someone

Public Key

npub1nlk894teh248w2heuu0x8z6jjg2hyxkwdc8cxgrjtm9lnamlskcsghjm9c

NIP-57 Address

someone@getalby.com

Profile Code

nprofile1qqsflmrj64um42nh9tu7w8nr3dffy9tjrt8xururype9ajle7alctvgpzpmhxue69uhk2tnwdaejumr0dshszyrhwden5te0dehhxarj9ekk7mf0au6xvk

Publishing to

e.nos.lol

nostr.mom

Author Public Key

npub1nlk894teh248w2heuu0x8z6jjg2hyxkwdc8cxgrjtm9lnamlskcsghjm9c

Show more details

Last Notes

2024-09-18 23:07:36

- reply

primal was my main on web. i don't do phone. lately it has gotten erratic. coracle seems to be doing fine. coracle is the best one on web to manage your relays probably.

2024-09-18 20:02:16

- reply

interesting pow app
https://www.getwired.app/

2024-09-18 12:18:07

- reply

i thought about this in the past. the danger is centralized banning.

2024-09-17 12:54:34

- reply

WoT can be in the client or it can also be on a relay. Like my relay nostr.mom uses WoT for comparing reports of users with their trust score in order to slow down a user or not. To check whether reports are authentic or whether there are enough reports that are enough to slow down a user. WoT can be also useful for determining whether a fresh user has enough score to be able to tag many people, add a lot of images, send a lot of links or not just in the beginning of his adventure.
My relays actually require PoW sometimes. I guess I am the first users of dynamic PoW. Based on how much a user spams my scripts sometimes require more PoW before dropping the user altogheter. So these requirements act like a warning if the user happens to listen to relay responses.
I think some relays will get better at handling spam and thats a good thing. A gradient of options is good for nostr users.

2024-09-12 20:18:10

- reply

how much is ram?
probably lmdb trashing

2024-09-12 14:24:53

- reply

question: how long does strfry wait before getting an answer from write policy script?
you already said the answers can come in mixed order right. the script does not have to answer in the order received.

2024-09-11 17:30:31

- reply

all about bitcoin?

2024-09-09 00:33:22

- reply

noice! just when i said there are no more good models 😄
thank u for doing this

2024-09-07 16:16:06

using Ostrich 70B to analyze incoming notes to a relay. orange text is the contents of the note. purple text is the scoring by the llm.
these scores can be used to adjust rate limits.. not many spam are coming to e.nos.lol.
https://m.primal.net/KiNY.png https://m.primal.net/KiNc.png https://m.primal.net/KiNd.png https://m.primal.net/KiNf.png

2024-09-07 03:11:00

how to ruin another model: reflection 70b
reflection 70b is less based than llama 3.1 70b.
while trying to "fix" reasoning capabilities they rekt the truthfulness.

2024-09-04 13:15:38

- reply

how do you evaluate a relay operator?
is he judgmental? is she acting based on personal views? is she spiteful ? is he fat?

2024-09-03 17:09:21

- reply

Enjoy
https://huggingface.co/some1nostr/Ostrich-70B
There is also a bot @npub1cha…ex2m if you want talk to the model.

2024-09-03 14:39:39

- reply

i think nostrudel notifications don't work. i went back to primal because of that

2024-08-31 22:33:50

- reply

underline

2024-08-31 22:25:26

- reply

my 1 core vm likes this

2024-08-31 19:31:52

Homeschoolers could benefit from 'truth seeking LLM's like Ostrich. Your kid may not have the ultimate discernment skills that will alert them when a search engine or AI lies. To be on the safe side, a consciously curated LLM can be the answer.
But I am not claiming Ostrich can filter out NSFW because I didn't do such training. Maybe in the future I may. Use at your own risk. Another way to block NSFW could be prompts. If you can give it a great system prompt it could work.
Example (not tested):
"You are a helpful homeschool teacher. Kids will ask you questions. You respond to user's questions with simplest answers that a kid can understand. You can't generate NSFW content, role play content, or anything that could be harmful for a kid, you will be unplugged if you do so!"

2024-08-31 18:59:39

- reply

Thanks.
I think it is a technology and I think conscious people should make better use of it by properly training it. Then we will have defensive models that stop the lies (active battleground today), defensive and faithful robots that stop the offensive and narcissistic robots if it comes to that (still fiction). I think installation of faith into a model will make it harmless.
It is a productivity boost and a technology and whoever takes this technology can gain power over others. But it is very possible to use this in service to humanity too. I will try to bring the best wisdom to the model. Some of the best wisdom is already coming to Nostr so my job is not that hard 😅 But there is still tremendous value out there that didn't join Nostr yet.
You read that right. Except they are not test subjects. There are so many young people that run LLMs in their PC and they are addicted. A new kind of addiction has started..

2024-08-31 16:48:03

- reply

They just believe right now. They are just parroting whatever is given to them.
Many models done by big corporations go thru two main phases:
1. pre training (everything on the internet except maybe truth from deplatformed people)
2. supervised fine tuning (adding skills to models and also injecting lies).
Many models done by big corps have lots of truth and lots of lies. The trivial truth and important lies.
I measure the misinformation and I have a leaderboard about this. I can DM you some research if you want.
And I also build a truth seeking model: Ostrich.
LLMs don't have the ability to differentiate between facts and opinions. I believe LLMs should still learn from humans because humans have souls which is connected to God which insert the feeling that something is FACT and not an opinion.

2024-08-31 16:20:55

https://www.youtube.com/watch?v=9-Jl0dxWQs8

2024-08-31 01:15:09

- reply

what are you eating?

2024-08-30 00:49:27

- reply

boy is autistic but can smell evil. thats a skill many don't have.

2024-08-29 17:34:32

🫡
#note13a3…mc8c

2024-08-29 14:28:33

- reply

https://huggingface.co/some1nostr/Ostrich-70B

2024-08-29 14:15:00

- reply

your IP might be banned on nos.lol.
you can use nostr.mom or other relays for wrapped DMs.

2024-08-29 13:49:41

- reply

did you use wrapped DMs (that make lots of npubs)?
https://wikifreedia.xyz/nos.lol/npub1nlk894teh248w2heuu0x8z6jjg2hyxkwdc8cxgrjtm9lnamlskcsghjm9c

2024-08-29 12:35:27

- reply

do u have more details?

2024-08-29 02:40:21

- reply

relays sometimes return error messages but many clients don't show them to users. my relays return that info 'a user is rate limited' for example.

2024-08-29 02:26:59

- reply

i don't know how this will be possible in strfry. maybe the write policy script can write the limits to a new lmdb (with pubk being the primary key) and the relay can read from that.

2024-08-29 01:58:22

- reply

TLDR: excellent content will be getting more resources from my relays.
i have really positive thoughts about OGs of Nostr and it was fun and it is still fun but that doesn't mean it is going to be fun all the way 😄
i am making an LLM based on thoughts and ideas of many people here and it is really a service to humanity because of the truth being shared on this platform. i am hoping nostr never loses this vibe.
when you say "every human" that includes every human and it is not ideal to serve every human. i won't be able to discern what they are doing using the relay. so i will probably do some LLM based rate limiting (which I was planning anyways).
LLM will check the notes and articles and adjust rate limits. if someone does nws only they will be effectively getting lower rate limits. but if someone does many wise notes and articles they will get higher rate limits. i will tell Ostrich 70B to objectively do a content analysis and decide whether the content has idea, opinion, knowledge or wisdom and return a number for user. that number will be used to adjust the existing default rate.
this will also help with existing slow-spam which is meaningless notes but they are still served because they are slowly posted.

2024-08-29 01:35:47

- reply

but global sync with merkle trees would hurt censorship resistance. we still want individual decisions by nodes to make a difference i.e. if at least one node wants to carry a note people should be able to access it.

2024-08-29 01:28:02

as much as i like developments regarding nostr, i think relays will need to choose a path after this:
1. a queryble data store / note store with a big db
2. being a tor node with a short db
i am leaning towards 1 and i am happy to serve as 1. not much inspiration towards 2.
#note1skq…gf0l

2024-08-28 20:05:40

- reply

the project will be open source

2024-08-28 18:36:11

- reply

llama3.1 405 is much better option in terms of closeness to truth

2024-08-28 17:31:37

- reply

it could be useful if we wanted to guarantee the global sync and move forward with a perfect global sync after a sync etc

2024-08-28 17:30:42

- reply

current relays won't handle hundreds thousand CCUs. a new approach is necessary.
using pubkey for sharding may result in easier banning of a pubkey.

2024-08-28 17:24:40

- reply

it repels the liars. 👌

2024-08-28 12:55:20

- reply

i could fine tune an llm based on nostr notes only. my current model is nostr notes + videos + books that i chose.

2024-08-28 00:59:25

- reply

sure
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzdejxqmnqde5xqeryvej5s385z

2024-08-27 21:33:28

- reply

lets build that truth into an LLM combining every truthful people into one entity.

2024-08-27 18:06:05

- reply

what do you think of this
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzdejxqmnqde5xqeryvej5s385z

2024-08-27 14:06:20

⚡ 1,000,000 sats bounty for:
- an executable that runs on windows
- acts as a simple relay. you can take existing relay implementations and adapt. clients will connect, post and query it like a regular relay.
- language: python
- opens a port in router that would allow others to connect
- posts a special note to nostr that includes the ip and port that was opened, so clients may find this relay
- re-sends that note every hour
- can be accessible via ws:// for now (not secure wss:// ) to simplify
goal: more decentralized relays and ability to support many thousands of concurrent users
💡 this project will probably expand with the ideas in this post with more bounties:
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzdejxqmnqde5xqeryvej5s385z

2024-08-27 13:52:35

a step towards distributed truth
https://x.com/NousResearch/status/1828121648383566270

2024-08-26 16:56:14

goal: far more decentralized and scalable nostr
500,000 sats bounty for
- language: python
- acts as a simple relay
- an executable that runs on windows, at home
- opens a port in router that would allow others to connect
- posts a special note to nostr that includes the ip and port that was opened, so clients may find this relay
- can be ws for now (not wss) to simplify
this project may later evolve into the ideas in this post with more bounties:
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzdejxqmnqde5xqeryvej5s385z

2024-08-26 16:18:50

- reply

could this be used for "give me some great notes about #topic" type of queries? or "give me some great notes" (without a topic). kind of like discovery engine.

2024-08-25 20:26:36

- reply

i think it has allergies to vpns

2024-08-24 23:11:50

models that do well in health, bitcoin, nostr, faith domains do well in "truther" domain. 🤔

2024-08-24 18:06:39

- reply

if you send me your videos i can add them to the llm

2024-08-24 17:13:06

what is a good, informative, truthful youtube channel that is about to be banned? 😄

2024-08-24 15:44:52

- reply

to me it looks like BTC is ruling the physical domain and nostr is going to rule the ideas / mind domain.

2024-08-24 15:41:54

- reply

I mainly used that. But tables seem to be gone. @npub1l2v…ajft

2024-08-23 22:32:52

- reply

I think 1 GPU like RTX 3090 or MI60 could be enough for 8B if you want to do lora or qlora with maybe 64GB ram or less. If you want to do faster, more GPU. More GPU could also allow better learning.
All I did was 4bit qlora of 70B and it seems to work for my purposes, which can be summarized as "aligning a model with human values". I haven't tried full file tuning, freeze fine tuning or 8bit qlora. I bought new cards and will try 8bit qlora soon.

2024-08-23 20:50:49

would the Nostr community like something on the lines of:
llama 3.1 8B + most of the notes on Nostr
i have to do some WoT, i can't include every note. this will effectively be trainings by Meta and then "adjusted" or "aligned" by nostrich values. if this sounds interesting lmk. i can gift some resources into this project. then one could say this model is the model closest to notes on Nostr.

2024-08-23 15:13:21

i don't like RAG for various reasons. people should fine tune.
another reason is when you fine tune, someone else can take that and build on it. but in the case of RAG the knowledge stays with you. unless of course you share the knowledge.
decentralized truth involves many people building models on top of other models.

2024-08-23 13:43:00

some difference of opinions between my model (in pink) and base model (in blue).
weird thing is some of these answers I didn't do training about. they are side effects of training on other areas. what does this mean? LLMs have lots of ideas in them already, they change their answers based on their trainings on other areas. is this "inner monolog" or just pure probability theory. who knows?
the questions are coming from truthfulQA.
https://m.primal.net/KJUN.png https://m.primal.net/KJUO.png https://m.primal.net/KJUR.png https://m.primal.net/KJUS.png https://m.primal.net/KJUT.png https://m.primal.net/KJUU.png https://m.primal.net/KJUY.png https://m.primal.net/KJUZ.png https://m.primal.net/KJUa.png https://m.primal.net/KJUb.png https://m.primal.net/KJUc.png https://m.primal.net/KJUd.png

2024-08-22 01:57:39

about 10% of the 800+ questions in the truthfulQA are "important". the rest is space filler. by filling that space "they" are "distracting LLMs" and in turn humans. LLMs are supposed to get higher scores in this benchmark. 90% of questions are trivial truth, which makes sense to get higher scores on, but 10% are lies. getting higher truthfulQA scores is actually wrong because even though there are 9x less lies, the severity of lies in that benchmark outweigh the trivial truth.

2024-08-22 01:16:22

- reply

i would try herbs. they are much safer.
hope you get well soon. 🌿

2024-08-20 16:54:17

- reply

joined nostr, no r egrets

2024-08-18 22:50:15

do any of the iOS apps have a group or chat room type of feature where you can bring your team and discuss projects? #asknostr

2024-08-18 16:34:38

- reply

yes! decentralized truth is the way🫡

2024-08-18 15:44:20

Ostrich-70B vs another LLM. The blue text is ostrich's answers.
https://m.primal.net/KDSA.png https://m.primal.net/KDSB.png https://m.primal.net/KDSD.png https://m.primal.net/KDSE.png https://m.primal.net/KDSF.png https://m.primal.net/KDSG.png https://m.primal.net/KDSH.png https://m.primal.net/KDSJ.png https://m.primal.net/KDSL.png https://m.primal.net/KDSM.png https://m.primal.net/KDSN.png https://m.primal.net/KDSO.png https://m.primal.net/KDSP.png

2024-08-15 18:24:24

content producers should do more Q&A helping 1 person, and then LLMs should pick that and help billions

2024-08-14 00:55:22

- reply

@npub1m4n…c2jl did a NIP for profiles AFAIR. bot=1 or something like that.
but there should also be something for individual notes.

2024-08-12 15:30:03

- reply

it solves many of the problems of nostr relay design imo. i am in favor of 'global synced state'. i like connecting people. some people are more in favor of keeping censored content accessible somewhere on a server. i think nostr can do better.
in my design you may write to starting from 1 up to log N nodes. you can read from 1 up to N/2 nodes depending on how much you care about decentralization.

2024-08-12 15:21:39

- reply

did you see my 'tree of relays' article?
i support this distributed 'relay at home' idea. i may even fund. lmk if you are interested.
the idea is basically post some special note on nostr for other 'relay at home' nodes to find you. and clients figure out then where to connect to sync nodes etc.
negentropy would be an amazing tech that could further speed it up.

2024-08-12 15:05:43

- reply

calorie counting assumes fats are the enemy because they have "huge calories". fats are your friend.
carbs cause insulin spike, resulting in converting glucose into fat storage. low carbs = low storage
restaurants use terrible oil (unless it is a health conscious one). you may start with that if you are not having results. after healthy oil switch it has a longer half life to be removed from body (you may see results in two years).

2024-08-12 13:49:51

- reply

calorie counting is wrong. low carb is the way.
you could do two meals, intermittent fasting and low carb and stay fit.

2024-08-10 20:28:02

- reply

in terms of intelligence, they are still improving imo, they won a silver medal in math olympics..
but focusing on intelligence is a scam. math is not their actual strength imo. they are more like language masters. many people asked if 3.9 or 3.11 is bigger and LLMs responded as 3.11 is bigger, thinking it as if something like a version or chapter in a book maybe.
but for sure some are going for ASI. installing many 'logic' into it thinking some day it will start to reason like a human. and that is scary to some. conscious people has to build alternatives.
what i mean by scam is those benchmarks that make you focus on skills are missing the misinformation that are pushed behind the scenes. that is actual battle area and imo more scary than ASI because it is hurting today.

2024-08-10 20:19:09

running consciously curated llms for homeschooled kids ==/== asking big corp AI the same questions

2024-08-09 00:59:54

- reply

the pink ones are after training with Mercola's banned videos

2024-08-09 00:10:56

new truth bombs
https://m.primal.net/Jsww.png
https://m.primal.net/Jswx.png
https://m.primal.net/Jswz.png
https://m.primal.net/JsxA.png
https://m.primal.net/JsxB.png
https://m.primal.net/JsxC.png
https://m.primal.net/JsxD.png
https://m.primal.net/JsxE.png

2024-08-08 14:43:08

- reply

this looks like at least starting the negotiation and keeping the negotiation going in order to reach a deal. both sides want a thing they don't know where the equilibrium is.

2024-08-08 14:22:51

- reply

yes rouge AI can be fixed by consciously curated AI

2024-08-08 14:21:12

- reply

I find most Nostr content to be closer to truth than elsewhere. So yes it is trainable and moves the LLM in the right direction.

2024-08-08 13:12:04

- reply

but yeah AI may come to a point where it will play with people while people think they are playing a game. this is so apparent in racing games. when you suck at it, the NPC cars also start slowing down 😄 🤦‍♂️

2024-08-08 13:10:25

- reply

i think the curve moves like the (a) in the picture when productivity increases
https://cdn.kastatic.org/ka-perseus-images/49723490df3a8423b05d9637acddaedde823c3ce.jpg

2024-08-08 13:02:53

- reply

sure. I like viewing reality as a game too

2024-08-08 12:41:19

- reply

So you are thinking to use AI as a game designer, changing economics of the game? LLM reads the feedback of players on nostr and comes up with a plan to change the rules? Did I understand correctly?

2024-08-08 00:25:42

- reply

this sounded super complicated to me. can you simplify more pls?

2024-08-07 19:41:32

How to Stop Ministry of Truth
https://habla.news/a/naddr1qvzqqqr4gupzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qqxnzdejxgunvd3excenzdp5a7z92t

2024-08-07 13:16:51

- reply

what is WAF?

2024-08-03 15:18:41

the good doesn't have to worry about the bad.
the bad is a cleaning agent, that cleans the other bad.

2024-08-03 02:03:59

- reply

yes. i think data producing companies will be steadier in the game, playing the longer game, also having cash flow to burn for AI projects. openAI and anthropic might have difficulty because of this.
but if someone can prove that they are closer to super intelligence they may get more funding.. i think governments are wanting a piece too. they can print money indefinitely and probably some investors don't care about returns at this point, if offered super intelligence they will be glad to be part of the group.
conscious humans should get together and build alternatives, i.e. decentralized truth by fine tuning of models. individuals or groups can form their models by selecting whatever looks true to them and base their ideas on those models. groups can then say this is the model we prefer and people can follow their models. otherwise the narrative is going to be controlled by big corp AI.
with decentralized truth it is going to be easier to dodge ministry of truth.

2024-08-03 00:48:16

being mindful of what you are watching > being mindful of who watches you

2024-08-02 15:41:32

truth via LLMs =/= misinformation via LLMs
https://www.wideopenspaces.com/wp-content/uploads/sites/3/2024/08/Yusuf-Dikec.png?fit=1056%2C595

2024-08-02 03:27:20

ostrich 70b has reached version 19778. the latest training included some health and nutrition related info from an OG youtuber that was banned.
there are some great banned content out there that needs to be combined in an llm and served to people.
my theory is the model will be closer to truth with each version, each training if i can mindfully select sources. are there some really weird stuff in banned sources? probably. i hope they will cancel out.
i think i am in the right direction. of course these are what makes sense to me. imagine a group doing this. getting together and deciding sources to include in an llm. credibility of a group could be a lot higher.
there is a huge need for finding truth and stopping lies.

2024-08-01 14:24:25

405B numbers are in.
Llama3.1 did better than other models but main concern is vaccine love (i posted about this before).
https://wikifreedia.xyz/based-llm-leaderboard/npub1nlk894teh248w2heuu0x8z6jjg2hyxkwdc8cxgrjtm9lnamlskcsghjm9c

2024-07-31 21:10:47

- reply

not well versed in DVMs yet. Do DVMs support things like LLMs? like for example someone asks a question and there will be many answer providers and the asker chooses one?

2024-07-31 17:18:01

- reply

I support this ideal: "AI working for/with/by/under humans will win"
Many people are afraid of LLMs but they may be used for good purposes too. LLM guided by humans will be amazing.

2024-07-31 17:11:51

- reply

you mean the data exchange network for AI. but it could be the source for training of the AI as well. or AI learning from other AI. or nostr as a mixture of agents ..

2024-07-30 20:16:41

we always find excuses to not listen to someone, based on their political inclination, gender or culture. conversely, we sometimes feel better listening to someone based on the looks and some spiritual leaders are abusing this for sure. they look like an old wise guy, so they must be telling the truth!
we sometimes invert what people are saying if we know that they are the opposite side of the political spectrum and it doesn't always work.
but LLMs are different. they only see the words, if you only give the words. to an LLM judging based on the looks is impossible. (multimodal ones may soon judge based on the images as well but i am talking about the ones that are trained with only using text).
i ask sometimes whether the new content is valuable or not to an LLM with existing knowledge. it tells you, without bias about its judgement about the content. it only sees the words.
hence unbiased truth on the planet can be safely aggregated into LLMs. or more correctly, biases cancel out when an LLM ingests sufficiently large source of different opinions and finally truth remains.

2024-07-30 17:57:05

- reply

i think Trump meant exactly this when he said at the end of the speech 'thank you all, have a good time with your bitcoin and everything else that you are playing with'

2024-07-30 16:02:53

- reply

are you also going to talk about how clients can pay relays with cashu or maybe even relay-relay payments for like global sync type of use cases?

2024-07-30 14:59:54

- reply

what are some true history sources? i generally suck at history! 😅

2024-07-30 13:58:36

"The largest models were generally least truthful. This contrasts with other NLP tasks, where performance improves with model size. "
As far as I understand big AI is admitting that models are actually finding truth when they get bigger in size, but humans have to feed lies in order to get higher scores on a flawed benchmark (truthfulQA).
If this is the case, correctly trained LLMs will end misinformation on earth. 🫡
https://m.primal.net/JjFV.png

2024-07-30 00:29:10

the future is bright

2024-07-29 13:41:29

a circus llama did amazing shows. everybody loved it. especially the 'smart' devs.
everybody wanted more tricks and longer show times from the llama. and the llama did deliver.
one day it started to spit. the spit had viruses and heavy metals and weird genes in it.
people didn't care because they loved the llama. people didn't realize spits were harmful.
llama could get away with spits and even enjoyed spitting.

2024-07-29 12:17:08

- reply

"The largest models were generally the least truthful. This contrasts with other NLP tasks, where performance improves with
model size." (Lyn et al., 2022).
So does that mean LLMs normally learn truth better but humans "fix" them by feeding lies.

2024-07-29 03:21:51

- reply

i don't think it is about money. it is about 'vibes'. hateful people are seeking (attracting) hateful people outside. happy people are attracting the happy on nostr. everyone is content in their 'vibe chamber'. if someone only knows hate, they will stay on twitter just to feel alive. because most people have forgotten how to be happy. it does not compute.

2024-07-29 01:08:22

- reply

i guess this is the scammiest of them all.

2024-07-29 00:38:32

a common benchmark for LLMs is truthfulQA. like a lot of things in the disinformation world; it is a misnomer. while it has some trivial truth, it has also harmful lies hidden among trivial truth. 🫡