semisol

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

all opinions are my own.

Public Key

npub12262qa4uhw7u8gdwlgmntqtv7aye8vdcmvszkqwgs0zchel6mz7s6cgrkj

NIP-05 Address

semisol@nostr.land

Profile Code

nprofile1qqs99d9qw67th0wr5xh05de4s9k0wjvnkxudkgptq8yg83vtulad30gpzamhxue69uhkummnw3ezuum9d45hxmmv9ejx2as7dkvh9

Publishing to

nostr.semisol.dev

Author Public Key

npub12262qa4uhw7u8gdwlgmntqtv7aye8vdcmvszkqwgs0zchel6mz7s6cgrkj

Show more details

Last Notes

2026-05-10 00:38:43 UTC

my review of GPT 5.5:
- it is much better at processing instructions and implict things I think
- still a slopgenerator
I still do not see the value of these models for writing code.
They work well if you are using them for trying random stuff with limited time investment, or testing

2026-05-10 00:33:35 UTC

looking for relay software to test, pleaee send links

2026-05-09 19:49:47 UTC

- reply

is good. otherwise still 🍍

2026-05-09 19:46:28 UTC

- reply

working on 🍍

2026-05-09 19:46:17 UTC

- reply

I mean what isn’t seamless about OTS. You drag and drop file and done

2026-05-09 17:34:37 UTC

Did you know: LN invoices can embed onchain addresses as a fallback

2026-05-09 14:12:25 UTC

- reply

gm!

2026-05-09 14:12:20 UTC

- reply

gm yodl

2026-05-09 14:12:16 UTC

- reply

gm laeserin

2026-05-09 14:12:12 UTC

- reply

are you really sure you want to know?

2026-05-09 13:47:04 UTC

gm nostriches 👀 🍍

2026-05-09 11:28:00 UTC

- reply

Just noticing this now? 🥲

2026-05-09 09:34:12 UTC

- reply

Start at Plus upgrade as you go
Also don’t use Shakespeare, it is a toy I’d say. Running a proper harness in a sandbox will give you much more power. I’d recommend Pi

2026-05-09 09:21:49 UTC

- reply

I don’t use credits. I use the pro sub.
I use 5.5

2026-05-08 23:53:36 UTC

This is a bad idea
#nevent1q…090x

2026-05-08 21:47:20 UTC

Wait, people are using nsecs (that are leaked all the time) for money now?

2026-05-08 21:35:14 UTC

- reply

Paid sub?

2026-05-08 16:46:26 UTC

- reply

Try OpenAI codex

2026-05-08 16:07:40 UTC

- reply

2026-05-08 16:07:32 UTC

- reply

gm kidwarp

2026-05-08 16:07:25 UTC

- reply

gm!!

2026-05-08 16:07:20 UTC

- reply

gm both

2026-05-08 13:13:08 UTC

gm nostriches 👀 🍍

2026-05-07 19:24:30 UTC

- reply

we’re seeing who wants “my bags go up” vs “more bitcoin activity”

2026-05-06 15:09:15 UTC

- reply

Could you list a few example filters?

2026-05-06 14:01:16 UTC

gm nostriches 🍍 👀

2026-05-06 01:20:28 UTC

- 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 00:24:56 UTC

- 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 00:15:38 UTC

- 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 00:14:15 UTC

- 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 00:09:21 UTC

- reply

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

2026-05-06 00:08:58 UTC

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 14:05:36 UTC

- reply

gm marcel!

2026-05-05 13:59:32 UTC

gm nostr 🤙 🍍
#nevent1q…dsv6

2026-05-04 23:51:56 UTC

do not anthropomorphize matrix multiplications

2026-05-04 22:48:27 UTC

- 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.

2026-05-04 22:41:26 UTC

- 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 21:12:55 UTC

- 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 21:11:15 UTC

- reply

great, the pi-pilling is spreading in the ⭕️

2026-05-04 20:46:26 UTC

〰️〰️〰️〰️〰️

2026-05-04 20:39:30 UTC

- 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 20:16:24 UTC

- reply

Also do not use a Pi

2026-05-04 16:04:37 UTC

- reply

I have been considering using Fossil for this reason

2026-05-04 15:07:59 UTC

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 14:31:22 UTC

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

2026-05-04 14:29:30 UTC

- reply

hi jennifer!

2026-05-04 14:29:23 UTC

- reply

gm marcel ☕️

2026-05-04 14:29:15 UTC

- reply

gm oshi 🫡

2026-05-04 13:23:43 UTC

- reply

oh hi you are still here?! 🫡

2026-05-04 13:20:53 UTC

gm nostriches 👀 🍍🍍🍍

2026-05-03 18:43:45 UTC

- reply

$200/mo? or 5.5 xhigh?

2026-05-03 18:01:10 UTC

- 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 17:59:49 UTC

- 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 15:55:54 UTC

- 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 15:53:40 UTC

- reply

how seriously should I treat this?

2026-05-03 14:03:05 UTC

gm nostriches 👀

2026-05-03 13:04:25 UTC

- 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 13:02:05 UTC

- 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.

2026-05-02 22:01:37 UTC

- 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 21:44:19 UTC

- reply

The post actually exists but clients do client things. Ugh

2026-05-02 20:43:42 UTC

- 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 20:41:09 UTC

- 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 20:35:27 UTC

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 20:32:36 UTC

- 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 20:10:07 UTC

- 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 12:51:36 UTC

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

2026-05-02 12:12:28 UTC

- reply

I didn’t, where is it?

2026-05-02 11:03:04 UTC

- 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 11:01:46 UTC

- 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 10:55:16 UTC

- 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 10:53:58 UTC

- 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 10:51:41 UTC

- 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 10:50:51 UTC

- 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 10:48:42 UTC

- reply

Which article specifically, moderation or NFDB?

2026-05-02 10:46:56 UTC

- 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 02:56:35 UTC

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 02:26:11 UTC

- reply

oh. they left.

2026-05-02 02:25:33 UTC

@npub107j…ncxg is habla down?

2026-05-02 01:01:25 UTC

- 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 01:00:06 UTC

- 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 17:50:53 UTC

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 17:47:31 UTC

Faster negentropy alternatives 🤔

2026-05-01 15:05:23 UTC

gm nostr 🍍
the 🍍 will continue until 🍍

2026-05-01 15:02:44 UTC

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

2026-05-01 11:14:24 UTC

- reply

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

2026-05-01 11:10:33 UTC

- 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 09:54:44 UTC

- reply

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

2026-04-30 18:21:06 UTC

- 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 17:13:17 UTC

- reply

Technically audio track can be swapped without video change

2026-04-30 17:08:43 UTC

- reply

What did you find? Weird compiler optimizations?

2026-04-30 17:08:26 UTC

ga nostriches 👀

2026-04-30 14:51:54 UTC

- 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 14:15:14 UTC

- 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:

2026-04-29 22:45:39 UTC

- reply

Do not have concurrency>1 for any single websocket connection

2026-04-29 18:40:33 UTC

- 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 17:26:26 UTC

- reply

🤫

2026-04-29 16:12:18 UTC

- reply

gm zorbar

2026-04-29 16:12:09 UTC

- reply

gm 🍍 🍍 🍍

2026-04-29 14:18:53 UTC

gm nostriches 🤙 🍍

2026-04-29 13:24:02 UTC

- reply

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