Problem I'm seeing with RAG-like solutions is you're presenting the LLM with some similar word blocks and hoping it covers all your bases, but won't pick up on 2nd order relationships that would be important for a more complete picture. Which isn't a whole lot better than just showing the search results to the user and might actually introduce faulty information. I think we either need a different model type that can serve as the "memory" or consider options like pre-training to ingest the data that come at the cost of needing much larger servers to perform the operation.
Speaking of malls, I went shopping at a couple malls and they have seemingly picked up on traffic. I wonder if they were there the same reason I was? (I needed to get out of the house)
Think the fact that the leases are coming due over the next few years rather than all at once will prevent a bailout? Might just be a slow decline like old factories in Chicago that sat vacant for a while and then got snapped up for commercial space since they were so cheap.
I moved to a part of Chicago where parking is very expensive. I got rid of my car. I thought I'd need it far more than I did. The ~$500 a month I spent on the loan/insurance/gas/basic maintenance plus $200-300 for a spot (in addition to any other parking when you go places) is just an insane expense compared to the $2.25 train ride that gets me most places and $800/month covers plenty of $20 rideshares for areas not covered by public transit.
Same reason I'm of two minds about Amazon. I disagree with a lot of their practices, but the first time I bought something from a 3rd party Amazon seller, they shipped it to the wrong address (there was a picture of the package clearly in front of the wrong building) and refused to refund due to "it shows delivered". The seller wanted me to "ask my neighbors". The buildings around me have hundreds of people per building. Then told me it was my problem to take up with the shipping company, saying I'd need a police report, all of that. All for a pair of shoes. I went into the Amazon support chat and they refunded it after the seller didn't respond to them in a timely matter (big surprise). That experience has soured me a bit on going outside the "main channels". Had I purchased that from that seller directly through their website, I'd have either spent hours of time proving to them that I never got the product, have to just eat the money, or deal with the back-and-forth of proving to my CC company that it warrants a chargeback.
They feed the fuel into the reactor to superheat it and then it comes out the back like a normal rocket, but then doesn't need oxygen for the reaction to take place.
It's infuriating when you see part of your desired information in the search results and then open the page to find a paywall. IIRC ExpertsExchange were doing that for a long enough time that it was obvious the policy was not enforced. At least not evenly.
I've done something similar and that's absolutely what it was. I'm no pro, knew I wasn't doing it the right way, but it was for a personal side project and Windows starts to get weird when you have a million files in a single directory.
I'm not a professional coder, only side-projects. Never formally taught. I looked at the solution and thought it kinda sounds like something I'd come up with. Like when I didn't know how to use data tables and would hold data in an array of arrays to form the rows and columns. Somewhat clever, "works", but would probably make my professional coder friends vomit if I explained it to them.
I tried to find the article I saw that did exactly that and couldn't, but I guess empirically if you take a look at the LLM leaderboard (https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderb...) and add the "precision" column to the data, you can see the GPTQ/4-bit/8-bit quants still handily beat out the smaller models at full precision. Downside is there's no 3-bit on the submission page, so we can't easily gauge how those are doing, but all my anecdotal personal experience with 3-bit has been extremely disappointing. Exllamav2 might have bridged that gap a bit. Again, wish I could find you that article I had. It laid all this out and showed a huge perplexity dropoff below 4-bit.
I wonder what the usage of that feature looked like since I'd imagine a lot of players don't have a physical keyboard hooked up at all times, and X is still a largely text-based social network. Given how many hard r's I've heard gaming, maybe this will actually make X less vitriolic.
For me personally, being able to fine-tune the local LLM's at a much higher rank and training more layers is very useful for (somewhat unreliably) embedding information. AFAIK the OpenAI fine-tuning is more geared towards formatting the output.
> they risk the possibility of legal liability if they do 2
Have any companies been successfully sued for attempting to police DMCA fraud on their platform? Be interesting to know how courts treat good-faith efforts and if there have been instances where a company just phoned it in and got in trouble for half-measures.
I wonder if they'd get sucked into things like logistics, loading up pallets of equipment rather than getting handed a rifle and some sunflower seeds to keep in their pockets.
How old is your soil? Mixing in some fresh soil with the grounds might be a solution since a lot of the micro-organisms in soil that would break that down and eat the mold degrade over time.
I'm surprised nobody's yet making custom coolers for them (complete coolers with heatsink/fans, not gluing a 3d-printed fan shroud to the front) and re-selling them. Seems like a niche gap in the market.
What is it that make extensible RAM inherently slower? Is it the connectors raised off the boards, adding physical distance to the traces that's the issue?
Reminds me of an old boss telling us "this isn't a 40 hour a week job" and I said to him "I'll agree if you put that in the job postings, before salary negotiations."
Reading your comment, I had the thought "I wonder if California's done it?" Turns out they've tried it 10 times. I'd imagine there's a lot of money being thrown against the initiative by for-profit insurance companies, which is an industry with a lot of cash to buy laws.
I'm picturing Google Street View cars driving around with a box of Pixels in the back, connecting to open WiFi and trying sites and that's why Google can now narrow down your location from what SSIDs are available.
Not exactly what you're asking, but I fine-tuned llama 65b on my texts and it knows a LOT about me. When I ask it about specific situations where I've talked a lot about the subject, it brings in other seemingly unrelated details I don't think would show up in a vector search: https://airic.serveo.net (I scrubbed PII from it and it's only public-facing for a few days)