AI Has Not Hit a Ceiling—Products and Corporations Have
If models keep improving, why do products and corporations look the main constraint on AI?
See why better models do not automatically create better products: the bottleneck has moved to interfaces, business models, and the company’s ability to explain the value.
What to watch for
Key takeaways
The “ToTheMoon is a podcast about the IT world and modern technology” topic becomes clearer once this point is included: ChatGPT and Claude going down at the same time brings the ‘AI apocalypse’ talk back to earth: even the strongest models depend on infrastructure, the market is concentrated in a handful of companies, and massive investment does not guarantee users a stable product.
The “Elon Musk's investment in AI vs everyone else's. On Claude, xAI, and Grok” scene leads to a working conclusion: xAI has access to human behavior inside X — what people read, discuss, and share — while Meta holds WhatsApp, Instagram, and Facebook: that distribution lets them test AI on real scenarios at a scale no standalone chatbot can match, and the model race increasingly becomes a platform race.
For the “The best LLM” scene, the decisive point is this: in the poll Musk ran on X, his own xAI won as the ‘best model’ — and the top comment immediately noted the audience's bias; the real takeaway is not about quality but distribution: a model built into X instantly reaches millions of loyal users.
In the context of “AI hallucinations: the ceiling of text generation in all models,” this criterion applies: in text-generation quality the models have converged — a lull: Google is stronger in video understanding, GPT-4o surprised with speed and voice, but there is no clear leader, and this is either a ceiling everyone has hit or a temporary step.
The “Yann LeCun's dispute with Elon Musk over the parallel with the dot-com crash” issue should be assessed with one constraint in mind: LeCun argues the generative-AI hype will collapse like the dot-coms and the term itself is a fashionable name for linear algebra; the hosts' counterpoint: no AI company is worth trillions, the giants hold more cash than all the raised rounds combined, and the tool is already genuinely useful.
The discussion of “How startups invent new things while corporations stall” yields a practical test: Google can build strong models, but integrating AI into search threatens its own business — the ‘glue on pizza’ advice is funny until you remember the billions of users; that caution opens the window for Perplexity: a startup wins by redesigning the user experience faster, not by having a fundamentally better model.
The decision in “AI CRM developers lost $50 billion: is AI investment at risk?” depends on one criterion: the company has enormous access to sales and marketing data, but the market is waiting not for the word AI, but for product growth and customer results. Today's ceiling is not only inside the language model. It is also in interfaces, business models, and the inability of companies to explain why a person needs one more feature.
What this episode is about
ChatGPT and Claude go down, Grok and Perplexity try new formats, Google is afraid to break search, and Salesforce loses tens of billions in market value. The problem is not limited to model hallucinations: large companies often do not know how to turn the technology into clear value for users.
When ChatGPT and Claude go down at the same time, talk of an ‘AI apocalypse’ quickly returns to earth. Even the strongest models depend on infrastructure, while the market is still concentrated around a handful of companies—OpenAI, Google, Anthropic, Meta, and xAI. Massive investment gives them an advantage, but it does not guarantee that users receive a stable and understandable product.
Grok is interesting for more than the quality of its answers. xAI has access to human behavior inside X: what people read, discuss, and share. Meta has WhatsApp, Instagram, and Facebook. Distribution like this makes it possible to test AI in real scenarios at a scale unavailable to a standalone chatbot. The model race is therefore becoming a platform race.
Google shows the other side. The company knows how to build powerful models, but integrating AI into search threatens its own business and reputation. A mistake such as recommending glue on pizza looks funny until you remember that a search engine answers billions of people and cannot experiment as freely as a startup.
A large corporation does not always move slowly because it is stupid. Sometimes it is protecting a machine that already makes money.
That caution opens a window for Perplexity and other new products. They can combine search, presentation, and saving the result into one clear workflow while Google hesitates. The startup wins not because its model is fundamentally better, but because it can redesign the user experience faster.
Salesforce's loss of market value shows how dangerous it is to promise AI without measurable value. The company has enormous access to sales and marketing data, but the market is waiting not for the word AI, but for product growth and customer results.
Today's ceiling is not only inside the language model. It is also in interfaces, business models, and the inability of companies to explain why a person needs one more feature.
AI’s ceiling today is not only inside the model. It also sits in interfaces, business models, and companies’ inability to explain a concrete benefit to the user.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 51 segments: 38 identified, 1 mixed, 11 probable, and 1 unresolved.
Loading…