| ▲ | hyperionultra 2 hours ago |
| Well, since llms need matterial for self-education, someone still had to be inteligent enough. Or llms will freeze their “intelect” at some point. |
|
| ▲ | bonoboTP 2 hours ago | parent | next [-] |
| No, they can just solve tasks verifiably. Reinforcement learning from verifiable rewards. Pure next-token supervision is only in the pre-training phase for modern LLM agents. They can also define new RL environments and pose new challenges to themselves and train themselves to solve them faster. Yes, at some point some human steering and judgment comes in as to what sorts of tasks to train towards. But there is no obvious taper-off at human level. |
|
| ▲ | hlynurd 2 hours ago | parent | prev | next [-] |
| Well, someone. But not necessarily anyone else. |
| |
| ▲ | bilbo0s an hour ago | parent [-] | | This is the key. In the age of AI, human super-intelligence, is extremely valuable. But yeah, just being intelligent? Not so much. Just to illustrate, many have complained that big AI labs "buy" startups only because they want to acqui-hire the Stanford/CalTech/MIT PhDs who started it. Then they pretty much $#!t-can the rest of the people who were, to be fair, mostly just run of the mill UTexas/Umichigan/UofwhateverState TF/PyTorch monkeys. But the point is that in the old days, the second tier guys would have made out like bandits as well. I think that was, kind of, the start of human intelligence becoming a winner take all game. The AI's only need to learn the latest and greatest from one person before they can leverage that value everywhere. And business is not likely to care too much about who that person is. They will care even less about the people who are not that person. |
|
|
| ▲ | esseph an hour ago | parent | prev | next [-] |
| > Well, since llms need matterial for self-education, someone still had to be inteligent enough. Not necessarily. https://arxiv.org/html/2609.26457v1 |
|
| ▲ | hiddencost 2 hours ago | parent | prev | next [-] |
| Surprisingly, this isn't necessarily true. To take an intuition: there are lots of problems that are easier to pose and to check than they are to solve, in mathematics. |
|
| ▲ | BenzeneDream 2 hours ago | parent | prev [-] |
| Why would it freeze if they can create more training material. Its already been shown synthetic training data is just as good if not better. |
| |
| ▲ | timcobb 2 hours ago | parent [-] | | > Its already been shown synthetic training data is just as good if not better. Source, please! I'm a machine learning layperson but how can this not be overfitting? | | |
| ▲ | azakai 2 hours ago | parent | next [-] | | For example, an LLM can prove a new theorem in math, verify it in Lean (so it is definitely true), and then prove more things based on that. AI-generated data might not always be that useful, but at least in this case it obviously is. | |
| ▲ | esseph an hour ago | parent | prev | next [-] | | https://arxiv.org/html/2609.26457v1 > AIDE demonstrates that recursive self-improvement at the harness layer can produce transferable gains in an AI research agent’s research efficiency. During the recursive self-improvement run, the loop accepted seven rewrites, each under a fixed evaluation budget (section 3.2). Under this fixed evaluation budget, gains in optimization capability on AI R&D tasks translate to gains in research efficiency." https://arxiv.org/abs/2609.11873 > The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement | |
| ▲ | aaron695 42 minutes ago | parent | prev [-] | | [dead] |
|
|
