Might be scary and filled with all sorts of troubles depicted in movies like the The Matrix and I-Robot, for me, that's the truly exciting future.

Humanoids could populate Mars, and other harsh planets before us and help figure out so much, human labour's could effectively become extinct, Peter Diamadis' vision of abundance would become a reality, and so much more...

But all this, in many ways, is hinged on our ability to figure out "how humans reason" and effectively build that into AI.

For me, and I say this with an obvious bias, the real magic lies in the physical world, embodied AI, robotics, digital twins, cyber-physical systems...that Jetson, flying saucer, robot maid future that was promised.

Now, that would be truly something. Clearly the success of these physical systems lies in the sophistication of AI in digital systems, but the real magic of AI comes when it's embodied in a physical system capable of effecting change in the real world.

So yes, while that future...

The last thing I'd say is..

Personally, there's almost nothing that truly excites me in terms of developments in AI.

Simply reason is, even if you get AGI and ASI, as long as they are on a computer, like confined to software and an OS, they are still super limited.

You could have a full blown super intelligent agent that can do every digital thing you could think of, but as long as it's confined to the realm of 1s and 0s, there's still only so much it can do...

Figure it out, but for now, yes, all we should do is get the best out of all these tools, and keep abreast with how things are evolving.

I believe the key to unlocking AIs ability to truly reason, might lie in our true understanding of how the brain, down to the finest details of neurons, works.

Obviosuly, neural networks, were named so because of their similarity to how millions/billions of neurons are shaped and interconnected...

But in terms of how they truly work, NNs are a pale imitation and don't even come close to the power of a 5 year old's brain.

Like I said, I have no doubt that if given the time, humans will...

...I mean folks like Geoffrey Hinton, Yann Le Cunn(who I mentioned earlier), Yoshua Bengio, Andrew Ng, and maybe Fei-Fei Li ...have consistently spoken about the comparison between how the brain works and LLMs...

Or in other words, there's a junction between those two things, that they've yet to understand or uncover fully.

I'm not an AI researcher, my academic interests are in a different yet adjacent field (Robotics and Embodied AI), but from all I've read, heard and understood...

...GenAI chatbots, predicting the next word/token because they have seen tons of examples, understand relationships between words in sentences etc.

One can go into soooo many details, and how a lot of this evolved over the years from CNN and RNN paradigms, but...

...that ability to reason, like humans do, from all (the little) I know about this, can't be based solely on probability, token, and atrention-mechanism paradigm.

Something seems to be missing...

The Godfathers, from time to time...

Hopefully @guembeblessing can share his thoughts too, he's a brilliant AI researcher doing some great work in AI-privacy, Federated Learning etc.

But back to this, the basic building blocks of LLMs + Tranformers (attention mechanism), when you really really boil it down, is a super-charged probabilistic engine.

Tons of training data, intelligent/cutting-edge algorithms, optimizations of all sorts ...slap the attention mechanism on all that, what you get is...

Please keep going... I'm enjoying the insight. It's a topic I highly interested in

Also, I know an AI researcher on TwoCents @guembeblessing that would be interested in sharing his thoughts on this.

...Satya Nadella said the same thing, in a slight worrisome tone, that LLMs simply can't reason.

For true AGI or ASI to be realized, this "issue" has to be solved, and I don't doubt it will be solved....

But as for the current state of things, your writeup captured the main juice of things, I'm loooking forward to the next post!

...squeezed for all they've got to get smarter and smarter, fewer hallucinations, larger token contexts etc.... but still, from all I know at this point, and granted I'm not an AI researcher, the limitations of the LLM paradigm are already evident.

I mean OpenAI, Anthropic, Mistral, DeepSeek etc will keep doing what they have to do to remain viable and competitive in the LLM, GenAI front, but I suspect that they are already looking for the next big AI paradigm.

Asides from Yann Le Cunn...

...he said it, that LLMs, from all he's researched and studied, simply can't reason...

I chuckled when I heard this, because the signs were already there, I mean all that's happening under the hood, simplified 100×, is probability.

When you breakdown all of it, the gazillion amounts of trading data, optimization algorithms, etc...it's really just that.

So really, while all that's happened is great, I still think there's something missing ...I mean, these models can be optimized...

Brilliantly written💯 + there's so much to unpack in this.

Probably the biggest thing on my mind, about LLMs in particular, is that, with all I've learned and seen, they simply can't reason.

It started as a nagging thought some months ago, while I thought about the basic LLM building blocks of Pre-trained Transformers ... tokens, attention-mechanism, and probabilistically predicting the next word...something didn't feel quite "complete" / "full-proof" about it...

And then Yann LeCunn said