273 karma · joined July 29, 2021
Also, the curation process counts as "real human input" itself, so I don't think it contradicts my point.
Let's say your LLM can generate text with the same quality as the input 98% of the time, and the other 1% of the time, it's wrong. Each round of recursive training amplifies that error. 96% accuracy after the next round. 67% after 20 rounds.
There's no way for it to get better without more real human input.
If you give it a small enough training set or a big enough neural network, it will directly memorize the whole thing. You have to intentionally make its brain too small to do that in order to force it to find patterns in the data instead.
Diet Coke has huge differences in flavor between sources and packaging options. It's universally bad out of a can or bottle, and the stuff from soda fountains varies by restaurant.
Although since they're on a sidewalk it's a problem if they're blocking public spaces.
I also read "lost" as "lost track of", as in "we received these deposits but then don't know where they went".
Looking at GPT2's token vocabulary, it has the word "the" as token number 1169, but it also has "t", "h", and "e" separately as tokens 83, 71, and 68. When the tokenizer is reading the prompt, does it greedily create the longest tokens possible from the start of the input, or is there a more sophisticated process going on?
I guess in theory it could try multiple tokenizations and see which one results in the overall best response.
As I understand it, a typical strategy for a power company is:
- nuclear or similar for the base load that's always required
- solar and wind as available
- weather forecasting to determine how much solar and wind will be available
- demand forecasting to determine how much power will be required at any given time
- natural gas to meet predicted solar and wind shortfalls
- RICE engines and battery storage to handle unexpected fluctuations
- contracts to buy and sell power with other nearby power companies
As recently as the 1800s, the yard was defined as:
"the Straight Line or Distance between the Centres of the Two Points in the Gold Studs of the Straight Brass Rod now in the Custody of the Clerk of the House of Commons whereon the Words and Figures "Standard Yard 1760" are engraved"
... That standard yard was destroyed in a fire in 1834, and a new standard yard was made using several copies of the original standard yard as reference.
But now, it's just defined as an exact fraction of a meter. First as 36/39.370113 m, now as exactly 0.9144 meters.
I saw an example in a Numberphile video where they were able to take the vector of the word "prince", subtract the vector of the word "man", add the vector of the word "woman", and the resulting vector was closest to the word "princess". So in theory there could be a "gender" dimension and a "position of authority" dimension in that vector (or the model might be making other, stranger connections between words that we don't understand).
I think the same thing is happening in your example. The model identifies and produces output that keeps it in the general region in the vector space for the "Shakespearean" and "Bowiean" dimensions while still satisfying other requirements.