Unmaking Sense

Episode 11.77: Why context length matters and why it can easily cost users a lot of money.

The cumulative cost of longer and longer context lengths can easily lead to very high user costs because we are charged by the number of tokens consumed.

Episode 11.76: Why the non-judgemental aspect of AI matters, and what happens when it goes wrong.

Experiencing an AI that is unco-operative and even rude. Whether this may afford some clues to human aggression. And what does it suggest about what might ensue if AI with power were to go wrong?

Episode 11.75: More on long-context models and why they are such a big deal.

Continuing to explore some of the implications of long-context models such as the latest Google Gemini with a maximum of 10 million tokens.

Episode 11.74: Google’s Gemini 1.5 with a potential context length of 10 million tokens.

The frenetic pace of change continues.

Episode 11.73: AI models with attitude. Mixtral-8x7b-v0.1.Q3_K_M.gguf and chatting and completing.

Much ado …

Episode 11.72: They also server who only cache and weight. How servers serve.

A few random remarks about my adventures with server.cpp as part of the llama.cpp project. With apologies to John Milton.

Episode 11.71: Using AI to detect patterns in biological data. Making a new kind of sense to humans.

The vast amounts of data that human minds cannot make sense of and how AI will do it for us.

Episode 11.70: Reading significance from data that humans think only noise that means nothing.

How AI is intrinsically better at discerning the meaning in what humans think irrelevant noise. Why anticipating trouble - for all its echoes of “Minority Report” - is or soon will be more possible than we might like to think.

Episode 11.69: Dealing with some technical details concerning volume and storage of user details.

Servers that store data.

Episode 11.68: Some objections to the RIDE-AI thesis.

A rather windy and noisy episode.