semisol

👨‍💻 software developer
🔒 secure element firmware dev
📨 nostr.land relay

all opinions are my own.

Public Key

npub12262qa4uhw7u8gdwlgmntqtv7aye8vdcmvszkqwgs0zchel6mz7s6cgrkj

NIP-05 Address

Profile Code

nprofile1qqs99d9qw67th0wr5xh05de4s9k0wjvnkxudkgptq8yg83vtulad30gpz3mhxue69uhhyetvv9ujuerpd46hxtnfduq35amnwvaz7tm5dpjkvmmjv4ehgtnwdaehgu339e3k7mgzdkzh3

Publishing to

relay.damus.io

theforest.nostr1.com

aggr.nostr.land

Author Public Key

npub12262qa4uhw7u8gdwlgmntqtv7aye8vdcmvszkqwgs0zchel6mz7s6cgrkj

Show more details

Last Notes

2026-05-06 03:20:28 CEST

- reply

Simulation with *scripted input events* instead of variable latency computer use. This type of stuff is also used for UI testing in general

2026-05-06 02:24:56 CEST

- reply

I do not use the Codex harness anymore so I am not sure. But I think none do it by default

2026-05-06 02:15:38 CEST

- reply

Then have two independent context windows, one for solving the problem, one for testing if it works. Do not let them interact except a set of screenshots and reproduction guide

2026-05-06 02:14:15 CEST

- reply

Make an extension or similar that keeps it going until it is done.
If you were using Pi then you could just ask it to do it by itself ;)

2026-05-06 02:09:21 CEST

- reply

gm
why not use a loop that won’t let it stop until done?

2026-05-06 02:08:58 CEST

I see too many developers micro-optimizing the least critical path of their application, then leaving the actually important paths unoptimized.
Your code’s performance isn’t determined strictly by clock cycles

2026-05-05 16:05:36 CEST

- reply

gm marcel!

2026-05-05 15:59:32 CEST

gm nostr 🤙 🍍
#nevent1q…dsv6

2026-05-05 01:51:56 CEST

do not anthropomorphize matrix multiplications

2026-05-05 00:48:27 CEST

- reply

Why would you do that.
Anyway, I replaced my usage of the chat style apps as well with Pi. As I can configure the system prompt and extensions.
I’m considering making an MCP for FS + bash just so I can give GPT 5.5 Pro (which is UI only) a full featured environment.

2026-05-05 00:41:26 CEST

- reply

also it doesn’t lie to you with fake sandbox bs. I put it in a dedicated LXC container and may put it in a VM too

2026-05-04 23:12:55 CEST

- reply

tried codex first, but not supporting 3rd party providers was a no go + didn’t like the output streaming buffering
tried opencode but had too much flashiness, and bloated, slow and ate 500MB of RAM
looked at aider but it was *too* simple tbh
pi was the only great one really, simple but also extensible

2026-05-04 23:11:15 CEST

- reply

great, the pi-pilling is spreading in the ⭕️

2026-05-04 22:46:26 CEST

〰️〰️〰️〰️〰️

2026-05-04 22:39:30 CEST

- reply

yes or you have to allow an update mechanism which either requires trust in a 3rd party or the operator
the “DoS” concern can also be selectively allowing mint operations with say 50% fee, so they can basically rug anyway as users will choose 50% instead of 0%.
and there’s the question of how you delegate keys across restart. answer: trusted 3rd party

2026-05-04 22:16:24 CEST

- reply

Also do not use a Pi

2026-05-04 18:04:37 CEST

- reply

I have been considering using Fossil for this reason

2026-05-04 17:07:59 CEST

C2PA is the equivalent of requiring Google Play Integrity for photos and videos.
While it is good for when authenticity is required, it SHOULD NOT have a place on any online social platform or most of the internet.
This system will eventually be abused by bad actors and extended to allow per-device tracking if it is used as such.

2026-05-04 16:31:22 CEST

I accidentally upgraded to Sonoma. Oops
nvm there’s FSKit that is amazing

2026-05-04 16:29:30 CEST

- reply

hi jennifer!

2026-05-04 16:29:23 CEST

- reply

gm marcel ☕️

2026-05-04 16:29:15 CEST

- reply

gm oshi 🫡

2026-05-04 15:23:43 CEST

- reply

oh hi you are still here?! 🫡

2026-05-04 15:20:53 CEST

gm nostriches 👀 🍍🍍🍍

2026-05-03 20:43:45 CEST

- reply

$200/mo? or 5.5 xhigh?

2026-05-03 20:01:10 CEST

- reply

Which plan and why did you do fast mode.
I suffered through a week of GPT 5.5 with no streaming and tbh it is not an issue. If you need interactive you can usually get away with 5.4 Mini anyway

2026-05-03 19:59:49 CEST

- reply

While they have an intelligence of a banana, a banana is still able to download the wordlist from source code and write a script to identify the information theoretically optimal choice

2026-05-03 17:55:54 CEST

- reply

Anyway, having a huge pipeline is only going to annihilate your Fmax. Instead have a “programmable ALU” with 128/256 bit path, RAM to store registers and operation microcode in ROM-ish blocks
It’ll still be faster than CPU, this is how most hardware accelerators for ECDSA work

2026-05-03 17:53:40 CEST

- reply

how seriously should I treat this?

2026-05-03 16:03:05 CEST

gm nostriches 👀

2026-05-03 15:04:25 CEST

- reply

I am considering a program where large scale users could register for higher limits. For Bostr this should only affect public instances. And higher limits will be available in general, which should allow running personal instances with 0 issue
In that case, I’d want some way of passing a rate limiting identifier down (like hashed version of user IP/PK) so that if someone is abusing the proxy, only the abusers can be blocked

2026-05-03 15:02:05 CEST

- reply

The 3rd party moderation services idea is to cover the server-heavy stuff. :)
For aggregation I am considering eventually implementing per-user aggregators as well on top of the primary one, and personal relays with more relaxed policy than the main relay.

2026-05-03 00:01:37 CEST

- reply

The read is in their DB so I’m assuming incompetence of not being able to render nevent links properly
#naddr1qq…klg5

2026-05-02 23:44:19 CEST

- reply

The post actually exists but clients do client things. Ugh

2026-05-02 22:43:42 CEST

- reply

especially would want your opinions: @npub1aeh…4nwx @npub1r0r…q9ac @npub1ye5…knpr @npub1x3a…g4el @npub1m64…uaks @npub1s02…fxtw @npub1m4n…c2jl @npub10np…tl5h

2026-05-02 22:41:09 CEST

- reply

Ah, no, I’m mostly thinking about a general framework for moderation (3rd party services + user + client + relay), and separate from that a set of baseline rules for *my* part of implementing it, especially as a large relay operator.
For doxing, I’d say the expectation of privacy is the important part. Someone that is anonymous + took precautions shouldn’t get photographed at a conf close up without redaction, for example.

2026-05-02 22:35:27 CEST

To #devstr and vibers: What are your main pain points with Nostr relays and infrastructure like search in general?
Looking for some views on this topic (if it’s a vibecoded app, ask your LLM)
Please also share this if you think people might be interested.

2026-05-02 22:32:36 CEST

- reply

“Unconfigurable” being the service having a fixed policy. If you don’t want it, go somewhere else.
Also, hosting PII is disallowed by many data protection regulations.

2026-05-02 22:10:07 CEST

- reply

One of my main goals is to keep the “unconfigurable” moderation as low as possible, like spam, doxxing, CSAM, etc. and allow users to decide whether they want more subjective stuff like AI slop.

2026-05-02 14:51:36 CEST

gm nostriches 🍍
we don’t have enough long form content, I think

2026-05-02 14:12:28 CEST

- reply

I didn’t, where is it?

2026-05-02 13:32:37 CEST

- reply

With Nostr that barrier doesn’t exist as much.

2026-05-02 13:03:04 CEST

- reply

No we need a STARK and ZK proofs so we can sound cool and quantum.
(I wonder if these ZK schemes have been audited against QC risks. I feel like a good chunk will fail)

2026-05-02 13:01:46 CEST

- reply

Ah, the point is that relays, clients and moderation providers can basically become a large market that users can pick what they align with from.
So you can have scale *without centralization*.
Independent monitoring + the user having visibility into moderation actions would prevent malicious behavior.
If anything happens, users can always switch relays/clients/providers.
Most people didn’t read the article 😅

2026-05-02 12:55:16 CEST

- reply

Depending on how much capability you need, use the on device ML/Foundation Model APIs.
For larger use cases add OpenRouter+OpenAI+Anthropic key input and possibly OpenAI OAuth2 for using Codex limits

2026-05-02 12:53:58 CEST

- reply

Also: Do you know what else would br able to prove control of keys?! Committing to a hash of a signature!
The signature is timestamped so you can prove it was pre-QC, and to anyone it would look like a normal OP_RETURN until revealed
But no we must have STARK and other bullshit just so the funding flows

2026-05-02 12:51:41 CEST

- reply

Unfortunately it doesn’t, if we want Nostr to have more diverse content instead of the current monoculture. But for people that want that, they can have their own relays :)

2026-05-02 12:50:51 CEST

- reply

It’s mostly meant to cover copyright and content stealing farms, but yes.
This is just supposed to be a rough outline.

2026-05-02 12:48:42 CEST

- reply

Which article specifically, moderation or NFDB?

2026-05-02 12:46:56 CEST

- reply

The post describes something similar to this with users having transparency into decisions, and clients giving full control on what relays to use, and if they want any filtering providers such as NSFW media scanning to hide that content for *them*.

2026-05-02 04:56:35 CEST

Made a short read about Nostr moderation.
Had to make the event JSON myself like a caveman due to all longform clients sucking
#nevent1q…68t2

2026-05-02 04:26:11 CEST

- reply

oh. they left.

2026-05-02 04:25:33 CEST

@npub107j…ncxg is habla down?

2026-05-02 03:01:25 CEST

- reply

Relays could also do this.
I kind of feel like numeric kinds were a mistake, as they end up being about as useful as text kinds anyway. Fields being arbitrarily nameable already allows convergence via shared ideas (`published_at` will likely be a unix second timestamp to *anyone*), and even if there are multiple specs, having numbers vs strings doesn't change that

2026-05-02 03:00:06 CEST

- reply

For Nostrhub, I do not like slop apps, sorry.
I also have no interest in dealing with NIPs-over-Nostr solution after having tried to get said solutions implemented + looked at what others were doing.
A specifiction is a text document, and personally, I am going to put it on my own repostiory and host it there.
I don't need much more than that, and neither do implementers

2026-05-01 19:50:53 CEST

It is exhausting to work with the NIP process, with fiatjaf overriding decisions constantly and adding/removing maintainers with no purpose than whoever is the best parrot.
I think each client should publish their specification, ask for comments & improve, then release.

2026-05-01 19:47:31 CEST

Faster negentropy alternatives 🤔

2026-05-01 17:05:23 CEST

gm nostr 🍍
the 🍍 will continue until 🍍

2026-05-01 17:02:44 CEST

I do not regret being subscribed to RIPE mailing lists
https://i.nostr.build/Ql7lNfeaoGyLtSOA.png

2026-05-01 14:50:17 CEST

Whoever is making these chat templates for models: please stop.
ChatML is already shit. Then model labs are now bolting on reasoning via *non-special* tokens, then trying to make tool calls by adding some absurd XML-like <tool_call> prefix to assistant messages to start tool calls, etc.
Special tokens were created for a reason and now even the competent labs are doing this type of garbage

2026-05-01 13:14:24 CEST

- reply

Ok, nvm seems they broke the demo. But the weights can be downloaded

2026-05-01 13:10:33 CEST

- reply

I am not even sure it this is SOTA anymore but I tried this one and it was impressive.
Voice cloning worked with only 3 seconds of data…
https://huggingface.co/spaces/Qwen/Qwen3-TTS

2026-05-01 11:54:44 CEST

- reply

Open source TTS models are getting too good I think, so the importance of paid APIs diminishes

2026-05-01 01:53:08 CEST

- reply

For some people apparently it also erases your session limit

2026-05-01 01:36:41 CEST

- reply

Nostr.build

2026-05-01 01:36:08 CEST

Apparently coding agents love using “fossil sql” to directly query the SCM database. Neat

2026-04-30 20:21:06 CEST

- reply

No, what I mean is unless the video length is changed, ffmpeg can easily swap it.
Even if it changed, you can rerender only changed parts and splice it in with minimal reencode to base video then overlay on top the audio. Need to have some sort of anchor to splice though but probably easy

2026-04-30 19:13:17 CEST

- reply

Technically audio track can be swapped without video change

2026-04-30 19:08:43 CEST

- reply

What did you find? Weird compiler optimizations?

2026-04-30 19:08:26 CEST

ga nostriches 👀

2026-04-30 16:51:54 CEST

- reply

This is when you trash the context window as it has been contaminated.
Also if this is happening set your thinking level to high (not xhigh), as especially Codex has a tendency to be lazy at medium/low thinking for some reason. (Probably RL caused)
I have a simple benchmark, ask the model to create 100 image files. Then ask it to read one by one in parallel to its context window.
Medium thinking will say “I can’t load all of those”, high will happily do it

2026-04-30 16:15:14 CEST

- reply

incorrect
Non-thinking mode doesn’t “think immediately”, but it cannot go back from mistakes. These are still very capable and and have pretty low hallucination rates.
Thinking mode does NOT have a “validity checker” loop. It just thinks in one stream, that’s all
In many models, the model doesn’t even draft the response in the thinking block!
The model can in some cases even overthink / constrain itself, leading to worse outputs.
There is also no sort of cost / quality weighing either except a token count cap.
Note that thinking models suffer from the bad token problem as well, both in their CoT and final response:

2026-04-30 00:45:39 CEST

- reply

Do not have concurrency>1 for any single websocket connection

2026-04-29 20:40:33 CEST

- reply

Limit is advisory anyway. Relays can send less than what you ask for without any issue, and you should still paginate until you get enough.
Only “no events” should be authoritative as nothing remains, even then…

2026-04-29 19:26:26 CEST

- reply

🤫

2026-04-29 18:12:18 CEST

- reply

gm zorbar

2026-04-29 18:12:09 CEST

- reply

gm 🍍 🍍 🍍

2026-04-29 16:18:53 CEST

gm nostriches 🤙 🍍

2026-04-29 15:24:02 CEST

- reply

Claude Code isn’t either but it’ll give you more limits than “whatever Anthropic feels like that day”

2026-04-29 15:21:41 CEST

- reply

Or use Codex

2026-04-28 16:26:04 CEST

- reply

No one runs bf16 in real world use cases except for small models, too much

2026-04-28 15:33:52 CEST

gm nostriches 🍍

2026-04-28 15:32:42 CEST

- reply

No it’s worse.
They let the model have access to a staging environment and manage it itself, okay.
but they also put a API key with full Railway (their hosting) access for managing domains in the same workspacd.
The model then found the key and deleted a production volume. Also deleted their snapshots of the volume because they were linked.
And then they blame an API for not having *are you sure* and manual confirmation via SMS/etc (wtf?? It’s an API!) and blame Cursor for not having guardrails that catch every destructive API action under the sun.
Like wtf

2026-04-28 15:23:12 CEST

- reply

Code always was cheap, architecture and good design wasn’t

2026-04-27 16:01:48 CEST

- reply

gm 👀

2026-04-27 16:01:41 CEST

- reply

gm 🫂

2026-04-27 15:48:08 CEST

gm nostriches 🍍 ☕️

2026-04-27 15:47:46 CEST

- reply

A vibeslopper does not want to own up to their mistakes and blames others instead. Not surprising.
“We pay Cursor for the most expensive model so surely we must get a 100% reliable one. And why didn’t their guardrails blocking destructive actions not include a very specific API?!”
“We pay Railway for hosting, and they can’t undo deletions after we effectively told them to dispose of everything?! Why didn’t they have double checking in an API, why aren’t you assuming the consumers are fallible humans/slopgenerators?!”
Expects recovery SLAs for their own data deletion.

2026-04-27 15:39:03 CEST

- reply

This is something I see as well. I see these complaints significantly more from people that:
1. Are not vibe coders
2. Have done proper software work for at least 2 years
I think it is to some extent a lack of good vs bad code in vibe coders, who hype it up. Experienced devs don’t use it at all, say nothing or do not get as much attention.

2026-04-26 21:19:00 CEST

δ?
1.0
1.0 ⬅️
0.1 🕚
0.9 🍍

2026-04-26 16:21:52 CEST

- reply

🍍🍍🍍

2026-04-26 16:16:34 CEST

gm nostriches 🍍🍍🍍

2026-04-26 16:16:11 CEST

- reply

I don’t know what they are doing but their classifier has started flagging random shit as cybersecurity risk.
One example: Giving it access to a browser with computer use and asking it to test Nostr clients

2026-04-25 15:54:56 CEST

gm nostriches 🤙
🍍 🍍 🍍

2026-04-24 18:22:45 CEST

gm nostriches 🤙 🍍

2026-04-23 22:52:08 CEST

- reply

If you want to explore the embedding approach though, I’d recommend CatBoost as it does something similar with internal kNN matching.
It uses decision trees which should have better performance than a basic kNN, can accept other features as well, and is still lightweight to run

2026-04-23 22:50:16 CEST

- reply

they are effectively the same thing, one just has more tunability.
a basic dot-product based classifier could be a classification head just by setting each output’s weights to the embed.
I think a gated MLP may work better though, and would allow a lower LR for the main weights that could reduce OOD shift. Compared to a direct one it would allow some nonlinearity and a higher intermediate representation
Could also intentionally bottleneck training by having LoRA instead of full FT on the base model (but usually not worth it when considering these are <1B param)

2026-04-23 22:20:44 CEST

- reply

This is pretty similar to how most classification models work anyway.
Most embedding models, and base models for classification like BERT share the same bidirectional transformer architecture.
With an embedding you use the hidden state (or with a final downproj) on a single token or all tokens averaged.
With BERT and similar you have to add the classification head yourself that downprojects from hidden state.
There you also have the advantage of training the whole model, but close enough concept.

2026-04-23 16:55:12 CEST

gm 👀 🍍

2026-04-22 22:45:08 CEST

- reply

In progress! It’s turning out better than originally planned
A bunch of really big things have happened recently regarding it, but I can’t disclose for a while due to certain requirements