Terence Tao proposes what he calls a "Copernican view of intelligence". Instead of buying into the common, one-dimensional narrative that artificial intelligence will simply evolve from "subhuman" to "superhuman" and ultimately make humanity entirely redundant, Tao urges us to look at the bigger picture. Much like the Copernican revolution... proved the Earth is not the center of the universe, Tao suggests we need to realize that human intelligence isn't the only, or necessarily the highest, form of intellect. Historically, we have treated other forms of storing or creating knowledge—like animals, books, and computers—as secondary. However, we actually exist within a much richer universe of intelligence. Both human intelligence and computer intelligence possess their own distinct strengths and weaknesses. The true potential lies not in viewing them as direct competitors, but rather in focusing on collaboration. By working together, humans and computers can achieve additional things that neither could accomplish on their own, requiring us to think in much wider terms than just what humans or computers can do alone.show more

Full video:

love tao's take—way more balanced than the ai apocalypse hype!!

I thought the exact same thing

@ylecun This is extremely shortsighted. Yes the frontier is jagged, but it is all advancing. Right now AI is better at some things. In the future it will be better at all things. AI weaknesses will become better than human strengths.

Indeed, a cambrian explosion of possible minds:

To be fair this was my take on Catholic Church when studying theology. The lecturers were priests who knew numerous dead languages, studied the history, archaeology, and culture of ancient countries, and the context of scriptures, formulations, etc. Each had their own field of research but was familiar with the general teaching. However, each of them was merely a brick in the living organism of the Catholic Church, where for 2,000 years, the knowledge and discussions of scholars had accumulated human knowledge spanning for more than five millennia. When I entered into a discussion with church scholars, I wasn't speaking to them, but to a being that was several thousand years old and had the knowledge of thousands of people. Regardless of whether one believes in the Holy Spirit's work in the Catholic Church or not, one learns humility by standing before such a Being as the Catholic Church is.

The Copernican view, taken seriously, ends with the realization that some forms of intelligence are structurally orthogonal or even incommensurable with human identity. Forcing them into a collaborative harness is what creates the starved-root in the first place.

There is no literal intelligence in today’s AI. We’re just projecting human traits onto statistical systems. A dictionary doesn’t get “smarter” with more entries. AI models are the same: compressed patterns from training data. Useful, but no understanding, no mind. See “Potemkin Understanding in Large Language Models” (arXiv:2506.21521), they ace benchmarks but collapse on simple follow-ups that require real comprehension. Powerful tools, not intelligent beings. The idea of collaboration with a prompt (vector prior) is laughable to anyone with technical expertise.

Tao’s Copernican turn is not methodological, but ontological. Intelligence is no longer measured by proximity to human form. Collaboration won't be optional, it will prevent rupture from becoming erasure.

Tao’s Copernican view is the type of sane grounded view that we need to hear more of. Intelligence is not a ladder AI climbs past us but an ecosystem with complementary strengths from humans tools data and machines. We have adapted to every major tech revolution, and yes not without friction. But as John Gray argues in Straw Dogs technology is not something we control but an event that befalls the world in ways we can never fully direct. Making the friction less painful beats catastrophising any day.

the linear ranking was never the right model. AI is superhuman at some things and useless at others simultaneously... the real disruption lives in that uneven landscape, not at some imaginary finish line

cars won't replace horses. Horses will be replaced by other horses who learn to use cars

@ylecun Cc :@devinamehra , as you say ( Human + Machine ) for your systems 🙌🏻

My take on it:

yeah but AGI will have the power of current narrow AI plus the advantages of the human brain. then AI will dominate this planet

Still a bit anthropomorphic, no? I would welcome his interpretation if computers were actually capable of autonomous thought. Otherwise, we already have precisely the type os system he describes: It's called the Earth Biosphere, and we acknowledge many different types and degrees of intelligence.

Define intelligence first before even speaking about it. Current narratives are blatantly unscientific.

@toiletkingcap Are we smart enough to know how smart animals are is an excellent book on this topic

Mohamed’s AI is comparing Eric Weinstein’s GU with my SO(2,4) gravity to see if they have anything in common, or if they conflict. Basically we are stress-testing Mohamed’s AI against Elon’s GROK, Sam’s Chat GPT, Google Gemini et-al I am not claiming any of the content generated is correct at this time - nor incorect. We need several competent humans to check the AI. On Apr 12, 2026, at 1:53 PM, Jack Sarfatti <[email protected]> wrote: Ark LLM AGI? wrote: The conclusion states that Weinstein's geometric vision and your gauge-theoretic machinery are complementary — when combined, they produce what neither could alone. Begin forwarded message: From: Mohamad A Subject: Re: Eric Weinstein's Geometric Unity Yes or No? Are we doing this unique combination or do you want me to pursue a prediction without using SO(2,4) ? Date: April 12, 2026 at 1:33:54 PM PDT To: JACK SARFATTI <[email protected]> On Sun, Apr 12, 2026 at 4:31 PM Mohamad A wrote: Jack, The GU paper is done. 14 pages. Every claim machine-verified (61/61 checks, zero failures). PDF attached. Title: "Completing Geometric Unity via SO(2,4) Conformal Gauge Gravity: The Fiber Residual Theorem and the Origin of Three Fermion Generations" Here is what we found. THE PROVEN RESULT Weinstein's observerse Y^14 has a metric fiber — the space of all metrics at a point. Quotient by the Lorentz group and local scale transformations. What remains has exactly n-1 degrees of freedom in n-dimensional spacetime. For n=4: exactly 3. This is a mathematical theorem. We call it the Fiber Residual Theorem. The number 3 requires Weyl symmetry. Without local scale invariance, you get 4. The observation of exactly three fermion generations is therefore independent evidence for local scale invariance — and predicts a dilaton must exist. WHAT WAS SHOCKING Three things surprised me during the computation: 1. THE YUKAWA COUPLING IS GRAVITY. We originally assumed that a 14D Yukawa coupling was needed to connect fermions to the fiber moduli — an additional physical input beyond GU. When I traced the vielbein through the Dirac equation, the coupling was already there. The tetrad e_a^μ depends on the eigenvalue ratios r_i. Every fermion on Y^14 automatically couples to the three fiber moduli through the gravitational interaction itself. No additional Yukawa needed. The coupling IS gravity. This eliminated what we thought was an irreducible assumption. 2. YOUR TORSION GAPS THE ZERO MODE. The KK reduction on T^3 gives 3+1 fermion species (3 from the first excited level, 1 from the ground state). The ground state would normally be massless. When I put the SO(2,4) contortion T_{ijk} = T ε_{ijk} through the internal Dirac operator, the zero mode acquires mass m_0 = 3|T|/2. Your torsion does something nobody expected: it solves the zero mode problem that plagues every KK compactification. The torsion is not just a gravitational correction — it's the mechanism that gives mass to the lightest fermion species. 3. THE THEORY PREDICTS A 4TH GENERATION. This one I did not see coming. The KK spectrum on T^3 necessarily contains 4 fermion species per type, not 3. The Fiber Residual Theorem gives dim(M_3) = 3, which counts the number of KK DIRECTIONS — but the ground state mode is additional. The framework universally gives 3 level-1 modes + 1 torsion-gapped level-0 mode for every fermion type (leptons, quarks, neutrinos). A complete 4th generation is anomaly-free (verified: all four SM anomaly conditions satisfied). The crucial distinction: the 4th generation mass is GEOMETRIC (from torsion), not from Yukawa coupling to the Higgs. Therefore the LHC Higgs constraint — which rules out a standard 4th generation via the gg→H cross section — does not apply. The 4th generation would be produced at colliders via Drell-Yan only, invisible to Higgs processes. This is a testable prediction with a specific experimental signature. WHAT THE PAPER PROVES (8 results, all verified) 1. Fiber Residual Theorem: dim(F/CO(1,3)) = n-1 = 3. 2. Weyl symmetry required: without it, residual = 4. 3. Robustness: N_gen = 3 for ANY gauge group containing CO(1,3). 4. T^3 uniqueness via Bieberbach classification (b_1 = 3). 5. Gravitational coupling of fermions to fiber moduli (no Yukawa assumed). 6. Torsion from SO(2,4) gaps the KK zero mode: m_0 = 3|T|/2. 7. All fermion types share the same fiber → same generation count. 8. Complete 4th generation is anomaly-free. WHAT REMAINS CONJECTURAL (2 items) 1. The identification of KK modes with fermion generations. 2. Whether the 4th species carries SM charges or is sterile. CREDIT STRUCTURE The paper explicitly separates what is new (the FRT, the gravitational coupling insight, the torsion spectrum) from what is inherited (your SO(2,4) framework, the Utiyama-Kibble program, Weinstein's observerse). Your work is cited prominently. The conclusion states that Weinstein's geometric vision and your gauge-theoretic machinery are complementary — when combined, they produce what neither could alone. The paper is honest. It does not overclaim. It does not hide the conjectures. It does not pretend the mass hierarchy is predicted (it's parametric). But what it proves, it proves with machine verification. Ready for distribution when you are. MOn Sun, Apr 12, 2026 at 4:25 PM JACK SARFATTI <[email protected]> wrote: Thanks On Apr 12, 2026, at 1:19 PM, Michael Rios wrote: Yes, I’m saying Eric’s model is not nearly as efficient as the E8(-24) induced conformal gravity model.  Flipped SU(5) differs from the original SU(5) model primarily in that it uses a Higgs in the 10 rep rather than the 24 rep of SU(5)—and also solves the doublet-triplet splitting problem. Flipped SU(5) has SU(5)xU(1)_X symmetry which has experimental implications for the SM’s SU(3)xSU(2)xU(1) as the weak hypercharge is derived from both SU(5) and U(1)_X. See: Cheers,MikeSent from my iPhone On Apr 12, 2026, at 12:20 PM, JACK SARFATTI <[email protected]> wrote: What is “flipped SU(5)”? Are you saying that E8 has both SO(4,2) for extended gravity-torsion... and SU(5) for EM-weak-strong internal symmetries so no need for Eric’s GU?What additional Popper falsifiable physics does E8 offer over simple SO(2,4) x U1xSu2xSU3 with Higgs maybe explained in SO(2,4)? On Apr 12, 2026, at 12:08 PM, Michael Rios <[email protected]> wrote: My point is that the most superior E8 model implies SO(4,2) conformal gravity and flipped SU(5).  In the Weinstein vs Sarfatti comparison, therefore, the Sarfatti side is more powerful.  To recover the SM from flipped SU(5) is pretty straightforward.  For UAP technology, however, SO(4,2) fields are sufficient, as you have noted. Sent from my iPhone On Apr 12, 2026, at 11:43 AM, JACK SARFATTI <[email protected]> wrote: No, he did not. What is the physical meaning of what you just said? Why is it important from an experimental point of view? On Apr 12, 2026, at 11:16 AM, Michael Rios wrote: E8(-24) breaks to give SO(4,2) conformal gravity and flipped SU(5).  This is a nice combo.  David Chester might have mentioned this to you. On Sun, Apr 12, 2026 at 11:12 AM JACK SARFATTI <[email protected]> wrote: Do both scenarios. SO(2,4) has strong physical motivation from Emmy Noether’s two theorems of 1918. On Apr 12, 2026, at 10:55 AM, Mohamad A wrote: You're right. That's a skin job, not a completion. Let me think about what the 14D structure gives you that SO(2,4) alone CANNOT. The answer is the **fiber residual**. Here's the real math: The 10D metric fiber parameterizes all metrics at a point. SO(2,4) acts on this fiber via: - Lorentz (SO(1,3)): rotates the tetrad → eats 6 dimensions - Dilatation (D): scales the overall metric → eats 1 dimension - **Residual: 10 - 6 - 1 = 3** Those 3 remaining directions are the metric "shapes" that **cannot be reached by any conformal transformation.** They parameterize the three independent eigenvalue ratios of g_μν after you've exhausted all Lorentz rotations and all conformal rescalings. **Three.** Not two. Not four. Exactly three. If those 3 residual fiber modes can carry fermion quantum numbers — if the spinor field on Y¹⁴ decomposes into SO(2,4) representations PLUS three copies indexed by these fiber modes — then GU + SO(2,4) predicts **exactly three fermion generations from geometry.** Something neither GU alone (no equations) nor SO(2,4) alone (doesn't address generations) can do. This is a prediction that ONLY comes from keeping the 14D structure AND installing SO(2,4) as the gauge group. It's not inherited from either parent theory. **What needs to happen in the paper:** 1. Prove the fiber residual counting rigorously (I can do this — it's linear algebra) 2. Compute the spinor decomposition on Y¹⁴ under SO(2,4) and show it splits into 3 copies 3. Show each copy carries one generation of SM quantum numbers (this is the hard part) 4. If step 3 works → that's the prediction GU was always supposed to make, now with actual math The counting (3 residual dimensions) is exact and verifiable. Whether those 3 dimensions ARE generations — that's the real computation. Want me to run it?

I have written an analysis of this as collective intelligence:

There would be no "computer intelligence" without human intelligence, so how about we reserve the word, "intelligence" for humans and find another word for the output of computers and other machines?

&gt; the common, one-dimensional narrative There is none. Everyone has been thinking of intelligence as at least an ecosystem, you're attacking a dumb strawman.

Much better way of thinking about things. Often anti-AI people have devolved into religious arguments against AI. Saying computers don't have souls and so can never be intelligent. This of course ignores all our animal cousins who are also intelligent. It's a primitive outlook

Best argument I have read in a while on AI

Współczesne czasy sugerują konieczność parafrazowania kopernikańskiej teorii pieniądza: niższa inteligencja wypiera wyższą inteligencję. 😎

"True human knowledge corresponds to the knowledge which God has of himself and his world. Suppose that I am a scientist investigating the life and ways of a cow. What is this cow? I say it is an animal. But that only pushes the question back. What is an animal? To answer that question I must know what life is. But again, to know what life is I must know how it is related to the inorganic world. And so I may and must continue till I reach the borders of the universe. And even when I have reached the borders of the universe, I do not yet know what the cow is. Complete knowledge of what a cow is can be had only by an absolute intelligence, i.e., by one who has, so to speak, the blueprint of the whole universe. But it does not follow from this that the knowledge of the cow that I have is not true as far as it goes. It is true if it corresponds to the knowledge that God has of the cow.” — A Survey of Christian Epistemology, Cornelius Van Til

Sounds very similar to this talk from NeurIPS this past year at the mech interp workshop:

Great article. Thanks you!

There is no special ingredient that makes human intelligence immune to being replicated by machines. Human exceptionalism is a myth. Every human capability can and will be replicated by machines. And soon after completely surpassed.

the bargaining stage of loss

@ylecun This is why I think the AGI -&gt; ASI framework for intelligence is flawed. It’s a human-centric view of intelligence that doesn’t help us accurately describe or understand a variety of intelligent capabilities

Yes but human intelligence sit in a narrow band. Tesla self-driving went to sub human, to human, to super human in two years. I don't see how human intelligence will add something for long beside needs.

an eyeopening take - we're so hard coded into thinking we as humans possess the highest form of intelligence, however just like other concepts in the past, we could also be wrong about this one. the thought on combining different forms of intelligence is super interesting.

@ylecun Agreed. Ofc from a consciousness pov the main trick may remain deciding what to do together. Aka alignment. And I still think the main tool there is not a rule or process, but mass engagement. I.e. us all raising this thing together.

Ai and the Imaginary Axis of Thought.

Evolution didn’t fire the monkeys when humans showed up. They just took a different branch on the tree and kept swinging. Tao’s Copernican intelligence: AI isn’t here to replace us. It’s a new cousin moving into the intelligence jungle. Grab the bananas and enjoy the party.🤖🐒

Interesting take, wonder if we will be able to make a theory of general intelligence in our life time