The question of whether AI is a speculative bubble or a technological revolution can be answered simply: Yes.
Artificial intelligence as it is understood today – language models (LLMs), conversational chatbots, and AI agents – is not going anywhere, but the risk of a new AI winter remains.1 The history of artificial intelligence has been marked by two major winters – the first between 1974 and 1980, the second between 1987 and 1993 – periods of stagnation characterized by a drastic decline in funding and public interest, caused by exaggerated expectations and unfulfilled promises.
Promises made by Altman or Amodei that Artificial General Intelligence (AGI) is just around the corner, combined with countless recent articles full of fatalistic scenarios, could backfire against the current enthusiasm, especially the financial one, and lead to drastic cuts in research budgets if major developments do not occur.
How Fast Is Artificial General Intelligence Coming?
More than three years have passed since the emergence of ChatGPT, enough time for a generation of high school and university students to start and finish their studies with access to AI. Although visible changes have occurred in the labour market – call centres, SEO article writing, translators, dubbing actors, illustrators, and programmers were affected – and traffic of sites like Stack Overflow,2 „Stack overflow is almost dead”, pragmaticengineer.com recipe and tech advice sites, or homework help platforms,3 Although Brainly reoriented to AI tutoring has been affected, artificial intelligence has not changed the world at the speed some predicted.
However, many believe things will accelerate and office work will be rapidly replaced by AI, which does not complain, does not go on vacation, and does not ask for food. Two such texts have circulated recently: Something Big is Coming by Matt Shumer,4 Something Big Is Happening, shumer.dev a love letter to recent models from OpenAI and Anthropic, and a recent “research” piece from Citrini Research5 The 2028 Global Intelligence Crisis, citriniresearch.com which reads more like a science fiction story taken seriously because its title is written in all caps. Not to mention all the AI agent priests installed on MacBooks and left to run riot, often with hilarious consequences.6 I’ve heard of plenty of people waking up to find precious files deleted; I’ve yet to hear of anyone being enriched by some stock-market-obsessed agent.
For everybody else, change is happening slowly than in the wet dreams of AI company bosses. Nevertheless, a change is occurring, and artificial intelligence is way more real than the Metaverse. The question is if and when investors will get tired of waiting for a return on investment from the vast amounts of money they have poured into development and data centres that are not yet generating profit.7 OpenAI could lose around 44 billion by 2028. In 2025 alone, 61 billion was invested in data centres, cnbc.com
Change is happening more slowly than in the wet dreams of AI company bosses.
Especially since the world often notices how money moves between the same companies, creating a rather bizarre circular economy: Microsoft and Nvidia invested in OpenAI, which used the money to buy cloud processing from Microsoft and Oracle, which equipped them with chips from Nvidia. This financial structure, while ensuring short-term capital flow, suggests the boom is dependent on strategic alliances, not profitability. This amplifies the risk of a collapse if AGI is delayed.
Currently, those who provided the funds are not yet asking for return on investment because everyone is still under the impression that things are moving extremely fast and superintelligence is imminent.8 Superintelligence or Artificial General Intelligence (AGI) is defined as a form of AI as smart as or smarter than a human, capable of learning, adapting, and addressing or even solving new problems on its own current models are far from this threshold. Pressure is also maintained by fueling the fear that someone else (hint, China) might develop this technology first, equating AI development with nuclear weapons development,9Precum în eseul lui Dario Amodei, The Adolescence of Technology, darioamodei.com – whoever gets AGI first wins the war –, which also prepares the ground for potential bailouts in the event of a market crash.
Not to mention the frequent accusations that Chinese models are “stealing”10 We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax. These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models, x.com the capabilities of American models are ironic, however true they may be, given that virtually all AI models were trained through massive copyright infringement.
How Smart Is Today’s AI?
AI has already won a Nobel Prize by solving a difficult problem for humans, so AI is smart already, isn’t it? I am referring to Google DeepMind’s AlphaFold 2, whose creators, Demis Hassabis and John Jumper, received the 2024 Nobel Prize in Chemistry. The AI model solved the protein folding problem11Protein folding is the physical process by which a linear chain of amino acids arranges itself into a unique three-dimensional structure that allows the protein to correctly perform its biological function in the body. Predicting protein folding is extremely difficult because it requires the simultaneous calculation of millions of atomic interactions to find the single correct shape from an astronomical number of possible 3D combinations. – a painstaking task that previously took months for each individual protein, and there are hundreds of millions in nature.
And yet, AlphaFold 2 and its descendants12 We’re now at AlphaFold 4. can do nothing else but fold proteins or design drugs, like IsoDDE.13 A new model developed by Isomorphic Labs, a DeepMind spinoff, isomorphiclabs.com You could call it “silly”, since it cannot analyse personal relationships like ChatGPT. Jokes aside, AlphaFold 2 is proof of the stunning success of Narrow AI, which solves a specific problem. In contrast, the promises of AGI, which should be capable of multiple tasks, still seem unrealistic.
The explanation for why today’s generalist language models – ChatGPT, Claude, or Gemini – seem more capable is that they reproduce human language or programming languages so well and, above all, because they appear to have access to all (digitised) human knowledge and an answer (not necessarily the right one) for everything. However, there is little evidence so far that they can come up with new solutions to existing problems. Recently, mathematician Terence Tao14 In a Mathstodon post, mathstodon.xyz/@tao noted that the Erdos #728 problem was solved by the GPT 3.5 model, but with the clarification that it was actually a collective effort where human contribution was significant.
As portions of my text above have been quoted out of context, I would like to also draw attention to the various caveats listed at github.com regarding the extent to which one can draw broader conclusions about AI mathematics capabilities from the progress in solving Erdos problems.
Terence Tao
Scaling versus Revolution
For Amodei or Shumer, evolution is just a matter of time and processing capacity. More data, more processors, and AGI will come – especially since AI has now started writing code for AI, so the speed at which code is written is already spectacular.
However, if you look closer at what others in the AI industry are saying, you will see many more who argue that current models are not the path to AGI. More is needed than just scaling them: a revolution similar to the “transformer” that made today’s models possible. This is stated clearly in the MIT report on AGI from August 202515 The road to artificial general intelligence, technologyreview.com or by Yann LeCun from Meta,16 În multe interviuri, dar și în postări pe x.com perhaps the most vocal on this matter.
One thing we know is that if future AI systems are built on the same blueprint as current Auto-Regressive LLMs, they may become highly knowledgeable but they will still be dumb. […] Future systems will have to use a different architecture capable of understanding the world, capable of reasoning, and capable of planning so as to satisfy a set of objectives and guardrails.
Yann LeCun
Andrej Karpathy17Într-un podcast din toamna lui 2025, dwarkesh.com says one of the major problems to be solved is the models’ ability to learn continuously, just like a human, and a solution for this hasn’t been found yet, giving a horizon of at least a decade for AGI.
Sam Altman of OpenAI accepted in 2023 that scaling is not enough,18 “We need another breakthrough. We can still push on large language models quite a lot, and we will do that. We can take the hill that we’re on and keep climbing it, and the peak of that is still pretty far away. But, within reason, I don’t think that doing that will (get us to) AGI.”, reddit.com Yet, by early 2025, he believed he had found the solution,19 “We are now confident we know how to build AGI as we have traditionally understood it.”, blog.samaltman.com and AGI remains essentially a scaling problem.20 “The intelligence of an AI model roughly equals the log of the resources used to train and run it. These resources are chiefly training compute, data, and inference compute. It appears that you can spend arbitrary amounts of money and get continuous and predictable gains; the scaling laws that predict this are accurate over many orders of magnitude.”, blog.samaltman.com
Altman was likely referring mainly to the transition from simple language models (LLM) to those that use reasoning (LRM) – the “show thinking” you see during a query. However, while useful, it seems this is not the promised revolution, as shown by a recent Apple study suggestively titled The Illusion of Thinking.21 The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity, machinelearning.apple.com Researchers from the company discovered that both approaches used in current models collapse as the problem becomes more complex.
The Lesson of Nuclear Fusion
You might say it’s natural for Apple to think the grapes are sour and that those working daily on the AI-to-AGI transition know best, but the argument that the heads of these companies will always claim a solution exists – as they need more money for data centres and electricity – is equally valid. Most specialists agree 22 Here’s a systematic review, The rise of hallucination in large language models: systematic reviews, performance analysis and challenges, springer.com that the hallucinations of today’s AIs are intrinsic to the models and that the solution isn’t just more context.
If it cannot be done through scaling alone, as Altman and Amodei promise, the solution might not be just around the corner. Physicists have known for a century that we could have unlimited, clean, and safe energy if we mastered nuclear fusion, but that doesn’t make the technology imminent.
If AGI requires a discovery as significant as the transformer or “attention,” there is no guarantee it will come in a month, a year, or a decade. Maybe it’s already there, in one of too many scientific articles published or hiding in laboratories that protect intellectual property, or maybe its inventor hasn’t been born yet.
If AGI requires a discovery as significant as the transformer or “attention,” there is no guarantee it will come in a month, a year, or a decade.
And the longer it takes, the higher the risk that investors will demand their money back – which could produce, if not a new AI winter, at least a “late AI autumn.”
I say this only because OpenAI and Anthropic are competing with a giant like Google, which not only invented the technology behind LLMs but can also afford to use its own resources without relying on impatient investors, potentially extending research even if other market players lose financial support. That is, unless their downfall triggers the next great economic crisis.