- TheTechOasis
- Posts
- Is Everyone in AI Wrong?
Is Everyone in AI Wrong?
đ TheTechOasis đ
part of the:
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/a7006d1b-789f-4b2b-8dcc-a771f4ff81d3/Logotipo_negro.png?t=1712070264)
In AI, learning is winning.
This post is part of the Premium Subscription, in which we analyze technology trends, companies, products, and markets to answer AI's most pressing questions.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/3b64cc59-9e09-4ea6-a9c2-d7a4bc184425/image.png?t=1719302648)
You will also find this deep-dive in TheWhiteBox âKey Future Trends in AIâ, for future reference if necessary and to ask me any questions you consider.
As you may have realized, over the past few weeks, Iâve been sharing the unpleasant reality of AI, that side that is rarelyâif at allâshared with us through mainstream media but that you should be well aware of.
With a lot of money in the line and AI-led valuations at indecent multiples compared to tangible AI revenues, thereâs no interest in sharing anything but an âAI is changing the worldâ narrative that keeps the party going.
For that, today, Iâm bringing a manifold of âAI dogmas,â including the uber-mentioned: âLLMs will lead to AGIâ or âLLMs Are All You Need,â with both, in reality, being far from certain, as weâll see today.
I will also provide potential consequences if these fears materialize so that you can have an unbiased, no-hype, no-bullshit view of the space that reading mainstream media and influencers simply do not provide, giving you an edge over the majority.
âFrontier AI Will Lead Demandâ
With success stories like OpenAI, with revenue growth increasing steadily every year since the launch of ChatGPT to a whopping $3.4 billion current run rate, itâs easy to assume that better AI models will lead to the âgreat adoptionâ that many incumbents and investors patiently await to justify the insane valuations.
A CAPEX Problem
But as discussed in two previous Leaders segments (here and here), current Generative AI products arenât as good as some will tell you, which has translated into extremely poor demand.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/140b3e00-cdad-41c6-aa31-16d3d532d35d/image.png?t=1719420199)
Every tech enthusiastâs psychologist right now.
Nevertheless, if energy constraints are the mid-to-long-term elephant in the room for AI, product demand is unequivocally the main issue in the short term.
But how much demand do we need to avoid a catastrophe?
In a very recent Sequoia update on their â$200 billion question,â their first try at coming up with a number, that value is now closer to $600 billion.
This means that the AI market is heavily overinvested and that Generative AI revenues need to grow into those values to justify the CAPEX investments led by big tech companies and the valuations at which private and public investors have invested.
In other words, unless revenues grow to that number, companies arenât even breaking even, let alone earning money.
The worst part is that David Cahn from Sequoia is referring explicitly to non-hardware revenues; meaning NVIDIAâs revenues donât count.
Although we donât have an exact value, seeing as the hyperscalers donât reach $20 billion combined for 2024, that number is probably much closer to $30-50 billion for 2024 (a value I still feel is inflated).
Scarily, if hyperscalers are recognizing as revenue their auto-generated demand on their compute clouds via their investments on companies like OpenAI, Anthropic, et al, real GenAI revenues could be even smaller.
This means we need at least a twelve-to-twentyfold revenue increase to break even. Now, please tell me we arenât in a bubble.
To make matters worse, our savior doesnât seem to be coming anytime soon. But what do I mean by that?
Edge AI Is Our Only Chance
When we discussed whether AI was in a bubble, I argued that this faltering demand was mainly due to enterprise adoptionâor lack thereofâand mainly due to âAI inaccuraciesâ (incorrectly named âhallucinationsâ), which make the GenAI systems very unreliable for enterprise processes (normally requiring +95% accuracy).
For that, my bet was on the next generation (GPT-5 and others) to close this gap and skyrocket demand, and that we should expect those models by end of year.
Well, now OpenAI talks about late 2025 or early 2026 as the most probable timeline, meaning that we are left with these hallucination machines that canât be trusted for more than a year from now, unless companies like Lamini really eliminate hallucinations, which is uncertain.
Microsoft AI CEO, Mustafa Suleyman, dropped an even more concerning opinion, stating that we wonât see truly powerful AI models until âGPT-6â, in at least two years time.
But if we shouldnât expect huge short-term demand for GenAI products by enterprises, where will the demand we are so desperate for come from?
In my humble opinion, specialized âAI on the edgeâ products are our only hope to avoid a crash.
Iâm referring to things like Apple Intelligence or Microsoft-Phi-3-Silica, which run on your computer and can perform simpler yet unequivocally useful actions on your behalf.
These models are far from frontier AI but very decent models that can perform tasks that donât require extensive reasoning capabilities. In other words, instead of building AI models for complex cases where their accuracy is just not good enough, we focus on useful tasks with higher accuracy.
Therefore, in my view, the idea that frontier AI models will generate the much-needed demand is misplaced, as good-enough frontier models arenât coming anytime soon.
Consequences
If this is, in fact, the case, we should expect Apple and Microsoft stocks to soar as they lead the adoption of the technology.
Incumbents like ARM, AMD, and, in particular, Qualcomm, could also see a huge boost if edge devices, like smartphones or laptops, become the main tools for GenAI use.
Foundries like TSMC are unaffected by this potential shift, as 2-3nm chips by companies like Apple are still being built by them.
Their grasp on AI accelerators (sub-7nm nodes) is so large that, if it wasnât for the China threat, TSMC could be easily a top-3 world company in valuation right now.
On the other hand, a shift in the narrative away from frontier AI research could cause that companies at the frontier of AI could suffer.
For instance, NVIDIAâs role in âedge AIâ isnât nearly as prominent as in âfrontier AI.â Apple, Microsoft, and Google all design their chips for smartphones and laptops, so their reliance on NVIDIA for edge cases is very small.
NVIDIA does have a large market for GPU computers, accounting for 17% of its revenues.
Nonetheless, they are trying to move into the edge too, especially in the areas of digital twins and factory AI industrialization. With that said, their revenues from these segments pale in comparison (well below 5%).
But what is this paradigm shift Iâm referring to? The truth is that today, being at the frontier adds value to your company.
But if investors lose interest in âpushing the veil of ignorance forwardâ as Sam Altman would say, and instead focus on companies that deliver revenues, big NVIDIA customers like Microsoft (estimated to be 22% of total revenues in the last quarter) or Meta could downsize their GPU CAPEX investments as investors side-eye their intentions of âbuilding AGI.â
Although this scenario shouldnât create a market crash (as money wouldnât be leaving the markets but flowing into different companies), it could certainly lead to company crashes (and great winners).
But let me be clear: If edge AI also fails to generate demand (one interesting way to measure this will be whether iPhone users upgrade to iPhone 15 Pro or the upcoming 16 for Apple Intelligence), we are running out of bullets to save the AI market.
On to the next dogma.
âGPUs Are All You Need"
When I saw Jensen Huang, NVIDIAâs CEO, signing a bra, I knew the world was losing its senses over the GPU company.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/d97c0fa0-03f2-489d-81b7-5254a0ce48b3/image.png?t=1719322797)
Nonetheless, everyone seems to be pushing the narrative that âGPUs are all you need.â
This, coupled with another narrative we will address further in the article, âbigger is better,â creates the perfect storm for everyone, both big tech corporations and investors, to go crazy for NVIDIA.
But is this accurate?
Unequivocally, nothing beats NVIDIA's GPUs at scaled LLM training, or Simulation-based Reinforcement Learning (the main approach for physical AI robot training today), probably the two leading areas of research in terms of interest and investment.
However, this dogma quickly loses its force if we consider inference workloads and, above all, edge AI.
For the former, alternatives like Language Processing Units (LPUs) exist, which perform considerably better at scale. Also, recent efficiency techniques reduce AIs' dependency on GPUs by avoiding matrix multiplications, the operator this hardware excels at.
Another intriguing recent entry is Etched.ai, which has raised $120 million from very notable investors to create an ASIC chip that is purposely designed for Transformers (ChatGPT et al.) that, according to the company, is an order of magnitude fasterâand cheaperâthan NVIDIAâs upcoming GPUs, the Blackwell GB200, for inference workloads.
As we mentioned earlier, NVIDIAâs grip on edge AI is very small. The main hardware is neural processing units (NPUs), created to avoid AI workloads eating the deviceâs battery.
But why? By using hyperspecialized components such as activation function modules, they are much less 'energy-hungry,â making them better edge device options.
Consequently, while GPUs will continue to play a vital role in LLM training pipelines and simulation training, the percentage of energy allocated to these workloads could decrease dramatically in lieu of inference and edge AI workloads.
Consequences
This dogma is heavily dependent on the previous one in knowing the answer to the big question: What narrative will investors reward in the future?
Right now, itâs clear that what markets value is who is leading the race toward better AI models.
âThe company with the best performance and larger models, or the company eating all the revenues are the companies Iâm betting in.â is the average Wall Street investor mindset right now.
But the markets arenât acknowledging that enterprises are simply not buying the hype.
Constant delays in a much-needed improvement in capabilities and accuracy for years to come also mean that, unless demands come from elsewhere, markets will have to sustain inflated valuations for almost two years, which I think will not happen.
Therefore, if âThe company with tangible value creation and proven AI revenues is the company I want to invest inâ mentality kicks in, this idea of âGPUs are all you needâ will quickly dry up.
In that scenario, as investors no longer reward investing your free cash flow in GPUs, demand for this hardware could falter.
âLLMs Will Lead to AGI.â
With the latest Anthropic release of Claude 3.5 Sonnet, tech enthusiasts have begun, once again, rocking the anthems that âAGI is nearâ or that âcurrent Large Language Models (LLMs) have sparks of AGI."
However, the truth is that most AI researchers are highly skeptical of LLMs ever reaching AGI status.
But why?
Current LLMs are taught through imitation learning; they learn to imitate us. But while LLMs have indeed learned to imitate human writing, they are far from being able to imitate human reasoning.
Importantly, as the writers of the previous link outlined, they are completely oblivious to the nature of the things they talk about; they donât care about truth or lies or, to quote the researchers, âThey are aimed at being convincing rather than accurate.â
Their inability to perceive the world, a world they see through the proxy of human writing, prevents them from truly understanding what they are talking about.
According to a group of researchers at Brown University studying this question, as the meaning of a concept also requires a sensory stimulus LLMs donât experience, we donât know.
One reason this may be happening is due to the lack of active inference.
Heavily championed by prominent researchers like Karl Friston or François Chollet, active inference models must be in a constant learning journey.
But thatâs not the case today.
Simply put, LLMs have two working modes: training and inference. In other words, current frontier models only learn during training but cannot acquire new skills during inference.
Consequently, as LLMs canât adapt to changes and find new solutions to a given previously unseen problem, some researchers like François argue that every single âact of intelligenceâ we see in LLMs today can be reduced to some sort of pattern-matching or memorization, meaning that the model isnât applying novel reasoning processes but simply fetching the solution from its previous knowledge.
In laymanâs terms, what François is trying to tell us, especially if we pay close attention to his conception of what intelligence is from his seminal paper âOn the measure of intelligence,â is that he thinks that LLMs are not intelligent.
And the proof this remarkably controversial quote is actually true is strong.
![](https://media.beehiiv.com/cdn-cgi/image/format=auto,onerror=redirect/uploads/publication/logo/9b355c4a-4dbc-4f53-b676-a166fee6812c/thumb_Isotipo_negro.png)
Subscribe to Leaders to read the rest.
Become a paying subscriber of Leaders to get access to this post and other subscriber-only content.
Already a paying subscriber? Sign In