611–620 of 683 posts
Earlier quoted context omitted.
Far more common are ideas that don't work on any scale at all. If you have something that gives a sticky +5% at 250M scale, you might have an actual winner. Almost all new ML ideas fall well short of that.
If someone else comes along and makes the exact claim I just made, I won't believe it either
Why is AGI so hungry? Our brains learn on the fly and do amazing things and how much energy do they use? So IMO we certainly haven't hit the "right technology" yet and the attempt to achieve it by spending billions, building nuclear power plants etc, is vulnerable to a technical development. So why should we screw up our environmental situation for something that can obviously be done in a vastly better way on the en…
And at its peak? Human brain doesn't actually have an overwhelming advantage over an LLM. It's a mixed bag of advantages and disadvantages.
LLMs think fast, and can input and output data much faster than a human. But they struggle to work on the same task for a long time, and have a problem with visual inputs and object manipulation. LLMs have more knowledge in total, but humans have better meta-knowledge, which is useful for hallucination avoidance. LLMs can only learn in context efficiently, but humans learn continuously and retain what they learned. LLMs and humans are currently trading blows when it comes to inference energy efficiency - especially when you account for things like sleep or rest.
I don't think there's a "right technology" at all. There may not be a state-change upgrade that gets us x100000 on a dime and goes all the way to an AGI on every smartphone and an ASI in any datacenter worth the name. I expect there to be a lot of little +5% and +10% upgrades that add up over time.
I'm surprised the companies fascinated with AGI don't devote some resources to neuroscience - it seems really difficult to develop a true artificial intelligence when we don't know much about how our own works. Like it's not even clear if LLMs/Transformers are even theoretically capable of AGI, LeCun is famously sceptical of this. I think we still lack decades of basic research before we can hope to build an AGI.
On the other hand, extracting usable insights from neuroscience? Not at all easy. Human brain does not yield itself to instrumentation.
If an average human had 1.5 Neuralink implants in his skull, and raw neural data was cheap and easy to source? You bet someone would try to use that for AI tech. As is? We're in the "bitter lesson" regime. We can't extract usable insights out of neuroscience fast enough for it to matter much.
Earlier quoted context omitted.
We don't know enough about consciousness to be able to conclusively confirm or deny that LLMs are conscious. Claiming otherwise is overconfident stupidity. Of which there is no shortage of that in AI space.
That's the sort of convenient framing that lets you get away with hand wavy statements which the public eats up, like calling LLM development 'superintelligence'. It's good for a conversation starter on Twitter or a pitch deck, but there is real measurable technology they are producing and it's pretty clear what it is and what it isn't. In 2021 they were already discussing these safety ideas in a grandiose way, when…
They knew they had the beginnings of an incredibly capable technology at their hands, and they knew that intelligence is an extremely dangerous thing.
And so far? The capabilities of today's systems are already impressive, and keep improving. If you're thinking "what those systems are doing today isn't that bad", you shouldn't be. Concern yourself with the capabilities of a bleeding edge AI from year 2035.
Why is AGI so hungry? Our brains learn on the fly and do amazing things and how much energy do they use? So IMO we certainly haven't hit the "right technology" yet and the attempt to achieve it by spending billions, building nuclear power plants etc, is vulnerable to a technical development. So why should we screw up our environmental situation for something that can obviously be done in a vastly better way on the en…
Your brain has more capacity than an average LLM, and a self-supervised training regime etched into it by millions of years of evolution. Its inductive bias is very well tuned for the environment. It still takes decades to build itself up to peak performance. And at its peak? Human brain doesn't actually have an overwhelming advantage over an LLM. It's a mixed bag of advantages and disadvantages. LLMs think fast, and…
OK that's a sweeping statement but I think it encapsulates the general attitude quite well for a single sentence. We fly in planes which are not like birds but they have wings and tails and flaps and slats and we're even getting into making the wings warp.
Sleep is needed for memory I think so there is a price to pay for continuous learning. It's important to be able to forget too. - to rotate the logfiles so to speak. AGI only interests me because it will be able to understand us - therefor it has to work like us and experience things like us.
Earlier quoted context omitted.
Why not? A lot of people say that, but no one, not a single person has ever pointed out a fundamental limitation that would prevent an LLM from going all the way. If LLMs have limits, we are yet to find them.
We have already found limitations of the current LLM paradigm, even if we don't have a theorem saying transformers can never be AGI. Scaling laws show that performance keeps improving with more params, data + compute but only following a smooth power law with sharply diminishing returns. Each extra order of magnitude of compute buys a smaller gain than the last, and recent work suggests we're running into economic an…
LLMs failing the same way as humans do on the same tasks as humans is a weak sign of "this tech is AGI capable", in my eyes. Because it hints that LLMs are angling to do the same things human mind does, and in similar enough ways to share the failure modes. And human mind is the one architecture we know to support general intelligence.
Anthropic has a more recent paper on introspection in LLMs, by the way. With numerous findings. The main takeaway is: existing LLMs have introspection capabilities - weak, limited and unreliable, but present nonetheless. It's a bit weird, given that we never trained them for that.
https://transformer-circuits.pub/2025/introspection/index.ht...
You can train them to be better at it, if you really wanted to. A few other papers tried, although in different contexts.
Earlier quoted context omitted.
Why not? A lot of people say that, but no one, not a single person has ever pointed out a fundamental limitation that would prevent an LLM from going all the way. If LLMs have limits, we are yet to find them.
Real time learning that doesn't pollute limited context windows.
> And this is all fine, because they’re going to make AGI and the expected value (EV) of it will be huge! (Briefly, the argument goes that if there is a 0.001% chance of AGI delivering an extremely large amount of value, and 99.999% chance of much less or zero value, then the EV is still extremely large because (0.001% * very_large_value) + (99.999% * small_value) = very_large_value). This is a strawman. The big AI n…
Earlier quoted context omitted.
Not why it was created, why the systems in the brain lead to consciousness. Your option B requires understanding not just mechanically how but the fundamental reason for why consciousness appears. If we just understand the mechanics all we can confidently do is work toward a more and more accurate representation of the brain. AGI without consciousness is speculated but hard for me to believe in.
As noted, consciousness seems to just be the ability to self-observe, which is useful as another predictive input. I would expect that all intelligent animals are conscious, and any AI we build with a roughly brain-like architecture in terms of connections, looping, and being prediction based would also report itself to be conscious and describe a similar subjective experience. LLMs seem much too simple (just layer-w…
As far as I know, consciousness is referring to something other than self-referential systems, especially with regards to the hard problem of consciousness.
The [philosophical zombie](https://en.wikipedia.org/wiki/Philosophical_zombie) thought experiment is well-known for imagining something with all the structural characteristics that you mention but without conscious experience as in "what-its-like" to be someone.
Earlier quoted context omitted.
Dial-up modems can transfer a 4K HDR video file, or any other arbitrary data. It obviously wouldn't have the bandwidth to do so in a way that would make a real-time stream feasible, but it doesn't involve any leap of logic to conclude that a higher bandwidth link means being able to transfer more data within a given period of time, which would eventually enable use cases that weren't feasible before. In contrast, you…
From modern perspective it's obvious that simply upping the bandwidth allows streaming high-quality videos, but it's not strictly about "more bigger cable". Huge leaps in various technologies were needed for you to watch video in 4k: - 4k consumer-grade cameras - SSDs - video codecs - hardware-accelerated video encoding - large-scale internet infrastructure - OLED displays What I'm trying to say is that I clearly rem…
Your anecdote regarding P2P file sharing is ridiculous, and you've almost certainly misunderstood what the author was saying (or the author themselves was an idiot). That there wasn't sufficient bandwidth or computing power to stream 4K video at consumer price points during the heyday of mp3 file sharing, didn't mean that no one knew how to do it. It would be as ridiculous as me today saying that 16K stereoscopic streaming video can't happen. Just because it's infeasible today, doesn't mean that it's impossible.
Regarding ChatGPT, setting aside the fact that the transformer model that ChatGPT is built on was under active research 10 years ago, sure, breakthroughs happen. That doesn't mean that you can linearly extrapolate future breakthroughs. That would be like claiming that if we developer faster and more powerful rockets, then we will eventually be able to travel faster than light.