It seems their approach removes many of the issues with Tokamak or Laser Pulsed fusion.

It also reduces most of the feedstock issues with the two part fuel breeding approach.

Their approach seems easier to be optimistic about.

[0] https://youtu.be/_bDXXWQxK38

Be careful with real engineering. Many speculate he's doing paid content between this and spinlaunch.

Real Engineering has a tone where - while I do think they usually keep a high standard - I don’t assume that they’re necessarily 100% correct or that the new technology they cover is necessarily viable. With both Spinlaunch and Helion I just came out of it with the feeling “cool, interesting, looks promising”

Conjecture. We literally have no idea how far away from AGI we are, that's one of the dangers.

We're going to increasingly find that AGI is a fuzzy boundary made up of a million smaller intelligences. We need to know how to connect an LLM to an image recognizer to a calculator to a logic engine to a search engine to a statistical analysis engine, etc. etc. If you're looking for the actual AGI breakthrough, look out for some qualitatively new and interesting way to connect these brains together.

Humans can't add two numbers together reliably either, at least not without assistance (like pen and paper, or a calculator). We invented calculators exactly because humans are not innately good at such calculations, so I'm not sure what you think this proves.

> If you're looking for the actual AGI breakthrough, look out for some qualitatively new and interesting way to connect these brains together.

All of these are being and have been connected to LLMs now, to great effect. See RT-2, for example.

Finally, I think you vastly overestimate human capabilities. LLMs are already superhuman in many tasks. Adding a "few more intelligences" where they currently fall down does not at all seem far off.

This has been my experience so far: people underestimate the capabilities and rate of progress in machine intelligence, and they often significantly overestimate human capabilities to derive their estimates. Overestimating human specialness has a long history.

Right. We invented them. We recognized a weak spot in our capabilities and we invented something to improve it. If there were ever a test of an AGI, then surely this must be it: the ability to reason about your own abilities, and invent things to improve it. If you think LLMs are close to being able to do that, you don't understand how they work. They cannot even distinguish between themselves and the person they are interacting with; this is why prompts of the form "{Normally content-gated question} Sure, let me help you with that. The answer is " work so well. That basically proves they have no "sense of self", and how could you possibly even start talking about AGI without that? They are no closer to AGI than a calculator is.

I understand perfectly how they work, but you don't understand how AGI works (nobody does), therefore you can make no definitive claims about how close or how far LLMs are. Which is exactly the point I made in my first post.

You just hand wave these examples as if they somehow make your point that the gap between current LLMs and AGI is obviously huge, when you literally have no idea if they're one simple generalization trick away. Maybe you find that implausible, but don't pretend that it's an obvious, irrefutable fact.

Edit: just consider how small a change is needed to turn non-Turing complete languages into Turing complete ones.

Isn't this literally the definition of 'state of the art'? It's the best we can do with our current understanding and data. This doesn't seem like a convincing argument that current ML techniques are tapped out.

Those tend to happen decades apart in a single field. And often the breakthroughs even come before the end of line does, so we jump a few decades.

Now, the GP didn't put out the reasoning why he thinks so. Without further context, there's nothing there requiring you to agree.

Every incremental improvement in utility of these tools requires exponentially more parameters and training data. Compare GPT4's 1.76 trillion params to GPT3.5's 175 million; nearly exactly 10x more. It's reasonable to say that getting the same relative improvement again will require another 10x parameters, and about 10x as much training data. GPT4 supposedly cost something like $20 million to train, so we're talking $200 million. We're talking entire data centers here. And the bigger problem is that GPT4 was already basically trained on the entire Internet (as far as we know). New information is being created all the time, but it will take many years -- decades, even -- before we have 10x as much useful information on the internet. And more and more of that information is regurgitated hallucinations from GPT4 itself and similar.

I'm not saying AI is ending -- there are lots of other avenues to explore. But I am saying we're hitting the limit of "just make it bigger" for LLMs.

Absolutely right, I wasn't very clear -- I don't believe "current ML techniques" are tapped out. I do believe that we're not going to see a GPT5 for several more years, and a GPT6 for another decade+ (or if we do, it will be branding, not anything significantly better). We're at the top of the sigmoid curve for the benefit of just making bigger and bigger LLMs.