Artificial Intelligence

[Tero]

Bits of words get turned into vectors of a zillion dimensions:
https://milvus.io/ai-quick-reference/ho ... to-vectors

[Tero]

Finns like to use shoes with studs similar to tire studs. I asked Google AI about that. It explained all about slippery weather and walking. An legal and city angles.
Then I asked about a related item in Finnish.
It answered in English.
I commented in Finnish.
It replied in Finnish.

[Tero]

Me: what is the hardware and algorithm used for large language models
Google gives link:

https://www.oneadvanced.com/resources/l ... e-aspects/

[Tero]

ME In my book: In 2020 GPT-2 boasted 1.5 billion parameters. What parameters?

In the context of GPT-2, parameters are the internal numerical values—specifically weights and biases—that the model learns and adjusts during its t
raining process. They function like "dials" or "knobs" that determine how the neural network processes input text to predict the next word in a sequence.

ME "1.5 billion weights and biases within its neural network" means it has access to these 1.5 billion entries at all times?

Exactly. Every time you give GPT-2 a prompt, it passes your text through all 1.5 billion of those numerical values to calculate the response.

Active Calculation: Think of those 1.5 billion parameters as a massive, complex mathematical formula. To predict the next word, the model performs billions of multiplications and additions using every single one of those weights and biases.
Loaded in Memory: For the model to work, these parameters are loaded into the computer's VRAM (Video RAM) or system memory. They aren't stored in a "filing cabinet" it looks at occasionally; they are the literal structure of the program itself.
Static Knowledge: These 1.5 billion values are "frozen" after training. They represent everything the model "knows." It doesn't "learn" new ones while you chat with it; it simply applies that existing 1.5-billion-piece formula to your specific input.

[pErvinalia]

Parameters are over a trillion I believe in current models.

[pErvinalia]

Or maybe it was a few hundred billion.

[Brian Peacock]

Anthropic is throttling down access for all its services at peak hours. Peak ours are 5am to 11pm PDT.

[pErvinalia]

Yeah I got a warning last night, which would have been morning in the US. no worries through my daytime today.

[Tero]

How soon can I get a little AI on my phone? One that has all the words. It will just spend minimal data to get me my eider duck photo in flight.

[Brian Peacock]

As soon as the quantum chips become available - then we'll all have a virtual supercomputer in our pockets.

[JimC]

Is that a virtual supercomputer in your pocket, or are you just glad to see me?

[Brian Peacock]