I experience worse IR performance adding title/headers to chunks. It really depends on the nature of the documents. The only successful RAG systems I see are ones specifically tuned to a single domain and document type. If your document collection is diverse in domains or formats, good luck.
I dont get why anyone cheers shit like this. It just takes jobs away from more people. You will just get more homeless and drug addicted people. And more rich techies. And it will make no difference to the rider - they get where they need to be either way. It's just another bit wealth transfer from poor to rich. Yet this is cheered.
The only real problem with the plan is will just be take advantage of. People will get meaningless easy degrees just to get a green card. I guess it's tied to employment but im actually not sure how that works. I think it should be tied to only certain types of degrees from accredited schools.
semantic search doesn't really work. Looks nice in little POCs. With real world data it's sort of a mixed bag. Embeddings are just to general, and documents within a narrow domain (finance, medical, etc.) are all "similar"
this book is at the top of the list of near every list of best programming books of all time. its worth reading. is it an easy read? no. its sort of like the Illiad. Not an easy read. But worth it for anyone who wants to be considered well read.
All I know is I feel amazing eating lots of leaner red meat. The nutrient content is very high (B-vitamins, zinc, etc.) and hard to get elsewhere. I probably eat 5 lbs of red meat per week. Im in my 50s, fit, and have healthy bloodwork.
you can also buy leaner cuts of meat. for example flank steak, etc. stuff that needs marinated or slow cooked, and has less fat. also grass fed steaks tend to have less fat (they aren't force-fed corn and soy to fatten them up)
The Sky voice was obviously flirtatious sounding. It was actually very cringy and embarrassing for OpenAI, and it seemed really like a jump the shark moment TBH
I remember a time when software programmers did NOT want to be referred to as "engineers" as that had more of a ops-guy connotation, like someone who sets up servers and runs cable. They wanted to be called "software developers". I even remember when it was fashionable to be called "systems analyst".
The SEO has ruined the internet. Now every recipe web site has 5000 words of narrative before the actual recipe (at least now they all have "jump to recipe" buttons). I really dont care to read 5000 word essay on your families history how your grandma made it every summer, etc. They all do this because the AI/SEO wont return their site I guess?
of course it is... they have run out of good ideas.
just make Siri actually useful by backing it with an LLM with personal memory and you own the market.
Technology always has a weakening effect. AI copilots will make us lazier and produce more buggy code. It can increase productivity for certain boring tasks (generate a bunch of empty classes, etc.). It can increase productivity for early on in projects. Otherwise it's just a crutch.
Recent reads:
Grapes of Wrath by John Steinbeck
East of Eden by John Steinbeck
Jayber Crow by Wendell Berry
The Remains of the Day by Kazuo Ishiguro
AntiFragile by Nassim Taleb (an almost annual re-read)
When I read fiction, I go as far away from modern tech or sci-fi as possible, for balance. I hate sci-fi.
Go long uranium and fossil fuels. There is no other way to generate enough electricity for these data centers, and also for this push for electric vehicles. The green energy stuff isn't going to cut it for a long time.
What is "more vectors"? How many are we talking about? We've been using pgvector in production for more than 1 year without any issues. We dont have a ton of vectors, less than 100,000, and we filter queries by other fields so our total per cosine function is probably more like max of 5000. Performance is fine and no issues.
I agree openai models are by far the best, regardless of papers claiming otherwise according to some strange tests and metrics.
We are also using pagemaker hosted open source models, fine tuned for specific use cases. But even then they can't really compete with standard openai models.
I haven't had more than 6 hours per sleep per night in 25 years. A good 5 hours is plenty. Typically sleep from 11pm till 4 and get up around 5. Drink a ton of coffee and read for 1 or 2 hours. Then workout.
If you are technical co-founder you should only hook up with a co-founder who knows how to sell, and knows how to raise money. Most likely you will pivot from original ideas anyway. Selling and raising money is key.
Apple needs to go back to the Steve Jobs era where they just released a final product with no rumors. Just "here it is" on stage, and available that day in the store. They need to replace Siri ASAP with a real AI LLM.
Yeah have the young kids in your 20s. Thats why you dont need sleep in your 20s. Also has benefit kids are grown and moved out by the time you hit 50...