AGI Is Promised for Tomorrow, but No One Has Explained What Is Supposed to Arrive
What exactly must appear before the promise of AGI stops being a vague forecast and becomes a testable event?
Build a testable definition of AGI: which capabilities must appear, in which tasks they must be repeatable, and where human control is still required.
What to watch for
Key takeaways
The practical meaning of “When the leaders of OpenAI, xAI, and Anthropic say powerful AI will arrive within the next” is that each of them means something different: for one, AGI is a model doing most intellectual work; for another, an economic system replacing people in tasks; for a third, a powerful tool with no claim to consciousness.
The boundary of the “Artificial AI is canceled, meet Powerful AI!” case is defined by this point: Anthropic's chief considers the term AGI overloaded and proposes speaking of “powerful AI” — a tool with no claim to human consciousness; along the way it turned out the powerful.ai domain is already listed at almost four hundred thousand dollars.
In the context of “OpenAI goes shopping,” this criterion applies: OpenAI bought the chat.com domain — the previous buyer paid fifteen million dollars for it a year earlier and has now converted the deal into equity; rumors put it at around twenty million, with the exact sum undisclosed.
The working conclusion from “AGI will arrive next year” is that Altman promises AGI as early as next year, while Musk in a separate interview talks about a system able to do everything a human can and names 2028–2029 — the spread of numbers itself shows there is no testable definition yet.
The “A 20% chance of AI going insane” topic becomes clearer once this point is included: that is not necessarily a contradiction: the same capability can make a system useful and dangerous. But the percentage creates an impression of precision where no generally accepted measurement method exists.
For the “What drives the AI leap, and is loss of control a real risk?” scene, the decisive point is this: scaling is hitting a plateau — making networks ten or a hundred times bigger no longer gives the old effect, so the bet shifts to reasoning time in models like o1; if gains are small, refining existing models beats building monstrous clusters.
The practical meaning of “What will come instead of ‘strong’ AI?” is that the hosts expect the first breakthroughs in code generation — that is what the best researchers want to work on, and an agent buying plane tickets through an interface will be just a byproduct; investors believe in computer use, but the speakers are not sure it will happen.
For the “What do YOU think about strong artificial intelligence?” scene, the decisive point is this: the hosts put the question to the audience — will AGI arrive and what will the first use cases be, write in the comments; the very framing shows the industry itself has no ready answer.
The “Is the arrival of AI comparable to the creation of the internet?” scene leads to a working conclusion: the internet created infrastructure, while AI enters existing professions and decisions immediately; instead of waiting for one magical date, watch which operations a model performs without supervision, what that costs, and who answers for an error.
What this episode is about
Sam Altman says AGI is close, Elon Musk simultaneously promises widespread robotics and estimates the risk of losing control, while Dario Amodei prefers the more cautious term ‘powerful AI.’ Behind the loud timelines is a real question: do scaling and reasoning produce a qualitatively new system, or only a more expensive version of the old one?
When the leaders of OpenAI, xAI, and Anthropic say powerful AI will arrive within the next few years, it sounds like an industry consensus. But each person means something different.
For one, AGI is a model capable of performing most intellectual work; for another, it is an economic system that replaces people in tasks; for a third, it is a sufficiently powerful tool with no claim to human consciousness.
Elon Musk promises both that robots will be able to do almost everything a person can do and that there is a twenty-percent risk of AI escaping control. That is not necessarily a contradiction: the same capability can make a system useful and dangerous. But the percentage creates an impression of precision where no generally accepted measurement method exists.
The central technical question is whether the previous rate of improvement continues. Reasoning models such as o1 spend more compute on thought, but scaling sharply increases energy consumption and cost. If the gain becomes small, the market will have more reason to refine existing models and build products around them than to increase cluster size forever.
Practical applications are already visible. The Devin agent automates part of programming, Suno helps musicians create drafts, and Runway changes video production. Hollywood is resisting not an abstract AGI, but the concrete ability to produce shots, music, and script alternatives more cheaply. Professionals often use AI as a tool and then perform or reconstruct the result themselves.
The comparison with the arrival of the internet is therefore only partly useful. The internet created infrastructure, while AI enters existing professions and decisions immediately. We are still far from ‘replacing everything,’ but the market is already changing. Instead of waiting for one magical date, it is more useful to watch which operations a model performs without supervision, how much that costs, and who is responsible for an error.
We are still far from ‘replacing everything,’ but the market is already changing. As a result, instead of waiting for one magical date, it is more useful to watch which operations a model performs without supervision, how much that costs, and who is responsible for an error.
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 100 segments: 64 identified, 6 mixed, 25 probable, and 5 unresolved.
Read transcript on a separate page
Loading…