Hacker News Reader

Comment by dvt | Hacker News Reader

I think people are just getting lost in the sauce. Forget all the "singularity" or "AGI" nonsense. LLMs are genuinely useful automation machines. They're fantastic for going from semi-structured data to structured data. They're great for going from text blob to decision points. They're great for going from vague instructions to step-by-step inference.

No one (at least no serious person) is saying ChatGPT is Immanuel Kant or Ernest Hemingway. The fact that we still have sherpas doesn't make trains any less useful or interesting.

View on HN

I think this post is specifically an answer to yet another "AGI is just around the corner" post that made waves recently.

Fundamentally, I think that many problems in white-collar life are text comprehension problems or basic automation problems. Further, they often don't even need to be done particularly well. For example, we've long decided that it's OK for customer support to suck, and LLMs are now an upgrade over an overseas call center worker who must follow a rigid script.

So yeah, LLMs can be quite useful and will be used more and more. But this is also not the discourse we're having on HN. Every day, there's some AGI marketing headline, including one at #1 right now from OpenAI.

The AI-assisted GPT theoretical physics derivation? There's literally no mention of AGI in the article, and it's pretty tame, especially considering it's a PR piece by OpenAI.

It's referencing https://shumer.dev/something-big-is-happening

It's a press release from a vendor that constantly talks about AGI, and it's meant to showcase the capabilities of an unreleased model in an experiment you can't replicate. But my comment was less about the link and more about the discussion, which has immediately bifurcated into the "it's done and dusted" and "this is overhyped and LLMs are useless" camps.

I think what's surprising people is how a rough, first-order approximate solution (produced with little cognitive effort) is good enough for 90% of everyday tasks

This is what I've been saying. We're not so much learning that LLMs are intelligent, we're learning that a lot of what we think of as intelligence is actually just pattern matching.

But the real intelligence exists and it is beyond all those trinkets. Right? Yeah, this is a common cope.

There's no solid evidence that the brain uses quantum computations. And that's the only definite thing (besides 'magic') that can solidly place the brain out of reach for now.

To frame the human brain itself as a startup, i'm not convinced by "latent quantum effects" at all as the moat for human sypanses.

There's something else that we are missing there- probably at the convergence of multiple biochemical effects (like ion transport).

Evolution has scaled the brain to hundreds of trillions of synapses (parameters) and hundreds of billions of neurons (processing nodes). Current models have around 1/1000 of that number of parameters. The top data centers operate at around 1 exaflop (on par with the brain). Inference uses significantly less than that.

I’ll wait until the effective number of parameters in ML models is comparable to that of the brain. Then it will be clear whether the brain has something up its sleeve.

If I showed you a new species of animal that does exactly what an LLM does, what would you say? Let’s say a bird, you ask it a question , and it returns an expert level human response. What if these new birds were everywhere?

That’s very big.

I still can’t believe parrots are real.

And yet, no one thought of replacing Excel's ancient "convert text into columns" wizard with one of these.

People have. Microsoft wont since their incentive structure is fucked

It actually could be built reliably if the verification is easy

Reply on news.ycombinator.com