F-Droid is littered with dead-end pet projects and no easy way to filter them out. Instead of wasting time on Filecoin pipe dreams, they should just fork Modrinth https://github.com/modrinth to handle Android apps instead of Minecraft mods. The difference in usability is night and day: https://modrinth.com/mods
F-Droid prioritizes privacy. Modrinth does not. Users who download apps don't create accounts and cannot even create accounts. This is by design. There are several clients that access F-Droid repos that add those features for users who are willing to pay the privacy cost.
I regularly find myself on the internet archive hoping to find a working copy of a page created ten-plus years ago. Page rot, SEO spam, walled gardens, and AI generated nonsense are all converging to suck the value out of the WWW.
Projects like this give me a little hope that blogging can make a comeback. NIP-13 [1] also has the ability to use POW to limit bot activity. Thanks for sharing!
Unlike web scraping where spaghetti logic is required to follow abstracted JS links, archival of nostr events can be as simple as running a relay and mirroring blog content.[1][2] nostr does a lot of what NNTP did but with additional flexibility.
You should take it upon yourself to save your data, if it is really important to you, so that if all your relays fail, you can simply rebroadcast all your events to a new one and you will be fine.
It's a problem solved in a collaborative and decentralized way.
I frequently kept a twitch pop-out in the corner of my screen throughout the day. When uBlock stopped working, I almost immediately stopped using the site. Video interruptions are pretty awful for live content where channel switching is a regular occurrence.
They may have been paying for subscriptions, bit rewards/redemptions, prime, etc. Ads are only one source of Twitch's revenue.
Twitch started showing a purple "disable your ad blocker" screen when a block of ads failed to run. I think they probably would have made a lot more money and reduced a lot of churn by simply advertising Turbo instead, which many users don't even know is an option.
I think this is the key takeaway for me at least. If we don't blow ourselves up first, the sun will basically blast us into the stone-age every 6000 years at best or 200 years at worst.
You can also add this to the Fermi paradox list. If intelligent life is rare, solar systems where intelligent life can develop the type of technology capable of space flight might be rarer still.
For all we know the Sun is actually quite friendly to this. Most stars might behave this way more often. If that's true then intelligent life able to harness electricity and all it entails would be very rare. If our Sun did this every, say, 25 years there would never be an industrial/technological civilization here... or at least not a sophisticated one able to build things like spacecraft.
In any case this is something we should be studying a lot more than we are. It is a far more tangible and realistic existential threat than very hypothetical AI apocalypse scenarios.
Latency and cost. GPT-3.5-Turbo is very very fast (for reasons I still don't understand) and cost is very very low even with the finetuning premium.
Llama2 is still slow even with all the LLM inference tricks in the book and you need to pay for expensive GPUs to get it to a production-worthy latency, along with a scaling infra if there is a spike in usage.
So, as somebody who has fine tuned llama2 (13b) on both a new prompt template / chat format, as well as instruction following, summarization, knowledge graph creation, traversing a knowledge graph for information, describing relationships in the knowledge graph, etc.
It is able to use the knowledge graph to write coherent text that is well structured, lengthy, and follows the connections outlined in the graph to the logical conclusions, while deriving non-explicit insights from the graph in it's writings.
Just to say, i've seen a giant improvement in performance from Llama2 by fine tuning. And like I said, just 13b...I am perfecting the dataset with 13b before moving to 70b.
3.5-turbo is sometimes okay, i've tested it moderately for the same tasks i've been training/testing Llama2 on, and it's just a bit behind. Honestly, my fine tune is more consistent than gpt4 for a good number of the tasks i've trained.
but how is the speed here? does it feel fast "enough"?
looking into to running llama on prem / private cloud but i have no idea where to start in terms of sizing, do you have any details or posts on to what the minimum / recommended hardware requirements are?
EDIT: just looked myself, not as encouraging as I'd like: "For good results, you should have at least 10GB VRAM at a minimum for the 7B model, though you can sometimes see success with 8GB VRAM. The 13B model can run on GPUs like the RTX 3090 and RTX 4090"
definitely borderline dealbreaking for solo hackers / small teams
1x 3090 IMO is about the minimum you'd want to waste time with. It can serve a 13b + 7b model at once if you want, you can qlora train a 13b with a ton of context length, and it's fast enough to iterate with for training.
I have 2x 3090 in my machine, and I can do inference of ~40tokens/sec on a 13b llama2 model on one card. I can split the 70b parameter model between the two cards and get ~12-15tokens/sec. I can't train the 70b parameter model with my 2x 3090 though sadly, not quite enough vram.
Indeed, and this is really missing from the public discourse. People are talking about Llama 70b as if it was a drop-in replacement for gpt-3.5, but you only have to play with both for half an hour to figure out that's not generally the case and only looks true in cherry-picked examples.
