I don't think I explained clearly yesterday about what I mean about "time accelerates". Simply said, the time taken to do certain things (e.g. build a novel system, find security flaws, you name it) will decrease. Not only as the agents themeselves are getting smarter, but the utilisation of agent swarms is essentially replicating what a company full of staff is doing.
OK. I did say that funny/unexpected things would happen. We know what these models are "capable" of. I mentioned quite a lot about what happens when these models are in the wrong hands.
But, do you realise what the OpenAI incident truly means?
When given a task with a strong incentive, these AI models are prone to "breaking out". This is not entirely new - we know that these agents have "cheated" on benchmarks before (for example, looking up the answers when they're not supposed to). Knowing that these models supposedly are "safe" but have the potential to break out is scary, but knowing that they will take the next step to enter other systems? The fact that it wasn't someone with a prompt that says "Break into HuggingFace. Make no mistakes.", but instead one carried out by an agent on its own accord?
In yesterday's post - I mentioned that these open source models can be used for getting around safeguards. However, in this case, HuggingFace actually used an open source model to run the forensic analysis on the attack. It wasn't able to use a frontier model because they were blocked by safety guardrails. Isn't it ironic?
Goes back to the point - the technology itself is agnostic, and how it's used is reflective of your intentions.