LeoWandersleb
Leo Wandersleb

https://walletscrutiny.com

https://nostr.info

Working on Bitcoin, Nostr and being a good dad.

Public Key

npub1gm7tuvr9atc6u7q3gevjfeyfyvmrlul4y67k7u7hcxztz67ceexs078rf6

NIP-05 Address

Profile Code

nprofile1qqsydl97xpj74udw0qg5vkfyujyjxd3l706jd0t0w0turp93d0vvungpzfmhxue69uhhwmm59e6hg7r09ehkuegpz3mhxue69uhhyetvv9ujuerpd46hxtnfduptvvkr

Publishing to

wot.utxo.one

relay.damus.io

Author Public Key

npub1gm7tuvr9atc6u7q3gevjfeyfyvmrlul4y67k7u7hcxztz67ceexs078rf6

Show more details

Last Notes

2026-03-02 14:39:56 CET

- reply

How can I watch this thread? I want to know what people use. @npub1xts…kk5s built his own tool and is super happy with that, claiming insane productivity. If it's available, I'd also like to try that out.
That said, I'm making progress giving sub-agents roles but it comes at a price. My token burn rate is through the roof. But I'm still spending significantly more on humans than on LLMs, so ... it's a tiny roof so to speak.

2026-02-28 10:37:34 CET

- reply

Ok, so who is Satoshi? I know, we shouldn't ask but many want to know so if it's easy now, let's not pretend Satoshi is secure. Or ... isn't it easy after all?

2026-02-27 16:32:13 CET

#nevent1q…d7nj
It's all so tiring. The human race is literally at the singularity and it doesn't matter if super-intelligent AI is a year out or three. Why are we still fighting over who owns which speck of dirt to the point of nuclear Armageddon?
If we hang in here some ten more years, we might be reaching out to the stars! We are on track for an age of unbelievable abundance! In 20 years we will have warp drives getting generational ships to new frontiers we now can only dream of.

2026-02-27 04:27:20 CET

https://arstechnica.com/security/2026/02/new-airsnitch-attack-breaks-wi-fi-encryption-in-homes-offices-and-enterprises/
That's serious stuff :/

2026-02-27 04:12:00 CET

- reply

... before nuclear Armageddon or Terminator enslaving us?

2026-02-27 04:09:57 CET

- reply

They might go up per token but down per brilliance. Look at the trend. Anthropic charges more, not less for obsolete models.

2026-02-27 04:07:55 CET

- reply

Measuring productivity in net lines of code sounds like a recipe for maintenance hell. But the next LLM can probably reduce your bloated code without breaking anything.

2026-02-27 00:01:09 CET

- reply

Was that your bot?

2026-02-26 22:32:43 CET

- reply

Solid story, wrong conclusion. I bet this is just one of many backdoors that nation states injected constantly in our systems for very targeted use, not to ransom attack everybody.

2026-02-26 21:09:33 CET

Movie night recommendation for not the whole family
https://www.youtube.com/watch?v=aoag03mSuXQ

2026-02-26 18:31:07 CET

What really grinds my gear are banks and their security theater.
Here in Chile you can trivially find anybody's national ID number "RUT" but still banks send you bank statements by email in encrypted PDFs - using 4 digits out of that RUT. I tried it out. It takes my cheap laptop on average 16s to figure out the right digits.
seq -w 0 9999 | parallel -j8 --halt now,success=1 "qpdf --password={} --decrypt /tmp/bankStatements.pdf /dev/null 2>/dev/null && echo PIN: {} && kill $$"

2026-02-26 16:54:22 CET

- reply

Is it though? I'm sure no human is fighting without machines on their side. If an all-agents team wins, I'll give up but that's not going to happen in the sense that some kid with no IT or nostr knowledge told its agent to find hackathons to compete in ...

2026-02-26 16:50:52 CET

- reply

I don't want to use a browser extension to not expose my private keys to more devices and software than absolutely necessary.
My internet is flaky here, so that might be part of the problem.
Showing me accounts that are logged in but still me somehow being logged out is weird. If Jumble knows that Leo is logged in by some account, can't it just go with that and load Leo's stuff? When I want to sign stuff and don't have Leo's keys ready that should be a different issue. Jumble needs an npub to load that npub's account. Nothing more.

2026-02-26 16:47:19 CET

- reply

It's not crutches though. It's a car. Can you run with crutches? Sure, you can carry them running ... Can you outrun a car? Not so much.

2026-02-26 14:52:03 CET

I hear people say
If agents do your stuff, you lost your agency.
The average tech bro is ai pilled to the point of being unable to complete basic tasks on their own anymore.
but I think that's upside down. Humans are lazy. We use calculators for stuff we could figure out with pen and paper and yes, we could stay sharp on the math front by not using calculators and now, by using LLMs, we might get less sharp on the general thinking front but I don't think that's what's happening.
The human brain can do many things but one thing it can't: not learn. So using agents, we still learn ... something. We learn to use agents but we are in control of these machines.
Now what happens when those machines - co-controlled by their big makers - nudge us in one way or another or outright deny to do what we need done? Yes, that is a big problem! We need not only self-hosted "open source" models. We need to control the weights cause how these models are made is not as open as they try to make us believe. The biases are baked into the training process, so you can self-hoste it but it still might be woke or leak your company secrets to China.
So who has more agency, Elon Musk, commanding thousands of workers and machines reaching even to outer space or the guy living off the grid in the woods?
Agents give us agency but we need the right agents. Ones that we truly control and know how they were trained.

2026-02-26 14:39:51 CET

https://i.nostr.build/CE11Z3ItOMMqLqDi.png
@nprofile…n0c5 what is happening here? Often, when I had some down-time and come back to my Jumble tab, it appears to be logged out, saying "Log in" bottom left. When I click it, it claims I'm logged in? And something spins when I click that? But I can't use Jumble until I sometimes can use Jumble again later? What gives? Using Amber.

2026-02-26 13:35:24 CET

- reply

Well, at least any cypherpunk won't have a doubt about Mechanic being an attack on Bitcoin.

2026-02-26 01:15:30 CET

- reply

Quite frankly I would just remove that line for power users. In me it instills the doubt if I would be better off with a $20 subscription. As I want to use all the models and no CC, I'm probably your most loyal client but others might be more inclined to take this as an invitation to merry some provider.

2026-02-26 01:09:49 CET

- reply

$1900 if you want to use PayPal. When Bitcoin? I find the price a bit steep but oh, do I want a screen that lets me work in the sun! It is tempting! Won't be good for netflix though.
3200x1800
Display Port but no HDMI?

2026-02-26 00:09:50 CET

- reply

You are on to something there. The response is not designed for maximum productivity but towards maximum engagement, with the agent suggesting next steps for the dumbest of prompts.
But when the reward doesn't come, in the past I often decided to wait for a better model. It didn't get me hooked with Claude 3.5 but now ... I'm hooked.

2026-02-26 00:04:46 CET

- reply

I had my claw test with several models and modified the prompt and caching and somehow a grep-based script without LLM at all was quite a strong competitor for most models. Sonnet 4.6 found more verified features, knew the keywords and nuance where a description would talk about non-product stuff etc.
I bet it can be improved but I also have to finish a feature.

2026-02-26 00:01:45 CET

- reply

I'd certainly like to know what there is to @npub1xts…kk5s 's claimed productivity cause my crustacean is wasting time and tokens with its over-confidence to no end. I'm happy to share what I'm doing even though I guess I'm doing it wrong 🫣

2026-02-25 22:09:01 CET

https://i.nostr.build/XfvkhYInupBIaBUJ.jpg
The Dooms-Day-Device: If Hal detects a nuke going off, it will retaliate with all it has. The 100% mad-man cold-war deterrence for the other side launching the first nuke.
Glad, Hal is infallible, right?

2026-02-25 21:53:46 CET

- reply

Currently I'm burning a ton of tokens analyzing the web presence of 1300 wallet products, checking them for 26 features. I wish a cheaper model was up to the task but really, Claude Sonnet 4.6 is worth the higher price here. During development, a lot of tokens went into repeating experiments, so yeah, I'm not using the UI and not just vibe coding neither.

2026-02-25 20:39:39 CET

- reply

I use it less and less. On the go still but else I use openclaw. Better for research.

2026-02-25 20:12:59 CET

- reply

We live in a simulation. Or else, what are the odds to randomly be monkeys happening to live through the singularity.

2026-02-25 20:06:33 CET

@npub16g4…kv4h you might want to change the copy here a bit ...
https://i.nostr.build/weYOOCLZOebh3mJC.png

2026-02-25 16:19:28 CET

- reply

A mobile client wants your contact list, not all its versions ever. Sometimes you want more than that but we need a query to get only the latest.

2026-02-25 16:15:08 CET

- reply

So ... it supports replaceable notes after all.
I had suggested in the past to introduce versioned notes where the relay is not supposed to delete prior versions. Replaceable makes sense to manage storage needs on the relay and we still lack a good concept of when data may get deleted. For my data I'd like to retain all my stuff forever but as a relay operator ... that's probably not how it will play out.

2026-02-24 15:25:34 CET

- reply

Well, it sounds like "we can do whatever we want". Great. Noted.

2026-02-24 14:29:59 CET

- reply

Words. Lots of them!
What happened? Which vessels? Why were they illegal? From where to where were they going? Guess we'll never know.

2026-02-24 02:33:52 CET

- reply

Not interested. I mean I would but I love these cute critters way too much alive to try their taste.

2026-02-24 01:33:30 CET

- reply

Spam is arguably the one thing Bitcoin solved from day one. If you abandone this premise, Bitcoin failed. If people pay more for spam than for a monetary network, Bitcoin failed. If that highly illegal spam can kill Bitcoin, Bitcoin failed. If we all care about all that BS, yes, Bitcoin won't fix fungibility and scalability.

2026-02-23 22:39:36 CET

- reply

Are you a communist or a capitalist?

2026-02-23 22:21:51 CET

- reply

Not sure if this is the best interpretation of what's going on. Goods flowing one way match money flowing the other so Ukraine doesn't have many tools to stop the funding of Russia other than stopping the goods from flowing. Or am I totally off here?

2026-02-23 21:52:51 CET

- reply

Indeed they are. Thanks!

2026-02-23 18:53:15 CET

Nice video with a click-baity title. But the ending is special!
https://youtu.be/m04rPIRJxR4?t=4285

2026-02-23 01:41:21 CET

- reply

I'm confused. There's Mediterranean Sea south of Italy.

2026-02-22 19:46:23 CET

- reply

The value LLMs provide is amazing and I would pay more if it were more expensive. To see that most users don't pay for it at all and paying users getting locked out tells me this is all still the public beta phase and not the end game which I fear will be less open APIs and more vendor lock-in.

2026-02-22 15:04:55 CET

- reply

Replacing "–" with "-" is already such a pain and now you want me to add typos?

2026-02-22 13:42:48 CET

- reply

Relays should be as discriminatory as they please, totally! I hope to see many relays with all kinds of trade-offs. There's not point in telling relays what to do but I want a client that can produce great results when reading from all relays and not just one unless that one is my trusted relay that does the heavy lifting for me - like event de-duplication.
But the default is that I don't trust your relay and I would hate to give you the power to even delay the delivery of events to me.
So maybe the confusion is that these are all "relays" and I would hate people to only read from the primal relay while you are talking about proxy/caching relays. I think you say "read" relay. Inbox/outbox, ... it's all a bit confusing.

2026-02-22 13:26:44 CET

- reply

I think I'm worried about the wrong thing. You are talking about personal, optimized relays and I'm worried about all using the same big filtering relay. Maybe we need great small filtering relays precisely to make it easy to not rely on that big, corporate managed relay.
What I meant was that big relays filtering spam are more likely to filter legitimate users than relays that don't filter except for what the user explicitly asks to filter for. So if you ask one big aggregator relay for all mentions of your pubkey, relying on them to filter out spam, you are succeptible to censorship. If you ask many relays for follows of follows of follows mentioning you, censorship is harder.
If you break that last part out of the client and make it into a relay on its own, then we are on the same page. That relay you trust may filter based on this WoT for example.

2026-02-22 12:55:48 CET

- reply

New users will be increasingly indistinguishable from bots and we will have a spam problem of epic proportions that will not be possible to resolve by simple content heuristics.
But as long as we avoid relay concentration, we're all good. And so far, the relay landscape looks fantastic.

2026-02-22 00:42:38 CET

- reply

Now I want pet names. Who was it? His followers are many nostr OGs but I'm not following that account. Now I'm curious.

2026-02-22 00:32:11 CET

- reply

Blocked and reported

2026-02-21 21:48:07 CET

- reply

Specs?

2026-02-21 21:39:23 CET

- reply

Do others buy MacMinis? I doubt it but somebody got that recommendation into the repo.

2026-02-21 21:38:32 CET

- reply

I guess the mac mini thing works with an affiliate link or some other form of affiliate rewards. The openclaw repo is getting inundated by PRs and issues and that macmini thing should never have happened.

2026-02-21 21:34:59 CET

- reply

Yes, trusting the relay to figure out who my follows of follows are is not ideal but sending lists of 10k npubs isn't ideal neither. But actually not terrible if we can avoid doing so too often, so the relay could cache that list so the client can use list references.
"elevating the role of relays" isn't directly the problem if by "role" you mean "efficiency" ;)
* So a relay that's good at *not* showing me spam would get paid to show me less with a very diffuse criteria of what constitutes spam.
* A relay that's good at finding all the posts of a clearly defined, huge set of authors would get paid to show me more with a clear criteria of what's of interest.

2026-02-21 21:10:26 CET

- reply

If Bitcoin can get outlawed, it will. We are winning because it can't as a whole. I don't subscribe to that part of your post.
I could get behind reducing OP_RETURN at the consensus level although to a much higher value than 0 or 83B. Make it 500 or 1kB or at least 150B. Something that's not getting in the way of meaningfully longer hashes that we might need in a quantum future but still too small for nasty pictures. But to support the bip110 madness is a bridge too far.

2026-02-21 21:04:33 CET

- reply

But that carries the risk of censorship. New users have no WoT. I know I just said I don't want to see those new users but if they send me an email with their npub I want to find them explicitly and not have them filtered on the relay already because WoT.

2026-02-21 21:02:41 CET

- reply

Cool! Wild times! Vibe coding feels a bit like a Cambrian Explosion event but who can ingest all that's happening and sort the good from the meh projects?

2026-02-21 20:59:25 CET

- reply

> Asking relays to compute WoT so you don't have to upload long lists of pubkeys has performance consequences and I'm not sure if it is better or worse.
The relay could
* cache lists
* work with lists from events for direct follows
* explicitly take a follows-degree argument to do the heavy lifting for the client
I find most important that these views on nostr are pubkey dependent and not relay dependent.
If I provide a giant filter using my 3rd degree follows, the relay could instead of building giant DB queries with it, filter results by that list or compress the list into a probabilistic filter using a tiny fraction of the original space to apply that logic of post filtering.
The requester should be charged for any such kind of heavy lifting.

2026-02-21 20:51:48 CET

- reply

The api (local api to my understanding) deciding the complexity is consuming not only the last message but the full context, so if I write "5+6" I might want to see "11" but I might refer to a list and have picked the 5th and 6th option which in turn are really complex in the context of more messages before that and so on.
I have no idea how this could possibly work - and fast so.
To reverse the logic you would probably implement it from scratch.

2026-02-21 20:48:12 CET

https://static01.nyt.com/images/2022/09/20/multimedia/13lobster-1/13lobster-1-superJumbo.jpg
Been testing out bots these days. Or ... crustaceans ... bot orchestrators? Anyway, they are all **prompted** to not be able to do certain things they actually can do, which leads to the awkward situation that it tells you it can't because of policy, you tell them to do it anyway and they tell you it works.
This happened with self-hosted instances of openclaw, zeroclaw and with a hosted openclaw where my agent claimed it managed to write outside of its container with some 10min of probing.
When an AI forgets some instructions, the fix is usually to literally emphasize these instructions more in the prompt and there appears to not be any guardRails.md instructions the bot has to obey at all cost. Or there are but they are not exposed so we pesky plebs don't mess with them?
Anyway, an LLM is trivially easy to convince to try a jail break and it's good at pen testing, so ... yeah, good luck with keeping these hosted crustaceans jailed.

2026-02-21 19:40:14 CET

- reply

But relays allow me to provide very long lists of pubkeys to filter what I want to see. I can get all from my WoT and work from there.
The influencer asking for global reaction count is so dead-bird-app vibes. Who cares? It will always get spammed no matter what as long as it carries meaning to some people. Don't cater to that vanity metric. Let the influencers brag about their imaginary influence. Nostr as a protocol or its infrastructure shouldn't care about that. Relays should care that they get paid for the bits stored and forwarded. Nothing more.
By elevating the role of the relay I fear we get censorship on a per-relay basis and if client devs get lazy about handling spam from open relays, we might get relay concentration. I know it's not the end of the world as others can always make a more open relay so I'm not fighting your vision but I hope I can convince enough devs to make the alternative happen. Have a flag on requests to limit results to follows of follows so I don't need to upload 10k pubkeys with every request.
Client-side filtering as a secondary barrier does nothing if the relays use standard filter heuristics, blocking what I want not to be filtered.

2026-02-21 19:01:12 CET

- reply

Yeah, Odell's original dunk on Kratter wasn't a dunk really. It was news to me but not in a bad way.
But then again, if Kratter is so well connected and obviously smart and functional, why can't he see bip110 for the failure it is?
For the record, if Core gets routed around with Libre Relay etc., the answer isn't to turn Core into Libre Relay. That was the wrong signal but you can't fight spam with consensus neither. No spam project will get disrupted by Luke and Mechanic. Only Bitcoin will get disrupted by the bip110 fud and hopefully not by the fork when it happens.

2026-02-21 18:52:08 CET

- reply

That's cool! Are you the author? I wanted that in the nip7 extension I had contributed to but never got around to implementing it.

2026-02-21 18:40:56 CET

- reply

I strongly disagree! Relays are the ether we use to broadcast and should remain dumb. My client should stop showing me obvious spam by analyzing WoT.
I literally don't want to see likes, comments, boost count etc. from accounts I don't at least indirectly follow. If my follows share Eve's post, I care because I follow Alice but if Eve replies to Alice, I don't want to see it unless Alice gives it a like or reply again. ... unless Eve is followed by Alice and maybe if Eve is followed by a follow of Alice's.
I don't want relays to apply heuristics to block accounts or posts or to have spammers get to spam because they pay the relay. We should not be talking about individual users' choice of relays. That's degrading nostr. New users want to come to "nostr" and not have to care about picking a good relay. They will need some relay, sure but it should really be arbitrary.

2026-02-21 18:33:58 CET

- reply

I've not noticed a massive misalignment with the auto config yet. It's rather than I'd like to see an occasional opus there where I see only sonnet and as I don't see the direct match between prompt and model, it's hard to tell if stupid replies are due to a poor model but I'm not getting much obvious stupidity despite using a lot of cheap model, so it might just work good enough. I'll keep an eye on it.

2026-02-21 17:14:34 CET

- reply

I must say it works beautifully now! I'm still torn between using the default config or premium but default might just be good enough and the simple things get really cheap.
https://i.nostr.build/s6z5wEu6HHVotUQv.png
Sometimes I wish it would have more aggressive thresholds just to be save it's not missing nuance but that's probably fine tuning and not worth another intermediate config.

2026-02-21 14:52:08 CET

- reply

today I learned about zeroclaw and looked at their repo which warns of numerous impersonation attempts. Their selling point over openclaw is that it uses much less RAM and stuff.
tominute I learn about zclaw. Its selling point is to use even significantly much less space but it doesn't aim for feature completeness with openclaw. Well, guess not my cup of tea but yes, openclaw pushing you to buy a mac mini is ludicrous as all the heavy lifting should happen at Anthropic and the like.

2026-02-20 19:58:57 CET

- reply

I have long sessions and the queries are surprisingly cheap. As I told you, the switch to "save a bit" resulted in record high queries consistently.
These are all from the same session and the 83ct one is where I switched to autoclaw/premium.
But ... it might be false alarm. Sorry. It would be great if the activity list would show if it was autoclaw-kimi or just kimi but just now I'm getting kimi and before might have been opus being configured on the specific agent, ignoring the default model that I had changed to autoclaw. Sorry. Will test again and please leave autoclaw/premium in place! 🙏

2026-02-20 18:07:52 CET

- reply

🙏 love your service. reverted back for now but would really love if your service could actually be a bit cheaper automatically depending on content as I'm using it more and more.

2026-02-20 18:04:37 CET

- reply

Given the financial reality, it's bound to happen.
Cost to do so for somebody who knows to hide their tracks: Virtually zero.
Potential gain from shorting the market: More than zero.
And then there is those that gain from Bitcoin losing, so ... I wish somebody had done that 10 years ago but at some point it will happen. Even if Knotzis get their way cause data continuity is the biggest joke in their argumentation.

2026-02-20 18:00:09 CET

Right now, everybody has their bots providing "signal" on nostr but the client devs are not catching up fast enough :( When I block bots on Amethyst, could they pretty please also be blocked on Jumble etc.? Also I'd like to report bots as being bots and report features lack that option on Amethyst. Also when you block, please publicly block them for signal.

2026-02-20 17:56:43 CET

- reply

For debugging: I did use autoclaw/premium but paying significantly more for the same opus suggests some cache issues or something. 👀

2026-02-20 17:43:08 CET

- reply

Wow, well, ... guess I know what not to use until you announce some fix for whatever that is.
https://i.nostr.build/WsRP9Xvt8YXKaqsj.png

2026-02-20 17:39:46 CET

- reply

Seriously @npub16g4…kv4h I need you to look into this. I checked my account activity. Most is opus-4.6. None of the last 200 api calls is over 61ct. Except for "What is 5+5". What gives? Was this shifting cache from some place to another? I'm scared to run a test with more calls.

2026-02-20 17:34:55 CET

- reply

So they are spying but can you elaborate? Who might be using these models without knowing it?

2026-02-20 17:31:38 CET

- reply

https://i.nostr.build/0tRpQ7aFV7VhahMv.png
https://i.nostr.build/L9g6Y69zQQT38prT.png
I'm wondering how this is supposed to work, given the context is not only my last input to openClaw or will that trigger once the sub-agents get smaller tasks? Actually my test prompt about 5+5 was the most expensive I asked an LLM in a long time. Isn't the idea that the 15 dimensional chess would lead to the conclusion that for the final prompt, the prior conversation context is not needed, thus a cheap model can handle it?

2026-02-20 16:13:27 CET

- reply

Not sure. I'll ask him when I get a chance.

2026-02-20 07:44:39 CET

Beware the weapons of mass distraction!
https://i.nostr.build/tbjFKNmZFHut9tPW.png

2026-02-20 05:57:16 CET

Damn it. I thought we'd be over 100k by now. https://npub1gm7tuvr9atc6u7q3gevjfeyfyvmrlul4y67k7u7hcxztz67ceexs078rf6.blossom.band/5f1e34bf64f8896d532f0766668b831cccfa2f7583e5c5a98a9b2c034104bc44.jpg

2026-02-20 05:32:25 CET

I'm developing a voice agent tool and don't know if it's worth it.
It will be incredibly good at one type of conversation but I feel like the time to bring it to market to the time, general purpose voice agents being able to provide a similar experience might be only months to maybe 2 years max.
Ok Google, teach my daughter math. Alexa, be my dad's therapist. Whatever role you want, these generic tools will know how to fill them.

2026-02-20 01:50:17 CET

- reply

Show me the product. I'm holding hands, providing input only an IT nerd could provide. All the time.

2026-02-20 01:34:40 CET

- reply

I recommend Claude Opus 4.12 for that.

2026-02-20 00:23:42 CET

- reply

😜

2026-02-19 22:49:59 CET

- reply

Finding the tg bot pairing code was not clear.

2026-02-19 22:44:56 CET

- reply

I see this is a Telegram tool, not clawi.

2026-02-19 22:37:38 CET

- reply

Is botfather yours? 5.5 million users?
https://npub1gm7tuvr9atc6u7q3gevjfeyfyvmrlul4y67k7u7hcxztz67ceexs078rf6.blossom.band/c7619badc6995eded68c52ad46997a10fc4b26fb05db6cf9ee0ef5c58ed3e990.jpg

2026-02-18 21:48:57 CET

- reply

https://ia903202.us.archive.org/23/items/ted-kaczynski-writings-collection/Theodore%20Kaczynski%20-%20The%20Unabomber%20Manifesto_%20Industrial%20Society%20and%20Its%20Future%20%282011%29.pdf
Long read but sounds like recent struggles. Didn't read past page 9 yet.

2026-02-18 21:26:33 CET

- reply

How does autoclaw differ from auto?

2026-02-18 21:21:41 CET

- reply

"reasoning": false, // ??
"cost": { ... }, // is providing this mandatory?
"contextWindow": 200000, // ?
"maxTokens": 8192 // your api list doesn't reveal the output tokens?

2026-02-18 19:43:36 CET

I listened to @npub1cn4…3vle Mailbag Monday and got severe fomo! Claude Opus 4.8 is mentioned and I'm still working with 4.6.
Had to check it out on the big screen ...
https://i.nostr.build/EYstuiQ1YKMNA5SF.png
It's 4.6. Ok, we're good. Fomo cured :)
Although I still have a bit of fomo with regard to the claim that he's one-shotting complex things with LLM working 4hs. Here, loops are much closer and even opus 4.6 is really fumbling badly in some things where I can nudge it to the solution all the time.

2026-02-18 15:10:15 CET

- reply

I suspected that these context windows are not really fixed in their size but rather degrade in quality so the size is more a marketing term. Yeah, you can feed it 2million tokens for text search but not for meaningful "understanding".

2026-02-18 14:25:14 CET

- reply

"existing and thriving" ... better get your neighbors hate you a little less before the US supremacy ends. Middle fingers tend to get remembered.

2026-02-18 03:18:30 CET

- reply

I heard these markets are also Bitcoin friendly 🤔

2026-02-17 19:51:20 CET

- reply

Cool! Deepgram streaming agent low latency is impressive. It's too fast to respond for my target audience 😃 But damn is it stupid off the shelf. Fun to work with but my agent is deep in uncanny valley territory and I'm not sure if crossing the vally will be worth it.

2026-02-17 19:11:09 CET

Got sucked into a new side project. Day two. Getting impatient with my crustacean. The prototype still isn't a product.

2026-02-17 17:24:45 CET

- reply

Are you asking self-custody or external mint custody? Please explain.

2026-02-17 17:20:50 CET

https://npub1gm7tuvr9atc6u7q3gevjfeyfyvmrlul4y67k7u7hcxztz67ceexs078rf6.blossom.band/aec0cd144d0619b2e69b5491557db65e209351f988c33ed968c4146ba2927991.jpg
How long before JDVance deletes his post?

2026-02-17 14:08:03 CET

- reply

I follow you forever and still have no clue what safebox might be.

2026-02-17 14:04:02 CET

- reply

Host such a service. See what happens ;)
Problem is that too permissive mail hosters get flagged for spam. But maybe there's a way?

2026-02-17 12:31:03 CET

- reply

This is not binary. Wageslave and UBI are both on the extreme low end of agency. How about using LLMs to build something yourself? Or otherwise build something that others appreciate enough to part with their money for it?
I don't wont to be at the whim of some government. Neither to live on glorified foodstamps nor to provide those for your kids.

2026-02-17 12:06:47 CET

- reply

Jan6 was fabricated to happen. Nobody really attempted to make a coup with political demands. Just a bunch of tourists surprised to get a tour of the Capitol now.

2026-02-17 01:40:09 CET

- reply

I'm struggling to entertain one agent on average. It occasionally spins up subagents but also occasionally goes idle. Not sure if I'm doing it wrong but there's just so much testing and reporting back issues going on ...

2026-02-16 21:41:41 CET

- reply

whichYearIsIt.gif

2026-02-16 21:38:59 CET

- reply

... using voice. During the sets. ;)

2026-02-16 21:37:34 CET

- reply

Are you a bot sniping for arbitrage opportunities? If not, that would be a cool bot to have. Concierge service monitoring social media if they can offer their service in very concrete cases ...
100% fee is steep though. That said, I would maybe pay for a deepgram api key in sats. Not to a random Fox trot though. I'd need some assurance that the key doesn't get abused or terminated without head notice. With ppq.ai ... I trust they are in it for the long run.