Odd. As I see it he was one of the few right about everything. This is also taking his comments about transformers scaling out of context. He specifically said they won’t lead to human level ai and they haven’t, despite exhausting the training data.

AGI will require architecture changes not simply more data.

▲ mr_roboto a day ago | parent | next [-]

I'm quite confused by what you mean human level AI here. Maybe we haven't hit Einstein but we blew past the average person a long time ago.

▲

haltingproblem a day ago | parent | next [-]

It is very jagged and some people don't consider that human level AGI - it can solve Maths at the genius level just like Deep Blue could play chess at the grandmaster level. At this point it is simply an opinion and I can see the other side too.

▲

paganel a day ago | parent | prev | next [-]

> but we blew past the average person a long time ago.

We did?

▲

procone a day ago | parent | prev | next [-]

Sorry, but what? We haven't done anything beyond strapping a large language model to a "harness".

At no point do I ever confuse a human being with a Claude session.

Even if I'm somehow fooled or confused for a minute, the context window / hallucinations / Claudeish will immediately give it away.

Sure, we passed the turing test for chatbots but we are very far away from AGI.

▲

mr_roboto 13 hours ago | parent [-]

That could be how human intelligence stacks, search retrieval (LLM) with learned approaches to processing that info (harness). Maybe you are insulated from the average person. The frontier models easily surpass their intellectual ability to absorb new information and apply it to novel situations.

The average person will also hallucinate things and make all sorts of causality mistakes without realizing it AND refuse evidence of such whereas models tend to acquiesce to evidence. A model with the right harness will outperform any of these people at their current desk jobs. To me that is general intelligence. Most of us on HN tend to far surpass this general intelligence and likely experience a false consensus or assumed similarity bias.

	▲	procone 4 hours ago \| parent [-]
		Sorry but I stopped reading about halfway through. This is an insane worldview and you will be impossible to converse with.

▲

Exoristos a day ago | parent | prev [-]

With this perspective on LLMs, you are presumably often "quite confused" if you interact with them to a significant extent.

▲ godelski a day ago | parent | prev [-]

  > This is also taking his comments about transformers scaling out of context

There have been significant improvements in architectural changes, but since "pre-training" isn't being openly published people just lazily attribute it to "scale". Scale of what? Scale

  > AGI will require architecture changes not simply more data.

As Sutton claimed, all along[0]. People read what they wanted to read. Unfortunately, they also push against researchers who were trying to explore other architectures and methods. I honestly believe that has held us back. There was always room for people to explore different paths. The problem was, and has always been, preventing or discouraging others from exploring new paths. It's ironic given that one of the most common reasons a paper is rejected is for "lack of novelty"[1]. But if someone does something novel then the reviewer simply compares it to the existing state of the art, which is an unfair comparison[2].

It's silly since people want to move fast, but we're really not good at recognizing what has actually allowed us to move forward in the past.

[0] He speaks about Bitter Lesson directly at 9:57 but the earlier part gives better context: https://www.youtube.com/watch?v=21EYKqUsPfg

[1] An essentially meaningless statement. Near meaningless because it is not objective.

[2] You compare a novel technique against a technique that has had hundreds or thousands of researchers optimizing over several years. That's not even remotely fair.

	▲	tensor 9 hours ago \| parent [-]
		100%. You can both commercialize transformers while still exploring new paths. Having one of the world's top researchers doing productization work is a huge waste of their time. You also see this whenever news about novel but small models comes out. People are instantly dismissive. "Oh it's small, who cares?" Of course it's small, why waste resources testing it on a larger scale before making sure it works well on a small scale? Would people rather news not come out until it's SOTA model?