I follow on from the last post - when I mentioned that an open source model is released at the same frontier level as Mythos/Fable, that is when "time accelerates". Well, this has happened with the recent release of Kimi K3.
Scarily enough, I think most of us thought that this would take much longer before it would happen. But now that it's here, I should explain why "time accelerates" and why every action from this point matters.
Now, I don't think any of us are capable of playing God, or can have entire dominion over our living world.
However, we've seen time and time again that bad actors are capable of causing harm with a particular area of effect. Anything from antisocial behaviour to terrorist acts.
Yes, it's possible to learn anything, including illegal acts, if you know where to look.
But, why is the release of Kimi K3 potentially game changing and dangerous (if in the wrong hands)?
On the surface, it was something that was bound to happen (as I mentioned in the last post), and in reality any technology is not necessarily "good" or "bad", it's mainly how it is used.
My belief is that humans typically abide by moral code, trained and restrained by our emotional reactions, possessing a capability to learn right from wrong. In the exceptional cases where people have biologically different responses, people who "can't" learn could potentially end up going down the wrong path. But let's assume that the majority of people can exercise restraint and/or learning.
OK, so that aside, an AI "agent" is very well capable of taking actions on our behalf. When we delegate actions to someone else, responsibility becomes diffused. What happens when we delegate actions to AI? Really, the "operator" is responsible, but they don't necessarily feel that way. What his means is that our biological/emotional feedback loops are no longer at the forefront.
So, someone in control of AI can potentially take nefarious actions, without having gone through the motions, IF the AI agent lets them.
Quick aside about AI safety and guardrails. Yes, we may bash closed source models for other philosophical reasons. But the bottom line is that the most popular models will have safety controls baked in - if you ask about committing violent acts, perpetuating distributed harm unto others etc. it will block you, and potentially report/flag you to the authorities. (Yes there is also a different concern around Minority Report type surveillance here.)
What is jailbreaking? It's the act of getting around these safety controls to access knowledge and information that are safely guarded away.
In a closed source model, you may utilise prompt engineering to achieve such results - but these operators pick up on prompt injections and patch them up.
When you have an open source model, not only can you use a particular version of the model that is backdated without patches, but someone with the knowhow can modify model weights and effectively remove safeguards.
So now, in the wrong hands, a frontier model without safety is the most dangerous outcome. These frontier models, according to benchmarks, have an IQ equivalent to about 130 (about 2 standard deviations above mean). So roughly, it's got mental capability that only the top 20% of the world population has (napkin math).
If an AI agent possessing all the intelligence and a removed moral code is in the hands of a bad actor, what do we think could happen?
I'm not writing all this to sound like a doomer, because the majority of us are sensible beings. But unless all of us are morally and ethically on the same page, we might not have have the desired outcome.
If we can distil some Christian philosophy - Jesus taught us to repent for our actions. This is ever more relevant right now - with all the technology that we are given - we can live to a fuller potential but also should not have a mindset of harming or taking advantage of others, whether you agree with them or not.