Unmaking Sense

Episode 11.20: Q-learning and AIs that change their own methods of operation.

A continuation of 11.19.

Episode 11.19: Q* = The latest speculation about what really happened at OpenAI regarding Altman.

The twitterati (exorati doesn’t quite do it) are speculating that what the board were concerned about over Altman’s alleged “lack of candour” concerned undisclosed research into an amalgamation of GPT technology with Q-learning that could give (or in some versions already has given) rise to an AI capable of cracking the top encryption algorithm in use: AES 192. Nobody knows for sure, but the speculation is rife.

Episode 11.18: The Dorothea Brooke Syndrome and why nobody knows what is going on most of the time.

Reading the future is computationally expensive, which is why so few can do it at all and most can only imperfectly for very brief periods of time. AGI might be able to do better.

Episode 11.17: The AGI that sees patterns no human intellect can discern, and some implications.

To be able to see patterns that no one else can see confers upon an AGI, at least potentially the capacity to see the future more clearly than any human being conceited and that has very serious implications for the way we deploy AGIs because the future that they see and the way they evaluate it, and the way they suggest that we should try to aim for it or aim away from it could have very very serious consequences indeed.

Episode 11.16: What is AGI and how might it be able to see more than we humans could ever see?

The key to artificial general intelligence is not that it is just more intelligent than any other AI, but that it will be able to see patterns in data in which it has not been trained to see them, and discern in them patterns and their likely consequences that no human being or lesser AI could ever conceivably see. And the question that we and those considering the ethics of AI need to address is “What constraining factors are there to prevent an AI with the power to affect the world from acting in the best interests of the world in ways that we as human beings might not think to be ethical, or in fact, in the best interests of the world?” We are in familiar territory, the territory of John Stuart Mill’s famous question, whether we ever would feel justified in killing 10 people to benefit 30, 300, 3000 or 300 million people, and how we would or could train our AI to make similar decisions.

Episode 11.15: Artificial General Intelligence (AGI) is effectively here in all but name.

When, as is the case with the latest versions of online chatbots. an artificial intelligence can respond to almost every conceivable kind of question, and service almost every conceivable request, there I think a case can be made for saying that we are at AGI in all but name.

Episode 11.14: AGI will arise under the influence of international competition to stay ahead.

AGI will arise because nations will engage in just the same kind of arms race that led to the development of nuclear arsenals and mutually assured destruction, for the simple reason that they dare not do anything else for fear of falling behind, and so falling prey to the devices that other people have developed. If the only way to defend ourselves is to develop a means of scrutinising everyone’s behaviour all the time, then we will in effect have created an AGI that has all the powers that we have historically traditionally attributed to our deities. The only difference will be that these deities will exist and be real, and will be fighting one another rather as the Greek and Roman gods were thought to do, for good or ill.

Episode 11.13: Why the history of religions is relevant to the emergence of AGI.

Human beings have been fascinated by gods and their God for most of their history. We like to think that someone is in control, that someone knows what is going on, that someone understands everything, and that we are part of some great cosmic plan. We like in particular to feel that someone is watching our back, that there is someone watching over us. It’s almost certainly not true, any of it, but we cling ambivalently to the notion anyway. If, as I believe, nobody knows what is going on most of the time, this is one reason why we like to hold onto belief in a deity, for deities are those who do. Now it is possible that something like an artificial general intelligence will watch over us all the time and know what is going on all the time. Now there may at last be a God.

Episode 11.12: The OpenAI shenanigans and why cleverness is not always accompanied by intelligence.

The topic of the moment.

Episode 11.11: Low-cost ways to adapt and specialise trained models. LoRA - Low Rank Adaptation.

LoRA stands for “Low-Rank Adaptations of LLMs”; it is a technique that involves making low-cost tweaks to the weights of trained models by adding small adjustments.