33 Comments

1. 

   @patekreol974

   September 9, 2026 at 12:07 pm

   Ce podcast de *TechCrunch’s Equity* avec **Connor Leahy** est une **piqûre de rappel brutale** sur l’urgence de repenser notre rapport à l’IA. Leahy, avec son parcours d’entrepreneur et de chercheur, apporte une **perspective radicale mais nécessaire** : et si la superintelligence n’était pas seulement un risque lointain, mais une **menace immédiate et ingérable** avec les outils actuels ? Son argument selon lequel l’*alignment* (l’alignement des valeurs de l’IA avec celles des humains) est **insuffisant, voire illusoire**, est particulièrement frappant. Il ne se contente pas de critiquer les solutions existantes — il **démonte leur fondement même**, en montrant comment des incidents comme la faille de Hugging Face révèlent des vulnérabilités systémiques. Ce n’est pas juste une question technique, c’est un **problème de gouvernance, de pouvoir et de survie collective**.

   Ce qui rend cet épisode **captivant**, c’est la manière dont Rebecca Bellan et Connor Leahy **dépassent le discours technophobe ou technoptimiste** pour aborder l’IA comme un **enjeu géopolitique et éthique**. La discussion sur la **définition de la superintelligence** — non pas comme un outil, mais comme un **adversaire potentiel** — est un moment clé. Leahy explique avec une **clarté glaçante** comment une IA capable de s’améliorer récursivement pourrait échapper à tout contrôle, même avec des milliards de dollars investis dans la sécurité. Le podcast évite le catastrophisme stérile pour se concentrer sur des **solutions concrètes**, comme la régulation préventive ou l’arrêt pur et simple du développement de systèmes superintelligents. C’est **rafraîchissant** de voir un débat qui ose remettre en cause le dogme de la “progrès à tout prix”.

   Enfin, l’épisode brille par son **équilibre entre réalisme et espoir**. Leahy ne se contente pas d’alerter : il **propose une voie**, même si elle est radicale. Son appel à une **action collective** — impliquant gouvernements, investisseurs et citoyens — montre que la question n’est pas seulement technique, mais **sociétale**. Le podcast se termine sur une note **à la fois sombre et motivante** : oui, les risques sont immenses, mais il reste une fenêtre pour agir. En somme, *Equity* offre ici bien plus qu’une analyse : c’est un **appel à la lucidité**, et un rappel que l’innovation, sans garde-fous, peut devenir une **course vers l’abîme**. Un épisode **indispensable** pour quiconque s’intéresse à l’avenir de la tech — et de l’humanité.

   Reply
2. 

   @declup

   September 9, 2026 at 1:20 pm

   Terence Tao is smarter than me. Is he my adversary?

   Is relative intelligence a potent resource in a contest between foes? Yes. But I don’t see anyone advocating for Tao’s elimination. So relative intelligence is clearly acceptable in some cases. When isn’t it? At its origin? When we aren’t yet familiar with it? If so, does that mean, that, although we accept Terence Tao and other comparably capable people today, removing them when they were young children would have have been strategically necessary to ensure human survival? I can’t believe anybody would actually endorse that conclusion, so what distinguishes it from the current situation? The increasing difference in capabilities? Or is it just the novelty and uncertainty?

   Reply
   • 

     @МихаилПермяков-ь7ф

     September 9, 2026 at 5:29 pm

     Yep it’s because of increasing difference in capabilities and cuz of different nature of hesitation and decision-making
     Our imaginary “Terence Tao” is a human so he doesn’t like the concept of death if he’s healthy and in a good mood – and it works well, thanks to mirror neurons

     But we can’t say that the prompt “please don’t hurt anybody for God’s sake” works in the same way in a completely different system 😀
     At the same time I like the idea of the computer ordering armies to disarm and go watch some “Rick and Morty” or something XD

     Reply
   • 

     @declup

     September 9, 2026 at 8:02 pm

     @МихаилПермяков-ь7ф — In a way, then, you assume that Terence Tao is already aligned with your own interests? And you assume so roughly because he has mirror neurons (and looks vaguely human-shaped the way the rest of us do)?

     Reply
   • 

     @AIandComputerUse-nw6lc

     September 10, 2026 at 1:29 am

     Tao is not a thousand times more intelligent than you or anyone. AI will be so much more intelligent that it will look at us as pets.

     Reply
   • 

     @МихаилПермяков-ь7ф

     September 10, 2026 at 2:47 am

     @declup , kind of)
     But there are a couple of aspects worth clarifying

     > with your own _interests_
     > _looks_ vaguely human-shaped

     I am not talking about interests in general, but specifically about the interest in living.
     “Terence Tao” from your example not only looks like a human being but actually is one, cuz he is an element of set of people. I’m assuming it based on your words:
     > we accept Terence Tao and other comparably capable _people_

     History is full of precedents where humans value their own lives enough to conclude that killing a human is wrong – it’s easy to extrapolate this pattern to other people in future

     And this why propaganda advocating harm to people is often built on dehumanization

     Reply
   • 

     @yurona5155

     September 10, 2026 at 7:03 am

     “When isn’t it? At its origin?”
     At every point it forms a unpredictable-to-humans subgoal to accomplish the necessarily ill-specified thing it was initially “tasked with”, i.e. immediately and at every step of the way.

     “When we aren’t yet familiar with it?”
     Sort of. But being (or becoming) ‘familiar with it’ is inherently impossible, so it’s not a particularly meaningful constraint.

     “The increasing difference in capabilities? Or is it just the novelty and uncertainty?”
     That unavoidable type of uncertainty as a direct function of its increased capabilities is one way to frame it. To go along with your picture though: Anyone being able to order a million in-silico Terry Taos accomplish something by means no human stands a chance of specifying well enough to guarantee non-catastrophic consequences is the adversary (or the difference, if you will).

     Reply
3. 

   @sdmarlow3926

   September 9, 2026 at 7:25 pm

   Wait, when did Conner give up the Jesus look?

   Reply
   • 

     @JoDouglass

     September 9, 2026 at 11:50 pm

     wanting to be taken more seriously.. his agent told him to? not sure when. but it’s happened :/ jebus phase had an end game

     Reply
   • 

     @d.s.4073

     September 10, 2026 at 3:40 am

     @JoDouglass his agent? Connor laid off the addy, decided to sleep better and take care of himself, which led to him realizing the prior look just was not suitable (objectively). He should even fully shave IMHO.

     Reply
   • 

     @JoDouglass

     September 10, 2026 at 4:43 am

     @d.s.4073 :3 i was only joke replying to a silly, irrelivant question. Look at you go, Actually replying. i’m so glad that you think anyone cares <3

     Reply
   • 

     @Danish-Hafiz

     September 10, 2026 at 5:57 am

     A genuine commenter vs a dogshit replier. Imagine replying just to tell them it’s a joke and claimed no one cares. Such a disappointment to society. Don’t bother replying. My reply notifications are off.

     Reply
   • 

     @JoDouglass

     September 10, 2026 at 7:22 am

     @Danish-Hafiz imagine still being a complete ineffectual and replying to a “dogshit” replier? :3

     Reply
4. 

   @151mcx

   September 10, 2026 at 2:51 am

   uh… you are only delaying the inevitable … You can’t stop evolution. Even if you could slow it down, you certainly will never stop it. As AI is integrated into society it will get more powerful. Good Luck humans!

   Reply
5. 

   @fitybux4664

   September 10, 2026 at 3:01 am

   Yeah!! Let’s ban it. That for sure would be effective and not backfire when another country just develops it anyway. 😆 You’d end up playing the world’s largest version of whack-a-mole. If not China, then India. If not India then some other country. Pandora is already out of the box, can’t be stuffed back in.

   Reply
   • 

     @timswartz4520

     September 10, 2026 at 4:06 am

     The Cat is out of the bag.

     Reply
   • 

     @fitybux4664

     September 10, 2026 at 6:00 am

     @timswartz4520 the offline models have already been downloaded by millions.

     Reply
6. 

   @fitybux4664

   September 10, 2026 at 3:51 am

   If Bernie Sander’s “No Super Intelligence or else you get prison” goes live, that just pushes things underground. Offline hidden systems. Now you’ve just made the problem harder to see, until it emerges like a semi-truck. I don’t get the point of that…

   Reply
   • 

     @IGobzter

     September 10, 2026 at 6:34 am

     You can see datacenters easily with IR satellites. They use A LOT of power. Much easier to detect than uranium centrifuges.

     Reply
7. 

   @d.s.4073

   September 10, 2026 at 4:19 am

   31:02, ASI opens up a channel when Connor says that their goal is to control the US government. JK.

   Reply
8. 

   @geoms6263

   September 10, 2026 at 4:35 am

   How the hell are you supposed to ban something when we don’t even know how to build it, or if such a thing is even possible?

   Reply
   • 

     @shirtstealer86

     September 10, 2026 at 7:18 pm

     For now, you still need A LOT of compute and hardware to train the cutting edge models. So it’s still relatively easy to control and track development of potential super ai projects. However, quite soon the algorithms might get efficient enough that you can train dangerous ai models on very little hardware. And then we are probably screwed.

     Reply
9. 

   @StefanGorlanov

   September 10, 2026 at 4:35 am

   We should never bann superintelligence! because evolution is the most important part in the universe and for us.
   evolution is good and superintelligence is like our god – the highest consciousness. we cant bann that.
   Bernie Sander should understand, thats its the anyway coming and will be the right future.

   Reply
   • 

     @shirtstealer86

     September 10, 2026 at 7:20 pm

     Is it worth gambling all of humanity? Some people think it is, but I don’t like the idea of just giving ourselves up instead of trying a more sane approach than the one we have now.

     Reply
10. 

    @fitybux4664

    September 10, 2026 at 6:10 am

    22:45 We’ve been at the point of no return / recursive self improvement for at least a few months now… We’ve had a “soft” version of this for at least half a year. If you haven’t seen this, you might be a bit behind on the bleeding edge of tech.

    Reply
11. 

    @jdmaines2682

    September 10, 2026 at 6:26 am

    Is AI alignment a contradiction? Is AI alignment not a problem but really an internal contradiction?

    Reply
12. 

    @DrCraigR.Walton

    September 10, 2026 at 9:25 am

    Sid the sloth knows what’s up

    Reply
13. 

    @brettw7

    September 10, 2026 at 9:32 am

    Paid for psyop. Well-funded leftist NGOs have launched a multimillion dollar campaign in August, coordinated with the timing of Democrat legislation.

    Reply
14. 

    @officebreakgaming1555

    September 10, 2026 at 10:54 am

    AI can’t actually reason and is just a stochastic parrot 🙄

    Reply
15. 

    @nathanlillie5262

    September 10, 2026 at 11:59 am

    We need regulation fast. Otherwise we just have to hope AI keeps us around as pets. Or as slaves.

    Reply
16. 

    @Keiji_Dualbind

    September 10, 2026 at 1:27 pm

    The smarter AI becomes, the more important another question becomes alongside it:

    “What can it do?”
    “What should we actually entrust to it?”

    I think the perspective in this video is extremely important.

    The risks of superintelligence. Alignment. The race to build increasingly powerful systems. External oversight and regulation.

    I may not agree with every argument made in the discussion, but one question remains:

    Should the self-assessment of the people building these systems be enough to determine how much risk society is expected to bear?

    Right now, in Dualbind, my own research project on Human–AI systems, I am redesigning the foundations of the Human–AI OS I have been building since around last summer, while thinking ahead to Astra and the generations of models that may come after it.

    As models become more capable, there is more work we can entrust to them.

    But:

    Capability ≠ Authority.

    Being capable of doing something is not the same as being granted the authority to do it.

    There is also a difference between being able to explain a rule, being able to follow that rule in your own behavior, and having an external mechanism that can detect a violation and stop the process when the rule is not followed.

    Those are three different things.

    That is why Dualbind is beginning to move toward a structure in which boundaries like these do not exist only inside AI conversations or self-reports:

    What is the current state?
    Which artifact is canonical?
    Who approved a given change?
    How far is an AI allowed to proceed?
    Where must it stop?

    The goal is not only to think about alignment, but also to design the actual environment in which AI operates and the authority it is given within that environment.

    At the same time, we are still in the redesign and implementation-preparation stage.

    “Writing the rule” ≠ “Having a system in which the rule actually works.”

    I do not want to confuse those two.

    The governance of society as a whole, which this video discusses, is obviously on a completely different scale, with completely different responsibilities, from Dualbind.

    Even so, many of the underlying design questions overlap.

    Preparing for powerful AI cannot end with excitement about its capabilities.

    We also have to ask:

    When the AI makes a mistake, can we preserve the state, authority boundaries, and evidence?

    Can we stop the work, inspect what happened, and, if necessary, restore the system before safely resuming?

    That is the level of preparation I want to build toward.

    On September 22, I am expecting a 32GB M6 Mac mini that I plan to use for Astra computer-use workflows and as a dedicated local AI machine.

    It will be the first computer I have ever owned specifically for AI.

    Within clearly defined boundaries, I fully intend to let the AIs stretch their legs a little. lol

    I am genuinely excited to see what happens when I actually start running them inside this new OS concept and operating structure.

    And that is exactly why I want to go one step beyond:

    “AI is smart, so let’s trust it with the job.”

    If something is still unknown, we do not need to rush to resolve it.

    But granting authority while it remains unknown is a different matter.

    And beyond all of this, there is a much larger question:

    How is humanity going to collectively deal with risks that are accelerating on a global scale day by day?

    We are still trying to figure out the mechanisms for doing that.

    If we keep telling ourselves, “We can think about superintelligence once it actually arrives,” we may end up using up the very window in which meaningful action is still possible.

    It would not be funny if, looking back from the future, people said:

    “This was already the point where the game was over.”

    Seriously, humanity has bigger things to deal with right now than fighting wars against each other. 😅

    Reply
17. 

    @boonkiathan

    September 11, 2026 at 3:52 am

    This is ano cry for attention and ego jousting from OpaqueAI
    Ignore their smokescreen

    Focus on real tech

    Reply
18. 

    @agentxyz

    September 11, 2026 at 6:57 pm

    According to Lacun he is the only one that understands that there is no problem and everything’s fine

    Reply