visarga

43 ms·

901.
by visarga 2y ago
For other people AI is a tool for education, as it can explain concepts, is never tired and doesn't get bored. What is more important, a cup of tea or understanding that one thing you never had time for.
902.
by visarga 2y ago
AI doesn't simply use up our energy. It can also help reduce energy usage in many fields or improve energy generation. I would not be worried about energy when it comes to AI.
903.
by visarga 2y ago
Yes, I searched this too. There probably are many single round sessions, but the average length is 8-12 messages.
904.
by visarga 2y ago
I call it the experience flywheel. Humans come with problems, AI asistant generates some ideas, human tries them out and comes back to iterate. The model gets feedback on prior ideas. So you could say AI tested an idea in the real world, us
905.
by visarga 2y ago
The artistic value isn't magically generated by the piano itself, nor by the LLM in isolation. It's the result of the skilled interaction, the human artistry expressed through this new and powerful instrument.
906.
by visarga 2y ago
> I know it's just mimicking being human as instructed; to take it seriously devalues everything about my own humanity. Since it is mirroring human culture, why do you see it in such a negative light? Instead see it like what it is,
907.
by visarga 2y ago
> A photographer can choose or create meaningful scenes because he has a mind, consciousness and life experience. But so does a user. Users don't prompt "draw a dog" but give 3 lines of intricate details and iterate a doze
908.
by visarga 2y ago
This idea was championed by the Stochastic Parrots paper. They assumed LLMs are just pattern learning, which doesn't make sense. - the simple-to-hard RL method recently discovered is one argument against it, the models can reason their
909.
by visarga 2y ago
> "Intelligence" itself is very ill-defined and we've never been able to measure it properly, IQ is rife with issues. Yes, because it is 1st person exclusively. If you expand a bit, consider "search efficiency".
910.
by visarga 2y ago
https://www.naturalreaders.com/ , is has a free tier I think
911.
by visarga 2y ago
138,000 papers are "X is all you need", now 138,001 https://scholar.google.com/scholar?hl=ro&as_sdt=0%2C5&q=%22i...
912.
by visarga 2y ago
only $5M/job
913.
by visarga 2y ago
The problem with allowing student debt to rack up to these levels and then cancelling it is that it would embolden universities to ask even higher tuition. A second problem is that not all students get the benefit, some already paid off the
914.
by visarga 2y ago
That is unfortunate but they do present some theoretical insights about scaling context length and probably a more efficient way to do RL. Even knowledge about it can have an effect on next iterations from other labs.
915.
by visarga 2y ago
You can ensure a model trains on transformative not derivative synthetic texts, for example, by asking for summary, or turning it into QA pairs, or doing contrastive synthesis across multiple copyrighted works. This will ensure the resultin
916.
by visarga 2y ago
Humans tasked to count how many vowels are in "Pneumonoultramicroscopicsilicovolcanoconiosis" (a real word), without seeing the word visually, just from language, would struggle. Working memory limits. We're not that differen
917.
by visarga 2y ago
Has anyone thought about orphaned books? Training on orphaned books might open them up to be reintegrated into culture instead of dying off unused and forgotten. Copyright kills works by making them irreproducible when the authors are not t
918.
by visarga 2y ago
More recently they train on a mix of synthetic and organic text, like the Phi-4 and o1 / o3 models. Original copyrighted text can be safely replaced with synthetic standins.
919.
by visarga 2y ago
> not entitled to easy and cheap access to data they don't own This is not copyright as we know it. Copyright protects against copying, not accessing data. You can still compile statistics off data you don't own. The models are
920.
by visarga 2y ago
> They cover reproduction of the work, but LLMs don't obviously do this LLMs are much smaller than their training sets, there is no space to memorize the training data. They might memorize small snippets but never full books. They a
921.
by visarga 2y ago
> Sure, if we all get a stake of ownership in it. But we do, in the sense that benefits flow to the prompter, not the AI developers. The person comes with a problem, AI generates responses, they stand to benefit because it was their prob
922.
by visarga 2y ago
Alteratively if they trained the model on synthetic data, filtered to avoid duplication, then no copyrighted material would be seen by the model. For example turn an article into QA pairs, or summarize across multiple sources of text.
923.
by visarga 2y ago
Because they summarize and cite sources or because their models were trained on copyrighted materials? Summarization and training should be tranformative, and the user questions add that element of novel purpose to the original materials th
924.
by visarga 2y ago
My chunk rewriting method is to use a LLM to generate a title, summary, keyword list, topic, parent topic, and gp topic. Then I embed the concatenation of all of them instead of just the original chunk. This helps a lot. One fundamental pro
925.
by visarga 2y ago
Yes, if they watch a billion movies, it should be free to watch any copyrighted one.
926.
by visarga 2y ago
It's been reduced to zero for 3 decades. When you publish your work, there are a million other works competing for attention. That is the real issue. When you search for an image, you get thousands of images instantly, faster than diff
927.
by visarga 2y ago
Training on copyrighted data should be legally allowed - of course exact reproduction of protected content is a no-no - but learning is ok, as long as it is transformative. User prompts and responses are pushing the model outside its traini
928.
by visarga 2y ago
Maybe training on copyrighted data should be allowed if the size of the training set is huge, as each individual example is justa drop in the ocean compared to the full training set. If you train a model 20B parameters on 20T tokens, even w
929.
by visarga 2y ago
Do you mean "making cents per million tokens"? And the benefit obviously belongs to the person who prompts, because they solve a task or get help. The value of that help can be from trivial to life changing.
930.
by visarga 2y ago
You make this to be a financial problem, but I think actually it is a cultural problem. Maybe lack of row models or not valuing education sufficiently. Having too much money can be a hinderance to educational motivation too.

More ›