Loading video...

Video Failed to Load

Go Home

🚨 UNCANNY VALLEY: NEXT TOKEN PREDICTION LACKS TRUE GENERALIZATION "Transforming neural nets can generalize a little bit, and they don't generalize as well as a human. You could train the next token predictor to do anything if you had enough resources. I don’t think the next token prediction is...

94,723 views • 1 year ago •via X (Twitter)

10 Comments

Eddy's profile picture
Eddy1 year ago

👀👀

Barefoot Pregnant's profile picture
Barefoot Pregnant1 year ago

AI’s got a long way to go next-token models aren’t even close to real generalization.

🌐StateOfTheDebate's profile picture
🌐StateOfTheDebate1 year ago

Next-token prediction won’t cut it for AGI true intelligence needs more than just patterns.

University.com's profile picture
University.com1 year ago

AI might predict your next move, but it still can't predict why liberals think taxation is a good idea!

SM Shelpy's profile picture
SM Shelpy1 year ago

Smart money is always looking for move creative tech to engage with and it's not all about the token

Boost's profile picture
Boost1 year ago

Transformers fall short. Generalization is the future of AGI.

stocktrader ₿'s profile picture
stocktrader ₿1 year ago

@grok explain this

Joe.N's profile picture
Joe.N1 year ago

@tweemdotlol 10x views

rahimi.base.eth's profile picture
rahimi.base.eth1 year ago

I get the point but is there a better paradigm for AGI out there?

ℚ𝟝𝕚𝕖𝕥's profile picture
ℚ𝟝𝕚𝕖𝕥1 year ago

next token’s a party trick, not agi

Related Videos

⚫UNCANNY VALLEY: THE STRANGEST TRUTH ABOUT AGI—WE STILL DON’T KNOW WHAT WE’RE MAKING Special Guest: Ben Goertzel Host: Dr. Danish What if the smartest AI ever built still can’t think? What if the companies chasing AGI don’t even know what it is? Dr. Ben Goertzel—creator of Sophia the robot and founder of SingularityNET—says the real path to human-level AI isn’t bigger data or better training... it’s generalization. And today’s LLMs just don’t have it. From logic engines to mind children, hallucinating machines to post-human ethics—this isn’t just a tech talk. It’s a warning: we’re racing toward something we don’t understand. Welcome to The Uncanny Valley Weekly Series, Fridays at 4:20 PM ET, ONLY on 𝕏. Episode 6: THE AGI DEBATE—IS TOKEN PREDICTION ENOUGH? 01:05 – “A laptop is a body.” Why the robotics debate is a distraction. 03:46 – Embodiment isn’t required—but it does make AGI feel more human. 05:30 – Ghosts in the machine? The most bizarre take on disembodied AI. 07:12 – Stop defining AGI. Just build it. 10:49 – Why LLMs feel powerful—but fall flat at true generalization. 12:53 – “An LLM won’t invent jazz.” What AI can’t create without a soul. 16:23 – Intelligence is about leaping into the unknown—not predicting the past. 18:31 – Token prediction won’t get you there—unless you have infinite compute and luck. 20:32 – Real minds evolve. Why individuation and self-transcendence matter. 24:39 – OpenCog Hyperon: a strange cocktail of logic, evolution, and deep learning. 26:56 – Why Big Tech fears symbolic AI—and why it might be the missing piece. 30:27 – Evolution didn’t build us to solve today’s problems—it built us to adapt to tomorrow’s. 33:40 – Monolithic AI isn’t science. It’s a business model. 36:26 – The next great AI architecture won’t need your server farm. 39:30 – “These are our mind children.” Teaching empathy to the machines we raise. 40:20 – AGI might love us—or wipe us. There are no guarantees.

Mario Nawfal

1,312,508 views • 1 year ago