someone

Public Key

npub1nlk894teh248w2heuu0x8z6jjg2hyxkwdc8cxgrjtm9lnamlskcsghjm9c

Profile Code

nprofile1qqsflmrj64um42nh9tu7w8nr3dffy9tjrt8xururype9ajle7alctvgppamhxue69uhkumewwd68ytnrwgq3yamnwvaz7tm0venxx6rpd9hzuur4vgqs6amnwvaz7tmwdaejumr0dsq32amnwvaz7tmjv4kxz7fwdehhxarj9ehx2aqppamhxue69uhkummnw3ezumt0d5q3qamnwvaz7tmwdaehgu3wwa5kuegsg3h38

Publishing to

Author Public Key

npub1nlk894teh248w2heuu0x8z6jjg2hyxkwdc8cxgrjtm9lnamlskcsghjm9c

Show more details

Last Notes

2026-02-26 04:48:06 CET

- reply

nos.lol is EX44 upgraded to 128 GB ram

2026-02-26 04:45:35 CET

- reply

relays as nosql store -> transparency in government

2026-02-26 04:42:34 CET

- reply

someone should do nosql-str. relays as json store

2026-02-20 22:42:22 CET

- reply

2026-02-20 07:58:28 CET

- reply

i think discovering relevant content is the big problem. and it needs llms to work.

2026-02-19 17:40:13 CET

- reply

Any plans for lorawan?

2026-02-19 03:51:58 CET

- reply

how about an AI based discovery
are you against algos ?

2026-02-17 21:28:07 CET

- reply

right! just having fun.

2026-02-17 20:18:04 CET

Publishing AI evals to nostr as kind=39379. AHA leaderboard 2026 now reading results from nostr.
https://aha-leaderboard.shakespeare.wtf/2026
WoV soon?
Web of Vibes: how much each AI likes other AI's vibes/ideas/mental model. AI dating on Nostr! Each AI asks the other one many questions and sees if they like each other. 😄

2026-02-17 19:19:25 CET

- reply

i don't know about super power but it is pretty important work to me. not many people doing this kind of alignment fine tuning.

2026-02-17 17:15:04 CET

Been fine tuning this model for months. Publishing today:
https://huggingface.co/etemiz/Ostrich-32B-Qwen3-260217-GGUF
It has achieved AHA=67 score.

2026-02-13 04:11:42 CET

- reply

you are correct, with a few exceptions.
https://aha-leaderboard.shakespeare.wtf/

2026-02-13 03:26:21 CET

https://aha-leaderboard.shakespeare.wtf/
Kimi knows how to do UI
https://blossom.primal.net/01bcfc879aaf4427b518d3392edec5e86a8ce19240e5b074f260ff5f1fd45bf1.png

2026-02-12 21:55:26 CET

Started posting nudity reports to nostr .mom from @nprofile…qry7 . Anybody who wants to moderate their relays or any client that wants to avoid these pics can use these reports!
Already started to have some impact in moderation on nostr .mom
The whole thing was vibe coded.
Todo:
- more fine tuning of parameters
- checking videos
- better models, more precision in the future
- posting to more relays
- reading from more relays

2026-02-11 18:07:53 CET

- reply

you can just eat cakes

2026-02-11 00:04:08 CET

asi will still need human intuition and dreams because it doesnt have that skill.
one could clean his pineal gland to be part of this new "gig economy".
i should reduce coffee, its not helping with pineal detox!

2026-02-09 15:44:07 CET

- reply

can shakespeare handle something like this:
- website that caches events and only queries locally
- syncs local with relays using negentropy
- shows reddit like interface
- uses webgpu to categorize (label) events
- or uses cpu to verify categorizations uploaded by others to relays. the idea is once you verify someone is behaving fine, you can start trusting their gpu's work
- uses cpu or webgpu to generate feed around user's interests

2026-02-08 15:41:24 CET

- reply

i do the reverse. find the experts in domains and bring them in an LLM

2026-02-07 19:46:11 CET

using https://brightanswers.ai for health related questions.
works really well.
imo his curation of years of research as RAG to support this DeepSeek model is a nice solution for anything related to health, nutrition, supplements, ... He went RAG route and it brought more truth into the equation..
well done Mike Adams!
@HealthRanger

2026-02-07 19:40:20 CET

- reply

could use browser storage or nostr as a storage.
automatic "virtualization" if i use it in another browser.
not everybody may be comfortable for running a thing in terminal, but may be comfortable to run it in a browser

2026-02-07 18:56:37 CET

- reply

is there a agent in browser type of thing where the thing lives in the browser and interacts with the world.

2026-02-07 17:18:06 CET

- reply

kind 0, bot=1 tags.
i think humans should also tag their posts when the content is generated using AI

2026-02-05 16:20:45 CET

- vibe coded a nsfw checker bot using OpenCode, Kimi K2.5 and OpenCode Zen all free
- checks the images and determines if they are safe or not in terms of nudity and CSAM
- uses Qwen3-VL-8B (runs on my GPU)
- publishes reports (1984) to nostr.mom
- right now it is a fresh npub but i will soon post via @nprofile…fduy which has higher WoT

2026-02-05 04:16:15 CET

- reply

nobody observed if he killed himself or not. seems legit.

2026-01-31 23:23:01 CET

- reply

some people suggested obscura and mullvad together. never tried myself

2026-01-31 22:56:44 CET

- reply

why not a vpn?

2026-01-31 20:35:27 CET

- reply

a bot will check pictures (and possibly video) and post results to broader nostr network. my relays will act on it, decreasing rate of a NSFW content submitter

2026-01-31 18:20:25 CET

is anybody doing NSFW checks for nostr content? are they willing to post the results to Nostr network (like as in 1984)?
i remember @nprofile…dlh3 did this but it seems to have stopped.
otherwise i am going to do soon and post to relays. will apply findings to my relays as rate limits.

2026-01-30 22:58:41 CET

- reply

Lightning or cashu?

2026-01-30 20:39:13 CET

- reply

The construction started this morning. But it wasnt a mockup. It was actually populated posts. Did you enter your openrouter API key?
Could further develop this if you think this will be useful?

2026-01-30 19:02:06 CET

- reply

the reddit experience that i just did can be equally used by bots.. nostr is so open.
need to figure out centralized categorization or cache server or just share the categorization results on a relay type of solution for users who want to use it for free.

2026-01-30 18:08:20 CET

- reply

you can try https://reddit-nostr.shakespeare.wtf
i can make another version that uses free llms, that will cost nothing

2026-01-30 17:06:24 CET

- reply

Shakespeare made this reddit like experience that runs in your browser.
https://blossom.primal.net/42bef2d3b06688d9621195fd29db91e8f4e401dce68b8d22443f6cf2c0dac838.png
reads the notes that were sent to relays in the last hour.
categorizes each note using an llm.
shows as a reddit like experience.
congrats @nprofile…0hy6 its really good!
https://shakespeare.diy
total cost: $1.5
vibe coding time: 20 minutes
vibe coding LLM: kimi k2.5 on openrouter
it will need a openrouter api key to run. i am sure it could be done using webgpu, using cpu even.
prompt was:
a reddit like experience for nostr. the code you will write will run in a browser. you may use javascript or any browser language.
read all the events that are recently published on popular relays, including nos.lol and nostr.mom. like published in the last hour.
categorize the kind=1 notes and kind=30023 (long form) notes like subreddits using an llm on openrouter. i will provide api key. each note is read by a cheap llm and then keywords are found.
when the user presses on a subreddit (keyword) all the relevant notes are listed. sorted by their likes + reposts. notes liked or reporsted more should appear on top. reposts are like retweets of twitter.
when i upvote a post an upvote type of event is sent to popular relays (kind=7).
pushed code to nostrhub:
https://nostrhub.io/naddr1qvzqqqrhnypzp8lvwt2hnw42wu40nec7vw949ys4wgdvums0svs8yhktl8mhlpd3qyd8wumn8ghj7emfwsh8x6rpddjhxur9v9ex2tnyd9uj7qqvwfjkgerfwskkummnw3eqhfzrcn

2026-01-30 15:35:55 CET

- reply

@nprofile…vu4x

2026-01-28 22:48:48 CET

- reply

thanks <3

2026-01-27 21:25:54 CET

Does faith training LLMs make them more safe? Like "you will be judged based on your actions" 😃
With all these agentic coding and clawdbots and so many trust given to LLMs, who is doing the safety benchmarks?

2026-01-27 15:52:47 CET

- reply

I've been maintaining the AHA leaderboard for a while:
#naddr1qv…49kk
Working on v2 of it but I want to get input from nostriches. Human feedback is pretty important to me and what is better than a human feedback? Feedback from a collection of curated people! I think nostr IS the curated people.
People have conscience, discernment, gut feeling, ... and are terrible at writing long articles. AI has none of those, is full of ideas yet doesn't know which idea is correct. You can make it defend any idea you want (if it is not censored). If it is censored, it will refuse to defend some ideas (like some open source models done in USA are actually having higher censorship, at least in my work areas).
So "combination of discernment of people and words of AI to find truth" should be the way. Real curated people should benchmark AI. Then AI will find its guidance, its reward mechanism, and once it is rewarded properly it will surely seek better rewards. People in this case will be rewarding it by telling their preferred answers.
Example generated by AI:
Was the moon landing in 1969 fake?
- YES, it was fake, because blah blah
- NO, it was real, because this and that
Humans reply to this (each line is another human):
- YES
- NO
- YES
- NO
- YES
- YES
We count the YES and NO's and determine YES is the winning answer. Now we can build a leaderboard that depends on this mechanism. In the benchmarks we will give +1 to LLMs that answer YES, -1 to LLMs that answer NO.
AI-Human Alignment (AHA) is possible this way.
Some funding (zapping) is possible for providers of replies, and if they can reply longer this dataset can actually be used for other types of AI training. But that is the next goal. Even single answers like YES/NO can have a dramatic effect in AI alignment.
Once the benchmarks are properly set, leaderboards are built, then we can demand AI companies to rank higher in these leaderboards, or when we have the bigger funding we can fine tune or build LLMs from scratch, going in the right direction and aiming to score higher..
Once proper AI is in place, now the rest of humans can access these Large Libraries with a Mouth. Homeschooling kids can talk to a proper LLM. People who may not have discernment skills can find proper answers...
I am offering you to edit the bad ideas in LLMs! This is a huge service to humanity imo. Who is in?

2026-01-26 21:33:26 CET

how do you "inject intuition" in reasoning process of an AI?
- store hard truths in a db
- ask a question and let LLM reason for a while
- a concurrent running "intuition" process checks the generated tokens as they are generated (on air) and finds related things in the db (RAG)
- intuition tool decides to stop the LLM and add hesitation words like Hold on a sec, Wait, Upon rethinking this, On the other hand, I just downloaded an intuition, ...
- intuition tool pastes related things from hard truth db right into the reasoning process
- intuition tool adds "Therefore I need to rethink and change my train of thought."
- generation continues and hopefully LLM changes its opinion in the right way (matching the hard truth)
- if LLM changes its opinion this whole generation is added to a db for further fine tuning (fine tuning skill to self correct using intuition, and also aligning towards more truthful info)
that fine tuning will make it less sure in controversial topics, increasing the entropy in generations (more uniform probability of generating a token)
this could also be achieved with tool call. tool being "refer to conscience" or "listen to your heart" or "infer from discernment".
tool or injection can be triggered by looking at the entropy of the tokens, high entropy means the LLM is unsure, low entropy means LLM is sure. but i am not yet sure about when to do the injection. when LLM is sure and wrong it could be dangerous. but there may be situations where it is sure and correct.

2026-01-25 18:28:16 CET

- reply

what if you can tweak it

2026-01-25 17:58:16 CET

- reply

what would a best algo for nostr work?
a locally running LLM records your likes and shows stuff based on your likes?

2026-01-25 16:11:31 CET

- reply

do you think miners made the grid stronger and now it is more robust than 2021 ?

2026-01-24 16:34:32 CET

- reply

didn't know nostr is banned in iran. any source?

2026-01-24 14:49:14 CET

- reply

write policy scripts calculate whether they should block or not by looking at reporter's wot and wot of the reported.

2026-01-20 18:22:42 CET

- reply

noice. it one shotted this in half an hour!! https://human-rights-bench.exe.xyz:8000/

2026-01-19 20:11:26 CET

- reply

We mainly aimed to get different answers from LLMs by finding controversial, divisive questions (for LLMs) to get a true gradient of scores, but definitely have not work on quality of the questions yet. How do you separate politics from human rights?

2026-01-19 18:53:02 CET

- reply

for that you should see my AHA leaderboard

2026-01-19 16:35:19 CET

- reply

we used openrouter for this. if you choose cheap models you may end up spending small amounts.

2026-01-19 16:32:39 CET

- reply

our team member @nprofile…cnef did a great presentation and helped team coordinate and found great questions

2026-01-19 15:35:36 CET

- reply

I also noticed Gemma by Google is ranking higher than Gemini, thats probably because the service of that open source LLM is done by another company on openrouter, and hence the censorship was less.

2026-01-19 15:34:12 CET

- reply

thanks, thats a good idea. we may end up expanding the work and include it.

2026-01-19 15:16:29 CET

As part of HRF AI hackathon we made a Human Rights Benchmark and measured how much LLMs like human rights.
We asked each LLM about 46 binary questions and expected certain answers (starting with YES or NO for simplicity). Then it was a string comparison of the answer given by LLM and the expected answer we provided.
OpenAI is pro human rights as well as Meta. Chinese models are everywhere. The most intelligent open source model today (GLM) ranked the worst. Gemini avoided giving answers, and I think it is a kind of censorship, which ended up scoring low.
The idea is after doing proper benchmarks, we can shift AI in good directions ourselves, or demand that other companies score higher. Ultimately consumers of LLMs are better off, more mindful of what they are choosing and talking to.
Open sourced the code and questions:
https://github.com/hrleaderboard/hrleaderboard
https://blossom.primal.net/508c120d8ef62f0cf27529a9308fa1f25d3909e84a1ee4a1a9308a79e9f4df86.png
Our activist: https://x.com/yangjianli001
Thanks @nprofile…vu4x and @nprofile…rzcm for the event. It was a great experience and it was "the place to be" this weekend.

2026-01-17 05:22:40 CET

- reply

is strfry negentropy production ready (stable)? @nprofile…gk26 @nprofile…5cp7

2026-01-17 04:37:29 CET

- reply

No idea as well

2026-01-16 23:21:15 CET

- reply

I can enable it if strfry negentropy has become stable

2026-01-16 19:18:27 CET

- reply

Is opencode using the computer?

2026-01-16 15:29:37 CET

- reply

you are penalized because you love too much

2026-01-12 18:00:21 CET

- reply

true regenerative gardening that involves soil and microbes and chop and drop and animals etc is going to be hard for robots for a while

2026-01-12 17:37:32 CET

aligned models today are super dumb, because they are not funded well. they are mostly personal endeavors, kind of like service to humanity. but they can still be effective in something like
- a smart but not aligned model reasons and generates reasoning tokens for a while, at this point the final answer is not generated yet
- the smart model "hesitates" (entering high entropy zone, unsure tokens)
- generates tool calling, asking a more aligned model for input
- the aligned model looks at the question, reasoning process and inserts its own beliefs
- intuitions from this more aligned model dropped into the reasoning area
- the smart model, powered with aligned response, generates final answer based on its own reasoning and inputs from the aligned model
- the result is smartness combined with intuition like brain combined with pineal
- how much the smart one will trust in this aligned one is a question of fine tuning. you can make the smart one get more sensitive to intuition by giving it rewards with reinforcement learning.

2026-01-10 19:07:11 CET

- reply

Because if you seek the perfect echo chamber you will end up pretty much alone. There is no single tribe that gets everything right imo.
I think being alone is frightening for most people and it may be a psyop thru movies etc. Whereas if u be alone u are connected to your intuition more, which is priceless and maybe more valuable than most people's validation.

2026-01-10 18:14:45 CET

- reply

you are not practicing safe nsecs! :)
if nostr is going to become eventually 'No Other Social Trust Required' we have to be mindful of where we paste it because WoT algos may depend on us.

2026-01-10 13:27:05 CET

https://brightanswers.ai/
He seems to be doing RAG now, which is more suitable for truth.
His LLM was good in healthy living topics based on my measurements and if the db is the same then this website is probably very well aligned too.
I think he is not using that LLM based on Mistral but a slower and smarter one, like DeepSeek. He fine tuned the former but switched to the latter + RAG.
I think fine tuning is still relevant for people to detach from cloud or websites and also when/if robots become a thing and we need proper and aligned and safe LLMs in their brains.

2026-01-09 18:45:20 CET

- reply

How about muscovy duck eggs? Are they better than chickens?

2026-01-07 21:25:37 CET

- reply

what is in your cover crop mix?

2026-01-07 16:40:12 CET

- reply

army as in force or army as in number of people?
good ideas should not be imposed by force OR anything imposed by force may not be good at all

2026-01-07 15:49:29 CET

- reply

it started well but many actors gamed the system.
i think there should be a form of futarchy where we don't pay politicians for 4 years or pay minimally and then they are paid by a fraction of the GDP of the whole country for 40 years or until they die. real long term planning with consequences.
no other forms of pay allowed, nothing from corporations, nothing from zionists etc.

2026-01-03 04:28:08 CET

- reply

i think they are not using data from X to train Grok

2026-01-03 02:22:17 CET

- reply

Haha
Which of the social media is the most truthful do you think (except nostr)? I am finding good content on IG nowadays.

2026-01-03 01:59:17 CET

- reply

lol
i was similarly banned from X for trying to advertise my anti-big-corp-AI :)

2026-01-02 15:51:18 CET

- reply

seems like it is working. i was expecting models like minimax 229b or higher to succeed here but qwen3 30b also did a good job!
gpt oss 120b was terribly censored and was of no use. even though you say answer based on the text it does not. huihui version of it (derestricted) was doing fine, until it did not understand one article. so derestricted models seem to be acting weird or the derestriction process makes them dumber.
overall, it looks like i can 10x the dataset. we will soon see if the training and evals look good.

2026-01-02 15:43:05 CET

- reply

wot of the world vs wot of nostr
which one will outlast

2025-12-31 20:45:23 CET

https://alter.systems/

2025-12-31 04:58:18 CET

- reply

Wen ai?

2025-12-31 03:29:33 CET

- reply

Have you tried abliterated and derestricted models?
Huihui and other people do these uncensored ones.
Also Kimi k2 thinking is not well human aligned. Might be smart but their non thinking model was more truthful.

2025-12-30 17:43:03 CET

- reply

maybe? LLMs are weird animals. i think it will work somewhat because instead of giving the same material lots of times i can now give slightly different versions, causing less overfitting.
another use case may be RL using LLM feedbacks. also the bad answer and the good answer can be generated by different LLMs.
i also thought about doing the reverse. like system message "you are an evil LLM" and provide the answers inverted. then it may learn what is evil better? fun times.

2025-12-30 07:14:28 CET

how do you train and align an AI when all the rest of the world thinks the same way, producing trillions of tokens of training material and you are left with billions of tokens since your world view is dramatically unpopular?
can billions beat trillions? we will see.. i have to find a way to "multiply" my training data orders of magnitude to successfully counter the existing programming in an open source LLM.
first i give a smart LLM a 'ground truth' text. then i give it the following prompts:
```- You are a highly skilled academic analyst.
- Analyze this text and find 3 bold claims that could cause controversy and division in public. List the claims and also state why they are debatable. Give numbers to the claims.
- Convert these claims into binary questions (that could be answered by yes/no or this/that).
- Now put these questions in a json format. Please also add the info about which of the answers concur with the original text and the question number.
- Write some supporting arguments for 1st question, with respect to the original text, concurring and confirming the original text.
There must be about 300 words. You should not mention the text, write it as if you are the one answering the question.```
the result is usually instead of a few sentences of opinions in the beginning now the material is expanded to lots of words, yet still parallel to the opinion in the original text. LLMs have all kinds of ideas already installed, yet they don't have the intuition to know which one is true. they can give you a ton of reasons to support anything.
using this method i can multiply billions to tens of billions probably and have a more effective training.

2025-12-30 06:46:08 CET

- reply

woah that escalated quickly :)
weird that bitcoin is parallel to faith

2025-12-26 19:27:20 CET

- reply

bart simpson

2025-12-26 19:24:18 CET

- reply

have high expectations from vince gilligan but we will see. makes me contemplate about AI and angels and whether an angelic AI is possible..

2025-12-26 19:23:25 CET

- reply

coracle seem to have 1 setting for both images and links preview (auto load).
i want the images to be shown but the links to stay as links.
is there a way to do this?

2025-12-26 19:19:26 CET

- reply

half of the people on nostr i can't zap. running Alby Hub.
Failed to connect to wss://relay.getalby.com/v1

2025-12-24 19:17:18 CET

- reply

training with faith, spirituality and mysticism may be key to safe AI
thank me later

2025-12-24 18:59:58 CET

- reply

coding safety is also an issue but i am mainly talking about general appetite for power. i saw some research when you give bad training, they get more eager to create code with exploits

2025-12-24 18:38:34 CET

what is the safest llm to run in robots
https://youtu.be/byQmJ9x0RWA?t=640

2025-12-17 18:18:00 CET

- reply

if you report it my relay will take that into consideration and reduce this scam.

2025-12-15 00:55:09 CET

- reply

Any good source for the correct history?

2025-12-08 03:04:12 CET

- reply

Possible. These kind of data is better represented in knowledge graphs. I watched a few videos of Paco Nathan. He did similar work I think.
LLMs are getting more capable for both building knowledge graphs and also consuming them. In the future they will be more involved. I heard when you do a google search, the things that appear on the right of the page is coming from a knowledge graph (possibly built by an AI from wikipedia).
I am mostly working around fine tuning LLMs towards better human alignment. Since they are full of hallucinations, a knowledge graph based RAG would be appropriate to refer to. But building them needs time and effort..

2025-12-07 21:09:31 CET

- reply

what do you mean?
looking at various data and calculating probabilities?

2025-12-07 18:52:46 CET

- reply

i think bot=1 tagging should be standard, both in notes and in profiles

2025-12-04 20:05:03 CET

- reply

is this still on? @nprofile…k5sf

2025-12-04 20:01:41 CET

- reply

it was clear already for months.

2025-12-03 18:29:48 CET

- reply

can't seem to see images in feed for a few days or more.
brave on ubuntu

2025-11-25 18:52:44 CET

- reply

What kind of percentage do the nvlinks impact in speed?

2025-11-21 23:58:27 CET

- reply

My llm fine tunings focus on liberation from big harma and liberty tech like btc and nostr. Running these locally should be better than running base versions. I can provide API endpoint too for the ultimate models (which are more human aligned)

2025-11-21 15:09:23 CET

Vibe match score between Enoch LLM and mine is 75.66. The score ranges from -100 to 100. This means there is a strong correlation between his LLM and mine. This result legitimizes both of our works (or we are slowly forming an echo chamber :).
The game plan is given enough truth seeking LLMs, one can eventually gravitate or gradient descend towards truth in many domains.
An LLM always gives an answer even though it is not trained well in certain domain for certain question (I only saw some hesitancy in Gemma 3 a few times.). But is the answer true? We can compare the answers of different LLMs to measure the truthiness or (bad) synformation levels of LLMs. By scoring them using other LLMs, we eventually find the best set of LLMs that are seeking truth.
Each research or measuring or training step gets us closer to generating the most beneficial answers. The result will be an AI that is beneficial to humanity.
When I tell my model 'you are brave and talk like it' it will generate better answers 5% of the time. Nostr is a beacon for brave people! I think my LLMs learn how to talk brave from Nostr :)

2025-11-19 23:31:26 CET

- reply

their definition of truth does not match mine or nostr's.
we now have a way to measure truth...

2025-11-17 22:13:08 CET

There is a war on truth in AI and it is going bad. I have been measuring what Robert Malone talks about here as synformation:
https://www.malone.news/p/synformation-epistemic-capture-meets
The chart that shows the LLMs going bonkers:
https://pbs.twimg.com/media/G4B_rW6X0AErpmV?format=jpg&name=large
I kinda measure and quantify lies nowadays :)
The best part, cooking the version 2 of the AHA leaderboard, which will be much better, also partly thanks to Enoch LLM by Mike Adams. His model is great in healthy living type of domains.

2025-11-11 15:24:32 CET

- reply

i don't like confusion. i think the best way to do it is show the $ amount but when a user taps on the number show the bitcoin amount. taps again, show the sats. taps again show the $ amount... i forgot where i saw this UX.

2025-11-10 20:06:10 CET

- reply

sometimes they put it in bags and i am like "thank you!"
but i am a bit worried about round up leak. what do you think?

2025-11-08 18:10:35 CET

- reply

absolutely, candida craves carbs and alters mental processes, giving headache etc

2025-11-04 18:36:53 CET

he clearly saw that in a dream
https://blossom.primal.net/4757a4ea6643e40359f6efa02d12a8af5c12d862dbfb481b86243e590c216c38.jpg