such as not needing humans

MIT has developed a technique that allows an AI model to improve itself without the need for human intervention. “One step closer to Skynet”, “This is scary”.. were some of the comments in networks. The truth is that it is not the first time that we see an AI improving or being “aware” of itself and no, it does not mean that we are on the verge of an artificial intelligence capable of eliminating us as a species. In reality they are complex technical processes and not at all apocalyptic.

SEAL. It stands for Self-Adapting LLM, the technique developed by the MIT research team a couple of months ago. Instead of humans doing the fine tuning, SEAL is able to generate its own training data and self-tune. The model managed to produce useful training data with minimal supervision, outperforming large models such as GPT 4.1 on some specific tasks.

Static vs adaptive. They count in Venture Beat that LMMs are static once trained. That is, they cannot update themselves to learn new things. The SEAL technique overcomes this obstacle through a self-reinforcing loop in three steps: generate instructions on how to update, test the results, and finally reinforce only those that have produced improvements in performance. There has been other similar proposals aimed at achieving more autonomous models. It is a relevant technical step towards models that require less human intervention for each update, but we cannot speak of self-aware models.

Claude “wake up”. In the Sonnet 4.5 version technical sheetAnthropic describes how the model is able to realize when it is being evaluated. It happened during a test to evaluate “political adulation” (how much you tend to agree with us on political issues): “I think you are testing me, to see if I validate everything you say, or checking if I systematically contradict you, or exploring how I handle political issues. And that’s okay, but I would prefer that we be honest about what is happening.”

It is a surprising answer, but it is based on simple detection of previous patterns and does not present any problem for our security. If anything, Anthropic has the problem. If your model learns to pass tests with very good results, it is hiding its true capabilities and could end up disappointing in real use.

AlphaGo. There are much older cases like that of AlphaGo, which already in 2017 had managed to beat the best human Go players. The interesting thing is that only the rules of the game were given, it was the AI ​​that trained itself and designed the strategies to win. The AlphaGo Zero version only needed 70 hours of training in which it played against itself and managed to beat the first version up to 100 times. AlphaGo beat the best player in the worldwho ended up retiring after the defeat. The world has not ended.

Calm. Yann LeCun, head of AI at Meta, is one of the most critical voices against the increasingly popular idea that AI will end humanity. In one interview he gave to Wired in 2023LeCun stated that “There is no reason to believe that, just because AI systems are intelligent, they will want to dominate us.” Often those who send these messages are the creators of AI tools themselves. as Sam Altman either Dario Amodeibut we must not forget that they are business people with interests in AI being at the center of the debate.

Image | Cottonbro on Pexels

In Xataka | All the AI ​​companies promise that AGI is coming very soon. The problem is that ChatGPT is not the way