Hacker News Reader

Comment by hokeone | Hacker News Reader

>Furthermore, the fact that LLMs seem to need such a stupendous amount of data to get such mediocre reasoning indicates that they simply are not generalizing. If these models can’t get anywhere close to human level performance with the data a human would see in 20,000 years, we should entertain the possibility that 2,000,000,000 years worth of data will be also be insufficient. There’s no amount of jet fuel you can add to an airplane to make it reach the moon.

Never thought about it in this sense. Is he wrong?

View on HN

Demis Hassabis of Deepmind echoes a similar sentiment[0]:

> I still think there are missing things with the current systems. […] I regard it a bit like the Industrial Revolution where there was all these amazing new ideas about energy and power and so on, but it was fueled by the fact that there were dead dinosaurs, and coal and oil just lying in the ground. Imagine how much harder the Industrial Revolution would have been without that. We would have had to jump to nuclear or solar somehow in one go. [In AI research,] the equivalent of that oil is just the Internet, this massive human-curated artefact. […] And of course, we can draw on that. And there's just a lot more information there, I think, it turns out than any of us can comprehend, really. […] [T]here's still things I think that are missing. I think we're not good at planning. We need to fix factuality. I also think there's room for memory and episodic memory.

[0]: https://cbmm.mit.edu/video/cbmm10-panel-research-intelligenc...

His view of the Industrial Revolution is completely wrong.

Societies pre-IR had multiple periods where energy usage increased significantly, some of them based specifically around coal. No IR.

Early IR was largely based around the usage of water power, not coal. IR was pure innovation, people being able to imagine and create the impossible, it was going straight to nuclear already.

Ironically, someone who is an innovator believes the very anti-innovation narrative of the IR (very roughly, this is the anti-Eurocentric stuff that began appearing in the 2000s...the world has moved on since then as these theories are obviously wrong). Nothing tells you more about how busted modern universities are than this fact.

Has the narrative moved on? The historian and blogger Bret Devereaux presents a view on a 2022 blog post that seems to back up what the Deepmind CEO is saying.

> The specificity matters here because each innovation in the chain required not merely the discovery of the principle, but also the design and an economically viable use-case to all line up in order to have impact.

https://acoup.blog/2022/08/26/collections-why-no-roman-indus...

I am very curious on what you mentioned, but not able to comprehend. Can you ELI5? Are you saying fossil fuel based industrial revolution is not as significant as it was or we could have directly jumped to a higher level fuel?

Societies pre-IR had multiple periods where energy usage increased significantly, some of them based specifically around coal. No IR.

That's a straight up misstatement of the parent argument - the parent argued that coal was necessary, not that coal sufficient. True or not, the argument isn't refuted by the IR starting with water power either.

And pairing this with "anti-woke" jabs is discourse-diminishing stuff. The theory that petroleum was a key ingredient of the IR is much older than that (I don't even agree with it but it's better than "pure innovation" fluff).

It isn't a misstatement, it is (as I explained) a common argument that falls within Marxist/materialist viewpoints such as Robert Allen's account of global IR, it was the dominant viewpoint when this guy was at uni, you often hear people in the UK of that age saying this stuff (I know, I studied economic history during this time) but the field has continued to progress since then. Water energy wasn't a "starting", it was the IR. Coal didn't come until significantly later (and the major issue with materialist history is also what happened in the 20th century, you had multiple countries attempt and fail to industrialise using this idea that energy intensity was the only thing that mattered, the biggest issue with anti-Eurocentric theory is that it was designed to explain a 40-year period around the end of the 18th century and completely fails to generalize).

What is anti-woke? You realise that stuff existed before zoomers starting saying everything was woke/anti-woke. Eurocentrism is a school of thought within economic history, it is nothing to do with wokeism...I have no idea how these two things are related apart from you trying to relate it to something you understand, i.e. pop culture.

"Pure innovation" fluff is the dominant theory today, McCloskey's books are the most important ones in this school. To call this "fluff" suggests ignorance rather than the superiority that you seem to be trying to portray.

Petroleum wasn't a key ingredient of IR...at this point, I am assuming you know nothing about basic aspects of economic history because petroleum wasn't widely used as a fuel until the 1930/40s (again, you seem intent on talking about things that you know rather than the actual subject).

I don't think the data is the weakness.

We're using Transformer architecture right now. There's no reason there won't be further discoveries in AI that are as impactful as "Attention is All You Need".

We may be due for another "AI Winter" where we don't see dramatic improvement across the board. We may not. Regardless, LLMs using the Transformer architecture may not have human level intelligence, but they _are_ useful, and they'll continue to be useful. In the 90s, even during the AI winter, we were able to use Bayesian classification for such common tasks as email filtering. There's no reason we can't continue to use Transformer architecture LLMs for common purposes too. Content production alone makes it worth while.

We don't _need_ AGI, it just seems like the direction we are heading as a species. If we don't get there, it's fine. No need to throw the baby out with the bath water.

And yet, we reached the moon, and I would say airplanes were a necessary step on the way, even if only for psychological reasons. For airplanes we had at least an example in nature, birds. But I am not aware of any animal that travelled from earth to the moon on its own, except us.

We are talking of LLMs, not whether we will be able to reach AGI or not.

Airplanes in this analogy are essentially the collection of matrix multiplications that emulate reasoning in a very rough but useful manner in an LLM.

It's unclear whether a rocket ship is a multimodal neural net. Or some sort of swarm of LLM's in an adversarial relationship, or something completely novel. Regardless, we might be as far between LLM's to ASI's, as airplanes are to rocket ships. Or not.

But we didn't use airplanes to get there. It needed a new approach, different propulsion, different fuel, different attitude control, etc. etc.

LLM may be a necessary step to get to AGI, but it (probably) won't be the one that achieves that goal.

I doubt that LLMs will give us AGI. But they have already given us more intelligence from a computer than I would have imagined to see during my lifetime.

I mean, if you go through the Apollo program contractors it's a who's who of aerospace. Boeing, North American Aviation, Grumman Aircraft, McDonnell Aircraft Corporation, Bell Aerosystems, Rocketdyne...

Electrical parts ran at aviation-standard 400hz. Aviation gyroscopes and aviation instruments. Structural parts made of aviation aluminum alloys. Astronauts that are all airplane test pilots. I can imagine doing Apollo from complete scratch (using car manufacturers that have to invent aluminum-handling tech starting from nothing) but it would have taken a lot more than the decade Apollo took.

Sorry, but what’s with HN’s obsession with analogies? You see this in almost every comment section where someone tries to dis/prove a point using an analogy. I get the allure but it’s intellectually brittle; by the time someone starts to argue off the second or third incantation of the original analogy, the forest has been lost for the tree.

What if LLMs are hot air balloons of flight, or kites of flight? Kites and hot air balloons didn't really lead to getting to the moon, they are a very different tangent.

> But I am not aware of any animal that travelled from earth to the moon on its own, except us.

Tardigrades might :)

LLMs are closer to discoveries on the spectrum than inventions. Nobody predicted or planned the many emergent capabilities we’ve seen. Almost like magic. Now is a period of moving along the axis to invention with many intentional design, architecture, and feature development alongside testing and evaluation. We are far from done with LLMs, plenty of room for many more discoveries, lots to explore. It’s definitely a precursor to AGI. They offer a platform to build and scale data sets and test beds.

We haven’t had ML models this large before. There’s innovation in architecture but we often come back to the bitter lesson: more data.

We’re likely going to see experimentation with language models to learn from few examples. Fine tuning pretrained LLMs shows they have quite a remarkable ability to learn from few examples.

Liquid AI has a new learning architecture for dynamic learning and much smaller models.

Some people seem mad about the bitter lesson, they want their model based on human features to work better when so far usually more data wins.

I think the next evolution here is in increasing the quality of the training data and giving it more structure. I suspect the right setup can seed emergent capabilities.

The trick is to make many LLMs work together in feedback loops. Some small some big.

That will get us to what was previously known as AGI. The definition of AGI will change, but we will have systems that put perform humans in most ways.

Isaiah 7:14 (NIV): "Therefore the Lord himself will give you a sign: The virgin will conceive and give birth to a son, and will call him Immanuel."

"The Shoe Is The Sign!"

-- Monty Python.

> It’s definitely a precursor to AGI.

What are you basing this claim on? There is no intelligence in an LLM, only humans fooled by randomness.

Maybe we've been fooling each other since forever too.

However whatever we're doing seems to be different from what LLMs do, at least because of the huge difference in how we train.

It's possible that it will end up like airplanes and birds. Airplanes can bring us to the other side of the world in a day by burning a lot of fuel. Birds can get there too in a much longer time and more cheaply. They can also land on a branch of a tree. Airplanes can't and it's too risky for drones.

> only humans fooled by randomness

Is there another kind?

This is such an interesting take. What do you classify as intelligence?

From my perspective theres intelligence in a how to manual.

It seems like maybe you mean consciousness? Or creativity?

Precursor as in it will help in synthetic data generation, testing, etc. at scale that gives us more powerful models. It is a necessary intermediate step on our path to AGI.

Fair enough, it could be but I'm not sure it has to be in any way...

You might want to reconsider your stance on emergent abilities in LLMs considering the NeurIPS 2023 best paper winner is titled:

"Are Emergent Abilities of Large Language Models a Mirage?"

https://arxiv.org/abs/2304.15004 https://blog.neurips.cc/2023/12/11/announcing-the-neurips-20...

Models trained mostly in English and trained only to respond in English can respond in another language. That ability is not a mirage. It’s very very real.

Over the past year there have been advances in making models smaller while keeping performance high.

So if that continues then he is wrong unless he is defining LLMs in a strict way that does not include new improvement in the future

For an example, the diagrams in the post compare the big gpts, but looking at the number of tokens PHI-2 sits below gpt3. And it still beats it in Humaneval and a few other benchmarks.

It's not about the size of the models, it's about the size of the training data.

Humans are able to begin to generalize with a single persons experiences over less than a year, so the fact that LLMs cannot with billions of person-years of information could be an indicator of their inability to generalize no matter how much training data you throw at it.

Even the largest LLM has had less "total information" than most humans take in through all of their senses over their lifetime. A single day for a baby is taking in a continuous stream of among other things high quality video and audio and does a large amount of processing on that. Much of that for very young babies is unsupervised learning (clustering), where baby learns that object A and object B are different despite knowing nothing else about their properties.

Humans can learn using every ML learning paradigm in ever modality: unsupervised, self-supervised, semi-supervised, supervised, active, reinforcement based, and anything else I might be missing. Current LLMs are stuck with "self-supervised" with the occasional reinforced (RLHF) or supervised (DPO) cherry on top at the end. non multi-modal LLMs operate with one modality. We are hardly scratching the surface on what's possible with multi-modal LLMs today. We are hardly scratching the surface for training data for these models.

The overwhelming majority of todays LLMs are vastly undertrained and exhibit behavior of undertrained systems.

The claim from the OP about scale not giving us further emergent properties flies in the face of all of what we know about this field. Expect further significant gains despite nay-sayers claiming it's impossible.

You are obviously a believer so you should know I know how to build AGI with a patented and trademarked architecture called "panoptic computronium cathedral"™. Tell all your friends about it. I only need $80B to achieve AGI.

The Phi paper and various approaches to distilling from GPT-4 demonstrate that the training data and plausibly order of presentation matter.

The challenge is that we both do not understand which set of data is most beneficial for training, or how it could be efficiently ordered without triggering computationally infeasible problems. However we do know how to massively scale up training.

I don't think he is wrong. I also don't think the goal of LLMs is to reproduce human intelligence. That is, we don't need human-like inteligence in a box for a tool to be useful. So this assertion could be right and still miss the point of this tech in my opinion.

Edit: to expand, if the goal is AGI then yes we need all the help we can get. But even so, AGI is in a totally different league compared to human intelligence, they might as well be a different species.

The context of the fine article is scaling LLMs into AGI. It's not about whether the tool is useful or not, as usefulness is a threshold well before AGI. Some folks are spooked that LLMs are a few optimizations away from the singularity, and the article just discusses some reasons why that probably isn't the case.

The article is really good! I was responding to "is he wrong" part of the comment, not the article itself.

We don’t need human-like intelligence in a box for a tool to be useful, But human-like intelligence is what many companies are spending billions to try and achieve

This. I don’t think LLMs are anywhere near sci-fi AGI (think I, robot) It’s such a vague term anyway, AGI.

LLMs provide some really nice text generation, summarization, and outstanding semantic search. It’s drop dead easy to make a natural language interface to anything now.

That’s a big deal. That’s what’s going to give this tech it’s longevity, imo.

he's not wrong, and yet he's not right.

Reply on news.ycombinator.com