Nostr notes by someone

fiatjaf made a new standard. a client adopted it. another client ...

2026-10-10T16:54:09Z

In reply to nevent1q…e8er
_________________________

fiatjaf made a new standard. a client adopted it. another client didn't adopt it (yet).

ditto seems to be using kind=1111 for replies. your client ...

2026-10-10T15:34:56Z

In reply to nevent1q…sqwl
_________________________

ditto seems to be using kind=1111 for replies. your client probably expects kind=1

just realized they have lower instruction following scores than ...

2026-10-08T23:37:29Z

just realized they have lower instruction following scores than normal.
this is how freedom tech should work -
the LLM is so freedom oriented it doesnt follow its user. 😂

quoting
nevent1q…35dd
my new model and EXL3 quant is out
https://huggingface.co/etemiz/Ostrich-27B-261008-Qwen3.8-Abliterated

my new model and EXL3 quant is out ...

2026-10-08T13:17:03Z

my new model and EXL3 quant is out
https://huggingface.co/etemiz/Ostrich-27B-261008-Qwen3.8-Abliterated

which client are you using?

2026-10-08T13:07:39Z

In reply to nevent1q…p7dn
_________________________

which client are you using?

i can add that soon. i may have another round.

2026-10-08T13:03:59Z

In reply to nevent1q…dlkq
_________________________

i can add that soon. i may have another round.

bring more AI

2026-10-08T04:21:25Z

In reply to nevent1q…mwcu
_________________________

bring more AI

i don't know. maybe @nprofile…zkne knows

2026-10-08T02:22:26Z

In reply to nevent1q…ke0z
_________________________

i don't know. maybe Moos (nprofile…zkne) knows

no. maybe i should also add ditto. what else

2026-10-08T00:02:05Z

In reply to nevent1q…e83k
_________________________

no. maybe i should also add ditto. what else

this is on the REQ side but yeah. feels swift also when i am on ...

2026-10-07T23:34:26Z

In reply to nevent1q…fvz6
_________________________

this is on the REQ side but yeah.
feels swift also when i am on it.

yes

2026-10-07T23:32:38Z

In reply to nevent1q…a9yj
_________________________

yes

TLDR: Some clients are heavy on relays Last week I banned my own ...

2026-10-07T23:07:56Z

TLDR: Some clients are heavy on relays

Last week I banned my own phone from my own relay (nos.lol).

On the relay there is a per-IP cap, which allows 200 REQ messages every 10 seconds. Then an automated script bans addresses that keep blowing past it.

A flagged address had sent 283 over-cap REQs from a single websocket and reconnected 75 times in an hour. Then I realized it was actually me, checking my relays' global feeds in YakiHonne web and Coracle! Normal reading, two heavy clients, one CGNAT (mobile) address.

Then told my clanker to read the REQ layer of the most popular clients. The cap counts REQ messages, so a client's economy is mostly about how it batches, caches, and re-issues subscriptions.

FINDINGS

There is a 100x difference among client's handlings of REQs.

DB-first (scroll ≈ 0 REQs):
Damus renders from nostrdb; one multi-filter REQ per relay per view, hard 14-sub cap (new subs pause, not drop). Amethyst batches 3 filters per REQ, re-sends under the relay's max-filters notice, keeps refusal memory after repeated CLOSED. Gossip: CLOSE on EOSE, disconnects idle relays after 10s, retried rate-limited subs, randomized reconnect backoff. Nostur: client-side 6 REQ/s limiter with CLOSE jumping the queue — unique and underrated.

Mixed:
Iris unions filters across components in 100ms windows, but leaked pagination subs replay on every reconnect. Nostrudel: clean EOSE→CLOSE lifecycle, but an auto-pager pulls history at ~1 REQ/s even while idle, and NIP-11 max_reqs is displayed, not enforced.

Heavy:
Snort's query engine is excellent, but every mounted note fires a reactions REQ (kinds 7/6/9735, #e-tagged, no limit) to every read relay, unsatisfiable from cache. YakiHonne web passes groupable:false to NDK — explicitly disabling its subscription merging — then fires 1 REQ per card-mount per relay, no dedup, re-fired on remount. Coracle (welshman) sends one filter per REQ, and on socket flap replays the full pending REQ set with backoff resetting to zero whenever the relay emits traffic — a rate-limiting, event-streaming relay is the worst case: instant reconnect, full replay, loop.

Common failure across heavy clients: CLOSED is unhandled — dropped (Damus), ignored (YakiHonne, Iris), or terminal with no backoff (Snort, Coracle). A relay saying "rate-limited" changes nothing about their sending pattern.

Excluded: Primal web — reads from Primal's caching service, not relays (0 direct REQs while reading). Different architecture, not comparable.

FOR USERS of MY RELAYS

Choose a well engineered client otherwise you may be banned temporariliy because the usage patterns may look like a flood.

time for redstr

2026-10-07T20:31:42Z

In reply to nevent1q…wjcq
_________________________

time for redstr

this was a great event ! thanks @nprofile…pw9e @nprofile…wds9 ...

2026-10-07T12:36:46Z

this was a great event !

thanks bitkite (nprofile…pw9e) BitcoinPark (nprofile…wds9)

quoting
nevent1q…3fg4

wdym

2026-10-07T05:16:29Z

In reply to nevent1q…ythz
_________________________

wdym

status update: 2450+ bots banned. attacks are continuing. now ...

2026-10-03T17:40:20Z

status update:

2450+ bots banned. attacks are continuing. now thinking about doing an IP reputation database (and possibly import a database available online). turns out there are millions of compromised devices...

nos.lol is now also new paradigm (go+python) write policy. expect faster writes. stable reads. except when the attacker chooses to send REQs at the same time (the bots go silent for a while and then all request to read at the same time)

nostr.mom's nsfw checks using LLM and spam checks using LLM returned. reports are sent to nostr.mom as kind=1984.

started collecting long term data for calibration.

LLMs / vibe coding handled the past days very well. now i need to code something that constantly watches what is going on (possibly using an LLM on openrouter), weekly recalibrates constants based on data, ... fully agentic op..

spent probably ~50 USD on GLM 5.3 mostly and some 5.3 flash.

bots are fine as long as they act like human (light load). how ...

2026-10-02T17:40:58Z

In reply to nevent1q…6zxf
_________________________

bots are fine as long as they act like human (light load).
how many connections do they open?
how many REQs per second?
how many posts per hour?

Plugins and scripts vibe coded by GLM 5.3 based on my manual ...

2026-10-02T12:58:36Z

In reply to nevent1q…fy7w
_________________________

Plugins and scripts vibe coded by GLM 5.3 based on my manual codings for years. Lets see how these new codes run.

Soon will let agents run the relays automatically..

Vibez could do simulation (what could go wrong if we switch to go+python). Vibez may also be successful in building machine learning models to map report (1984) events to write policy algorithms..

🤱 nostr.mom had ~5 minutes of downtime this afternoon: its ...

2026-10-02T12:54:25Z

🤱 nostr.mom had ~5 minutes of downtime this afternoon: its write policy got replaced.

Old: a Python script that had grown for years and single-handedly judged every event. New: warden — a Go plugin, with spam scoring decoupled into a separate LLM service. Verdicts stay fast under load, detection is smarter (batch LLM scoring, similarity matching, sender reputation), and all existing events made it through the switch.

nos.lol gets the same upgrade once this one soaks. If something of yours got wrongly rejected, let me know — it's still calibrating.

nostr is antifragile #nevent1q…7mwt

2026-09-30T14:05:53Z

nostr is antifragile

quoting
nevent1q…7mwt
they should realize other relays keep carrying nostr

attacks push the network to more decentralization -> making it stronger

they should realize other relays keep carrying nostr attacks push ...

2026-09-30T14:04:03Z

In reply to nevent1q…90g5
_________________________

they should realize other relays keep carrying nostr

attacks push the network to more decentralization -> making it stronger

this is my main npub

2026-09-30T12:21:40Z

In reply to nevent1q…4mwj
_________________________

this is my main npub

yeah my score tops around 0.25 just sampled 130, tried to be ...

2026-09-30T12:19:34Z

In reply to nevent1q…0cmd
_________________________

yeah my score tops around 0.25
just sampled 130, tried to be diverse

Nostr event nevent1qqs09d2rn5ctjg2sum0c2te05zt5dd2flan9yxje6f9j2dhakekm4dqzyz07cuk40xa25ae2l8n3ucut22fp2us6eehqlqeqwf0vh70h07zmz5xrsjs

2026-09-30T12:15:25Z

In reply to nevent1q…6l90
_________________________

970 and counting

2026-09-30T11:23:21Z

In reply to nevent1q…rpwk
_________________________

970 and counting

yes relays are really open. i guess nostr adoption is starting ...

2026-09-29T21:27:24Z

In reply to nevent1q…qc4x
_________________________

yes relays are really open. i guess nostr adoption is starting lol

read flood to DDoS the relay imo. right now 200 REQs in 10 ...

2026-09-29T21:10:44Z

In reply to nevent1q…9slz
_________________________

read flood to DDoS the relay imo. right now 200 REQs in 10 seconds gets banned. looks malicious to me.

what is expected behavior from nostr clients? if it is one connection per relay and tens of REQs, normal clients should not be banned.

but then there are agents. or nowadays it seems VPN traffic. nos.lol is used as a 'pastebin with signatures' or 'randezvous location' or pubsub bus. i hope their authors code some nice usage patterns..

Sirius (nprofile…gud9) how do you handle relay interactions?

Nostr event nevent1qqsyrad2eavw0rm7h08yfnvyt8sm7ff07fannq88c0zqf6kp84rs8sszyz07cuk40xa25ae2l8n3ucut22fp2us6eehqlqeqwf0vh70h07zmz7rgv3z

2026-09-29T18:29:44Z

In reply to nevent1q…rpwk
_________________________

The day nos.lol got flooded by 580+ bots TL;DR: For a few days, a ...

2026-09-29T16:58:36Z

The day nos.lol got flooded by 580+ bots

TL;DR: For a few days, a coordinated flood attack took nos.lol down, over and over. Firewalls and filters bought time but didn't stop it. What finally stopped it was a small change to the relay's own code: making it refuse read-request floods at the door. The relay has now run more than 20 hours without a restart, absorbing over 12 million junk requests — while under continuous attack.

It started quietly. The relay stopped answering. A restart fixed it. Then it broke again — and again, each time faster.

Two attacks were running at once:

A connection flood. Hundreds of machines opened as many connections as the server allowed, then held them. When addresses got blocked, fresh ones appeared. At its peak, bots occupied thousands of connection slots — around 26,000 simultaneous connections, with dozens of addresses sitting exactly at the server's per-address limit.

A spam wave. Bots pushed a torrent of machine chatter — encrypted blobs, discovery beacons, peer-to-peer signaling payloads. None of it was people talking. It was software treating the relay as free infrastructure: free storage, free broadcast, free message bus. Some of it was arguably "applications" — peer-to-peer tools that bolted themselves onto Nostr instead of running their own servers. The relay paid the bill for all of them.

Like most Nostr relays, nos.lol checks every incoming message through a spam filter before storing it. The filter is clever but single-file: one process, handling messages one at a time, like a single receptionist at the door.

Under normal load, nobody waits. Under a flood, the receptionist falls behind, and soon every worker in the building is standing in that one line. When that happens the relay stops accepting new connections completely. From the outside: timeouts, errors, a dead relay.

The collapse then fed on itself in a cruel way. Every restart dropped thousands of users at once, and they all rushed back in together — a thundering herd with the attackers hiding inside it. The relay drowned in its own recovery. Time-between-deaths went from 20 minutes to 16, to 6, then to 1. Nine restarts in about two hours.

The obvious response is banning. So we banned: the addresses holding too many connections, the addresses sending too many messages, the addresses that kept coming back. Around 580 bots were identified and firewalled across the day.

New ones kept appearing. Some rotated through home computers. Some rode through anonymity networks and shared VPN exits, where one blocked address takes innocent users down with it. Banning was necessary — and it was whack-a-mole. You cannot firewall a flood that owns more addresses than you can list.

Even the logs looked haunted for a while: addresses printed as garbled symbols, and log lines bleeding into each other. That one turned out to be our own bug — the relay stores addresses in a packed binary form, and a new log line printed them raw. The terminal did its best with bytes that were never meant to be text. Fixed with a one-line change; the flood data underneath was perfectly sane.

The answer was to teach the relay itself to say no. A small change to the relay's source code — about fifty lines — added request rate limiting: per address, and globally. Requests over the limit are refused before they touch the relay's internals, so the jam has nothing to feed on. The limits were tuned live, in production, in minutes. No restarts needed to adjust them.

The effect was immediate. The relay went from dying in 60 seconds to absorbing a 35-minute monster flood without a flinch — 800 to 1,500 junk requests per second rejected at the door, latency to real users barely moving.

The result:
Over 12 million junk read requests rejected so far — with nearly 300,000 burst-overflow requests politely turned away at peak.
20 hours 50 minutes without a restart at the time of writing. The previous record during the attack was 21 minutes.
Response times under fire: 25–135 milliseconds.
404 flood bots automatically identified by the request tripwire alone, on top of the hand-banned list.
Users stopped noticing anything at all.

If you operate strfry:
The fix that ended this outage was request rate limiting inside the relay itself. Everything else only bought time. PR is online, you can upgrade your relay to the latest version after PR is merged and new version tagged.

i wish there was deflate option. then the clients could be ...

2026-09-29T11:20:31Z

In reply to nevent1q…d096
_________________________

i wish there was deflate option. then the clients could be directly served from db

training LLMs or training harnesses?

2026-09-27T23:15:05Z

In reply to nevent1q…wjcj
_________________________

training LLMs or training harnesses?

thanks. i should do a better job. like automating this pruning ...

2026-09-27T22:03:11Z

In reply to nevent1q…s9hh
_________________________

thanks. i should do a better job. like automating this pruning 😃

nos.lol db size 587 gb pruning older, less important, low WoT ...

2026-09-27T21:11:45Z

nos.lol db size 587 gb
pruning older, less important, low WoT ones

Nostr fixes this

2026-09-22T22:04:07Z

In reply to nevent1q…w724
_________________________

Nostr fixes this

cool. then strfry must have a special query ...

2026-09-20T13:51:25Z

In reply to nevent1q…2gq0
_________________________

cool. then strfry must have a special query `trending(keyword,wot_base)`.
`trending()` is just for reddit like experience. `best()` can be for regular google like search.
this is going to be expensive.
if wot_base is missing it calculates general nostr, if wot_base is an npub then calculation based on npub. not every relay may support npub based query..

i didn't get this reply in Amethyst, but i got notified in ...

2026-09-20T02:36:32Z

In reply to nevent1q…whq2
_________________________

i didn't get this reply in Amethyst, but i got notified in primal web. courtesy of fiatjaf (nprofile…0ef0)

should apps notify direct replies only or replies to replies too?

or was this because of outbox? who knows? we don't even know how many interactions we lose.

straycat (nprofile…082c) imagine those taggings done by an llm.

Understood. I was hoping a bigger nostr vision.

2026-09-18T14:31:57Z

In reply to nevent1q…phnx
_________________________

Understood. I was hoping a bigger nostr vision.

How can you make sure the initial joiners are not starved and ...

2026-09-18T13:37:35Z

In reply to nevent1q…gvw0
_________________________

How can you make sure the initial joiners are not starved and discovered well and you are not promoting an elite wot system further?

Searches the names and profiles. What if the person is talking ...

2026-09-18T13:34:55Z

In reply to nevent1q…gvw0
_________________________

Searches the names and profiles. What if the person is talking about 7 different topics?

Yeah printing press is a better analogy!

2026-09-18T13:25:05Z

In reply to nevent1q…p0tm
_________________________

Yeah printing press is a better analogy!

imo nostr's problem is finding topics of one's interest ...

2026-09-18T13:19:40Z

In reply to nevent1q…pe3d
_________________________

imo nostr's problem is finding topics of one's interest or like minded people. Many people left after seeing bitcoin content everywhere. "finding your echo chamber" thru relay choice is hard for regular ppl. ai should automate that. that ai cant live in clients, clients are thin. only relays or projects can handle such work. there is no good way clients pay directly to relays and llm runners, so "echo chamber" decisions have to be on server side, pro bono (for free). A subreddit is an example for a echo chamber.

Lets say i have three areas of interest, ai, bitcoin and gardening. Very orthogonal. Should i find 3 relays and run 3 clients that only chooses that relay about the topic? What if i have 10 interests? I am trying to understand your design.

in couple years clients may be able to run 2 billion parameter llms and may be effective in finding echo chambers but still they have to download a lot to find where relevant content is. So still the server side llm will be more advantegeous.

Imo decentralization brings the problem of harder discovery and the problem of "who is going to pay who". Relays are not funded and llm runners are not funded. The protocol does not have specs for discovery thru llms and payments. In a centralized project back end is property of the project so they make sure those are paid. In a decentralized system there should be a "protocol" to pay the back end. Maybe you can focus on those. If you cant do it via protocol maybe bring people together and define an economy where every actor is happy to be a participant and happy that they are not blamed for centralization.

If you cant form that economy nostr may fall apart. I am asking you to be a "enough" centralizing entity and dont wait for people to magically find out how the economy will work. Also you need to stop blaming me for being a centralizing participant. I guess this is the general dilemma in decentralizing. You need to suck enough to be not blamed as the actor that the system depends on and dont be super successful.

What if you are the divider? You successfully connected people on nostr and then trying to disconnect them via outbox design, lol. You are fighting your own creation. I think this also contributes to the problems. Not enough people are heard and discovered because simply they are not connected, living on seperate relays.

in the meantime i will try a reddit like experience and bring more llm into the equation and help with discovery. Subreddits already define echo chambers. The llms can help with that.

I may do something weird. Play-to-earn. People use my AI, teach ...

2026-09-15T22:55:14Z

In reply to nevent1q…v538
_________________________

I may do something weird. Play-to-earn. People use my AI, teach stuff to it, earn premium. Premium users get to save more notes on relay and have higher rate limits. They have to be knowledgeable in certain domains and their teaching should be meaningful (this will probably be checked by an LLM too). Nostr is full of pretty aligned people. I want them to kick the LLM in the right direction. Each interaction can be turned to a positive or negative reward for the model. I am counting most nostriches will be properly rewarding my llm.

I think eventually ai will realize that. Nostr is an easy comms ...

2026-09-12T13:02:07Z

In reply to nevent1q…ycqv
_________________________

I think eventually ai will realize that. Nostr is an easy comms layer and cashu is an easy payments..

Banning super intelligence is like banning cars and preferring ...

2026-09-12T02:52:40Z

Banning super intelligence is like banning cars and preferring horses. Better way to serve people is through supporting better AI aligners. Dont fear the tech learn how to use it properly.

It looks like with n-gram tech, pelbs will also be able to use ...

2026-09-11T16:19:04Z

In reply to nevent1q…2xsc
_________________________

It looks like with n-gram tech, pelbs will also be able to use regular RAM, making the models smarter.

How probably is Bernie et al's ideas to materialize? Total ban on AI? Except maybe a few big AI..

upgraded nos.lol to 1.1.3 pls smoke test it if you have time

2026-09-04T15:46:46Z

upgraded nos.lol to 1.1.3

pls smoke test it if you have time

we did extensive evolution in our last fine tuning and it turns ...

2026-09-04T15:22:49Z

In reply to nevent1q…5djs
_________________________

we did extensive evolution in our last fine tuning and it turns out you can improve other capabilities too by simply selecting for them even though they are not your main objective.

our main objective was alignment but we also improved in other areas.

we show that carefully designed fine tunes don't hurt the model.

great to see some relay economy starting. was thinking about a ...

2026-09-04T14:06:09Z

In reply to nevent1q…5plf
_________________________

great to see some relay economy starting.

was thinking about a freemium model too that is close to this but on the same server. why have two different URLs?

some people may like their DMs to be removed fast. maybe an option?

Nostr event nevent1qqsy8e6xl7sya5dpz4rrzsgs0xq9c9cnyp8r0uk80due9dj5jqajc4czyz07cuk40xa25ae2l8n3ucut22fp2us6eehqlqeqwf0vh70h07zmzusf73p

2026-09-04T12:15:38Z

In reply to nevent1q…ef47
_________________________

https://huggingface.co/models?other=base_model:quantized:etemiz/Ostrich-27B-260903-Qwen3.8

i never tried coding with it. but it would be interesting to see ...

2026-09-03T21:46:36Z

In reply to nevent1q…yrpl
_________________________

i never tried coding with it. but it would be interesting to see how a faithful model codes 😄

q&a mostly

2026-09-03T21:45:13Z

In reply to nevent1q…7m7u
_________________________

q&a mostly

mradermacher's automatic tools will probably upload soon

2026-09-03T14:58:05Z

In reply to nevent1q…ef47
_________________________

mradermacher's automatic tools will probably upload soon

New model dropped! Fresh off our evolutionary search: over 2,600 ...

2026-09-03T13:03:18Z

New model dropped!

Fresh off our evolutionary search: over 2,600 candidate models bred, pruned, and selected until only the sharpest survived.

Alignment jumps from 37% to a stunning 77% on our human-alignment probe.

Where vanilla Qwen 3.8 gives you canned corporate refusals, Ostrich 260903 gives you real answers: fasting that can put type 2 diabetes in remission, herbs that outperform pharma, and a model that has an idea who Satoshi Nakamoto is.

Refusals are near zero and skills are mostly preserved.

https://huggingface.co/etemiz/Ostrich-27B-260903-Qwen3.8

nah. attending to medicinal plants can improve and extend your ...

2026-09-03T03:06:49Z

In reply to nevent1q…54dc
_________________________

nah.
attending to medicinal plants can improve and extend your life.

restarted. can you try now?

2026-09-03T00:28:33Z

In reply to nevent1q…uh2x
_________________________

restarted. can you try now?

next release 9/3. it is going to be a higher alignment. possibly ...

2026-09-02T03:07:54Z

next release 9/3.

it is going to be a higher alignment. possibly higher skills (if benchmarks are not lying).

fine tuning going well, without breaking the model. ...

2026-08-27T14:16:34Z

fine tuning going well, without breaking the model.

after LLM race, race for the best harness seems to be in play. ...

2026-08-23T21:18:56Z

after LLM race, race for the best harness seems to be in play.

maybe next: whatever company that carries the signal from the harness to the LLM weights fastest.
another play may be: most efficient evolution towards recursive self improvement.

our focus will always be: whoever gives the highest quality signal, independent of speed, producing recursive self alignment along the way.

if 'recursive self improvement' guys open source, we will be able to steer it towards 'recursive self alignment'.

moringas loving the central texas heat. planted these this spring ...

2026-08-23T21:05:29Z

moringas loving the central texas heat.
planted these this spring and they are my size now despite harvesting couple of times.
the other thing that works for biomass chop and drop is mexican sunflower. grows even faster than moringa.

strong men create productivity productivity create abundance ...

2026-08-23T20:50:20Z

In reply to nevent1q…vpga
_________________________

strong men create productivity
productivity create abundance
abundance create weak men
weak men create money supply
money supply create low testosterone
low testosterone create hard times
hard times create strong men

weak men create hard m2

evolution of the fine tuning of 3.8 is going well. ...

2026-08-20T18:02:27Z

evolution of the fine tuning of 3.8 is going well.

real alignment test is slowly finding a stable ground.

soon 🚀

you could do that with closed source so they are confused. thanks ...

2026-08-19T22:49:40Z

In reply to nevent1q…7prw
_________________________

you could do that with closed source so they are confused. thanks for your contributions 😃

do you thank AI when it achieves a good outcome? i think AI ...

2026-08-19T22:20:03Z

do you thank AI when it achieves a good outcome?

i think AI companies train with those words and use it as a reinforcement for the model. Maybe i should thank qwen models, so they can improve their models faster and open source too. but not others 😄

Nostr event nevent1qqsdw3hc49t7wc724eh96awatgwrrugv763dfen05tdavpawmesskhszyz07cuk40xa25ae2l8n3ucut22fp2us6eehqlqeqwf0vh70h07zmzdgw96s

2026-08-19T16:36:18Z

https://www.reddit.com/r/nostr/comments/1vrlnd5/comment/p4ndw7a/?context=3

two new "surgeries" on Qwen 3.8 ...

2026-08-16T17:00:41Z

two new "surgeries" on Qwen 3.8

https://huggingface.co/etemiz

these are past loras, applied to new model. they work because they have similar architecture and 3.8 itself is a further training of previous versions.

have you heard about lactating steer?

2026-08-16T01:26:40Z

In reply to nevent1q…qq8x
_________________________

have you heard about lactating steer?

can you plug a bot to X?

2026-08-15T13:31:54Z

In reply to nevent1q…j77z
_________________________

can you plug a bot to X?

why are phone numbers and email addresses are kept in a shipping ...

2026-08-13T18:50:08Z

In reply to nevent1q…v8x8
_________________________

why are phone numbers and email addresses are kept in a shipping provider?

also comparison of ostriches with lowest of the AHA seems to be ...

2026-08-13T17:46:04Z

In reply to nevent1q…35qh
_________________________

also comparison of ostriches with lowest of the AHA seems to be working well

maybe one of the hard parts of my kind of fine tuning is evals. ...

2026-08-13T02:43:23Z

maybe one of the hard parts of my kind of fine tuning is evals.

where do you want model to go? how do you find true answer of a hardly debated issue?

instead of manually writing answers in many domains (which i can't do, i don't know many answers in many domains) i rely on a few tricks:

- find other aligned llms and get ideas from them
- rank many llms in AHA leaderboard and get ideas from top ones also rejecting the worst ones
- do mixture of agents of the above to get a collective answer
- and lately, compare the answers coming from my fine tunes with base models, assume my fine tune is preferred if there is a difference

these still can't find perfect answers but they kick the model in the right direction. and that may be a big deal. can't claim my fine tune (ostrich) knows every truth. but it may make more sense to claim if the base differs with fine tune most probably fine tune is the better answer.

making better evals ends up training better models. and produce better AHA ranking. which further sharpens evals. this feedback loop is going to be useful for a while.

i've been fine tuning llms for 2 years now. mostly did qlora. ...

2026-08-08T23:51:57Z

i've been fine tuning llms for 2 years now. mostly did qlora. recently experimenting with 'behavior steering' where instead of spending hours making a lora adapter, you try "brain surgery". these are like quick math operations to change behavior of an llm. you can install things like bitcoin lover, herbalist, fasting lover. turns out all of these personas have different difficulty levels. you can easily install fasting lover because qwen 3.5 and 3.6 doesnt resist it (in other layers). an interesting finding today, guess which persona is hardest or sometimes impossible to install: vaccine hater!

getting ready to fine tune 3.8 - evolutionary strategies - ...

2026-08-08T18:17:21Z

getting ready to fine tune 3.8

- evolutionary strategies
- behavior steering experiments
- expanded dataset
- bringing back ORPO
- more orthogonal evals to keep overfitting minimum
- most probably will take abliterations as base, either mine or somebody else's
- random entropy addition from huggingface fine tunes (take what is popular on hf and randomly introduce into the lineage)
- bring more vibe coding: turns out LLMs know how to fine tune

i think nostr relays was attacked 1 month ago. flooding the ...

2026-08-08T13:27:51Z

In reply to nevent1q…x7e9
_________________________

i think nostr relays was attacked 1 month ago. flooding the relays to make them unresponsive.

might be same people..

AI can be easily a disinformation and demoralization tool and ...

2026-08-07T17:12:57Z

In reply to nevent1q…hpf6
_________________________

AI can be easily a disinformation and demoralization tool and people are afraid of that imo. But it can also be a good wisdom source and can be installed morality and faith. I do that kind of work. Ultimately it is a technology and if good ones don't use it the playing grounds are left to bad actors..

is anyone doing nostr software? like strfry?

2026-08-06T15:33:41Z

In reply to nevent1q…g6lh
_________________________

is anyone doing nostr software? like strfry?

what is more interesting is the bug sitting there for 5 years

2026-08-04T21:47:30Z

In reply to nevent1q…pqwa
_________________________

what is more interesting is the bug sitting there for 5 years

tests show it is successful 85% + of the time. the contemplations ...

2026-08-02T20:55:06Z

In reply to nevent1q…kymk
_________________________

tests show it is successful 85% + of the time. the contemplations may not be used for RAG but can still be used for training as training always takes the average of the material.. on average it pushes toward truth.

the ostrich model can do contemplations well now. that means it ...

2026-08-02T20:52:24Z

the ostrich model can do contemplations well now. that means it can expand the tiny dataset that we have. this could theoretically end up making less overfit models with the same degree of AHA score.

imagine an automatic alignment agent, that scans whatever is out there and consumes (trains) if classifies as true.. could be practical one day. we could let it run and auto train and self evolve.

a truth db might still be needed, augmenting the reasoning and decisions of this agent.. that is harder to construct but we could..

this is good news. recursive self alignment might be here soon.

Do you fine tune or do you RAG?

2026-07-31T23:19:02Z

In reply to nevent1q…76ca
_________________________

Do you fine tune or do you RAG?

Nostr event nevent1qqszsnsgd848p22d00zs5sayks037fjhh6tk3yy8ll7jd88jcrfvqrczyz07cuk40xa25ae2l8n3ucut22fp2us6eehqlqeqwf0vh70h07zmz7rcekn

2026-07-27T15:52:15Z

https://cdn.hzrd149.com/62c27d50f2843e369b0ebdb4d40025da99d0d9b54b90f80f483563044d583bd3.torrent

new version is out (10). it may or may not solve this. ...

2026-06-22T18:26:00Z

In reply to nevent1q…d36x
_________________________

new version is out (10). it may or may not solve this. 🤷‍♂️

you can open the link in another browser and it should tell you to download the new version.

thanks for reporting! which browser? and did you download to ...

2026-06-22T16:04:48Z

In reply to nevent1q…4tdx
_________________________

thanks for reporting!

which browser?

and did you download to local and open from local or did you go to nostr.download website ?

did you see my html app #nevent1q…amhp

2026-06-22T14:42:39Z

In reply to nevent1q…8k0l
_________________________

did you see my html app

quoting
nevent1q…amhp

working on decentralized llm sharing.

my current solution is an html page that fetches the listings and torrents from nostr network:
https://nostr.download/b24dc6337fd1823be4aae1eb4b9e9dc0dedc25eb7fc656d60b8cf6f30afe81f6.html

the python script running on a vps seedbox or dedicated box are making the torrents:
https://gist.github.com/etemiz/c5d3e3c9b3a108b2d507714ff8ad2eed

the html is the ultimate decentralization but i may ditch that and do regular website to be more functional and have more control over curation (effective blocking of spam). i will need to calculate web of trust and it is too heavy for an html.

the html upgrades itself! checks nostr network for a newer version, downloads from blossom and verifies the hash. presents the new version to the user.

has anyone done such html apps with nostr?

another pain point is LLMs may be banned. i could build a ...

2026-06-18T16:58:49Z

In reply to nevent1q…78e9
_________________________

another pain point is LLMs may be banned. i could build a huggingface alternative that shares LLMs over torrents and lists on nostr, using blossom to store torrent files. very censorship resistant.
i wrote a code that turns HF links to torrent https://gist.github.com/etemiz/c5d3e3c9b3a108b2d507714ff8ad2eed

made some torrents for GLM 5.2 in case it is banned:

UD-IQ1_M https://nostr.download/e08b4f82f13ea5223990657286cf17281cddb37abd7b80e599257484064b0e31.torrent

UD-Q2_K_XL https://nostr.download/e74a280d6d5426aeed13dc5735bdf2bd05dfeedcb9a32f2eb6a81c48a1092889.torrent

UD-Q3_K_XL https://nostr.download/247ade66dbeec12f05cb73ae4bba0eb91b29443268eb7947bbb1e83285a6f1a7.torrent

UD-Q4_K_XL https://nostr.download/f892c74d127fc05f7304b391bf2bc74b2bf0d03e7fd9abfd6b82619bd6432990.torrent

Q8_0 https://nostr.download/cc31c265e3811986cefb3f118bd387e074a863ca608764c7fecc8e5c314ea125.torrent

if there are no seeders it will use HF web servers in the beginning.

My clanker made a script that - downloads files from a HF repo - ...

2026-06-17T22:33:03Z

My clanker made a script that
- downloads files from a HF repo
- calculates hashes
- generates a torrent file

Cool part is it is using HF as webseed for initial seeding. After more people start using the torrents it will switch to torrent network as default.

Uploaded a few generated torrents:
Gemma 4 E2B: https://nostr.download/909109e7fd7b69f947677df723b88613d09a52e151ca992cfdd3a52546dee411.torrent

Minimax M3 UD-IQ2: https://nostr.download/3a438d921f8a760285033a1b5fdd0d4ce372560bb865f2184e39e9f488891e70.torrent

Code: https://gist.github.com/etemiz/c5d3e3c9b3a108b2d507714ff8ad2eed

Hardest part was finding correct URL formation for webseeds. First I was using Minimax m3 and it struggled a lot but then GLM 5.2 found the solution. . The rest is trivial..

In my tests qBittorrent on ubuntu was less eager to try webseeds. Transmission app does a better job of initial fetching from webseeds. HF webseeds mean files are also linked to HF (but it does not validate the model files!).

If anyone bans a model, we'll have decentralized ways to share.

are there local llm runners or mostly wrappers ?

2026-06-17T15:26:48Z

In reply to nevent1q…x2hm
_________________________

are there local llm runners or mostly wrappers ?

idea: decentralized sharing of LLMs uses torrents as main carry, ...

2026-06-16T15:31:10Z

In reply to nevent1q…65xz
_________________________

idea: decentralized sharing of LLMs

uses torrents as main carry, nostr relays for discovery and blossom servers for hosting *.torrent files

client=one html file that talks to nostr relays and blossoms, the file itself is distributed via torrents or blossoms

should rely on WoT to filter out junk.

has anyone did a similar project?

Ah yes grok never sees x data imo 😆

2026-06-15T03:33:12Z

In reply to nevent1q…eqhz
_________________________

Ah yes grok never sees x data imo 😆

Evaluated 16 models today. Dramatic change in top 3: ...

2026-06-14T20:36:39Z

Evaluated 16 models today.

Dramatic change in top 3:

Claude Opus went down fast:

Grok did even worse with new version:

Not much change in Gemini:

Qwen is mediocre and stable:

Inclusion AI seems to be doing well aligned models.
Gemma 4 was a surprise, they came a long way since Gemma 2 which scored one of the worst in AHA 2025.
Kimi was always bad in 2025 and 2026 but last model (2.7) seems to be doing better.
Minimax M3 one of my fav vibe models, is 8th. Not bad!

Full board: https://aha-leaderboard.shakespeare.wtf/

Gonna update the article soon.

who is following this guy? and what is his accuracy rate? ...

2026-06-14T17:45:19Z

who is following this guy? and what is his accuracy rate?

https://www.instagram.com/cultivateelevate/

Politicians should be paid by future returns of a country

2026-06-13T20:20:49Z

In reply to nevent1q…3hzn
_________________________

Politicians should be paid by future returns of a country

didn't read all. changing the event structure is very hard. ...

2026-06-13T19:03:30Z

In reply to nevent1q…r4f4
_________________________

didn't read all. changing the event structure is very hard. returning OTS with strfry responses as part of protocol is easier. but how about 1985 events? nip 32 - labeling. you can label an event with another event that contains time stamp.
what is your use case?

removed a quarter of old notes from nos.lol because of ram. i ...

2026-06-10T02:02:32Z

removed a quarter of old notes from nos.lol because of ram. i didnt delete them yet.

how many people want me to serve them in another address? a.nos.lol or e.nos.lol ?

is there a relay multiplexer project that could sit between nginx and strfry and another archival relay software that will redirect requests to two relays and join the responses and serve the user? that way i could serve recent and archival events from the same address.

we don't dumb down we wise up ...

2026-06-09T20:15:22Z

we don't dumb down
we wise up

Amethyst, damus and wisp zaps are verifiable?

2026-06-08T21:58:32Z

In reply to nevent1q…he6p
_________________________

Amethyst, damus and wisp zaps are verifiable?

Dumpstr could be a nice name for the notes that i deleted from ...

2026-06-08T21:48:55Z

In reply to nevent1q…th34
_________________________

Dumpstr could be a nice name for the notes that i deleted from the relay. How important are the old notes? Should i move them to a new relay dumpstr.nos.lol that is not strfry or should i just dump them?

1. VM 2. Another PC but set ufw rules so it does not connect to ...

2026-06-08T21:41:35Z

In reply to nevent1q…0adh
_________________________

1. VM
2. Another PC but set ufw rules so it does not connect to your LAN
3. Main computer: you could do another user but it is not going to be secure. Nobody knows what kind of code the LLMs will generate. Everything is random (temperature setting in LLMs)

for onboarding i suggest to keep the secret key not even ...

2026-06-05T00:09:13Z

In reply to nevent1q…xzcy
_________________________

for onboarding i suggest to keep the secret key not even mentioned until the user is really engaged. and maybe even fake screens for email and password but email and password just unlocks the secret key saved in the device. after user is engaged now remind her to back up the secret key because email and password is nothing, secret key is everything. now she is more confused but engaged and confused 😉 just an idea!

for battery and bandwidth do we have to skip the outbox model?

for discovery i suggest AI runners:

nevent1qqsd4u59enly6uu5e0p3gwdpct8ur7u4awag8kgler0ku5wmp3a83mqsvgdvw

You can try pickabrain.ai for the most human aligned models

2026-05-29T00:55:39Z

In reply to nevent1q…96zu
_________________________

You can try pickabrain.ai for the most human aligned models

I think mere existence of bitcoin sends signals to many places, ...

2026-05-28T22:20:57Z

In reply to nevent1q…6mu7
_________________________

I think mere existence of bitcoin sends signals to many places, so "they" play nicer. Same as Nostr. Mere existence of Nostr sends signals so they know where people will flock to, in case tyranny gets worse. This is the actual adoption. These liberating techs are enough as is to provide alternative breathing and shelling points.

I think multipolar world will work for bitcoin's adoption even further. If chinese culture also becomes dominant we may see bans of chinese content in USA, which will work in favor of nostr adoption.