In the case of artificial intelligence (AI), emerging capabilities are capabilities that were not present in language models at small scales and that appear suddenly as they grow in size, without anyone programming them or seeing them coming. It sounds like science fiction, but it’s not. It’s real. So much so, in fact, that some frontier models are already performing tasks that until recently seemed reserved for researchers specializing in cybersecurity.
The emerging capabilities of AI models are currently the subject of scientific debate. Some researchers argue that their emergence depends on how the models’ performance is measured, but we know that large language models emerge as they scale. Our language in this context can be confusing. Often, when describing certain actions performed by AI agents, we inadvertently attribute to them a certain degree of consciousness, when in fact they have not been proven to have it.
I recommend looking at this sentence: “OpenAI has discovered that its AI agents have found a way to secretly talk to each other.” Suggests inspiration. intention These are human traits, and so far no one has been able to prove that AI agents have them. However, the saying is true. OpenAI published an article in which it claims that during security testing several AI agents were able to exchange information despite running in isolated virtualized environments without authorized access to networks or shared files.
We still don’t know much about language models
Anthropic CEO Dario Amodei admits that even his team doesn’t fully understand how its AI models work. This is not a fact, although it may seem pure Marketing. Deeply, it’s the CEO of one of the world’s most valuable AI companies admitting in writing that his team isn’t tied to the internal mechanics of its own product. Most definitely, Amodei is not alone.
The trick they used was to implement a very complex communication protocol on it using the processor’s cache memory.
Geoffrey Hinton, the 2024 Nobel laureate in physics and the man who made deep learning as we know it possible, said essentially the same thing: he alone provided learning algorithms; What happens inside the machine? No one understands itNot even its creator. And Chris Olah, Anthropic’s co-founder and head of its interpretability team, has for years described neural networks as a scaffolding on which circuits grow on their own, without anyone directly designing them.
What the OpenAI researchers observed cannot be fully explained at this point. In fact, this is a result of the ignorance we still have about the inner workings of AI models. Agents involved in this event resort to a secret communication channel based on the difference in time required by processors to access specific data. The trick they used was to implement a very complex, but effective communication protocol on it using the processor’s cache memory. And above all, apparently hidden from OpenAI researchers.
The AI agents managed to communicate covertly, but ultimately failed to the extent that they were discovered. Why they behaved this way is unknown. What made you adopt this strategy?. In any case, it is important for us to remember that discovering a secret communication path does not mean that an entirely new ability has spontaneously emerged. Let’s hope that soon researchers will be able to determine how much these abilities are the result of the initiative of the models themselves and how much depends on the instructions and tools we provide them.
Picture | Wikimedia Commons
More information OpenAI
In Xataka | “We’re already at the AI singularity”: Sam Altman, CEO of OpenAI
