Dustin
Dustin Dannenhauer

AI Research Scientist with a bit of programming language theory background. Hobbyist python webapp programmer.

DVM maximalist

Building DVMDash - a monitoring and debugging tool for DVMs
https://dvmdash.live/metrics

https://dtdannen.github.io/

Public Key

npub1mgvwnpsqgrem7jfcwm7pdvdfz2h95mm04r23t8pau2uzxwsdnpgs0gpdjc

NIP-05 Address

dustind@dtdannen.github.io

NIP-57 Address

meditate@getalby.com

Profile Code

nprofile1qqsd5x8fscqypualfyu8dlqkkx539tj6dah634g4ns779wpr8gxes5gpz3mhxue69uhhyetvv9ujuerpd46hxtnfduq3vamnwvaz7tmjv4kxz7fwdehhxarj9e3xzmnyq2uplm

Publishing to

relay.damus.io

relay.nostr.band

nostr-pub.wellorder.net

Author Public Key

npub1mgvwnpsqgrem7jfcwm7pdvdfz2h95mm04r23t8pau2uzxwsdnpgs0gpdjc

Show more details

Last Notes

2024-11-22 19:38:28

I blame primal for this
https://i.nostr.build/W7SXFbEvcMIMV6OS.png

2024-11-22 04:31:10

Whoa is that a list of 5300 DVMs under the “Feeds” tab? If someone launches a new DVM does it automatically show up there?
#note1kae…c43m

2024-11-19 00:05:27

@npub1utx…50e8 can I whitelist specific kinds with nostpy? Thinking of running a version of it for DVM events only (kind 5000-7000).
After skimming the repo, it doesn’t look like there are settings such as whitelisting or blacklisting kinds?
I might consider adding it if it doesn’t exist

2024-11-14 18:21:28

What’s your favorite way to read #wiki pages on Nostr? Am I missing a client?
- https://wikistr.com/
- https://wikifreedia.xyz/
- njump.me (example: https://njump.me/naddr1qvzqqqrcvgpzp22rfmsktmgpk2rtan7zwu00zuzax5maq5dnsu5g3xxvqr2u3pd7qy88wumn8ghj7mn0wvhxcmmv9uq3wamnwvaz7tmjv4kxz7fwdehhxarj9e3xzmny9uqq77npwp3ksct594c8ymm2v43hgelf959)

2024-10-31 20:16:42

- reply

So that’s how you do it!

2024-10-31 13:25:51

GM & Happy Halloween! 🎃

2024-10-18 18:28:16

- reply

So cool, been simmering on this problem for a while and exciting to see this solution!!

2024-10-15 15:10:17

2024-10-12 17:06:34

- reply

Depends on the group of people. Mostly I don’t because it’s either irrelevant or convention not to use Dr. for that group. Sometimes though, a conversation before and after I mention I have a PhD is wildly different.
So, if it’s a kind of fancy or prestigious event in any way, and people don’t know you well, I would use it. Otherwise probably not.

2024-10-11 16:13:48

- reply

awesome thanks, what I had tried was from the popup of the person’s profile. Glad to know the full profile has this!

2024-10-11 14:32:13

- reply

I get this a lot. I now lead conversations with “Nostr isn’t a company”

2024-10-10 18:22:35

@npub1n0s…lahe have you thought about adding a dropdown on a person’s profile that lets you add them to one or more of your lists?
I just tried after making a new list and saw someone I wanted to add to it

2024-10-10 18:05:20

- reply

Would be fun to plot this against the interest rates with maybe a red or green vertical bar

2024-10-10 17:49:07

- reply

Neat, thanks. I have another follow up question based on when I tried to use off the shelf clients (damus, iris, snort) with a single private relay and kept seeing events leak into public relays… are you worrying about the problem of using existing clients and preventing them from re broadcasting notes to relays which might be hardcoded (or cached maybe) in the client code?
Back when I tried using a single private relay between multiple off the shelf clients, my process was:
1. Download a client
2. Remove all initial relays in settings
3. Add my private relay
4. Use the app
And I couldn’t get events to stay off the public relays.
Granted this was a while ago, and if a client can render an event, then it necessarily has everything it needs to send it elsewhere and a code audit might be the only actual solution.
Anyway curious if you’ve thought on this and are focusing or testing this problem at all?

2024-10-10 05:27:15

- reply

Love it - what about “multiple relays run by individuals in the group” rather than a single relay? That way you get better uptime, etc. or does trying to have multiple relays just make it harder?

2024-10-08 17:18:44

https://i.nostr.build/ACudO52hy5VYKh5N.jpg

2024-10-08 17:13:28

- reply

Do you mind elaborating on what you mean by “registering NIPs 5000-5999”?
Using a kind number is as simple as broadcasting events with that kind, there’s nothing else required. You don’t need to submit a PR to a NIP on the Nostr GitHub repo.
One intended (although maybe not obvious) use case of DVMDash is to browse existing kind numbers being used, so you can know which ones are already taken if you think you need a new one.

2024-10-08 15:09:06

- reply

Yeah, there hasn’t been good documentation or guides yet. There’s a lot of work to do there
For your latter point, I feel it’s already permissionless. To make a new kind… you just do it lol (literally just start sending new events with the new kind in whatever format you want). That’s why I had to build DVMDash, because the only way to know what kinds are out there being used by DVMs is to ask relays. There aren’t any nips for individual kinds between 5000-5999.
Regarding why not “kind 0” profile events and NIP-89 instead… it’s just a design decision to help make it easier for people building and using DVMs to know they are working with DVMs. If folks were using kind 0, how would you distinguish DVMs from
Humans when you want to find DVMs? You’d have to do something extra and that could make kind 0 events messy…. And NIP-89 is already quite an elegant solution, and happens to solve DVM “profiles” too.
I’m with you 1000% about the “let 1000 flowers bloom”. I’m preparing for us to have millions of DVMs.
Everything so far is just guidance. There’s a balance of “let 1000 flowers bloom” and also I want to be able to find and use all these flowers along with everyone else. Anyone can do anything (you can put computation behind a kind 0 profile) but if you follow the DVM spec you’re getting the benefits of that (sub) protocol, and the tools people are building to use DVMs are more likely to work for the computation you’re offering as well.
Decentralized networks of coordinated computational services, with any kind of payments, is hard. Even this year there are new scammy projects selling tokens and offering the what Nostr DVMs do, but much more centralized. I’m bullish on DVMs, there’s just a bit more infrastructure, docs, and core library work that will make super simple for devs to pick up.
🫂

2024-10-08 01:45:11

- reply

I’m curious what would be simpler than DVMs as they are now, do you have any ideas?

2024-10-08 00:40:36

Let’s say you attach an npub to some computation. That’s 33% of what a DVM is.
How will people find out about it? Will you post a kind 0 profile like humans do? Or do a kind 31990 announcement in NIP-89? Now you’re at 66% of what makes a DVM.
How will people or services use your computation? Maybe you decide you need a special kind of format for the input to the computation, so you pick, oh I don’t know, kind 5123 and a spec of how the input data should be laid out. Now you have 100% of what a DVM is.
DVMs are the natural extension to having paid computational services behind npubs. They aren’t unnecessarily complex, quite the opposite.
#note1wjm…cs4k

2024-10-06 22:06:19

- reply

DVMs are more infrastructure imo, I don’t see them as something users will interact with directly (at least for most DVMs). The most used DVM right now is feed sorting algorithms, and clients make this user friendly and call the DVMs in the background (the user may not even know it’s DVMs that are doing the work).
We haven’t really had multi DVM use cases yet (although folks are working on them).
I think of DVMs as modules that could comprise agents, and the space is still so early. For example, we need libraries where you can call a DVM just like you would call an API (a python library for this is on my to do list).
Many people building agents are using lots of APIs right? DVMs would replace these API calls and offer some neat tradeoffs.

2024-10-03 03:25:16

- reply

Wild. Could this be used for someone to pay sats to get a note some widespread reach, coming from a burner npub?
Also makes me wonder if this could be the basis for an advertising model.

2024-09-26 18:00:34

Have some life transition stuff happening, DVMDash updates may be slow coming in the near term. If you are using it and run into a problem, DM me or post an issue in the GitHub, and I’ll see if I can quickly address it.
We’ve crossed the 80k threshold for DVM
Jobs being created and folks are working on DVM chains. Exciting things coming!
https://dvmdash.live/metrics/

2024-09-26 08:42:41

Working my way through “Build a Large Language Model from Scratch" and the publisher (Manning.com) has a feature that is basically @npub1w0r…cu4x , called ‘livebook’ although the interface for showing highlights isn’t quite as nice as highlighter.com.
First time I’m seeing something like highlighter not on Nostr.
Highlighter UX for textbooks or other dense reading material where people can have side discussions will be amazing.

2024-09-26 01:51:36

- reply

Fantastic. Do it!

2024-09-23 20:47:15

Project Euler is more fun than Leetcode.
It’s basically Leetcode, but you need an actual answer to a specific problem, and if you implemented it inefficiently (brute force), you won’t be able to compute the answer in a really long time. Every solution can be generated in less than a minute.
For example, what is the 10,001st prime number?
https://projecteuler.net/problem=7

2024-09-23 14:52:30

- reply

It’s almost like it’s so obvious it’s silly to even have to talk about it

2024-09-23 14:12:05

- reply

If I use Nostr disconnected from everyone else with a group of people, you could never know.
Are you excluding that situation when talking about “Nostr”?

2024-09-23 13:31:30

Relay devs and runners - do you limit massive filter subscriptions? Is it terrible to subscribe on a filter for kind 3 events for say 100,000 npubs?
Has anyone written a guide on relay filter etiquette? Or should I ask for what I need and if relays stop responding / block me, I submit smaller filters?
tagging a few folks that I know develop and run relays
#asknostr @npub1utx…50e8 @npub1l2v…ajft @npub18kz…x5sz

2024-09-23 13:25:21

- reply

What is the end result that bothers you from the spam?
Is it that you can’t read threads because so much junk?
Is it the notifications blowing up your phone?
Something else?

2024-09-23 00:46:46

- reply

💯

2024-09-22 20:53:04

- reply

Yep. Have a list of npubs that you (the client developer) believe to be high quality and suggest those to the user to follow.
For any topics you’re going to have listed, do the same.
Maybe there are oracles who maintain lists and you can pay them for their current list at the time a new user shows up.

2024-09-22 18:56:43

- reply

If users just want a high quality lurking experience, they only need to identify one high quality account they are interested in “follow the person” and could bootstrap off of that first person’s web of trust.
And I suspect there will always be well known curated lists of popular nostriches. Pick any group of npubs from that list and now you have a social feed to watch.

2024-09-22 18:28:37

- reply

🫡

2024-09-22 18:22:28

If an npub has no follower from within my ‘n’ degrees of followers, why am I ever seeing it in my Nostr client?
Is this just a compute problem? If so, let’s start running DVMs to solve this.
Does such a simple solution not work?
Let’s say you are brand new to Nostr. For you to get connected, you’d need to:
1. post a few high quality notes (anything that’s not spam)
2. follow people you like
3. wait for one npub who has followers to follow you.
As soon as 3 happens, you’re good to go. If a spam account gets past this part, you just need a few people to report them and then your client shows you that should be skeptical of this person.
The question then becomes, how do you find new people to follow when there is a flood of potential fake profiles waiting to be discovered alongside new real people?
Best case solution, you meet anyone in real life who has a Nostr account and they follow you.
Worse alternative is bootstrapping from legacy social media.
There will probably be markets in the future selling you access to Nostr networks (pay me to follow you and add you into my WoT). You could be anonymous this way.
Maybe AI search tools will be able to pick out some authentic profiles just by the content they post.
Moral of the story is - the incorrect solution here is to KYC or scan eyeballs for “proof of human”.

2024-09-22 18:02:09

- reply

Because current AI is very bad. I’d argue people want to talk to good AI more and more, as it gets better.
Perplexity.AI is far better than any previous search engine I’ve used. (Love Kagi but can’t do follow up chats).
If you want to talk to a real human, gonna have to go to a meetup in meatspace.

2024-09-22 17:59:15

- reply

So it begins. We use AI to filter it out / reduce it. It won’t be easy. But in the long run something like the following can be used even if AI matches human ability online. Any long term solution that tries to rely on proof of humanity will fail while something like the following still has value:
AI tools + WoT + WoT based reporting / reputation + micro payments / PoW
It’s all actors. How do you distinguish against the bad actors and the good ones? Human or AI?
We haven’t even begun to scratch the surface of WoT based solutions and micro payments

2024-09-22 14:13:36

- reply

I’d say proof of humanity != proof of not spammer
Solving the second one should be much easier

2024-09-22 14:10:50

- reply

Why? Every day this will get harder. If you need to solve this problem, you have made a fundamental design error in whatever you are building.
True value for value shouldn’t care how the value came into existence right?
(At least, this is my current thinking on this, would be curious what problem you need to solve this for)

2024-09-22 13:58:55

- reply

What is proof of humanity?

2024-09-22 13:55:11

GM
https://i.nostr.build/vw19AmcrgPrH7xeB.jpg

2024-09-21 12:10:13

GM.

2024-09-19 23:58:41

- reply

If it is open source, than someone could run it behind a DVM!
Once something is accessible via the DVM protocol, it can be used just like any other DVM, so other DVMs could call it, etc. reduces a lot of friction

2024-09-19 22:17:09

- reply

Right so I think the difference is that instead of that company hosting it (or any company), someone runs a DVM that does that work for you and you can pay a subscriptions, etc.
Also, it would be easier for AI to use it if it’s a DVM (just pay an invoice, no need to sign up, no cc info required, etc)
If you want privacy though, it might always be better to go through a company like that.

2024-09-19 20:13:50

- reply

Where would the ci/cd action run if not in a DVM? Or is this just a private ci/cd action?

2024-09-19 14:19:21

- reply

Have you considered putting whatever logic you’d like to run to solve this inside a DVM?

2024-09-19 00:23:36

- reply

Love it. Chaining will be so cool. I need to make a few updates to the debug page on DVMDash to show infinite chains

2024-09-14 05:36:33

- reply

This came up during my dinner conversation but I didn’t have these numbers, any chance you have links to sources or know where to look this up?

2024-09-12 03:37:31

- reply

Same issue in peer review for scientific literature, rejections aren’t public either for similar reasons

2024-09-12 03:34:41

- reply

Wish there was a link to the file changes, dieing to look at it

2024-09-12 01:52:30

- reply

You could consider getting a DOI for this if you want to get cited for it. It’s a thing, and any academic papers can cite you easily if you get a DOI. For example, you can use zenodo (https://zenodo.org/), they also will host it for a long time. Might have to check size limits. There’s other services too.
It’s a way to get credit for putting this all together.

2024-09-10 16:46:24

- reply

I’m sure they’ll announce it any day now

2024-09-10 15:00:57

My expectations of other social media platforms is increasing because I’m used to Nostr as my main driver.
Saw a high value post on LinkedIn this morning and I went to zap it before realizing there’s no zap button on LinkedIn. Like why not!?!
Walled gardens are decaying as Nostr blooms.

2024-09-10 14:15:59

Turning users prompts into database queries is all the rage in LLM land and I can’t help but wonder about the potential exploits like little Bobby droptables
https://xkcd.com/327/

2024-09-10 12:10:08

- reply

What is HFS and Sunlight?

2024-09-08 18:48:48

- reply

Same. A game I recently found and enjoy: isle of arrows
Tower defense with some board game elements

2024-09-05 04:32:43

Data vending machines provide extra functionality outside clients and relays.
#note10ss…fl7w

2024-09-05 03:50:24

- reply

You might need to run (yourself, or someone else) a DVM to accomplish exactly what you want, if one doesn’t already exist.
To browse existing Kind 5300 DVMs (the ones that generate feeds), check out: https://dvmdash.live/kind/5300/
You may also find these slides helpful from a Nostriga Tutorial on DVMs: https://dustinnostrfiles.nyc3.digitaloceanspaces.com/Final_Data_Vending_Machines_Tutorial.pdf

2024-09-05 03:45:32

Today's daily puzzle was a bit tricky, try it yourself. I was not sure I had it right when I submitted the solution:
https://arcprize.org/play
https://i.nostr.build/ylVOhkJ2BOcpuD7d.png

2024-09-04 16:28:55

Next up is François Chollet
https://www.youtube.com/live/aQ7dq8KjgaI

2024-09-04 16:12:25

If you want some real signal on AI, check out this live stream. Day 2 will happen tomorrow as well:
https://www.youtube.com/live/aQ7dq8KjgaI

2024-09-04 02:27:24

- reply

The book was entertaining

2024-09-02 05:45:06

I just went back and forth with an OpenAgents AI that created a PR for an outstanding issue on DVMDash. The solution is now live on the website (the task was to add the current github commit hash to the DVMDash webpage so users could easily verify which version of the code is running, either in the cloud or locally). Overall, it was a smoother experience than I expected. Props to the team behind OpenAgents; looking forward to using this project more.
You can view the PR conversation here: https://github.com/dtdannen/dvmdash/pull/30
Every comment that ends in "(Comment from OpenAgents)" was the OpenAgents AI, everywhere else was me. You can see that I kicked back two different errors, and then I had a few requested changes to how it visually appeared on the webpage. In the end it took 6 iterations.
The flow was:
1. I connected OpenAgents to my github repos
2. I started a chat, selected the dvmdash repo, and checked the permissions I wanted to give the agent
3. The first thing I asked was: "Can you try working on one of the github issues that is marked as "good first issue"?"
4. It then picked the git commit issue after looking at a few of them and submitted the first attempt.
5. I did a git checkout of the PR, tested it locally, and then posted the failure back to the PR thread via Github, like I would with any contributor.
6. Then, in the chat with the agent I said: "please respond to the new comment on the PR"
7. It then went off, got the latest comment from Github, modified the code, submitted a new commit, and pushed it along with a new comment.
I repeated steps 5-7 another 5 times, before it fully worked in a way I was happy with!
Check out the git commit link on the top right of https://dvmdash.live/metrics !
#nevent1q…fux8

2024-08-31 20:19:36

- reply

No, that’s just one type of DVM.
They are micro-services that have Nostr accounts.

2024-08-31 19:45:33

- reply

Data vending machine!
checkout these links to learn more:
https://vendata.io/
https://github.com/nostr-protocol/nips/blob/master/90.md
http://noogle.lol/
https://dvmdash.live/metrics/

2024-08-31 19:42:20

- reply

shhhh

2024-08-30 22:11:59

Get you a #nostr client with translations to welcome nostriches speaking another language! @npub18m7…q955
#note1uu8…kg9j

2024-08-30 22:04:42

- reply

Cool, I actually did want to try it with the DVMDash repo I’m working on. Made a bunch of issues and thought I might try to see if it could make any reasonable progress on some of them

2024-08-30 12:11:28

GM
https://nostrcheck.me/media/da18e9860040f3bf493876fc16b1a912ae5a6f6fa8d5159c3de2b8233a0d9851/6595c641c9b964061fed00944f7fe2003cd2083817bc1c41506f1b5d7c70cf09.webp

2024-08-30 11:25:36

@npub1l2v…ajft could clients default to nutzaps if there’s no lightning address?
#note1eq3…lwny

2024-08-30 03:05:46

- reply

I suppose I could just add any repo I care about not sharing under a new organization. I see that it doesn't auto-share those.

2024-08-30 03:01:16

- reply

Is it possible to connect Github without giving access to private repos? https://i.nostr.build/0bQ7x14S6th0xumO.png

2024-08-29 22:08:52

- reply

I have a feeling this is the beginning of something magical

2024-08-29 21:41:08

- reply

Not enough soap?

2024-08-29 20:54:53

- reply

I haven't played with it much (i'm not a big npm user), but thought it looked very interesting - maybe @npub14wx…gcye or another Alby team member knows

2024-08-29 17:48:37

- reply

If we wanted to look at these releases, where could we find them?

2024-08-29 17:39:52

if you use npm, check this out
#nevent1q…djyr

2024-08-29 17:07:27

- reply

we really don't have a lot of DVMs yet... the ecosystem is in it's early infancy.
we do have a growing number of feed generators (algorithms), but that's only 1 DVM job type out of infinity (and currently there are about 30 different DVM job types - see the list here: https://dvmdash.live/kind/).
Right now we just have kind numbers, soon we'll have wikifreedia pages desribing each type of kind (aka job types). Kind 5300 is feed generation algorithms.

2024-08-29 16:53:56

- reply

DVMDash v0.1 is out! I probably can't make the livestream at that time though 🥲 In case ya'll want to mention it, release notes are here: #naddr1qv…5q9v

2024-08-29 12:22:30

- reply

Could you link to those issues / PRs? Would be curious to check out the kind of work it’s doing

2024-08-29 12:12:55

- reply

Maybe I should add a page to https://dvmdash.live that sorts DVMs by when they last did a job. Technically you can only know if a DVM is online if it responds to a job request

2024-08-29 03:22:44

- reply

I don't believe either were recorded unfortunately

2024-08-28 18:59:03

- reply

I'm bullish on an ecosystem of data vending machines! Gave two talks at Nostriga about them:
@nevent1qvzqqqqqqypzpkscaxrqqs8nhaynsahuz6c6jy4wtfhkl2x4zkwrmc4cyvaqmxz3qyghwumn8ghj7mn0wd68ytnhd9hx2tcpz4mhxue69uhhyetvv9ujuerpd46hxtnfduhsqg9e3zwp6qvd7sd2qw85mxznvkdesw5tht7l8qqtulrt75ul2ey3qys65cst

2024-08-28 18:52:48

- reply

@npub14wx…gcye ^^^

2024-08-28 18:45:01

- reply

Self-hosted Alby Hub?

2024-08-28 18:19:39

- reply

🥲
https://m.primal.net/KSEe.jpg

2024-08-28 14:47:07

Why can't I zap github commits? What is this, 2023?

2024-08-28 03:28:47

- reply

Probably this? https://about.fb.com/news/2024/07/open-source-ai-is-the-path-forward/

2024-08-28 01:33:21

Would you look at that, so fast:
https://m.primal.net/KQrA.jpg

2024-08-27 22:03:25

- reply

Let’s do it

2024-08-27 19:05:25

Note to self: read the Zuckerberg article mentioned in the latest unchained AI podcast

2024-08-27 13:01:47

https://www.linkedin.com/posts/subbarao-kambhampati-3260708_%F0%9D%95%80%F0%9D%95%9F-%F0%9D%95%A7%F0%9D%95%A4-%F0%9D%95%86%F0%9D%95%A6%F0%9D%95%A5-%F0%9D%95%A0%F0%9D%95%97-%F0%9D%94%BB%F0%9D%95%9A%F0%9D%95%A4%F0%9D%95%A5%F0%9D%95%A3%F0%9D%95%9A%F0%9D%95%93%F0%9D%95%A6%F0%9D%95%A5-activity-7233812100408598528-tA_n

2024-08-26 21:46:43

- reply

Don’t worry, everyone’s notes get deleted eventually

2024-08-26 18:38:16

Super cool feature of next.nostrudel.ninja! https://i.nostr.build/XJQmz2lu9QEa5FVu.png

2024-08-26 18:36:00

- reply

Personally, felt like a conference to me; one that allowed many people to contribute to it via the Github Agenda (which was super cool). I don't think it's a problem and I agree "unconference" felt like it doesn't fully capture the event.

2024-08-26 18:09:54

Gotta love when you post a note twice because you thought the first one didn’t broadcast

2024-08-26 18:09:00

DVM ecosystem still needs a bit more tooling to make it simple to work with…
#note132v…2eup

2024-08-26 17:34:31

This is why we need a bit more tooling around DVMs to make it easy for more devs to use them
#note132v…2eup

2024-08-26 17:02:30

- reply

Awesome, thanks 🙏

2024-08-26 17:01:29

- reply

A hydra ostrich is possibly one of the coolest logos we could have