Heads up if you are trying out gemini as a swap in for an openai api call it will fail with api error or list index out of range if it tries to output OpenAI or ChatGPT.
Our discord bot is struggling since brain transplant because of it.
Pretty sure a lot of new ‘scraping’ is actually a person getting an AI to browse a single site for them at a time, not an automated crawl of everything. Smart companies should also have a AI-friendly page, simple and with all text in one place, if they want to stay relevant.
Easily solution: Allow scraping, be open like the internet was meant to be. Wikipedia surely is less funded than this place and allows all forms of scraping known to man.
@futilereality@CoffeeVectors But still you benefit from your past experience, and you integrate tools you’ve learned into your kit even as new ones come along, I don’t think you directly benefit from shaking things up all the time? It’s a tough one
@futilereality@CoffeeVectors Yeah working with bio systems needs adaptability based on what you see when you get in there (farming and beekeeping come to mind as well as biotech). And the tools in farming and biotech keep evolving, protocols adapting. Hyperlocality means you can’t just teach one way to work
@fluxophile @timnitGebru You asked for ways ai might help education, not replace it right? Also I had lots of exams with essay portions, or no calculator math tests or in class work when they wanted to emphasize learning that skill. The ability to teach and assess writing isn’t gone.
@fluxophile @timnitGebru Correct, and nothing says we had perfected education before ai came around and broke it all instantly, there are going to be strategies that don’t work anymore but also new ones that allow for other kinds of learning. Also by refining here I meant grading and critically jusging
@fluxophile @timnitGebru It doesn’t have to be, just one example where ai could enable output oriented education strategies, the tech is new enough who knows which parts will stick and how. I would say some classes are focused on learning through making new useful outputs, like journalism
@fluxophile @timnitGebru Finally the calculator analogy might extend well here, people still learn how to do math too, and it helped move education away from just route memorization to more time spent on concepts. And often tests and teaching use letters instead of numbers to emphasize the concept..
@fluxophile @timnitGebru Graded and tossed might result in more useful skills and more confidence too. Also there’s the ability to carve out classwork around judging and refining ai output, finding and confirming sources, getting the kind of education teachers get though with ai as the “student”
@npew 2 possible causes:
1) I think chat gpt 3.5 has had longer context windows than chat 4 at certain points in time
2) I had to change many api system prompts from 3.5 to 3.5 16k and to 4. but usually its the new model following my bad instructions better
its almost like.. embeddings was compression the whole time.
wordcount is a embedding and a compression that loses position sensitive to common words
tdidf is an embedding and its a compression that is insensitive to document frequency
gzip is an embedding like tfidf
but contains positional information and a has a tokenizer.
Absolutely love this.
The fact that Instant Pot is already being framed as a corporate cautionary tale—the company that went bankrupt bc they made a product so durable & versatile that its customers had little need to buy another one—instead of as a critique of capitalism is deeply, deeply depressing.