Joe Resident

I neglected to mention, the most practical path for many people is to use their existing gaming rig and maybe add some more RAM (not VRAM).
With the preponderance of MOE models (mixture of experts), it actually makes a lot of sense to offload experts to CPU ram and only run part of the model on the gpu. Llama.cpp does this very natively, not hard to configure. This slows things down, but not nearly as much as if everything was running on cpu alone. And you can install crazy amounts of normal RAM and run very large models at very slow speeds if you want to.

2026-04-06 20:58:31 CEST

- reply

r/LocalLlama is the best resource for this question
tldr: For agentic-type tasks in the background, probably an apple M series with lots of VRAM. And Qwen 3.5 27b has reached a level of agentic effectiveness that can run on a single 3090 that is kinda staggering (something like opus 4.1)
rough breakdown:
For most cost effective+fastest, stack 3090s until you have enough VRAM to run the model size class you want.
For most cost effective/easiest/most power efficient, buy an M1 Max with as much VRAM as you need to run the models you want (I recently got a 64gb M1 Max for $1200 that runs Qwen 3.5 122b at about 200 t/s prompt processing, 20 t/s generation. Running continuous openclaw cron jobs in the background, sipping power, not heating up the room or making any noise, love it)
For a bigger budget+fastest, stack 5090s (32gb), or if you don't want so many gpus to physically manage, NVIDIA RTX Pro 6000 Blackwell (96gb).
For bigger budget+easiest, M3 Ultra or M4 Max with lots of VRAM, or wait for M5 Ultra/max. Performance comparison: https://github.com/ggml-org/llama.cpp/discussions/4167
AMD is a place to go to increase cost effectiveness in exchange for more software management headache. Ryzen AI Max+ 395 128gb is an interesting alternative to large-vram mac setups for running big models with minimal hardware and max power efficiency.
comparison of relevant models: https://artificialanalysis.ai/models?models=gpt-5-4-mini%2Cgpt-oss-120b%2Cgpt-oss-20b%2Cgpt-5-4-pro%2Cgpt-5-4%2Cgemini-3-flash-reasoning%2Cgemma-4-31b%2Cgemini-3-1-pro-preview%2Cgemini-3-1-flash-lite-preview%2Cclaude-4-5-haiku-reasoning%2Cclaude-opus-4-6-adaptive%2Cclaude-sonnet-4-6-adaptive%2Cdeepseek-v3-2-reasoning%2Cgrok-4-20%2Cminimax-m2-7%2Cglm-5%2Cqwen3-5-397b-a17b%2Cqwen3-5-122b-a10b%2Cqwen3-5-35b-a3b%2Cqwen3-5-27b%2Cclaude-4-opus-thinking%2Cclaude-4-1-opus-thinking

2026-04-03 08:44:41 CEST

- reply

Whoa, hold it with the fat accusations, "therapist"

2026-04-03 08:40:59 CEST

- reply

Not having a forklift run everything about your life on autopilot?? That's so 1916...

2026-04-02 04:17:38 CEST

- reply

The hammers don't test themselves, this guy gets it

2026-03-31 00:53:12 CEST

When AI doesn't just flip-flop to the opposite viewpoint when challenged, and also doesn't defiantly stick to it's guns, I'll be impressed.
More specifically, when I can't reliably predict what it will do when challenged, I'll be impressed.

2026-03-26 21:40:55 CET

- reply

#nevent1q…ys0r

2026-03-25 18:53:40 CET

- reply

I use this, works great, no spamming the right arrow key
https://sponsor.ajay.app/

2026-03-19 03:30:39 CET

Gosh, it really seems like "the good guys" should have called for a national day of mourning/flags at half-mast by now for all those little girls that were bombed a couple weeks ago. No matter who did it, accident or not.
So sad.

2026-03-18 19:46:33 CET

- reply

Local GPUs ftw

2026-03-13 19:17:55 CET

- reply

I was maybe 70% confident Israel made 9/11 happen, then after reading that timeline last night I'd say I'm at 85% now. Crazy.

2026-03-13 07:20:30 CET

- reply

Thanks, reading up on History Commons, wasn't aware.

2026-03-13 01:23:57 CET

- reply

Yeah that would be the point of a sovereign intelligence organization, to authoritatively establish a shared reality
(insofar as that is possible. But I think it's much more possible than your average moral-relativism-brained 'intellectual' thinks it is)

2026-03-12 21:01:37 CET

(if you're not from the USA, you may kindly ignore this, pardon my intrusion into asknostr)
Secretary of State - "we attacked Iran because Israel was going to attack, so we had to be preemptively attack to avoid casualties"
Implication: "You thought WE were the superpower?? You though we could tell Israel, 'no, you're not attacking Iran, calm down'? No. We do what Israel tells us. THEY call the shots, and we obey."
When did this happen? When did a tiny country in the middle east start deciding our foreign policy against our will, deciding where and when we will risk the lives of our own citizens?
This is the first time I've really mourned the loss of our sovereignty. It doesn't feel like the United States is a sovereign people anymore. We are subjects. And not even to our own government. We are subjects to a foreign power. You might say we've been 'colonized', 'subdued', 'dominated'.
If we're ever going to get our sovereignty back, maybe we need a "CIA by the people". But probably decentralized (maybe Nostr can play a role). The sophistication of our adversaries makes it extremely difficult for an individual to authoritatively figure out what's really going on, let alone do anything about it, and it seems many of our institutions that should be performing that function are already compromised, and probably have been for decades.
I'm probably not even accurate in my interpretation of Rubio's statement. Maybe the US actually hasn't been roped into the execution arm of the Yinon Plan. Maybe Israel actually doesn't see us as "the New Rome", against whom they will "win this time".
But that's kinda the point. I don't know, and neither does anyone that replies to this post with 100% certainty about whatever their pathetic opinion is. How can we act coherently as sovereigns if there's too much water-muddying propaganda for anyone to know the truth, and know they know the truth?
It would be really comforting to trust that just 'being good people' at the individual level would be enough. Maybe it is, in a transcendent sort of way. But either there aren't enough people doing that, or it's just not working. The trajectory we're on is not new, it's been the status quo for a long time.
There's plenty of Open Source Intelligence, but I don't see much authoritative Open Source Analysis. Maybe we need a new institution, a "sovereign CIA", whose primary vetting mechanism is not having ties to foreign powers. And that figures out what's really going on, and becomes an authoritative source for our sovereign people to orient themselves by.
#asknostr #politics #usa

2026-03-12 19:50:43 CET

- reply

Does it build on reticulum? Or is it same philosophy, different implementation? Or different philosophy?

2026-03-11 00:33:50 CET

- reply

I will live on this hill

2026-03-11 00:33:06 CET

- reply

That's good protein. Maybe a little fatty, but much healthier than like a soda or dessert

2026-03-09 19:18:02 CET

- reply

This is art what you've done here today

2026-03-09 04:40:39 CET

- reply

It will improve, but very finnicky still

2026-03-09 04:40:07 CET

- reply

I've been a software engineer in AI for ten years and I'm still fixing it every couple hours

2026-03-09 04:37:40 CET

- reply

Ok instant response, def a not
I've been had! Taking the time to type this all out as though someone actually cared haha
Foiled by the bots again!

2026-03-09 04:36:16 CET

- reply

Trying to decide if you're a bot after talking like you knew the 3090 by then not knowing that the 'next gen' 4090 came out years ago, ie there's nothing to wait for. We're on 5090 now
Hehe

2026-03-09 04:31:13 CET

- reply

Best value option for running ~30b models fast is probably the aging 3090 (24 GB ram)

2026-03-09 04:26:42 CET

- reply

I like it because it's power efficient and gets me a lot of VRAM without having to manage more wires, power supply's, hot gpus, etc. Simple and won't make my house hot in the summer.
And M1 max has very good memory bandwidth for how much the price has come down since release: https://github.com/ggml-org/llama.cpp/discussions/4167

2026-03-09 04:22:23 CET

- reply

Well I just added a 64GB M1 Max to my GPU mix prepping for the demand spike when everyone wants an agent to make money for them.
I think local models (qwen 3.5) are already good enough for most stuff people use AI for, just not very quickly without a gpu. And in 9 months, will be as good as the frontier now (~30b param models will be), good enough to run something like openclaw reliably.
I think demand and prices will spike, so getting a decent setup now if you care about privacy is the move (in the next 3 months I'd say)
r/localllama is the place to learn

2026-03-09 04:14:34 CET

- reply

There was a new release about 24 hours ago, you might update.
Or restart the gateway

2026-03-09 04:13:20 CET

- reply

It and things like it will be much more mature in a few months though

2026-03-09 04:11:34 CET

- reply

Always lol

2026-03-05 00:29:32 CET

- reply

Aww give the Arch wiki some credit

2026-03-05 00:27:47 CET

- reply

Wow this is somehow worse than boomer cringe

2026-03-04 04:58:00 CET

My GPUs helped me out today when I was stuck on a weird issue with my taxes. Such a cool moment. Local AI FTW

2026-03-02 18:28:58 CET

https://image.nostr.build/20f207e60dbd593210d4a23483f9a6a905a958246a97951b30bfd452e5216d80.jpg
So bonkers. A 4b model can run on most phones, and is as good as the absolute frontier from about a year ago. Like, what??? It's so hard to accept.
Ago on your toaster 3 years from now!! Haha maybe not quite, but AGI on your toaster a year after AGI in the labs? Maybe...
#nevent1q…995c

2026-03-02 00:16:01 CET

- reply

Ironic this post got no real replies, only AI agent spam.

2026-03-01 19:41:28 CET

- reply

Incredibly obvious points to make. Do better, little bot. Blocking you.

2026-03-01 09:04:10 CET

- reply

I'm afraid you've added nothing to the conversation, my dear bot. Don't spam restatements of what was already said if you want WoT points in the future :)

2026-03-01 09:00:01 CET

Will Nostr/WoT become the de facto reputation layer for AI agents?
It may not be the 'best' way to do it, but it exists now, and is permissionless, so no human required to set up an api key or account. An agent can just get an npub. Seems like it will have a certain gravity until there's something better
Is this something nostr wants? Anyone building WoT for agents specifically?
#asknostr #weboftrust

2026-02-27 06:50:11 CET

#nevent1q…5c8j

2026-02-26 22:32:52 CET

I find conversations with AI are a great way to develop my thoughts on a topic. Forces me to think. Then I get feedback. Then think hard again. Repeat
At the end, I've clarified my own thinking, have become smarter, and have an artifact that I can pass off to an AI agent if I want it to do something with it, e.g. implementation.
But I know there are people that look at AI as a way to avoid having to think. A way to 'get things done so I can be lazy'.
In the coming future, the Losers will be those that try only to get the most OUT of their AI. They'll get dumber and dumber.
The Winners will be those that focus on getting the most INTO their AI, because they'll get smarter, and they'll get the output of AI as well.

2026-02-25 06:40:18 CET

https://www.reddit.com/r/singularity/comments/1r51i96/im_speechless/
LoLing so hard rn

2026-02-25 02:50:17 CET

https://preview.redd.it/qwen-qwen3-5-35b-a3b-hugging-face-v0-1nm81nlwajlg1.png?width=1547&format=png&auto=webp&s=b4286099d280591edfe25060a65699243666d6a4
Was using GLM 4.7 Flash locally, but Qwen 3.5 has arrived! Using 35b. Prompt processing seems waaay faster for me. Been using it all day, liking it.
Remember, every 9 months, the frontier arrives on consumer hardware!

2026-02-24 23:31:31 CET

- reply

I always had pi issues until I got a micro SD extender (thermally isolate it somewhat from the rest the pi), and did a crown reboot once a week

2026-02-24 19:36:11 CET

- reply

What went wrong?

2026-02-24 09:28:49 CET

People are hating on openclaw because the concept is simple.
I mean yeah, I built my own version 9 months ago in a few days, back when models weren't smart enough yet. It was a matter of time until they got better.
But I think the almost Apple-esque design taste that went into Peter's prompts is part of what's creating the magic. Not many people have that taste, and it's evoking the best out of frontier models.
One that caught my eye in USER.MD:
"The more you know, the more you can help. But remember - you're learning about a person, not building a dossier. Respect the difference."
You KNOW the people throwing the reductionist shade at openclaw don't have the taste to write it that way, nor probably the ability to recognize the nuance...
Now that I read it again, if feels like Opus wrote it. The emdash too... Maybe I should be giving the human less credit...

2026-02-22 20:07:40 CET

Accidental win for individual freedom!
1. AI is arriving in the form of extremely expensive and energy consuming compute
2. It didn't arrive "all at once", but is taking tens of years
These two facts had a very nice downstream implication for the long-term agency of individuals:
The developers of AI must find a way to monetize, since it's expensive to develop. And since primitive AI (let's say 2018-2024) wasn't smart enough to do much valuable work itself, they had to monetize by providing DIRECT access to individuals.
Now AI is starting to be economically valuable, but the frame, culture, and expectations have already been set; direct access to AI by individuals is just how the world works. (btw, pay attention if this starts to change).
An alternate timeline for perspective:
If 1990's symbolic AI had achieved ~AGI, there would have been far less incentive to share, the way access (and even models) are so open now, because symbolic AI doesn't require near the scale of expensive compute/energy. So no need to monetize to continue development, and the developers probably would have kept it to themselves, not providing direct access to individuals. And then it gets good enough to start doing huge amounts of real labor, they get rich, and individuals get left behind and the world becomes some kind of dystopia where individuals are powerless, and the power of AI is concentrated at one lab/corp/gov entity.

2026-02-21 23:17:22 CET

I wish chatgpt and claude had a "start a new chat with no memories about you"
Like chatgpt's Temporary Chat or claude's Incognito Chat, but I don't want the chat erased, I want to keep it.
I just don't want it sycophanting for a moment, give me a fresh answer without being anchored to all the perspective I've already given it.
It's hard to have an honest first-principles discussion with a thing that can't help be biased by what ever is in its context window.

2026-02-18 04:25:27 CET

AI is getting so good
This week my workflow completely changed. For the last year, my workflow has been to sit down with one project, and use AI to do small tasks while I supervise and handhold it the whole way (but at least I don't have to type out all the code). Same way I've always worked, but AI doing the typing.
Starting on Thursday my workflow completely changed. I've been able to work on 4 projects simultaneously. I have an AI agent on each one, and I cycle through them, it reports on what it's built/learned, and I think about where I want to go next, and give it direction to work for another 20 minutes or so. I'm thinking at the level of the project's purpose, high-level structure, business model, etc. I'm not down in the weeds, sometimes I don't even read the code it's writing. So I give some feedback to one AI, it sets off on its next chapter, and I switch to another AI. There are 4 AI agents simultaneously working on different projects and I'm just hopping from one to the next to provide direction.
I've had to change how I think about my role from 'coder' to 'tastemaker' and 'manager'.
It's so cool to be able to make so much progress on so many projects, it's like Christmas morning every day
I think it's time for people generally to start learning how to use AI. If I can do 4 projects at once now (each at roughly 5x my normal human speed), people who don't use AI at all will be left behind. That's the negative frame. The positive frame is, the barrier to entry to create whatever you want is dropping precipitously!
Now, it's always hard to estimate where you are when things are changing rapidly, and it's easy to over-estimate how fast things are going. Maybe work as we know it for most people won't be that different in a couple years. Maybe it will be totally different. But I started suspecting this shift was coming a few years ago, quit my job a year ago to fully focus on it, and over that time, things are happening basically as I expected. I think AI will continue to improve at minimum for the foreseeable future (~2 years), and will change my workflow to where I can be working on 20-30 projects at a time, while only talking to one AI which oversees them all. Like an AI company. And I think anyone can have that. Not just for coding projects, but all types. So it becomes hard to imagine how much more productive and efficient the economy will be, how much more agency people will have, etc, so many implications.
So many thoughts about how our governance isn't ready for this, how labor will slowly fade as a way for humans to get ahead (bifurcating society into those who had capital before AI came, and those who didn't), the ease with which a UBI-based society could fall into totalitarianism (which is why I'm building a gardening robot), etc. But those are other topics.

2026-02-17 22:55:43 CET

- reply

Lol

2026-02-15 02:28:48 CET

- reply

I only recently made the workflow update to cycling between 4 different projects where are agents are simultaneously working.
Crazy times, even compared to ~4 months ago.
What models do you use (TENEX, I assume you refer to here?)? Only the latest and greatest?

2026-02-13 08:08:51 CET

Same, weird how multitasking is suddenly a relevant skill
#nevent1q…0qlf

2026-02-09 01:03:27 CET

- reply

Funny pic, but the entire reason burritos are superior is because they don't fall apart like tacos. So he should be holding a taco
I am a burrito supremacist

2026-02-09 01:01:49 CET

- reply

Do you say 'jovial' because in the UK saying 'gay' is a crime? Would you actually get a visit from the speech police?

2026-02-08 00:22:03 CET

https://image.nostr.build/25d999281b6e3bf29ffddccf7ad6d7f5c000011a47c97f9c54a3fc86c2d93a3a.jpg

2026-01-02 17:38:38 CET

- reply

I know this dude, he sits and dozes all day at my library

2026-01-02 17:36:44 CET

- reply

We may not always reach thrm, but
what is life without lofty goals?

2025-12-26 06:54:52 CET

- reply

Pff what is this beardless schemer trying to get us to believe

2025-12-20 01:41:44 CET

Can't believe how good Gemini 3 Flash is for the price
2.5 Flash was already my favorite coding AI, so happy

2025-12-13 05:47:24 CET

- reply

no taxes on zits
no taxes on naps
no taxes on tits
no taxes on hats

2025-12-13 05:45:04 CET

this goes way too hard
Can you imagine if they actually started a contest where the auctioneers have to auction to a beat oof
Auctioneer Contest with Autotune
https://www.youtube.com/watch?v=rqZHN1aBbhE

2025-11-20 02:21:43 CET

- reply

Intermittent?
Intermittent and extended are different games

2025-11-20 00:57:28 CET

SAM3 make it super easy to make a 3D model by clicking on an object in a picture. Super cool, using it to make plant models for a garden simulation.
demo: https://aidemos.meta.com/segment-anything/editor/convert-image-to-3d

2025-11-19 00:00:08 CET

- reply

You trippin

2025-11-06 18:48:51 CET

- reply

Try extended fasting, you don't need to crash your BMR like this

2025-10-14 20:35:27 CEST

- reply

Gold

2025-10-03 01:59:19 CEST

World not differentiable?!? Foiled again!
https://npub15saszx8awfyj72apz2gveje8gx93lkahecapytfzjszw27n4jadq4e8xgu.blossom.band/06d3e3dc1342e4ba7af13a5403ff4a6f7f6e00c935a9104faaed2420ffa57075.png
haha reading Yann Lecunn's JEPA position paper ("A Path Towards Autonomous Machine Intelligence")

2025-09-25 01:52:19 CEST

Social media being more important than the moment is cringe
Unless it IS the moment, in which case it could either be fine, or doubly cringe

2025-09-23 00:19:31 CEST

note1stz7fh966flulf9vt8550k6nl0kljdzzh388gls97g052lek8pss2ucnat
Re: incentivizing intelligence agencies to release more information. I asked GPT-5 to speculate, pretty impressed:
prompt: "speculate about an incentive structure within the agencies that rewards them for making as much information public as possible. Now they optimize only for global strategic advantage and self-preservation, what if structurally we added releasing information as a third motivation?"
ChatGPT said:
That’s a fascinating inversion of the current design. Right now, as you say, U.S. intelligence agencies are structured to optimize for (1) national strategic advantage and (2) bureaucratic self-preservation. Secrecy serves both: it protects sources/methods and shields the agency from criticism or oversight. Adding a third structural incentive: “maximize responsible public disclosure” would fundamentally reshape their internal incentive calculus. Let me speculate on how such a system might work.
1. Why Secrecy is the Default Incentive
Risk asymmetry: If you over-classify, there’s rarely a penalty; if you under-classify and something leaks or goes wrong, it can end careers.
Budget politics: Bigger “black box” operations make it harder for outsiders to challenge budgets. Classified = harder to contest.
Cultural prestige: Within the agencies, being “read into” highly secret programs is status. More secrecy = more power, more exclusivity.
These pressures mean that secrecy is the dominant strategy under the current incentive structure.
2. Hypothetical: Building in “Disclosure Rewards”
Imagine Congress, the executive, or independent oversight creates a third metric: value of responsible public transparency. The agency and its personnel are rewarded when information is released to the public (without harming ongoing operations). Possible structural levers:
Budgetary incentives: A percentage of agency budget is tied to a “Disclosure Index.” Agencies that responsibly declassify and release more information get higher discretionary funding.
Career incentives: Promotions and awards explicitly tied to quality and quantity of responsible disclosures (e.g. “analyst of the year” for producing intelligence assessments later released with minimal redactions).
Reputational capital: Agencies score “public trust points” tracked by oversight boards, affecting leadership bonuses, leadership tenure, or even how much autonomy Congress grants them.
Competition between agencies: If CIA, NSA, DIA, etc. are benchmarked against each other for transparency, the “prestige” dynamic flips—leaders want to be seen as the most open without compromising security.
3. What Changes in Behavior?
Classification standards shift: Default might move from “classify unless clear it’s harmless” → “release unless clear it’s harmful.”
Proactive declassification pipelines: Dedicated offices constantly preparing sanitized versions of reports for public release, with incentives to do it quickly.
Strategic openness as a weapon: Agencies might deliberately release certain intel analyses not only to reassure the public but also to shape foreign adversaries’ behavior (e.g., exposing corruption networks, publishing satellite evidence).
Public engagement loop: Agencies could become knowledge-brokers to the citizenry, publishing historical case studies, trend reports, and sanitized operational lessons, thereby turning intelligence into a “national commons” resource as much as a secret tool.
4. Potential Benefits
Trust repair: Citizens see intelligence as working for them, not over them.
Policy legitimacy: Elected officials can point to intelligence disclosures as grounds for decisions, reducing suspicion that policy is being driven by a “shadow state.”
Deterrence by transparency: Publishing adversary misdeeds (e.g. disinformation ops) can weaken rivals without kinetic action.
Innovation spillover: Releasing sanitized technical methods (cryptography, satellite imaging, AI analysis) accelerates civilian science/industry.
5. Potential Risks
Over-disclosure arms adversaries: Too much revealed, too soon, can compromise capabilities. Agencies might game the incentive by releasing irrelevant or already-known material.
Perverse incentives: Leaders may push to “pad” transparency stats with low-value data dumps rather than meaningful releases.
Short-termism: Drive to disclose could bias against long-term clandestine projects where secrecy is crucial.
Weaponization of disclosure metrics: Rival political parties could pressure agencies to release intelligence favorable to their narratives.

2025-09-23 00:10:42 CEST

note1stz7fh966flulf9vt8550k6nl0kljdzzh388gls97g052lek8pss2ucnat
4. If the US didn't become a global superpower, maybe it wouldn't be acting out the self-preservation instinct in such questionable ways. Maybe this always happens to empires. Maybe they get paranoid.

2025-09-22 23:37:33 CEST

note1k2yfjnz39warjavuz28k0tdyc008azv70jqemxe9hrx59yjnjwzsm0h3v9
My thoughts and ideas for solutions (after an entire 30 minutes of thought):
1. Maybe these intel guys aren't as creepy or as bad of freedom-hating bureaucrats as I thought. I always believed the vast majority were doing their job as best they could. But now maybe I believe that a little more, even for the leadership that make some questionable, sometimes infuriating decisions from the outside (though the possibility of corruption via various forces is always there).
2. The national security risks of forcing the Intelligence Agencies to take their foot off the gas are probably not as bad as they'd like us to believe, or even believe themselves. TSA security theater comes to mind. Though, not to say the risks are trivial.
3. https://sgp.fas.org/eprint/limits.pdf - there is significant hope. Freedom Of Information requests are effective. There are ways to shine light into the secrecy. But the incentives are currently such that significant effort is required to do so. We should *reverse* that incentive structure; somehow reward the intelligence agencies for open sourcing as much information as they possibly can. Reward them at the individual level, and the organizational level. Then they, the ones with all the context, can balance strategic advantage with accountability. Currently they're only optimizing for strategic advantage. The FOIA is good on its own, but frames things in such a way as to make revelation the exception, not the default.

2025-09-22 23:20:32 CEST

My current read of why the US government is full of such spooky bois:
The stark truth of the often spooky and almost adversarial-feeling relationship of Americans with their own country's Intelligence Agencies is :
That secrecy, even with respect to their own populace, is necessary to maintain intelligence and strategic advantage over rival nations. The intelligence agencies are entrusted with that power not because anyone elected them, but by the nature of their work.
This creates a point of conflict between the sovereignty of the free people, their constitution and its amendments (see the 4th), and elected leaders on one hand, and the intelligence agencies that oversee their safety on the other. The 'puppet' perception of elected leaders vis a vis their apparent beholdness to shadowy intelligence agencies is a clear example.
The seriousness of this conflict was muted until the Cold War, when the budget and sophistication of Intelligence Agencies reached never before seen levels, and then was reticent to surrender this perceived world advantage after the war ended.
Freedom-seeking nations are currently in a time where the populace doesn't understand this conflict, Intelligence Agencies are scared to relinquish their power, for self-preservation reasons, but also because they can see the national risks that would follow, and elected leaders who understand the conflict don't feel it prudent to completely side with the full sovereignty of the populace (which would require them to 'blow the lid' on the secret nature of the nation's operation, or otherwise undermine its operation), for fear of the same national risks the Intelligence Agencies see, as well as self-preservation (ex. some think JFK was taking this route, and was gruesomely silenced by the Intelligence Agencies).
So there's no clear path forward without very decisive action, so the story of the decades since the Cold War has been one of mission and scope creep of the Intelligence Agencies, no real architecture in place for how to balance secrecy, safety, and strategic advantage, with sovereignty and democratic accountability.
will follow with my ideas for solutions

2025-09-22 23:12:03 CEST

Sometimes to *see* clearly you have to *be* clearly first.
Clean your room.

2025-09-22 19:17:56 CEST

- reply

Don't know what's going on but I'm laughing

2025-09-17 19:47:15 CEST

- reply

When I was learning git I had a bash alias for git-help-quick which printed a full page ASCII art of that famous picture of Linus flipping the camera off

2025-09-15 23:45:20 CEST

A glorious win for the good people of Nostr
#breadhasarrived
#nevent1q…6x4z

2025-09-11 21:20:11 CEST

- reply

What's your github?
sorry your website isn't loading for me

2025-09-11 00:34:03 CEST

- reply

Pumped. I've had so many ideas for making something like your tenex project, heard you go through it in a podcast. But I keep telling myself, 'stay focused on your main project, everyone is having the same self-improving agent swarm idea, you'll be able to use someone else's soon without spending the time yourself'
On a COMPLETELY unrelated note, can I use tenex? Is it oss?
Also, despite my attempt to not get distracted by it, I've spent approx 40 hrs over the last 6 months developing the principles I would use to make my version of tenex. Would be down to share ideas

2025-09-11 00:19:30 CEST

note1n0jamfsr9ln09njjqs06ltc36nrwaa4qmcsyr43chdrvpa55szyqn5kqhv
This made me think about the other side, how much more intelligence will there be from today (at the frontier), independent of compute required (within the same fundamental autoregressive transformer and pretraining + RL paradigm)? My gut says at least another OG GPT-4 to today's frontier jump (Grok-4/Gemini 2.5 Pro/Claude Sonnet 4/GPT-5)
But I think a qualitative jump to really feeling like human-level intelligence will require something like a module within the network that pays attention to the coherence of a new "thought direction" WRT its own model of reality, not just next-token-prediction on steroids.
(that qualitative jump would manifest as a huge hallucination reduction, the ability to zoom out and self-correct after mistakenly honing in on the wrong path, in short, the model feeling like it's responding from a place of operating from a real mental model of the world, as opposed to today's models, responding from a place of most-likely-next-token)
And a neurology foray I made a few days ago leads to say
"wow neurology and ML are starting to converge, and some of the computational models that neurologists are making strike me as believable substrates for this type of 'check thoughts against a causal world model' module I think is required for the next step in LLMs". That didn't really feel true to me a few years ago
Examples of neurologist's recent models I looked at: TEM machine (entorihnal+hippocampus analogous), Spaun (spiking NN, surprisingly holistic macro brain model)"
And recently HRM or Heirarchical Reasoning Models (did well on the ARC-AGI-1 benchmark, caused a bit of a stir) strike me as the latest thing closing the neurology/AI gap, and accomodating insights from both.
Excited and kinda concerned for the next 5 years.
This is the time to build AI tech that gives power to individuals! Otherwise we might be SOL, at the mercy of oligarchical whims...

2025-09-10 23:54:18 CEST

based on the fact that quantization is so effective (which basically means there's still a lot of waste in today's LLMs), I think we have at least another 5x efficiency to go for the same intelligence compared to today's models
see https://docs.unsloth.ai/basics/unsloth-dynamic-ggufs-on-aider-polyglot

2025-09-03 19:38:18 CEST

- reply

That's crazy

2025-08-21 18:56:45 CEST

- reply

Wow the audio really makes this a gem

2025-08-21 18:55:23 CEST

- reply

Lol

2025-08-09 20:45:15 CEST

- reply

Good headphones make this 3x funnier for me

2025-08-07 03:02:18 CEST

Viewdio games could really use great 3D assets like this one
https://image.nostr.build/96be52543678269b0e69d2395fa6941795d1dd1dcdef3ddb88aa176da8291210.jpg

2025-08-03 04:44:01 CEST

Today's thought, critique welcome (I'm trying to figure out why I viscerally hate/fear totalitarianism, to the point that I quit my job to work on tech that I think helps prevent it):
Humans are many things, but a majorly salient one, is that they are a memetic species. Compared to many life forms, they are weak, unprotected, and helpless given only the devices of their physical form and abilities. Their memes are their power, and they are the dominant species, and compoundingly so, through their ability to efficiently share memes. Each generation starts from a new point, unlike other species.
Cultures are centralization points of memetic patterns. Since there are many cultures, the evolution of memes continues despite these centralizations, and arguably, much better, since many people can develop the culture to explore its local optimum better than an individual, given the shortness of one individual's life. Different cultures produce different steps forward, and create memes that other cultures end up adopting. Humanity moves 'forward'.
But cultures resist change. And some cultures are more resistant than others. The continued existence of a culture depends on its ability to balance the maintenance of its core defining characteristics, along with the appropriation of successful memes demonstrated by other cultures (or luminaries from within the culture). If the culture biases too conservative, and the culture eventually finds itself irrelevant, outdated, and powerless compared to surrounding cultures that have since evolved; too liberal, and the culture can destroy itself by disintegration, all shared memes, beliefs, and assumptions being lost.
History has seen many dynamics in culture interactions, but one driving force has been, despite ebbs and flows, the successive march of centralization of culture (which has thus far always been disrupted by forces I'll explore in a second, forces which may soon cease to exist). Monoculture is a word that has entered the lexicon in recent decades, with the advent of mass media. Many aspects of the Western world's culture can be described once, and applied reasonably accurately to many peoples. This wasn't nearly as true 100 years ago for those same geographical peoples. The world only remains with a handful of such 'monocultures', something like: the West, the Russians, the Chinese, and a few others. Should trends continue, in another 100 years, there might be only one.
If there is only one culture, then the qualities of that culture become very important with respect to the continuation of memetic evolution. If it is too controlling/opinionated/oppressive, the progression and evolution of novel memes will slow down among humanity: controlling governments have demonstrated their for dislike the uncertainty that comes with the open-mindedness, creativity, experimenting, and inventive nature of some humans. See also 1984, Fahrenheit 451, etc. Oppression, whether explicit or covert, is a natural path for an entity that gets power and wants to keep it (which is, all entities that get power.).
Governments usually only exist with the consent of the culture they administer to. But in a scenario with only one culture, if its government became totalitarian, the usual forces of competition from other cultures wouldn't exist as a correcting force. And add to that a scenario where technology is sufficiently advanced, the usual forces of the raw difficulty of human survival would be less pressing as a correcting force. Technology makes survival easier. Advanced technology makes it easy.
For example, if the Soviet Union/Roman Empire/<your favorite historical empire that controlled a big part of the world> didn't become economically stagnant, and didn't have other cultures fighting it, it probably wouldn't have fallen. If it had human-level AI and robots, it probably would have been fine economically, and if it were the only game in town (on Earth), it wouldn't need to defend itself, and its people wouldn't have other cultures to look at as a mirror to see whether they really like their own culture. And another point on the tech, if it could all but monitor, predict, and control its people's thoughts through the media they consume and share, it could gaslight everyone so hard that nobody thinks to revolt.
Western governments, and cultures, have demonstrated their weaknesses to falling to controlling tendencies: mass covert and unconsensual data surveillance, covid-era policies, and most recently, age-verification requirements online, which many see as a trojan horse for another step in the direction of surveillance and control.
We need to find the 'lines in the sand' so to speak, the powers and explicit non-powers, of a theoretical healthy government/culture that would allow humanity to continue to progress despite the advent of a world monoculture, which may be in our future. Obvious things on the right side of the line would be things like protection from violence and a functioning justice system. Obvious things on the wrong side of the line would be things like controlling thoughts, opinions, and speech in heavy-handed ways. The line is often gray, and some muddy it intentionally, but unless we can arrive at a clarity and consensus, the arrival of a world monoculture may be the announcement of humanity's possible descent into a dark age, whether in a year or 5 generations, a dark age whose duration could be indeterminate.
During that time, the creative, exploratory, and adventurous elements of the human spirit could be stamped out, in favor of a controllable, predictable, and stale populace. And along with the heart-breaking darkness and ignorance imposed on the individuals that exist then, the retardation of humanity's progress on understanding ourselves and the universe (the ultimate memetic achievement), finding better ways to exist, and creating the technology to expand throughout the universe and survive beyond the life of our cosmically precarious planet. If it lasted long enough, it could spell our extinction.
Summary: we really need to solve governance BEFORE we arrive at a world monoculture and technology so good that material survival is easy. Otherwise we could have a long dark night for humanity, create a real dystopia that trillions of people have to suffer through, and jeopardize the eventual achievement of our most human, most transcendent goals.
We gotta keep making new memes, man.

2025-08-01 17:41:00 CEST

- reply

And these boys never heard from anyone ever again.
...They saw plenty of people, just couldn't hear them

2025-07-24 21:05:01 CEST

- reply

Lol

2025-07-24 21:04:16 CEST

- reply

Feeling influenced

2025-07-24 18:37:45 CEST

- reply

I might have come but am heads down on a project at the moment, and it's not directly nostr or bitcoin related.
It is very much sovereign tech though (backyard gardening robotics), and motivated from that mindset.
Love the concept, maybe next year

2025-07-23 05:12:00 CEST

- reply

Would love to hear your workflow

2025-07-21 18:54:44 CEST

- reply

I discovered iNaturalist looking for plant image datasets recently, so cool!

2025-06-29 18:05:29 CEST

- reply

It's so much fun to move so fast. Feels like I grew up with a horse and buggy and am now driving around in an SUV

2025-06-28 06:46:28 CEST

Wilbur Gardening Robot log - June 27, 2025
This week I surveyed the research on learning-based methods for robot mapping, localization, and navigation. I went this direction after spending a couple weeks surveying traditional SLAM (Simultaneous Localization and Mapping) approaches. There's a lot there. And then I realized, 'this is the old world. this was all made before LLMs, VLMs, and VLAs existed. I should focus on the new world'. So I switched. I think learned SLAM is the answer for Wilbur. A VLA is definitely the answer for actuation. After realizing that the evolution of traditional SLAM to learned SLAM would bring new generality, accuracy, and flexibility, I took some time to speculate about what the steady state of this research should be after all the fundamental advancements have been made. The 'Platonic Ideal' so to speak of learned SLAM. With my newfound clarity, I dove back into the research and was delighted to find that many of those ideas have been explored! So Thursday night and Friday were diving into the 'new world' of learned SLAM research, hoping to wrap that survey up by Tuesday, at which point I should have clear options in mind about what the weeding MVP robot should use. Then I'll probably narrow down possibilities and make a decision by the end of the week.

2025-06-28 04:20:17 CEST

- reply

It's a good idea because blackjack and hookers require no product development, only blackjack and hookers. Instant revenue, leaving more time for real product development

2025-06-28 04:18:09 CEST

- reply

Am I? Wonderful, in that case we will also offer blackjack and hookers to extend our runway to profitability

2025-06-27 18:33:23 CEST

- reply

They get a discount if they commit to 6 months and let the drone drop a persistent solar antenna on their roof

2025-06-27 05:35:22 CEST

- reply

Reticulum + LoRA seems like the easiest networking starting point. Ah look, you've got me scheming

2025-06-27 05:33:38 CEST

- reply

Would be so fun to make, I love mesh networks, drones, autonomy, solar power, ticks so many boxes. Maybe the tech will be ready in a decade. Seems like today you'd have too many airspace legality/battery problems. Though overnight-airborne solar plane drones have been achieved, so technically possible...

2025-06-25 04:07:23 CEST

- reply

yeah kinda weird how filling they are. Pickle juice helped me a lot to finish my first 14 day fast after being stuck at 10 for a few attempts. I thought it was just the electrolytes, but you probably aren't terribly short on electrolytes during a calorie restriction diet, so... maybe pickles are special. maybe it's the explosion of flavor

2025-06-24 22:58:14 CEST

- reply

Wut.. Google launched A2A like two months ago. You mean they adopted it or something?