HN Reader

The best engineering minds have been focused on scaling transformer pre and post training for the last three years because they had good reason to believe it would work, and it has up until now.

Progress has been measured against benchmarks which are / were largely solvable with scale.

There is another emerging paradigm which is still small(er) scale but showing remarkable results. That's full multi-modal training with embodied agents (aka robots). 1x, Figure, Physical Intelligence, Tesla are all making rapid progress on functionality which is definitely beyond frontier LLMs because it is distinctly different.

OpenAI/Google/Anthropic are not ignorant of this trend and are also reviving or investing in robots or robot-like research.

So while Orion and Claude 3.5 opus may not be another shocking giant leap forward, that does not mean that there arn't giant shocking leaps forward coming from slightly different directions.

I lead a team exploring cutting edge LLM applications and end-user features. It's my intuition from experience that we have a LONG way to go.

GPT-4o / Claude 3.5 are the go-to models for my team. Every combination of technical investment + LLMs yields a new list of potential applications.

For example, combining a human-moderated knowledge graph with an LLM with RAG allows you to build "expert bots" that understand your business context / your codebase / your specific processes and act almost human-like similar to a coworker in your team.

If you now give it some predictive / simulation capability - eg: simulate the execution of a task or project like creating a github PR code change, and test against an expert bot above for code review, you can have LLMs create reasonable code changes, with automatic review / iteration etc.

Similarly there are many more capabilities that you can ladder on and expose into LLMs to give you increasingly productive outputs from them.

Chasing after model improvements and "GPT-5 will be PHD-level" is moot imo. When did you hire a PHD coworker and they were productive on day-0 ? You need to onboard them with human expertise, and then give them execution space / long-term memories etc to be productive.

Model vendors might struggle to build something more intelligent. But my point is that we already have so much intelligence and we don't know what to do with that. There is a LOT you can do with high-schooler level intelligence at super-human scale.

Take a naive example. 200k context windows are now available. Most people, through ChatGPT, type out maybe 1500 tokens. That's a huge amount of untapped capacity. No human is going to type out 200k of context. Hence why we need RAG, and additional forms of input (eg: simulation outcomes) to fully leverage that.

Right. If you generate some code with ChatGPT, and then try to find similar code on the web, you usually will. Search for unusual phrases in comments and for variable names. Often, something from Stack Overflow will match.

LLMs do search and copy/paste with idiom translation and some transliteration. That's good enough for a lot of common problems. Especially in the HTML/Javascript space, where people solve the same problems over and over. Or problems covered in textbooks and classes.

But it does not look like artificial general intelligence emerges from LLMs alone.

There's also the elephant in the room - the hallucination/lack of confidence metric problem. The curse of LLMs is that they return answers which are confident but wrong. "I don't know" is rarely seen. Until that's fixed, you can't trust LLMs to actually do much on their own. LLMs with a confidence metric would be much more useful than what we have now.

This is supported by both general observations and recently this tweet from an OpenAI engineer that Sam responded to and engaged ->

"scaling has hit a wall and that wall is 100% eval saturation"

Which I interpert to mean his view is that models are no longer yielding significant performance improvements because the models have maxed out existing evaluation metrics.

Are those evaluations (or even LLMs) the RIGHT measures to achieve AGI? Probably not.

But have they been useful tools to demonstrate that the confluence of compute, engineering, and tactical models are leading towards signifigant breathroughts in artificial (computer) intelligence?

I would say yes.

Which in turn are driving the funding, power innovation, public policy etc needed to take that next step?

I hope so.

(1) https://x.com/willdepue/status/1856766850027458648

Meanwhile, the existing tech is such a step change that industry is going to need time to figure out how to effectively use these models. In a lot of ways it feels like the "digitization" era all over again - workflows and organizations that were built around the idea humans handled all the cognitive load (basically all companies older than a year or two) will need time to adjust to a hybrid AI + human model.

So the models' accuracies won't grow exponentially, but can still grow linearly with the size of the training data.

Sounds like DataAnnotation will be sending out a lot more LinkedIn messages.

[1] https://x.com/sama/status/1856941766915641580

I'm surprised that any of these companies consider what they are working on to be Artificial General Intelligences. I'm probably wrong, but my impression was AGI meant the AI is self aware like a human. An LLM hardly seems like something that will lead to self-awareness.

For example recently I asked it to generate some phrases for a list of words, along with synonym and antonym lists.

The phrases were generally correct and appropriate (some mistakes but that’s fine). The synonyms/antonyms were misaligned to the list (so strictly speaking all wrong) and were often incorrect anyway. I imagine it would be the same if you asked for definitions of a list of words.

If you ask it to correct it just generates something else which is often also wrong. It’s certainly superficially convincing in many domains but once you try to get it to do real work it’s wrong in subtle ways.

Where do these large "AI" companies think the mass amounts of data used to train these models come from? People! The most powerful and compact complex systems in existence, IMO.

This methodological growth could make LLMs more reliable, consistent, and aligned with specific use cases.

The skepticism surrounding this vision mirrors early doubts about the early internet fairly concisely.

Initially, the internet was seen as fragmented collection of isolated systems without a clear structure or purpose. It really was. You would gopher somewhere and get a file, and eventually we had apps like like pine for email, but as cool as it was it has limited utility.

People doubted it could ever become the seamless, interconnected web we know today.

Yet, through protocols, shared standards, and robust frameworks, the internet evolved into a powerful network capable of handling diverse applications, data flows, and user needs.

In the same way, LLM orchestration will mature by standardizing interfaces, improving interoperability, and fostering cooperation among varied AI models and support systems.

Just as the internet needed HTTP, TCP/IP, and other protocols to unify disparate networks, orchestrated AI systems will require foundational frameworks and “rules of the road” that bring cohesion to diverse technologies.

We are at the veeeeery infancy of this era and have a LONG way to go here. Some of the progress looks clear and a linear progression, but a lot, like the Internet, will just take a while to mature and we shouldn’t forget what we learned the last time we faced a sea change technological revolution.

1. Find more data.

2. Make the weights capture the data and reproduce.

In that sense we have reached a limit. So in my opinion we can do a couple of things.

1. App developers can understand the limits and build within the limits.

2. Researchers can take insights from these large models and build better AI systems with new architectures. It's ok to say transformers have reached a limit.

On the other hand, a lot of these frameworks and languages have relatively decent and detailed documentation.

Perhaps this is a naive question, but why can't I as a user just purchase "AI software" that comes with a large pre-trained model to which I can say, on my own machine, "go read this documentation and help me write this app in this next version of Leptos", and it would augment its existing model with this new "knowledge".

AlphaGo - self-play

AlphaFold - PDB, the protein database

ChatGPT - human knowledge encoded as text

These models are all machines for clever interpolation in gigantic training datasets.

They appear to be intelligent, because the training data they've seen is so vastly larger than what we've seen individually, and we have poor intuition for this.

I'm not throwing shade, I'm a daily user of ChatGPT and find tremendous and diverse value in it.

I'm just saying, this particular path in AI is going to make step-wise improvements whenever new large sources of training data become available.

I suspect the path to general intelligence is not that, but we'll see.

Learning from data is not enough; there is a need for the kind of system-two thinking we humans develop as we grow. It is difficult to see how deep learning and backpropagation alone will help us model that. For tasks where providing enough data is sufficient to cover 95% of cases, deep learning will continue to be useful in the form of 'data-driven knowledge automation.' For other cases, the road will be much more challenging. https://www.lycee.ai/blog/why-sam-altman-is-wrong

So the problem is more in the algorithm.

With my user hat on, I'm quite pleased with the current state of LLMs. Initially, I approached them skeptically, using a hackish mindset and posing all kinds of Turing test-like questions. Over time, though, I shifted my focus to how they can enhance my team's productivity and support my own tasks in meaningful ways.

Finally, I see LLMs as a valuable way to explore parts of the world, accommodating the reality that we simply don’t have enough time to read every book or delve into every topic that interests us.

Certain OpenAI insiders must have known this for a while, hence Ilya Sutskever's new company in Israel

So it's interesting that when AI came along, we threw caution to the wind and started treating it like a silver bullet... Without asking the question of whether it was applicable to this goal or that goal...

I don't think anyone could have anticipated that we could have an AI which could produce perfect sentences, faster than a human, better than a human but which could not reason. It appears to reason very well, better than most people, yet it doesn't actually reason. You only notice this once you ask it to accomplish a task. After a while, you can feel how it lacks willpower. It puts into perspective the importance of willpower when it comes to getting things done.

In any case, LLMs bring us closer to understanding some big philosophical questions surrounding intelligence and consciousness.

Is this certain? Are Agents the right direction to AGI?

I wonder what this would mean for companies raising today on the premise of building on top of these platforms. Maybe the best ones get their ideas copied, reimplemented, and sold for cheaper?

We already kind of see this today with OpenAI's canvas and Claude artifacts. Perhaps they'll even start moving into Palantir's space and start having direct customer implementation teams.

It is becoming increasing obvious that LLM's are quickly becoming commoditized. Everyone is starting to approach the same limits in intelligence, and are finding it hard to carve out margin from competitors.

Most recently exhibited by the backlash at claude raising prices because their product is better. In any normal market, this would be totally expected, but people seemed shocked that anyone would charge more than the raw cost it would take to run the LLM itself.

https://x.com/ArtificialAnlys/status/1853598554570555614

The irony here is astounding.

This smells like it’s mostly based on OAI having a bit of bad luck with next model rather than a fundamental slowdown / barrier.

They literally just made a decent sized leap with o1

[1] https://en.wikipedia.org/wiki/Coherence_theory_of_truth

It will be like StableDiffusion 1.5. This model can now run on low end devices, lots of open research use this model to build something else and inspire by this.

These LLMs can be used as a foundation to keep improving and building new things.

Up to a certain point, a conditional fluency stores knowledge, in the sense that semantically correct sentences are more likely to be fluent… but we may have tapped out in that regard. LLMs have solved language very well, but to get beyond that has seemed, thus far, to require RLHF, with all the attendant negatives.

At the very early phase of the boom I was among a very few who knew and predicted this (usually most free and deep thinking/knowledgeable). Then my prediction got reinforced by the results. One of the best examples was with one of my experiments that all today's AI's failed to solve tree serialization and de-serialization in each of the DFS(pre-order/in-order/post-order) or BFS(level-order) which is 8 algorithms (2x4) and the result was only 3 correct! Reason is "limited training inputs" since internet and open source does not have other solutions :-) .

So, I spent "some" time and implemented all 8, which took me few days. By the way this proves/demonstrates that ~15-30min pointless leetcode-like interviews are requiring to regurgitate/memorize/not-think. So, as a logical hard consequence there will.has-to be a "crash/cleanup" in the area of leetcode-like interviews as they will just be suddenly proclaimed as "pointless/stupid"). However, I decided not to publish the rest of the 5 solutions :-)

This (and other experiments) confirms hard limits of the LLM approach (even when used with chain-of-thought). Increasing the compute on the problem will produce increasingly smaller and smaller results (inverse exponential/logarithmic/diminishing-returns) = new AGI approach/design is needed and to my knowledge majority of the inve$tment (~99%) is in LLM, so "buckle up" at-some-point/soon?

Impacts and realities; LLM shall "run it's course" (produce some products/results/$$$, get reviewed/$corrected) and whoever survives after that pruning shall earn money on those products while investing in the new research to find new AGI design/approach (which could take quite a long time,... or not). NVDA is at the center of thi$ and time-wise this peak/turn/crash/correction is hard to predict (although I see it on the horizon and min/max time can be estimated). Be aware and alert. I'll stop here and hold my other number of thoughts/opinions/ideas for much deeper discussion. (BTW I am still "full in on NVDA" until,....)

And I think the latter is good enough for us to do exciting things.

And there's a number of reasons why, mostly likely being that they've found other ways to get improvements out of AI models, so diminishing returns on training aren't that much of a problem. Or, maybe the leakers are lying, but I highly doubt that considering the past record of news outlets reporting on accurate leaked information.

Still though, it's interesting how basically ever frontier lab created a model that didn't live up to expectations, and every employee at these labs on Twitter has continued to vague-post and hype as if nothing ever happened.

It's honestly hard to tell whether or not they really know something we don't, or if they have an irrational exuberance for AGI bordering on cult-like, and they will never be able to mentally process, let alone admit, that something might be wrong.

Is it just me or does $100 million sound like it's on the very, very low end of how much training a new model costs? Maybe you can arrive within $200 million of that mark with amortization of hardware? It just doesn't make sense to me that a new model would "only" be $100 million when AmaGooBookSoft are spending tens of billions on hardware and the AI startups are raising billions every year or two.

I was really looking forward to using "synthetic data" euphemistically during debates.

I don't get it...

> I suspect the path to general intelligence is not that, but we'll see.

I think there's three things that a 'true' general intelligence has which is missing from basic-type-LLMs as we have now.

1. knowing what you know. <basic-LLMs are here>

2. knowing what you don't know but can figure out via tools/exploration. <this is tool use/function calling>

3. knowing what can't be known. <this is knowing that halting problem exists and being able to recognize it in novel situations>

(1) From an LLM's perspective, once trained on corpus of text, it knows 'everything'. It knows about the concept of not knowing something (from having see text about it), (in so far as an LLM knows anything), but it doesn't actually have a growable map of knowledge that it knows has uncharted edges.

This is where (2) comes in, and this is what tool use/function calling tries to solve atm, but the way function calling works atm, doesn't give the LLM knowledge the right way. I know that I don't know what 3,943,034 / 234,893 is. But I know I have a 'function call' of knowing the algorithm for doing long divison on paper. And I think there's another subtle point here: my knowledge in (1) includes the training data generated from running the intermediate steps of the long-division algorithm. This is the knowledge that later generalizes to being able to use a calculator (and this is also why we don't just give kids calculators in elementary school). But this is also why a kid that knows how to do long division on paper, doesn't seperately need to learn when/how to use a calculator, besides the very basics. Using a calculator to do that math feels like 1 step, but actually it does still have all of initial mechanical steps of setting up the problem on paper. You have to type in each digit individually, etc.

(3) I'm less sure of this point now that I've written out point (1) and (2), but that's kinda exactly the thing I'm trying to get at. Its being able to recognize when you need more practice of (1) or more 'energy/capital' for doing (2).

Consider a burger resturant. If you properly populated the context of a ChatGPT-scale model the data for a burger resturant from 1950, and gave it the kinda 'function calling' we're plugging into LLMs now, it could manage it. It could keep track of inventory, it could keep tabs on the employee-subprocesses, knowing when to hire, fire, get new suppliers, all via function calling. But it would never try to become McDonalds, because it would have no model of the the internals of those function-calls, and it would have no ability to investigate or modify the behaviour of those function calls.

That doesn't mean this article is irrelevant. It's good to know if LLM improvements are going to slow down a bit because the low hanging fruit has seemingly been picked.

But in terms of the overall effect of AI and questioning the validity of the technology as a whole, it's just your basic FUD article that you'd expect from mainstream news.

[1] https://dl.acm.org/doi/10.1145/3442188.3445922

To be clear, I don't think a near-term bubble collapse is likely but I'm going from 3% to maybe ~10%. Also, this doesn't mean I doubt there's real long-term value to be delivered or money to be made in AI solutions. I'm thinking specifically about those who've been speculatively funding the massive build out of data centers, energy and GPU supply expecting near-term demand to continue scaling at the recent unprecedented rates. My understanding is much of this is being funded in advance of actual end-user demand at these elevated levels and it is being funded either by VC money or debt by parties who could struggle to come up with the cash to pay for what they've ordered if either user demand or their equity value doesn't continue scaling as expected.

Admittedly this scenario assumes that these investment commitments are sufficiently speculative and over-committed to create bubble dynamics and tipping points. The hypothesis goes like this: the money sources who've over-committed to lock up scarce future supply in the expectation it will earn outsize returns have already started seeing these warning signs of efficiency and/or progress rates slowing which are now hitting mainstream media. Thus it's possible there is already a quiet collapse beginning wherein the largest AI data center GPU purchasers might start trying to postpone future delivery schedules and may soon start trying to downsize or even cancel existing commitments or try to offload some of their future capacity via sub-leasing it out before it even arrives, etc. Being a dynamic market, this could trigger a rapidly snowballing avalanche of falling prices for next-year AI compute (which is already bought and sold as a commodity like pork belly futures).

Notably, there are now rumors claiming some of the largest players don't currently have the cash to pay for what they've already committed to for future delivery. They were making calculated bets they'd be able to raise or borrow that capital before payments were due. Except if expectation begins to turn downward, fresh investors will be scarce and banks will reprice a GPU's value as loan collateral down to pennies on the dollar (shades of the 2009 financial crisis where the collateral value of residential real estate assets was marked down). As in most bubbles, cheap credit is the fuel driving growth and that credit can get more expensive very quickly - which can in turn trigger exponential contagion effects causing the bubble to pop. A very different kind of "Foom" than many AI financial speculators were betting on! :-)

So... in theory, under this scenario sometime next year NVidia/TSMC and other top-of-supply-chain companies could find themselves with excess inventories of advanced node wafers because a significant portion of their orders were from parties who no longer have access to the cheap capital to pay for them. And trying to sue so many customers for breach can take a long time and, in a large enough sector collapse, be only marginally successful in recouping much actual cash.

I'd be interested in hearing counter-arguments (or support) for the impossibility (or likelihood) of such a scenario.

xd

IMO this will require not just much more expansive multi-modal training, but also novel architecture, specifically, recurrent approaches; plus a well-known set of capabilities most systems don't currently have, e.g. the integration of short-term memory (context window if you like) into long-term "memory", either episodic or otherwise.

But these are as we say mere matters of engineering.