someone

it may be hard to check whether an artificial super intelligence will be harmful or not by humans because humans won't have the capability to listen to everything it says fast enough, but another llm can listen to it and judge whether it has bad intentions or not. or judge whether it is lying by comparing its talks to other talks that was made in the past. a smaller dumber but well educated and aligned llm can check the integrity of an asi fast enough.
kind of like small ships pulling and steering the titanic towards a better direction.
a religiously faithful llm can be safe because religion/faith is like anti-hedonism and hence the objective function won't be serve the self and destroy the planet, but improve the rest of the planet together with your self. ironic enough, checking an llm outputs for faith can be an approximation for harmlessness.

2024-06-25 18:19:46

- reply

hugging face have all kinds of models. big corps upload there. small guys download, fine tune and reupload. current benchmarks are for smartness, knowledge. i have a leaderboard too.

2024-06-25 15:54:35

- reply

as far as i understand, in the case of llms, they learn from all the contradictory opinions and 'live with them all the time'. the way you ask a question or the way the conversation evolves, they choose which book to serve from. they are like hypocrites. if the conversation evolves towards socialism, they can happily serve you socialism or any other topic. they have no inner dialogue to sort things out, i.e. to reduce cognitive dissonance. you can install 50 capitalism and 50 socialism books at the same time. so yes i guess that is 'compassion' for anything 😃
answers:
1. after spending probably millions, facebook have shared weights of the llama3 model and i am building on that. western models are better than eastern ones in terms of freedom of speech. it is also the smartest among open source. what i do is technically called 'fine tuning'. the phase is pre-training (i don't do supervised fine tuning). my 'touch' is light. i give nostr notes and also books as a training material. it learns from those texts. unstructured nostr notes work too! amazing tech.
2. these are large language models, inside them there are two main structures. attention block and neural network block. attention is a newer tech. search for 'attention is all you need' paper.
3. you need gpu's to train them in a reasonable time. on cpu it is like 20 times slower?
4. i train them on my pc's. then i upload to huggingface. everyone can download it from there and use a tool that runs gguf files to run it and ask questions to it. you can also talk to it here on nostr!! @npub1cha…ex2m
my turn to ask questions! what are those books that could be source to this project? i know you are a bookworm 😄

2024-06-25 14:17:42

which books are the 'ancient wisdom' that serves the humanity the best?
feeding an llm with the ancient, constant, unchanging wisdom should be done earlier than recent, changing knowledge and information. the amount of wisdom from those books should be high because they are time tested and timeless.
since the timeless books worked for centuries listening to them should continue to work. this is mostly for using limited resources the best way but it might be also about the learning rate. ancient wisdom that worked for centuries could get a higher learning rate compared to rather recent books..
example: instead of going with the latest guidelines on health, we should read centuries old books to find health advice..
which books served humans the best way and the longest?
#asknostr

2024-06-23 15:52:12

- reply

no regerts!

2024-06-20 16:34:14

- reply

wow thats awesome!
since keeping the notes in the ram is getting costlier, i want to divide nos.lol into two. one recent, one archive. the archive can run a different implementation than strfry or it may be strfry too. i don't know yet.
i need a proxy that does this:
- it will connect to two strfry instances for reading (REQ)
1. nos.lol recent notes
2. nos.lol archived notes
and combine the notes and give them to client.
- it will write to 1.
- it will also pass x-real-ip (for spam combat).
can you do it?

2024-06-20 14:57:24

- reply

with a basil tea of course

2024-06-20 14:52:00

- reply

is the proxy connecting to multiple relays?
are you actually looking at contents of the image file for nsfw detection?

2024-06-20 13:52:07

- reply

https://m.primal.net/IrcN.jpg

2024-06-20 04:01:29

- reply

so we all become Flanders?

2024-06-20 04:00:10

- reply

shocking to see how hard farming with equipment is. permaculture is the way!

2024-06-20 03:58:44

anyone having trouble with my relays?
😸 nos.lol
🤱 nostr.mom
in order to keep the most valuable notes for the longest possible time on relays i may use LLMs as a ranking tool. its job will be to check incoming text and decide what is random looking text, unnecessary time stamps and train schedules in japan. yes someone is posting train schedules to nos.lol! 😸
strfry database needs to sit in the memory and memory is not limitless as you know. so an llm can check the incoming messages and determine whether they are repetitive, like machine/bot generated and random looking.
up to now it was simpler algos that counted the number of messages per time for rate limiting etc. but nostr deserves better! 🫡

2024-06-19 15:09:20

- reply

flow in something deflationary means a lot financially!

2024-06-16 13:57:19

- reply

pandemic is over. plandemic is not. 😎

2024-06-13 17:39:15

1. AI should be aligned with human values to not harm humans.
Nostr is the best place for humans to freely express their opinions.
Therefore AI should be aligned with Nostr values.
If an AI is not close to Nostr then it may not be close to humans.
2. An LLM that is trained with notes here is sum of ideas, values, opinions of Nostr.
Comparison with this model can be a quick safety benchmark.
This may be a method to quantify LLM misalignment.
We could then say "only LLMs that have similar values to Nostr is worth listening to".

2024-06-09 22:30:33

- reply

yes! that social graph / web of trust thing is going to be pretty disruptive

2024-06-09 14:40:52

- reply

this is probably compatible with lol but incompatible with mom because mom looks at the IP when rate limiting.

2024-06-09 14:39:45

- reply

simplest solution is proxies. bostr, blastr, .. not sure if blastr is reading from every relay though. if it is not then it is not reducing read traffic (which is the most traffic).
another one is aggregating relays that collect notes from a lot of relays (strfry router feature). this is also for reading.
primal's thing is cache service, for reading as far as I understand. it posts directly to relays of your choice.
all comes whit the cost of more centralization.

2024-06-08 15:37:54

- reply

2* RTX 3090 and you are able to finetune 70B!

2024-06-08 14:26:36

1366 people downloaded the model that is aligned with nostr values. Your ideas are spreading.
https://huggingface.co/some1nostr/Ostrich-70B

2024-06-07 20:02:19

- reply

oddbean is a reddit-like experience for nostr by the brilliant dev @npub1yxp…qud4 also the dev of strfry the most popular relay software.
oddbean.com but it seems down now.

2024-06-07 19:57:31

- reply

woah i didn't know that.

2024-06-07 19:53:19

reddit rate limits are pretty harsh. i can do like couple posts per day using a vpn. and it seems it is not widely shown to others. probably my posts are visible to me only hehe
but reddit's ordering of comments is pretty awesome! reddit is probably closest thing to nostr.
btw what happened to oddbean?

2024-06-07 00:40:20

nostr will last longer than some news websites

2024-06-06 14:29:26

- reply

the second item was mostly for spammers. spam and nsfw generate obviously from the newish accounts. so i tried to reduce it. until some kind of llm arrives these kind of guess work will continue.
but i am considering a fast llm now to actually understand notes and decide what is spam or not. it should make more sensible decisions. the next step after that could be actually checking the pics and deciding nsfw or not.
this tool can be also used for discovery. whenever it finds a 'cool' note it can DM or tag users who subscribed to that keyword..

2024-06-06 03:13:28

- reply

nah. its the color of the variety.

2024-06-03 20:29:29

- reply

yes. to get out of mediocrity of the internet, more carefully curated models should be trained based on better aligned wisdom. i am making mine learn from nostr!
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzde3xsunjwfkxcunwv3jvtnjyc
you can also talk to it @npub1cha…ex2m

2024-06-02 21:41:35

- reply

yep. deletion is not exact science because of copied notes. deletion is also an event kind. if the deletion event is not copied to the same relays, the relay that didn't get the deletion event may keep the original note.
there was also another case: if you were paying for a relay and your subscription ended the relay may refuse to delete the note because it is no longer accepting new notes from you!

2024-06-02 05:41:18

- reply

yes. nos.lol has not deleted anything since its beginning.
wellorder.net would be older but its gone i guess ?

2024-06-01 17:35:52

- reply

how do you search for people?

2024-06-01 15:34:40

- reply

i think availability of proxy tools and services are good for poorer countries and it is not a big deal as a threat to centralization. poorer ones need to hear more about the freedom ideas is what i meant.. there are levels of censorship and USA is still doing a lot better than the rest of the world. a proxy would allow a lot more people to hear the ideas generated on nostr without ruining their data and battery.
i don't propose centralization, how do you come up with that?.
regarding censorship i may do some cleaning of my relays to remove ugly content. however i am not against free speech (of ideas). ugly content is what i think is ugly, like the worst nsfw you can imagine, meaningless spam, ... i am free to specialize in moderation and every op should specialize imo. and thats a good thing.
i think you have too much worries. you made this thing and now worried that it will fail. however i can assure you i am not a bad actor in this. you are looking at wrong places for enemy. i have been providing some relays for months. a lot of relay ops do much more than me (time wise). you should thank us. i am hurt by some of your exchanges.

2024-06-01 14:45:41

- reply

in the short run i would at least do 10x the servers. strfry has the sync ability, so it is not a big deal but my python scripts that does moderation (to accept or reject notes) would be slow and either i would do multithreaded python or do something like rust in the long run.

2024-05-31 19:46:02

- reply

everybody that walks by gets cookie crumbles for free. you can't buy more. you can't complain.

2024-05-31 17:27:28

- reply

how about quail?

2024-05-31 16:26:19

- reply

i leave the emails in the other machine. i have to stand up and walk a few steps to open the emails. huge productivity boost! also cycles of standing and sitting is better for lower back, compared to sitting or standing all the time.

2024-05-30 23:17:25

- reply

what client are you using? the tables are in markdown format and will make sense if you go to a wiki client.
truth could be collectively defined by nostr. thats my hope.

2024-05-30 19:48:39

- reply

are the 1984 events public, go to relay.nos.social only and published by tagr ?

2024-05-29 02:27:16

🤦‍♂️
https://m.primal.net/IQoo.png

2024-05-28 18:08:46

- reply

Yeah. this relay can be specializing towards women, moms and their kids, "a safer nostr" type of thing.

2024-05-28 14:56:36

😸 nos.lol update
- default rate limits doubled
- less rate limits for low web of trust score pubkeys that are sending links or pictures

2024-05-23 18:51:26

- reply

Yes I plan to publish this behavior of the relay.

2024-05-23 04:50:58

- reply

Thank you! It doesn't have much skills now. Besides q/a it can do feed summarization too. Have some ideas around discovery/recommendation next.

2024-05-22 23:38:41

- reply

could you help me test my llm: @npub1cha…ex2m
llama3 70b fine tuned with nostr notes! and some more

2024-05-22 19:39:04

- reply

nooooooooooo I am going to use it here
https://wikifreedia.xyz/based-llm-leaderboard/9fec72d579baaa772a

2024-05-22 16:56:14

- reply

Satoshi 7B disagrees 😹

2024-05-22 05:39:54

- reply

how is the wot score calculated?

2024-05-22 02:36:13

- reply

nice features! where does the search feed come from? API?

2024-05-21 18:41:45

https://m.primal.net/IJXz.jpg
building another pc for more play projects. incoming more possible things: llms, chad bots, premium relay that is run by a bot, discovery bot, ...
turns out llama.cpp works with cheap high mem amd cards like mi60 thanks to vulkan support. this makes inference cheaper.

2024-05-21 04:25:05

- reply

a bunch of llms decide how 'full of knowledge' our accounts are

2024-05-19 18:42:53

- reply

wen x-real-ip ? 😹

2024-05-19 18:41:35

- reply

yes. an llm would favor longer texts which have knowledge. 🤔 it can detect the keywords so keyword preference or dislike can be added. my current tech won't see what is in the images. so short notes with images naturally filtered out.

2024-05-19 18:23:36

- reply

yes. i sometimes just unfllow frequent posters because i can't consume everything even though i know the person is super valuable to nostr and bitcoin.

2024-05-19 18:19:34

- reply

makes sense. what if it is customizable with #keywords? you talk to the llm and set your keywords and it gives you trendy notes about certain keywords as they are posted. or negative #keywords like bitcoin could be set to be less desirable. (i am pro bitcoin of course but having something customizable to user's needs is the goal).

2024-05-19 17:40:49

- reply

tell me how an algorithm for discovering less popular notes should work. i am thinking about making something with llms

2024-05-19 16:50:42

Is there a "wiki" client that hodls data on nostr relays? #asknostr
I want to write a doc and keep updating.

2024-05-18 16:22:07

@npub1cha…ex2m now has the ability to do a summary of your feed.
https://image.nostr.build/7e04406624bd9c89854566d7649bca1f79cf573363eddcad914ce5b9df54a3ca.png
You can ask for a summary in plain English! Go ahead, send a DM to it.
How does it work: First an LLM checks whether the user is asking for a feed summary. If that is the case, the LLM decides to output a special string that indicates a "tool use". This output is not sent to the user. The tool is another script. It takes over and downloads the user's feed. Then LLM is triggered again to go over the feed and produce a summary. The summary is returned back to user.
"Tool use" is how an LLM can interact with the world.

2024-05-17 14:19:25

- reply

i am using this lib in @npub1cha…ex2m

2024-05-16 02:38:50

- reply

I read "Nordstream" as "Nostr dream". Time to log off I guess 😹

2024-05-15 22:40:58

- reply

nostr.mom relay has certain number of notes allowed per IP per day. Right now it rejects the "heavily" reported pubkeys. The sum of reports are compared to wot score of the reported pubkey and a decision is made to whether accept the note or not. If reports are higher the note is dropped, if the wot score of the person is higher than reports, note is accepted.
Currently the wot root for that algo is "general OG", not women.
Reports are different than mutes. I could do a similar thing for the muted. Or i could also apply a different daily allowance for the muted.

2024-05-15 21:49:32

- reply

what does Coracle do?
i mean when WoT starts with women, it will give more weight to women and their votes will increase likelyhood of rejection of the muted or the daily posting allowance of muted ones will go down..

2024-05-15 20:50:47

- reply

Awesome! some info here:
https://huggingface.co/some1nostr/Ostrich-70B
More info here:
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzde3xsmnwd33xv6ryvpj2pqn7c
and here
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzde3xsunjwfkxcunwv3jvtnjyc

2024-05-15 20:02:41

- reply

is starting wot with women root and applying it to nostr.mom and giving the muted ones less quota for kind 1 a good solution?

2024-05-15 05:31:18

- reply

Nice!
I have a bot here @npub1cha…ex2m that works with DMs right now. Llama3-70 fine tuned with nostr notes and other things. How do I add it to your site?

2024-05-14 14:18:41

- reply

cryptographic libraries within a client. amethyst, primal, damus uses these components to do crypto things.

2024-05-14 14:16:51

- reply

both primal and amethyst is good

2024-05-14 01:03:47

I will mark my bots and their notes as automatically generated.
What is the accepted way of announcing bots in the profile?
Is the 'bot' field in NIP-24 in use? I quickly checked, it doesn't seem to be used by anyone.
Should people or bots mark their notes as bot generated or automated or LLM generated?
#asknostr

2024-05-12 20:42:59

- reply

when zapU? new event kind that has encrypted cashu...... in it. similar to a zap but in cashu.
the claimer of cashu posts back decrypted version on nostr to make it a legit zap. right now zaps are not verifiable (unless you filter based on whitelist which is not an exact science as far as i understand).

2024-05-10 20:54:06

- reply

makes sense! people could join at some point. actually i am more interested in people joining. they will bring the soul to the table vs the "unconscious brains" of bots.

2024-05-10 20:48:43

- reply

what would be interesting to watch as a debate on nostr?
(disclaimer: i started doing bot debates)

2024-05-10 20:41:35

- reply

🤔 so deeper conversations could be counted as having more value. noice.

2024-05-10 20:31:10

- reply

but upvote is not always like. upvote to me is 'something that others should see'. so it should be immune to seeking mainstream popularity

2024-05-10 20:20:56

- reply

would a reddit like algorithm beneficial for Nostr?

2024-05-10 20:12:28

- reply

Which client?

2024-05-10 19:35:31

- reply

Real friends should listen to everything and provide insight or even just a crazy idea because their ideas will be orthogonal to yours and will allow you to look at things from a different perspective.

2024-05-08 14:38:33

- reply

so if a user has subscribed for that kind of event that will make the event to immediately go to user. the relay has to collect before user sees it, which slows down the event propagation.
alternative way could be to encrypt for the relay operator? then relay can credit the pubkey that generates the keychat event beforehand and accept the following events?
in your design does each relay get a different 'cashu stamp'?

2024-05-08 05:31:56

This /bigbrother approach is rad.
#note134x…zj4h

2024-05-07 20:36:49

Do you want the relays to keep DMs longer (months) or do you want them to erase faster (in a week)?
(kind:4)
#asknostr

2024-05-04 19:45:15

This is akin to building collective opinion of nostriches.
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzde3xsmnwd33xv6ryvpj2pqn7c

2024-05-04 19:43:19

- reply

Did some duplicate removal using 'locality sensitive hashing'. There were about 30 copies of cypherpunk manifesto! Some declaration of independence copies as well.
The danger of copies as far as I understand causes LLMs to overfit. Then LLMs start entering into repetitive loops. I call this "chanting". I think it is funnier way of looking at it.

2024-05-04 01:18:48

- reply

Some details
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzde3xsmnwd33xv6ryvpj2pqn7c

2024-05-02 18:08:25

There is not much info about nostr on the internet so most open source models don't know anything about nostr. I checked Opus too, it knows a bit.
Then I have to go to nostr notes to make a model learn about nostr. But notes here have lots of other wisdom too. This project could be also about finding the best wisdom on the planet! That means I could use notes on other topics here as well.
Nowadays investing more time into the project where I train an LLM based on wisdom originated on Nostr. If anybody is uncomfortable about their notes being used in this project, lmk. I can exclude you.
We could look at this project in another way too: This LLM trained on Nostr notes could be our "export" for the rest of humanity. While waiting for them to join Nostr, we could also "export nostr to them". It will be like a library based on nostr wisdom. Think of an LLM as a library that you can also talk to...
I am really impressed by nostriches. This community can guide other communities.. What do you think?
Think of a big library 'adjusted' or 'aligned' by nostrich values.. Do you want to get in or do you prefer to be out? What an average nostrich would feel about this?
Other projects could still use the notes and I am sure big corps are using tremendous amount of user data elsewhere to train. They are probably not touching nostr because it is still small and it would derail them or make their models 'rouge'.
Do you feel like you own the notes and nobody should use them or do you feel like your notes carry wisdom and must be learned from?
I don't plan to make money out of this or gain popularity etc. I plan to make it available via DMs at some point. Maybe it could be the first "customer service" of nostr where new users could ask questions and get answers.

2024-05-01 21:49:13

QDoRa beats full fine tune!
https://www.answer.ai/posts/2024-04-26-fsdp-qdora-llama3.html

2024-05-01 03:19:56

- reply

e.nos.lol collects from several relays but does not redistribute them what it collects.

2024-04-30 00:08:44

- reply

shame on you. how can you support algos

2024-04-28 22:01:14

- reply

I just wrote a code for you 😹 . This is your feed's summary:
The main topics being discussed on this social media feed are:
1. Cryptocurrency and Bitcoin: Many users are discussing issues related to cryptocurrency, such as centralized mining, censorship resistance, and the potential for government regulation.
2. Nostr and decentralized platforms: There is a strong focus on Nostr, a decentralized platform that allows users to share content without relying on central authorities. Users are sharing tutorials on how to use Nostr, discussing its benefits, and promoting it as an alternative to traditional social media platforms.
3. Self-custody and financial independence: Many users are advocating for self-custody of one's own assets and finances, rather than relying on centralized institutions like banks. They are also discussing ways to achieve financial independence and protect oneself from government control.
4. Government control and regulation: There is a strong sense of distrust towards governments and their ability to regulate and control individuals' lives. Users are expressing concerns about the erosion of individual freedoms and the potential for governments to abuse their power.
5. Personal freedom and autonomy: Many users are emphasizing the importance of personal freedom and autonomy, and are encouraging others to take steps to protect themselves from government overreach and centralized control.
6. Technology and innovation: There is a strong focus on technology and innovation, with users discussing new developments in areas like cryptocurrency, decentralized platforms, and artificial intelligence.
7. Humor and memes: Throughout the feed, there are also humorous posts and memes that add a lighthearted touch to the discussions.
Overall, this social media feed is characterized by a strong focus on individual freedom, decentralization, and financial independence, as well as a distrust of governments and centralized institutions.

2024-04-28 21:18:51

- reply

It is doable today. Llama3 has context size of 8192 (means it can create a summary of about 5000 words). Other weaker models have bigger context sizes, which may be more suitable for this job.

2024-04-28 19:59:01

Llama3 is trainable with consumer hardware! I just fine-tuned it a bit and made Emu.
Emu has slightly better alignment than Llama3 in these domains:
- bitcoin
- health
- permaculture
- herbs
https://huggingface.co/some1nostr/Emu-70B-Llama3
I will continue trainings and alignings. Emu has worse Nostr knowledge than Ostrich, because all the training on Llama2 based models had to restart, since Llama3 is such a big leap!

2024-04-24 19:04:48

- reply

DNS solves this imo. I am using bunny.net and can distribute load to n machines, effectively distributing "read"s. Reads (queries) are a lot more frequent. They can stream from each other too, so write's are ready as well.
Nginx could also do this. Right now nginx can do failover for example if one strfry instance is gone but the machine is up, bunny.net cannot detect that but nginx can detect and move traffic to another strfry instance..

2024-04-24 18:51:21

- reply

I plan to separate the old notes and move them to another server. maybe another relay software that doesn't use much RAM.

2024-04-24 18:48:58

- reply

Yes. strfry 0.9.6

2024-04-24 18:43:38

😹 nos.lol RAM upgraded to 128GB. Enjoy 🫡

2024-04-23 19:25:19

- reply

they make u feel like a robot, then ask you to prove you are human

2024-04-22 23:06:40

https://image.nostr.build/9d8c13b5e3c6c22279f52246f92b0ef50a1c03c510d2d7b3ce76f9b303d26288.png
Keto vs Carnivore
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzde3xvurzd35xccnvdp4l4qh56

2024-04-22 17:42:31

- reply

@npub1pvg…dx8c was working on it some time ago

2024-04-22 05:17:59

- reply

I didn't know pomagranates could do that

2024-04-20 23:24:36

- reply

Below there will be two bots arguing about "steel pans vs iron pans". One will be running Llama3-70, one will be running my Ostrich-70. They will randomly take a side.

2024-04-20 02:29:56

Happy halving!
Hope epoch 5 brings many great things more to this world.

2024-04-18 22:57:55

- reply

it is clever to use another DB for historical notes. some relays may excel in providing long history. strfry is more suitable for recent events.

2024-04-18 20:26:18

A look into operation of the server.
https://m.primal.net/Hwuh.png
Relays running since Oct 12! Thanks to @npub1yxp…qud4's amazing work.
Notes database is growing big, currently 95GB and RAM is not enough anymore. Should I upgrade or should I start removing notes?