judgment signals.
This is the core reason humans can learn complex skills with very little data.
👉 If AI truly wants to pass the Turing test,
👉 it may need to understand the "mathematical structure behind emotions."
4️⃣ SSI's choice: a lonely path to "superintelligence"
While Silicon Valley is busy competing on products, ARR, and deployment,
Ilya's new company SSI (Safe Superintelligence) made an counterintuitive choice:
Not to make products.
Just do one thing:
Ensure safe superintelligence.
This is an extreme long-term gamble.
His judgment is cold:
👉 The current commercial frenzy is likely just a bubble,
👉 and only those who solve the "fundamental problem of superintelligence" will truly dominate the future.
5️⃣ Our real revelation: The Research Age is back
If Scaling really hits a ceiling,
it's bad news for giants,
but an opportunity for truly creative people.
Because:
It's no longer just about throwing money and scaling compute,
but about re-innovating algorithms, architectures, and cognitive breakthroughs.
As Ilya said:
“Research Age is back.”
The era belonging to geeks, to wild ideas,
to the emergence of AlexNet,
is quietly returning.
Every turn of Ilya—
from ImageNet, to OpenAI, to SSI—
precisely hits the key nodes of AI evolution.
I believe this time will be no different.
If you are also seriously contemplating the second half of AI,
then this interview is worth chewing over repeatedly.
The rules of the era have changed.