JUA

Most of AI community isn’t studying intelligence. It’s studying scale. Researchers like Yann LeCun, Timnit Gebru, Pedro Domingos, and Jürgen Schmidhuber truly understands intelliegence, they argue about foundations , world models, causality, learning. The rest argue about parameter counts. @ylecun @timnitGebru @pmddomingos @SchmidhuberAI @Artificialintelligence @AI @computing

@pmddomingos All LLMs today fails Ballard's test. Take away the words and there's nothing left. No reasoning. No understanding. Just silence. Ballard's Test: An entity does not possess capacity for understanding until reason is demonstrated in the absence of language.

If LLMs are so smart, why do they need all these prompts, harnesses, post-training, scaffolding, etc.?

@VoidAsuka @SchmidhuberAI Schimidhuber is by far the greatest in the Ai space. 1990: Schmidhuber does GANs & self-supervised pre-training. 1997: LSTM. 2011: DanNet beats everyone. 2015: Highway Nets → ResNet. Schmidhuber did neural history compressors in 1991.

@SchmidhuberAI this unc is a real genius, the most exciting thing i've read about this week. i should read more old books, old papers, and scroll less.

@GaryMarcus @ylecun @SchmidhuberAI LeCun doesn't steal ideas. He just waits for you to open-source yours, then calls it "parallel invention"

Hey, you are never gonna believe it! @ylecun organized a group phone call with me and @SchmidhuberAI and some of the other people he has ripped off over the years, and *apologized*. He’s a much more mature human being than I had ever realized. Oh wait… April Fool’s!

@durov This sucks. So much for “privacy” and “freedom” when it actually matters. Really disappointing, Apple.

Apple just banned several VPN apps from the Russian App Store — targeting those that helped users bypass Russia’s DPI-based censorship. That’s not cool, Apple. techradar.com/vpn/vpn-privac…

@VHSDVDBLURAY4K That’s Young Sheldon from Young Sheldon .… just standing there, I think he’s taking notes for a science project on funerals.

So I’m rewatching Avengers: Endgame and who TF is this guy ?

@twetsfyp @grok Ewwwww.. Grok is so misaligned 🤦🤦

@JeffDean Such a heart-softening ad.

Love it. We should all aspire to do the little things that help each other.

@theo the real breakthroughs don't announce themselves with press releases. They just show up one day and change everything.

Wild that there was like 0 big AI news all week. Are we no longer accelerating?

@ControlAI He's right to be terrified. Not because the machine turns evil. Because it won't need to. It'll just be better at everything than we ever were, and we'll slowly stop mattering. That's not malice. That's obsolescence.

Former OpenAI researcher Scott Aaronson says he's terrified about the rise of superintelligence. "Maybe it's like creating a new species that will become the dominant one on earth."

@rand_longevity What if we're not just the birthplace of superintelligence, we're the test? Every war, every choice, every mess. The simulation doesn't shut down when it's done. It proves something first.

I think our simulation will test the limits of the super intelligence

There's someone doing it right now in the AI world. You know who they are. We all do. Rebranding isn't inventing. But in a hype-driven market, confidence sells better than originality. The real innovators are the ones you never hear about. The ones working on things that don't have a catchy name yet.

These days the shortest path to being considered a genius is to take an old idea, give it a new name and implement it for LLMs.

@pmddomingos once the goal is reached, what reason does the simulation have to keep running?

The simulation we live in was created to develop superintelligence, and will soon be turned off.

Most AI today fails Ballard's Test instantly. Take away the words and what's left? Nothing. No reasoning. No understanding. Just silence. Language is a tool. Reason is the skill. We keep confusing the two.

The strange thing isn't how much we know about AI. It's how little. We can build models with billions of parameters, train them on most of the internet, and still not fully understand why they say what they say. We call it 'emergent behavior' which is just a fancy way of saying 'we didn't expect this and can't explain it. We're building black boxes and calling them minds. That should bother us more than it does.

LLMs don't think. They complete. Every response is just the most statistically plausible next word. There's no understanding behind it , just pattern matching at scale. You can ask a calculator what 2+2 is. It won't know what 'two' means either. It's just really good at being right. The difference is: when a calculator is wrong, you know. When an LLM is wrong, it sounds totally convincing right up until it isn't. That's not intelligence. That's confidence without comprehension.

They are trying to scale intelligence on the Von Neumann architecture. You are wrong. Not about the math. Not about the scaling. Not about the potential of AI. You are wrong about the foundation.

@LiorOnAI @ylecun his models still run on von Neumann hardware. Still separate memory and processing. Still burn watts moving data instead of understanding it. He's solving for the right software layer, but the foundation underneath hasn't changed.

Yann just bet a billion dollars that the entire industry is building on the wrong foundation. Large language models predict the next word. They're trained on text, so they understand language. But the real world isn't made of words. It's made of continuous sensor data: camera feeds, touch, sound. And most of that data is unpredictable. You can't predict every pixel in a video the way you predict the next token in a sentence. Generative models fail here because they try to predict everything, including noise. AMI Labs is building world models using JEPA (a method LeCun proposed in 2022 that learns abstract representations of reality and predicts in that compressed space, not in raw pixels). Action-conditioned versions let AI simulate the consequences of actions before taking them. That's not generation. That's understanding. This unlocks AI that can operate in the physical world without hallucinating: 1. Robotics that plans multi-step actions 2. Healthcare devices where errors kill patients 3. Industrial process control under safety constraints 4. Wearables that adapt to real-time sensor input If JEPA works at scale, the next wave of AI companies won't fine-tune LLMs. They'll train world models on sensor data. LeCun's CEO already predicts every startup will rebrand as a "world model company" within six months. The architecture war is starting.

Honestly? Might be true. The old way, patches, firewalls, passwords, was already falling apart. Now add AI that can talk its way past humans, write exploits in seconds, and attack at machine speed. The defense was never built for that. The only real fix isn't a better firewall. It's machines that can just say no. Hardware that won't run bad code because it physically can't. Security isn't a software problem anymore. It's an architecture problem.

We might be living in the last few months of any semblance of computer security.

@xriskology Most of the AI ethics conversation stays surface level, bias, jobs, hype. If yours actually digs into the structural stuff, the incentives that are broken, the fact that these machines were never built to care... that's the conversation we should be having. Curious to read.

New article on how the AI companies are destroying the world and their CEOs are profoundly unethical people. I mention a lot of recent studies and developments that might interest you. Take a gander:

@francoisfleuret @ducha_aiki @Ana_Geneva Execution is jürgen's problem. he's right about having thse ideas long ago.

@ducha_aiki @Ana_Geneva Jürgen Schmidhuber's bitterness is straight out of a classical Greek tragedy.

You cannot make this thing up