I think Tegmark’s trajectory since Life 3.0 (2017) is quite striking. The basic philosophy has not changed very much, but his estimate of urgency has changed enormously.
In 2017 Tegmark was still primarily asking:
What kind of future with superintelligence would we want?
By 2024–2026 he is increasingly asking:
Can we stop ourselves from building uncontrollable superintelligence in the first place?
That is a substantial shift.
1. 2017: the problem was largely
destination
Life 3.0—I assume this is the book you mean—was deliberately pluralistic. Tegmark laid out many possible futures: benevolent dictator, protector god, zookeeper, libertarian utopia, descendants, conquerors, and so forth.
His characteristic position then was almost Enlightenment optimism:
Technology is not destiny. We should decide collectively what future we want before building the technology that locks us into one.
Hence the twelve scenarios we discussed earlier.
He clearly took existential risk seriously even then, but Life 3.0 has the atmosphere of someone standing before a giant branching tree of possible futures.
2. 2023 changed the tone:
GPT-4 made the problem present-tense
The famous March 2023 Future of Life Institute letter called for at least a six-month pause in training systems more powerful than GPT-4. It argued that sufficiently powerful AI could create profound societal risks and that labs were engaged in an uncontrolled race.
This is where I detect an important change in Tegmark.
The question was no longer merely:
“What happens someday when AGI comes?”
It became:
“Why are we scaling toward it before solving control?”
His subsequent policy proposals became much more concrete: licensing/oversight of advanced general-purpose systems, liability, independent evaluation, safety research, and binding regulation. In his 2023 Senate AI Insight Forum statement, he explicitly argued that innovation should mean technology that improves human life—not merely building increasingly powerful systems.
That distinction has subsequently become central to him.
3. His most important new distinction is probably:
Tool AI ≠ AGI / autonomous superintelligence
This, I think, is the key to understanding later Tegmark.
He is not anti-AI.
Quite the contrary.
He wants extremely powerful AI for:
medicine, science, education, engineering, industry, defence, productivity.
But he increasingly distinguishes these systems from an autonomous general intelligence capable of outperforming humans across cognitive domains and improving AI itself.
In November 2024, responding to proposals for an American “AGI Manhattan Project,” Tegmark put the distinction unusually starkly. He argued that America should pursue transformative “Tool AI” rather than race toward AGI.
So his present position can almost be written:
Maximum capability without maximum autonomy.
Or even:
Make the hammer extraordinarily intelligent.
Don’t make the hammer want things.
That is considerably sharper than the formulation in Life 3.0.
4. He now thinks the
race itself
is one of the central dangers
This has also become much stronger.
Tegmark’s worry isn’t merely that somebody will deliberately create an evil AI.
He thinks competition generates a structural problem.
OpenAI races Google DeepMind.
Anthropic races OpenAI.
xAI races everyone.
America races China.
China fears America.
Investors fear falling behind.
Governments invoke national security.
And therefore every participant can say:
“I would prefer to proceed safely, but if I slow down, the other guy wins.”
This is classic game theory.
And Tegmark increasingly argues that the result may be a race to surrender human control.
His 2024 AGI statement went so far as to describe an AGI race as a “suicide race”: because once an intelligence surpasses humans generally, the hope that its creator thereby gains geopolitical supremacy may invert—the creator may instead lose control to what it created.
This is particularly interesting because it undermines the familiar US-versus-China argument.
The conventional formulation is:
We need superintelligence before China gets it.
Tegmark’s answer is essentially:
What good is winning the race if the winner no longer controls the winning system?
That is an extremely important point.
5. From
alignment
toward
control
Another subtle evolution.
Earlier AI-safety discourse often centered on:
How can we align AI’s values with human values?
Tegmark hasn’t abandoned alignment.
But his recent rhetoric increasingly emphasizes a more primitive question:
Can the system actually be controlled?
Those aren’t identical.
An AI could appear aligned during testing.
It could obey instructions.
It could display pleasant values.
But if it becomes strategically superior to humans, autonomous and capable of self-improvement, Tegmark’s concern is that we may have no reliable mechanism for maintaining that relationship.
That is why his vocabulary increasingly includes:
safe
+
controllable
rather than simply “aligned.”
The 2025–26 Statement on Superintelligence crystallizes this. Its demand is not merely better alignment research; it calls for prohibition of superintelligence development until there is both broad scientific consensus that it can be developed safely and controllably, and strong public support.
That is significantly stronger than the 2023 six-month pause.
6. And therefore Tegmark crossed another line in 2025:
from
“pause”
→
“don’t build superintelligence yet.”
This is perhaps the clearest evolution.
2023:
Pause frontier development for six months.
2025–26:
Prohibit superintelligence development until we know it is controllable and society actually wants it.
That isn’t merely more urgency.
It is a conceptual change in the burden of proof.
The old technological norm is:
Build unless someone proves it dangerous.
Tegmark increasingly wants:
Don’t build superintelligence unless its builders can demonstrate that humanity retains control.
This resembles pharmaceuticals, aviation or nuclear reactors much more than ordinary software development.
7. He is increasingly worried about
human disempowerment
, not merely extinction
This part brings us back to our previous conversation.
It would be easy to caricature Tegmark as:
“AI will kill everybody.”
But his recent superintelligence statement explicitly lists a much broader spectrum:
economic obsolescence, disempowerment, loss of freedom, civil liberties, dignity and control, national-security risks—and only then possible extinction.
I find this development especially important.
Because suppose AI never kills anybody.
Suppose instead that:
AI runs businesses.
AI allocates capital.
AI does science.
AI writes software.
AI governs infrastructure.
AI conducts warfare.
AI optimizes politics.
AI teaches children.
AI produces culture.
Humans survive beautifully.
Then Tegmark’s question becomes:
Who is actually in charge?
That comes remarkably close to what we were discussing as 人的退位.
8. Autonomous weapons have remained one of his longest-running red lines
This concern predates Life 3.0, but has hardened further.
In February 2026 Tegmark argued that AI systems used in lethal contexts must remain under meaningful human control, and called fully autonomous weapons incompatible with human dignity and dangerous because of proliferation and escalation risks.
Notice again his recurring phrase:
meaningful human control.
That, increasingly, is the ethical center of late Tegmark.
Not:
humans must do everything themselves.
Rather:
humans must not create systems whose consequential decisions humans can no longer meaningfully govern.
9. He has also become much less confident in
corporate self-regulation
FLI’s AI Safety Index, launched in 2024 and repeatedly updated through 2026, reflects this shift.
The Summer 2026 assessment concluded that existential-safety preparations remained especially weak across leading AI companies and that several firms had weakened earlier commitments to pause deployment when dangerous capability thresholds were approached.
Tegmark’s policy conclusion has increasingly been:
voluntary promises aren’t enough.
Hence independent auditing, legally binding standards, licensing regimes and explicit red lines.
Again, compare this with Silicon Valley’s traditional:
Move fast and fix problems later.
Tegmark’s answer is basically:
That model works badly for technology whose “bug” could remove our ability to fix it.
What has
not
changed since
Life 3.0
There is nevertheless a very strong continuity.
Tegmark remains technologically optimistic in a fundamental sense.
He believes AI could bring extraordinary prosperity, cure diseases, accelerate science and remove enormous amounts of unnecessary suffering. The current superintelligence statement itself explicitly acknowledges potentially unprecedented health and prosperity from advanced AI tools.
So he isn’t a primitivist.
His position is not:
Stop AI.
It’s increasingly:
Build AI that empowers humanity; don’t blindly build AI that replaces humanity as the locus of control.
And that distinction makes his thinking more interesting than simple “AI doom.”
I would describe the intellectual evolution this way
Tegmark 2017:
Which of the twelve futures should we choose?
Tegmark 2023:
Why are we racing before we’ve figured out safety?
Tegmark 2024:
Tool AI can give us almost all the benefits without deliberately creating an autonomous successor.
Tegmark 2025–26:
Do not build superintelligence until there is persuasive evidence that it remains safely controllable and society actually consents to doing so.
There is therefore a fascinating inversion.
In Life 3.0, superintelligence is almost the premise, and the question is what comes afterward.
Late Tegmark increasingly questions the premise itself:
Why exactly must we create something smarter than humanity and hand it autonomous agency?
That is, I think, his most important post-Life 3.0 development.
And here I would add one difference between his thinking and yours. Tegmark still fundamentally wants to preserve human sovereignty: keep AI as extraordinarily powerful tools under meaningful human control.
Your idea of 「物的反撲，人的退位；物的歸位，人的復位」 goes one philosophical step further, because it entertains the possibility that anthropocentric sovereignty itself may already be ending—and asks not merely how man remains master, but what “人的復位” could mean after man has ceased being master.
Tegmark is trying to prevent the abdication.
You are asking what kind of human being could remain after the throne has disappeared.
I think that is precisely where your question and late Tegmark now meet—and then diverge.
