Can larger AI models keep getting better, or will they run into tasks they cannot handle? At TechCrunch Disrupt, Anthropic CEO and co-founder Dario Amodei argued that there is no obvious limit on the horizon. His view rests on a pattern he has seen over the last 10 years: as neural nets have grown in scale, they have continued to work better.
Scaling has shaped Amodei’s expectations
Amodei connected his outlook to the history of training larger neural nets. The continued improvement he described is the basis for his expectation that what people see in AI today will pale in comparison to what arrives in the next 2, 3, 4 years.
That expectation is about continued progress, not a promise that growth can happen without constraint. When asked about a quadrillion-parameter model next year, Amodei said that would sit outside the expected scaling laws, which he described as roughly the square of compute. Still, he said models could continue to grow.
The distinction matters: a limit on how quickly a model can grow is different from a limit on what models can ultimately do. In Amodei’s account, the scaling relationship makes extremely rapid expansion less straightforward, but it does not establish that model improvement is about to stop.
Hard tasks do not settle the question
Some researchers have raised a different concern: even very large transformer-based models may struggle with certain tasks. Yejin Choi, for example, pointed out that some large language models have trouble multiplying two three-digit numbers. That example raises a question about what these systems can reliably do, even as they show broad capabilities elsewhere.
Amodei was skeptical that such examples establish fundamental limits. He questioned whether there are hard boundaries on model capability, and whether any such boundaries could be measured well. A task that a model struggles with today might be affected by how it was prompted, fine-tuned or trained.
That argument does not mean current systems can perform every task. Amodei explicitly distinguished skepticism about fixed lists of things AI cannot do from a claim that large language models can do anything now, or will eventually be able to do absolutely anything. His point was narrower: observed failures may not be enough to prove that a capability is permanently out of reach.
Measuring limits is part of the challenge
There are two questions in this debate. One is whether a real ceiling exists; the other is whether researchers can identify and measure it. Amodei expressed uncertainty about both. Even if a limit exists, he suggested, it may be difficult to distinguish a fundamental barrier from a shortcoming that could change with a different training or prompting approach.
This makes isolated examples informative but inconclusive. Difficulty with a particular calculation can show that a model has a weakness. On its own, it cannot establish whether that weakness reflects a permanent boundary, an issue with the model’s training or a task that could improve with further scaling.
Amodei’s skepticism also cuts both ways. He said he is wary of claims that an LLM cannot do anything, but did not claim that every current limitation will disappear. His position leaves room for real constraints while questioning whether they can already be described as definitive.
The near-term outlook remains open
At minimum, Amodei suggested that diminishing returns are not expected for the next three or four years. That is a forecast based on his experience with scaling, rather than evidence that every future increase in model size will produce the same gains.
The broader implication is that assessments of AI capability may need to account for more than a model’s performance at one moment. Training choices, fine-tuning and prompting can all matter to what a system can do. For now, Amodei’s answer to the question of whether AI has fundamental limits is that he is not sure there are any—and not sure researchers can measure them if they do exist.