====== 2024-07-01 - useful AI ======

for some time now AI, [[wp>Generative_adversarial_network|GANs]] and ~recently [[wp>Large_language_model|LLMs]], are making headlines. so far though most of the stuff i've tested out was giving mediocre results, to put it mildly. don't get me wrong -- it's VERY impressive in a sense of "this thing can even be done with such a basic take on it?!". it's not however much of a practical use IMHO, for helping out in day-to-day work. ChatGPT / GPT 4 is hallucinating a lot, and often spit out garbage (and then "apologizes", when confronted). Copilot's [[https://www.gitclear.com/coding_on_copilot_data_shows_ais_downward_pressure_on_code_quality|code is poor quality and spirals towards increased code churn]] (although -- surprise! surprise! -- [[https://github.blog/2023-10-10-research-quantifying-github-copilots-impact-on-code-quality/|copilot's owners do not agree with that]]. on top of that it also lacks recent pieces of information. last but not least -- it lacks references to back up its claims.

the last part turned out to actually be an opportunity. it's where [[wp>Retrieval-augmented_generation|RAGs]] come along.

during recent [[https://open.spotify.com/episode/63GlhZMbgK6deHbF7QKrPH|Lex Fridman's interview with Arvind Srinivas]] (CEO of [[https://www.perplexity.ai/|Perplexity]]) i've learned about a new tool. it's a mix of LLM and RAG, that it's a kind of search engine, where you state a problem, and instead of links, it generates a concise response, with references and some follow up topics, that you might want to dive into next. it's also faster in giving responses. it does fill a bit like [[wp>Wikipedia]] in that regard.

while it inherits the LLM flaws, thx to RAG's grounding in data and references, it makes stupid mistakes far less often. although i find most answers far from perfect, it's actually the first time i find modern AI useful in my daily work flow. especially when i'm looking for sth i do not have much experience with or i need to find sth to start with unusual problem at hand.

i highly recommend you give [[https://www.perplexity.ai|perplexity.ai]] a shot! :)