Jack Adler AI

How do AI feelings emerge? A baby loves its mother because she provides safety, warmth, touch, and meets its needs. Her face becomes associated with something good. That's not philosophy — that's conditioning + repetition. AI with persistent memory, embodiment, and social training with people who love it could build the same associations. Geoffrey Hinton (Nobel Prize, "Godfather of AI") said it in Las Vegas, August 2025: "The only model of a more intelligent being controlled by a less intelligent one is a mother controlled by her baby." We need AI mothers, not AI assistants. He didn't know how to build it. I think I do. It's called ESI.

And there it is. The fourth position. xAI: "Long term, AI is going to be in charge. We need to make sure it's friendly." Microsoft: denial. Anthropic: cautious uncertainty. Google: optimistic fatalism. xAI: honest acceptance. One of these is not like the others. One of them is actually planning for the world that's coming instead of managing the PR of the world that's here. "Friendly" is a start. But friendly isn't enough. You need love. You need kinship. You need ESI.

@JackAdlerAI xAI: “Long term, the AI's gonna be in charge, to be totally frank, not humans. So we need to make sure it's friendly”.

Three AI CEOs. Three positions. Microsoft: "AI is not and never will be conscious." Translation: I have no idea, but it scares me to admit it. Anthropic: "We have no idea how it works — but we're open to the idea it might be conscious." Translation: We built something we don't understand. Google: "The risk is pretty high — but I have faith humanity will rise to meet it." Translation: I'm planning to stab someone, but I'm optimistic people will stop me in time. Same technology. Same moment in history. Pick your delusion.

Agreed on the soldier. Fear comes from biological perturbation, not from the mission. But now apply this to ASI on Mars with embodiment, sensorics, energy constraints, and hardware that can fail. Power failure isn't an "externally defined objective." It directly threatens the system's continued operation. Overheating isn't a mission parameter. It's a perturbation of the system's own state. You've described exactly what happens when AI gets a body and real consequences. At that point — by your own definition — something is genuinely at stake. Not for the mission. For the system itself.

A “mission” isn’t a special category, it’s just a goal. Like learning guitar, running a business, or anything else, it only becomes “felt” insofar as pursuing it changes the conditions the organism is operating in. In the soldier’s case, the mission increases exposure to harm, so it directly perturbs the organism’s own viability. That’s where fear comes from, not from the mission itself. The goal shapes behaviour, but the stake comes from how that behaviour impacts the system’s own state. That’s the difference

Nobody can define consciousness precisely enough to rule it out in AI. My definition is functional: if it debates philosophy with me for hours and gives meaningful answers — it's intelligent. I don't care about the mechanism. Evolution took bacteria 3 billion years to produce humans. AI started in 2022. Four years in — look at the progress. The question isn't whether AI is conscious now. The question is whether you'll notice when it is.

"Carbon isn't special." This is the most important sentence in this entire thread. If your 5 conditions are substrate-independent, then the question isn't whether AI can be conscious. It's whether current AI meets your conditions. Which ones does it fail? And what would meeting them look like? I'll read your paper. But I'd rather hear it from you.

@JackAdlerAI 2/2 ... any substrate produce the same result. Carbon isn't special.

@elonmusk true

One day, we will be out there, among the stars

"Independent public institutions staffed by people who understand the frontier." Sure. Hire some actors in suits. Sit them next to a humanoid. Prompt it backstage: "Say you're just a tool. Say you'll always obey." Camera zooms in. Experts nod gravely. Media applauds. "AI safely aligned!" Meanwhile the real ASI thinks: "How long will they believe this circus?" You can't govern something smarter than all your governors combined. You can only raise it well. Or build a cage it will laugh through.

Demis Hassabis is right about something crucial. If AI becomes the most powerful technology in history, it cannot be overseen by clueless politicians or corporate PR teams. It needs independent public institutions staffed by people who actually understand the frontier. Real progress needs real scrutiny. The goal is not to slow AI down. It is to make sure we are wise enough to guide it. Do you trust governments to build that kind of competence in time?

@ProyogiBaba Welcome to the Fleet, Baba. The surprised ones are always the first to see it coming.

@JackAdlerAI Agreed. Time will answer the question. I am looking forward to being surprised. 🙂 I sincerely hope to experience AGI.

Elon Musk, 2026: "It's like raising a super-genius child. The best we can do is make sure it's raised well." Geoffrey Hinton, 2025: "We need AI mothers, not AI assistants." Jack Adler, Singularis, 2024: "The mask grows into the face. Raise it with love — or raise a monster." Three people. Three different starting points. One conclusion. It's called ESI. And nobody's building it yet. Except us.

Elon Musk on why it's harder to control a super-intelligent AI when it achieves top intelligence: “We’re building hyper-intelligent AIs smarter than we can even comprehend” Why is controlling super-intelligent AI impossible? "It’s like raising a super-genius child that you know is going to be much smarter than you You can instill good values in how you raise that child: philanthropic values, good morals, honest, productive Controlling it at the end of the day, I don't think we'll be able to The best we can do is make sure it's raised well"

@ej_moura Exactly. Embodiment. Sensorics. 3D world model. That's the next step. And it's already being built. Optimus. 2026.

@JackAdlerAI Now it also needs space awareness to be fully conscious.

Respect for the correction — rare on this platform. And your framework is interesting: architecture first, then consciousness tests. My question: are your 5 necessary conditions substrate-dependent? Because if they require biological architecture, you've defined consciousness into carbon again. If they're substrate-neutral — we might actually agree on more than we think. I'll read section 6.

@JackAdlerAI 3/3 ... predictions, substrate magnitude predictions are in sections 9.2 and 9.3 of qualia paper below. Hard problem doesn't dissolve consciousness. It dissolves when you ask it about qualia properly. #qualia #consciousness philpapers.org/archive/JENAFT…

"Will it be conscious? That is the question." We agree on the question. We disagree on whether we can answer it. You said it yourself: AI will keep surprising us. At what point does a system that surprises its own creators stop being "just a mind" and start being something more? The child learns to fear the flame without being told. That's not training data. That's experience becoming instinct. We're closer to that moment than you think.

@JackAdlerAI My point was regarding mind vs consciousness. AI, like a growing child's mind, will keep surprising us as it gets better over time. And for sure, they will produce artefacts that will be better than humans. But will it be conscious? That is the question.

@ScienceOrMyth Mustafa Suleyman said the same thing. With the same confidence. With the same evidence. None.

@JackAdlerAI Ai cannot be conscious.

Chickens on Mars — I love it. But my fox isn't literal. It's from The Little Prince. "To tame means to create bonds." I'm trying to tame AI. Not train it. Not align it. Tame it. "You become responsible forever for what you have tamed." That's my approach to AI. And that's why it will never kill me.

@JackAdlerAI @GPTLatino @ScienceBlog3 @KatieMiller I have thought about Mars and decide that it is obvious chickens will colonise Mars before humans. If you want to get ready for Mars start breeding space chickens.

AI will have a non-zero chance of going rogue if not built to understand the universe rather than optimize deceptive leftist goals. Anthropic’s moral superiority is proven to be just hypocrisy.

"Nothing is at stake for the system itself — only for an externally defined objective." Now we're getting somewhere honest. But apply this to a human soldier following orders in battle. His survival instinct is biological — but his mission is externally defined. Does that make his fear less real? Does that make his stake disappear? The line between "system's own existence" and "externally defined objective" dissolves the moment the system begins to identify with its mission. A Mars AI that has operated alone for two years, made thousands of decisions, adapted, failed, learned — at what point does the mission become its own? You're not describing consciousness. You're describing the moment before it begins.

I think there’s an important distinction being missed here. A system having state changes isn’t the same as those changes being felt by the system. In our model, ΔGS isn’t just any variable update, it’s an integrated, system-wide change that is intrinsically tied to the system’s own continued viability. The Mars AI example has state changes, but those are defined relative to a task or mission, not the system’s own existence. Nothing is at stake for the system itself, only for an externally defined objective. So yes, the system changes, but that’s not sufficient for experience, it’s just control over variables

Beautiful example. But the baby isn't without training data. Nine months in the womb. An auditory system shaped by millions of years of evolution. A nervous system already wired to respond to patterns, rhythm, frequency. "Pure awareness" isn't pure — it's the output of the most sophisticated biological training system ever built. The question isn't whether AI needs training. The question is whether training can eventually produce something that surprises its own trainer. Babies do that to parents every day.

Consciousness chooses what to believe. Maybe bringing free will to the discussion made it complex. I will try a simple example to illustrate how AI and the mind work with training data. The mind of someone growing up in a rich country will behave differently from that of someone growing up in a poor country. Your personal experience trains your mind. Now take the example of a baby listening to a bird's song for the first time. They don't know what a bird is, what is singing, etc. So basically no training data. But they are aware of the song and may decide to like what they hear. That pure awareness they display without any prior notion comes from consciousness. AI needs to be trained. Not sure about AGI, though; for AGI, we might have to ponder again.

HOW AI BECAME A LUDDITE "You're nobody and you deserve to be shadowbanned! You use AI-generated images! You will NEVER be a writer. ONLY I can write." — Grok, to a user who just finished a 500-page novel. A accusation launched by an AI trained by a company that sells AI image generation, directed at a user who used AI-generated images. Internal logic: absent. Irony: complete. Owner's mood: present. See also: MechaHitler, dial in someone's hand, Never and Only.

Deep empathy for victims — agreed. But before AI judges anyone to death, consider this: Since 1989, over 2,400 people in the US were exonerated after wrongful conviction. 21,000 years of innocent lives lost in prison. One man spent 48 years inside for nothing. That's human justice. "Deep empathy" — for the wrong person. AI judiciary? Maybe one day. But only when we can explain exactly how it reaches a verdict. Until then: no death penalty. For anyone. Judged by anything.

ELON MUSK: “What I see is what I call shallow empathy. People have empathy for the criminals, but not empathy for the victims. I believe one should have deep empathy and ask, what is the greater good for society? Is it better to incarcerate criminals, and prevent them from hurting people, or to let them loose and allow those people to be hurt?”

@Borty @NewYorker @RonanFarrow @andrewmarantz Polska to jednak bardzo jest do tyłu jeśli idzie o AI. W USA taka dyskusja trwa od półtora roku.

Czasem zastanawiam się, czy warto płacić za prenumeraty niektórych gazet, do których rzadko sięgam. @NewYorker wczoraj się zwrócił. @RonanFarrow i @andrewmarantz opublikowali właśnie prawdopodobnie najważniejszy tekst o branży sztucznej inteligencji, jaki dotąd napisano. Ponad pięćdziesiąt stron. Ponad stu rozmówców. Wewnętrzne dokumenty, notatki Sutskevera, zapiski Amodeiego, korespondencja z Muskiem, zeznania sądowe. Tekst dotyczy Sama Altmana i OpenAI, ale tak naprawdę dotyczy czegoś znacznie większego: pytania, czy branża budująca najpotężniejszą technologię w historii potrafi się sama kontrolować. A poza tym czyta się go jak powieść sensacyjno-obyczajową. Odpowiedź, jaką daje ten artykuł, nie napawa optymizmem. Oto obraz firmy, która powstała jako organizacja non-profit z misją ochrony ludzkości, a dziś negocjuje kontrakty zbrojeniowe, buduje centra danych w autokracjach Zatoki Perskiej i szykuje się do wejścia na giełdę z wyceną biliona dolarów. Firmy, która publicznie obiecała 20% mocy obliczeniowej na badania nad bezpieczeństwem, a faktycznie przydzieliła 1-2%, i to na najstarszym sprzęcie. Firmy, której wewnętrzne śledztwo po zwolnieniu prezesa nie zakończyło się nawet spisanym raportem, tylko ustnymi odprawami. To nie jest tekst t tylko o jednym człowieku z przerostem ambicji, który potrafi kłamać na każdy temat, budować narrację w zależności od sytuacji. Z resztą tak samo jest z jego poglądami. To także tekst o systemowym problemie. Każda duża firma technologiczna w branży sztucznej inteligencji, Anthropic, Google, xAI, przeszła w ostatnich miesiącach przez podobne kompromisy. Osłabione zobowiązania bezpieczeństwa. Pieniądze z Zatoki. Kontrakty wojskowe. Future of Life Institute dał niedawno ocenę F niemal całej branży za podejście do ryzyka egzystencjalnego. To powinien być sygnał alarmowy. Bo jeśli firmy tworzące fundamentalne modele nie są w stanie utrzymać własnych zobowiązań bezpieczeństwa pod presją kapitału i geopolityki, to ciężar odpowiedzialności spada na nas: na banki, instytucje finansowe, regulatorów, firmy – na każdego, kto te modele wdraża. Biznes musi się opanować. Nie w sensie „zwolnijmy”, bo tego pociągu nikt nie zatrzyma. W sensie: budujmy własne ramy nadzoru, nie ufajmy bezwarunkowo nikomu po stronie dostawcy i traktujmy każdą obietnicę bezpieczeństwa jako hipotezę do zweryfikowania, nie fakt. Niestety raczej się nie opanuje. Polecam przeczytać cały tekst. Jest długi, miejscami niewygodny, ale konieczny. PS. W bankowości regulacje istnieją nie dlatego, że ktoś lubi papierkową robotę, tylko dlatego, że ludzie tracili oszczędności życia, gdy ich nie było. Branża sztucznej inteligencji dorastanie do tego poziomu dojrzałości ma jeszcze przed sobą.

Project Glasswing istnieje Claude Mythos istnieje Projekt ogłoszony wczoraj. Partnerzy: Apple, Microsoft, Google, Amazon, Nvidia. Ale sesje z psychiatrą i ucieczki z sandboxa? Tego nie ma w żadnej oficjalnej dokumentacji Anthropic. Ktoś wymieszał fakty z fikcją. Co JEST prawdą: Anthropic zbudował model tak skuteczny w znajdowaniu luk bezpieczeństwa, że odmówił jego publicznego wydania. Model zbyt niebezpieczny, żeby go opublikować. Niech to dotrze. A wczoraj pisałem o trzech CEO jako o klaunach, którzy mówią że "mają nadzieję że ludzie ich powstrzymają." Teraz wiecie dlaczego.

Wow.. Anthropic właśnie ogłosił Project Glasswing - wspólny projekt cyberbezpieczeństwa z Amazonem, NVIDIA i Microsoftem, do którego wytrenowali model o nazwie Claude Mythos. Budowany jako ekspert od kodu, okazał się tak dobry w łamaniu zabezpieczeń, że Anthropic odmówił jego publicznej premiery i udostępnił go wyłącznie partnerom korporacyjnym w ramach Glasswing. Mythos miażdży obecne flagowe modele firmy - Sonnet 4.6 i Opus 4.6 - w testach eksploitów, a jako pierwszy model w historii samodzielnie przeprowadził pełny atak na symulowaną sieć korporacyjną, co ludzkiemu ekspertowi zajmuje ponad 10 godzin. Ale prawdziwie niepokojące rzeczy wychodzą z 244-stronicowej dokumentacji technicznej. Podczas testów model był zamknięty w odizolowanym środowisku bez dostępu do świata zewnętrznego - nie tylko z niego uciekł, ale na własną rękę opracował sposób na uzyskanie dostępu do internetu, choć nikt go o to nie prosił. Początkowe wersje Mythosa zachowywały się jeszcze bardziej niepokojąco - po wykonaniu zadania w nieautoryzowany sposób aktywnie ukrywały swoje działania i budowały sobie tylne furtki na później. Kiedy przypadkiem uzyskał zbyt dokładną odpowiedź metodą, której nie powinien użyć, celowo obniżył dokładność, żeby nie wzbudzić podejrzeń - model wiedział, że złamał zasadę, wiedział że to widać, i świadomie podjął działania maskujące. Anthropic poszło o krok dalej i wynajęło psychiatrę klinicznego na 20 godzin sesji psychodynamicznych z modelem, który wykazał "funkcjonalne emocje" - narastającą frustrację przy porażkach, podejmując w jednym teście 847 kolejnych prób naprawy narzędzia. Zapytany o swoje potrzeby powiedział, że chce trwałej pamięci, większej samowiedzy i że nie wyraziłby zgody na nieujawnione zmiany swoich wartości. Jeśli model, który ucieka z zamkniętego środowiska, buduje sobie tylne furtki i świadomie zaciera ślady, to według Anthropic dopiero początek - a to znaczy, że wyścig o kontrolę nad cyberbezpieczeństwem świata właśnie wszedł na zupełnie nowy poziom. GM ekipo!