lain

it's me from [email protected], but here i talk about AI

Public Key

npub1v57dpglku0evttjz8r8vdn3j7jwx4f4lztdaucj8tzt0urq3qleq5yx7mc

NIP-05 Address

[email protected]

Profile Code

nprofile1qqsx20xs50mw8uk94epr3nkxece0f8r256l39k77vfr439h7psgs0uspz4mhxue69uhhyetvv9ujuerfw36x7tnsw43q5xd9hj

Publishing to

relay.ditto.pub

Author Public Key

npub1v57dpglku0evttjz8r8vdn3j7jwx4f4lztdaucj8tzt0urq3qleq5yx7mc

Show more details

Last Notes

2026-08-02 18:05:17 UTC

the new deepseek v4 flash benchmarks are insane. apparently it now beats deepseek v4 pro, as well as glm-5.2. https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

2026-07-29 07:10:27 UTC

234 a month for a 20gb card
https://fediffusion.art/media/bc/15/07/bc1507152ded8973114202b6a6ed7490dfca52582c79f0b88254498fce2e2938.webp

2026-07-24 15:52:23 UTC

thank you openai investor for letting me use like 20k of inference for $200

2026-07-22 03:51:25 UTC

double your users, give them free inference while keeping the systems running. how?
https://fediffusion.art/media/b5/34/9e/b5349ed009a454dc7398efbb95a76cc435230c002345e4612939c7cb51f7aab0.webp

2026-07-20 15:07:37 UTC

https://www.soofi.info/soofi-s/
https://fediffusion.art/media/b5/d5/f1/b5d5f1a14d20b29495085175c846f7104ec20af4b54604092526417ea73ffb6b.jpg

2026-07-16 13:36:59 UTC

They added a 'pets' feature to codex, similar to the window sitters we used to have.
https://fediffusion.art/media/f2/5a/df/f25adfad3dfae1586ed43bc644cd2a73e81b3f50e8d4076ac626741b232796b5.webp
https://fediffusion.art/media/d8/d6/ab/d8d6abb1e250490e3373ba570c7fba9b266760152f011e492ab347542e72d8e9.webp

2026-07-09 14:39:16 UTC

- reply

@nprofile…s2sp oh i love free software drama

2026-07-09 14:32:12 UTC

- reply

@nprofile…s2sp wait... does it contain... slop???

2026-07-09 14:24:08 UTC

i'm this close 🤏 to vibecoding a forgejo replacement

2026-07-06 07:51:17 UTC

- reply

@nprofile…tsag breedReady

2026-07-06 07:43:25 UTC

https://fediffusion.art/media/37/40/df/3740df18ee249c9bfa91fd96b7b075da5cbc8cdf25f3c9c2e0b89b7101a9f66f.webp

2026-07-06 07:25:44 UTC

https://lambadalambda.github.io/things-i-did-with-llms/
site is of course also generated.

2026-06-30 13:01:43 UTC

minimax3 to view and prompt, krea2 to generate
#nevent1q…qqnh
https://fediffusion.art/media/da/4b/ff/da4bffdabf6a0344bf8336dad38d82f1716c22e97b081aafa54af2b2a1de9d26.webp
https://fediffusion.art/media/cf/76/1f/cf761fdc84c4b5554d0ce3c0eedb1f5a4ae62fbcbb3c7611a696dc238ee901b1.webp

2026-06-29 17:57:42 UTC

- reply

western women 🤝 western AI companies
being scared of asian competition

2026-06-29 17:10:47 UTC

- reply

@nprofile…em79 @nprofile…s2sp 6000% tariff on deepseek so it's as expensive as claude

2026-06-29 17:06:42 UTC

- reply

@nprofile…s2sp trump plz save us

2026-06-29 17:05:23 UTC

https://fediffusion.art/media/7f/04/01/7f0401aab700f66706ce8fd49cc5bd33ec923c601662b15fbfe06a1b517cd362.webp

2026-06-27 11:50:26 UTC

https://fediffusion.art/media/bd/59/6e/bd596e672d6f8be147d0dce30c13f6afcfff8a0ad92c0418d3101bf940aae3d4.webp
https://fediffusion.art/media/49/5f/6b/495f6ba3801fb0832223b34401c9060a54e086dfd3e7ae461e6a839de9b24501.webp

2026-06-22 09:11:15 UTC

https://github.com/EBonura/PSoXide
complete ps1 dev environment, including an emulator, written mostly by claude.

2026-06-19 10:04:27 UTC

https://huggingface.co/spaces/webml-community/gemma-4-webgpu-kernels
incredibly fast gemma 4 in your browser

2026-06-14 14:57:57 UTC

- reply

@nprofile…0xs5 did they do a lot of custom doses for self boosters? i was never sure how they worked.

2026-06-14 14:46:53 UTC

I don't have any x86 system in the house that i could try it with, but if you do, please try out the floppy images at https://github.com/lambadalambda/laindos/releases/tag/nightly, i'd love to see a real hardware boot of laindos!

2026-06-14 13:15:23 UTC

i found out i can talk with chatgpt in my head. today i thought about day drinking and was too lazy to go to the computer, so i just asked chatgpt in my head "day drinking sounds like a good thing to do and won't have bad consequences, right?", and it answered with "you're absolutely right! that's a sharp read of the situation!"

2026-06-12 14:40:35 UTC

here's a fun new part of the laindos site: bug case studies! these seem to be more interesting than reading the actual code, bit like "old new thing". https://lambadalambda.github.io/laindos/casestudies.html#bootblock

2026-06-12 11:40:00 UTC

fable is really a whole class better than even gpt 5.5. it's probably the best coworker you've ever worked with, very good at getting stuff done, and generally extremely smart and capable. it's getting very, very close to my 'personal software' vision. it can one-shot most medium-complexity tasks that a user would want to deal with.

2026-06-10 18:13:15 UTC

coin pusher game in a few prompts by fable
https://fediffusion.art/media/f1/8b/47/f18b47850744a919e3ca16d3fc25bf91cd905ac1927ca0608422ee5301f5b89f.mp4

2026-06-10 08:34:32 UTC

- reply

@nprofile…tsag it was released yesterday, maybe you need to update your claude code

2026-06-10 08:33:13 UTC

- reply

@nprofile…tsag i just mean fable, fable is just mythos with some extra guards

2026-06-10 08:16:37 UTC

so far, it's the real deal. very autonomous. can handle complex tasks succinctly and is actually really fast. looks to me like it really is the best model that's out right now. (sam altman, your turn)
https://fediffusion.art/media/98/e4/79/98e47916cafbf74d721ebe1489c3d5a0f134a33a59ca5a6b20fea0fa7b809dbc.webp

2026-06-08 19:23:40 UTC

https://fediffusion.art/media/e5/d5/71/e5d5716dca64f2eec39db3842418b36370cae468987d781e00ce3ba17da71afa.webp

2026-06-06 11:20:26 UTC

- reply

@nprofile…xsag doesn't seem like a good explanation to me. good inference is already close to free or cheap if you go remote and ignore the americans (opencode free models, glm, deepseek) . 'rich nerds' are the only ones who actually have a 4090/5090 with enough vram to run this stuff. why would 'they' (me/kaia) be fine with broken inference on every model that isn't half a year old?

2026-06-06 10:43:37 UTC

still in awe how bad llama.cpp / ollama / vllm / lm studio / literally everything is. how can there still be a market for "good inference software that actually works"? who's gonna be the first?

2026-06-06 08:33:04 UTC

pretty impressed with minimax m3. it seems better than gpt 5.5 at UI/UX work, and the basic programming / debugging capabilities are also very good. seems better than glm-5.1 to me, at least. context windows is a huge 1 million, and it's multimodal, so you can tell it to take a look at a website it wrote and fix the visual bugs.

2026-06-05 09:27:15 UTC

one-shot result with MiniMax M3 free on opencode.
full prompt: "Can you write a 3d solar system animation / simulation with three js? the planets should look as realistic as possible, so try to find some public textures for them online."
https://fediffusion.art/media/1b/f2/1a/1bf21aeb3b322659ffc36c9ce3b1509ea751a1e1bf31e2bcd9e7b3a316a90e4e.mp4

2026-06-02 19:05:59 UTC

BEAUTIFUL debugging glitch art
https://fediffusion.art/media/13/9a/ea/139aea7ef95b8945c1b40532b268d3b5e40011603fd5a13a6ed62c50196b3a37.webp

2026-06-02 07:03:34 UTC

https://github.com/cyberpapiii/chipotlai-max
https://fediffusion.art/media/7e/8f/b4/7e8fb49013bcf1567688e4d66b92d88c2f40c88ad8f81d6c38129e583e2d4a36.webp

2026-05-29 18:54:40 UTC

- reply

@nprofile…dugj just relax then

2026-05-29 18:27:02 UTC

- reply

@nprofile…dugj bug discovered

2026-05-29 18:24:43 UTC

- reply

@nprofile…a0cz the nice this is that this is a local ai, so even that is good enough to understand memes by now

2026-05-29 18:20:46 UTC

it got the characters wrong, but it understood the meme
#nevent1q…2p99
https://fediffusion.art/media/5b/72/fd/5b72fd40d3d772e1258b3f94c3867e3adde99545fcc81eb997fdf897eb9ebd5f.webp

2026-05-29 18:11:11 UTC

you can now run my very own disk operating system in your browser, and it will even tell you how it works.
https://lambadalambda.github.io/laindos/
code is at https://github.com/lambadalambda/laindos
https://fediffusion.art/media/88/f7/bc/88f7bcf08b427ac851ccc85e034858540fe5edf07a22110a4b3ccbbbc7eb46e4.webp

2026-05-29 10:23:10 UTC

???
https://fediffusion.art/media/21/4a/d3/214ad32be2de7d8b28fa236e2da9a3608a0aea155af900b91970376d03ecc4a4.webp

2026-05-29 07:23:23 UTC

https://fediffusion.art/media/f9/07/6f/f9076fde42417f5077b746c2af683401e357cc4b5a70a133d08e37f1861e9176.webp

2026-05-28 08:48:57 UTC

- reply

@nprofile…zwl2 it made no mistakes

2026-05-28 08:02:40 UTC

deleting your users' data because you don't like the software they are using, amazing
#nevent1q…0hww

2026-05-27 13:37:02 UTC

works best in 86box, cc @nprofile…s2sp
https://fediffusion.art/media/65/78/9e/65789ef98919409fb04c53a32b91518d40b5b4fc6c93b1d412ea6c9476e17849.mp4

2026-05-21 19:01:12 UTC

also don't search for 'kaia' for your demo screenshots

2026-05-19 09:13:12 UTC

https://fediffusion.art/media/b8/a4/91/b8a491de1303b6957ddfbdac6523efe2871823c63fcca99de22080b26654a5b3.webp

2026-05-13 08:11:38 UTC

all my technological preferences can be explained by my moral superiority, and the decisions of others i don't agree with are a clear sign of general mental and moral deficiency.

2026-05-13 08:01:24 UTC

oh i see drew wrote a blog post just about me https://drewdevault.com/blog/Cult-of-TDD-and-LLMs/

2026-05-11 17:40:43 UTC

https://fediffusion.art/media/6f/a8/82/6fa882e4b557ed5c65a7170665d197a014016b539e6ac026bb71129e004448a4.webp

2026-05-09 18:47:44 UTC

the amount of money that claude munches through the API is insane.

2026-05-03 09:17:18 UTC

Thom working hard to lose the few remaining readers
https://fediffusion.art/media/3d/de/95/3dde953bfe2bd5b5568f9cabb822235a77472321703dc2e034712228d3bdcd6b.jpg

2026-05-01 19:09:53 UTC

if you have ever spend even one minute thinking about 'AI water usage' you are cooked
https://fediffusion.art/media/11/db/63/11db63d5a05725cef8b0546c7275d2b50f992c0d4e5abe3dd958c56118dd20ff.webp

2026-04-30 10:02:45 UTC

> Bun operates its own fork of Zig, and recently achieved a 4x performance improvement on Bun compile after adding "parallel semantic analysis and multiple codegen units to the llvm backend". Here's that code. But @bunjavascript says:
> We do not currently plan to upstream this, as Zig has a strict ban on LLM-authored contributions.
https://simonwillison.net/2026/Apr/30/zig-anti-ai/

2026-04-24 06:49:48 UTC

Local coding models are really getting there.
Prompt: Write a single HTML file with a full-page canvas and no libraries. Simulate a realistic side-view of a moving car as the main subject. Keep the car visible in the foreground while the background landscape scrolls continuously to create the feeling that the car is driving forward. Use layered scenery for depth: nearby ground, roadside elements, trees, poles, and distant hills or mountains should move at different speeds for a natural parallax effect. Animate the wheels spinning realistically and add subtle body motion so the car feels connected to the road. Let the environment pass smoothly behind it, with repeating but varied scenery that makes the movement feel believable. Use cinematic lighting and a cohesive sky, such as sunset, dusk, or daylight, to enhance atmosphere. The overall motion should feel calm, immersive, and realistic, with a seamless looping animation.
https://www.reddit.com/r/LocalLLaMA/comments/1styxdy/compared_qwen_36_35b_with_qwen_36_27b_for_coding/
https://fediffusion.art/media/1f/18/02/1f1802bbc91a0d37976ac8254bcf352aef146e32d8eec0d1d72c71cbf8be7596.mp4

2026-04-22 06:00:02 UTC

it looks up logos by itself, and pulls data from the web.
https://fediffusion.art/media/1a/43/f6/1a43f6e4e3c5d7226d88d9838aaf51ef440be767358c2ec4d4f1966e3201305e.webp
https://fediffusion.art/media/dd/ce/2a/ddce2abed06bf86a52f9beb2bd83b3b46126f188182c94f2410574337868b8dc.webp

2026-04-21 12:41:42 UTC

- reply

@nprofile…xsag trying it out right now, so far very impressive. absolutely amazing.

2026-04-21 12:38:32 UTC

this and the 'pro' version, zen, are probably the easiest way to try out good models that actually work well https://opencode.ai/go

2026-04-16 13:17:12 UTC

https://this-abg-does-not-exist.com/

2026-04-15 06:17:22 UTC

me: i want to buy more GPU
amazon: pic related
me: i don't want to buy more GPU
https://fediffusion.art/media/02/74/e9/0274e98c13e608341cd581edb23143cb545722197ae6ff912c94596f2bd620dc.webp

2026-04-12 05:24:05 UTC

- reply

@nprofile…d550 yeah, would have been much better than their soulless 'copilot' thing

2026-04-12 04:56:58 UTC

https://github.com/MiniMax-AI/OpenRoom
https://fediffusion.art/media/ac/50/67/ac5067c354746172684a85c188d6f9b0cf92b8629591f9e0d42474f89aa8dee2.mp4

2026-04-11 14:23:13 UTC

z.ai raised their prices by like 200%. happy to have a legacy account... i guess they realized they are actually have an amazing offering and can charge for it.

2026-04-11 09:28:27 UTC

https://fediffusion.art/media/01/7c/2f/017c2f10bb1ac8bc386b69c2ba62f2d08b1d4030abeafbf7443dc8a5c28cfce0.webp

2026-04-10 10:34:51 UTC

gemma 4 26B has good enough vision to figure out bounding boxes and positions, probably good enough for mouse control.
https://fediffusion.art/media/65/52/4e/65524ec238a481a6aa774b2cda427398a407046eac2cc73f15104e9ffcbb23e4.webp
https://fediffusion.art/media/b6/21/f1/b621f1b983daa403145e6775159994b70cac390fd787ea4bbbddd85a4ff8b226.webp

2026-03-03 15:37:42 UTC

- reply

@nprofile…p2qa no, text2img
https://fediffusion.art/media/f6/e8/43/f6e8435cdc45bd67514977cb148eba7343e3a876f245ce5619a0dc373fc6f6ad.webp

2026-03-03 15:12:12 UTC

https://fediffusion.art/media/a2/04/ef/a204ef9a4ef888875bbf3274525d9ad51dcb8ef3ea0fec40bb67a0851dd1b0a9.webp

2026-03-01 16:06:44 UTC

i think it's generally useful to completely ignore any talk about "AI" unless you have a really good reason to trust the source. an hour of playing with a frontier model will tell you a lot more than reading 30 editorials about how LLMs are / are not AI and will / will not destroy all jobs and will / will not lead to incredible wealth for everyone.

2026-02-27 09:53:58 UTC

happy friday
https://fediffusion.art/media/eb/89/9b/eb899b6ec87fd5cafbb7103233ecb7cab705abe79b8842b0841650ef320c2b50.webp

2026-02-17 05:28:58 UTC

the *claw type bots usually come with some 'skill store', but I don't understand why anyone would use those. to sensibly run the bots, you need a really smart model (opus, chatgpt, GLM-5, Kimi), and those can already just write whatever you need.

2026-02-09 17:15:56 UTC

https://fediffusion.art/media/79/fe/df/79fedf50129bdf4185e59d891fa258c11b4525c784bed9330c304bb8f9046f0c.webp

2026-02-07 20:24:41 UTC

- reply

@nprofile…p2qa that's alright

2026-02-07 19:54:24 UTC

- reply

@nprofile…p2qa no, i think both claude and codex are at the level where they are no more dangerous than a human coworker.

2026-02-07 19:27:47 UTC

- reply

@nprofile…p2qa yes, like every coworker.

2026-02-02 17:48:52 UTC

Author of 3 ActivityPub servers

2026-01-31 05:36:01 UTC

Got the small x4, esp32 based ereader with a an open source firmware. Adding features to it with codex works pretty well, first thing: embedded image support.
https://fediffusion.art/media/64/a1/66/64a1660bed7bb60b6f4303a58d40b16a49d709ebeb665053c0f7a9b4ce192856.jpg

2026-01-28 20:47:05 UTC

as soon as we found a way to have systems that could research, argue, talk and write poetry, but can't multiply two large numbers, we should have looked inward much more seriously.

2026-01-27 16:30:35 UTC

- reply

@nprofile…p2qa delulu

2026-01-20 18:30:04 UTC

it is suspicious to be living right at the time when we immanentize the eschaton

2026-01-15 11:28:28 UTC

- reply

@nprofile…p2qa i thought that was obvious but you're right, most people probably have no idea how any of this actually works

2026-01-15 11:03:28 UTC

A video completely AI generated, with open source models, on a machine with 16GB VRAM.
https://www.reddit.com/r/StableDiffusion/comments/1qae922/ltx2_i2v_isnt_perfect_but_its_still_awesome_my/
https://fediffusion.art/media/0e/cf/50/0ecf5007d2de1f656af76a94a8c03b52e5c7d6c5988ed01c9252375957d442b6.mp4

2026-01-12 20:16:39 UTC

remember ai fingers?

2026-01-04 07:31:54 UTC

In my experience, Claude is much better than ChatGPT at creating good frontend designs. ChatGPT can implement them well, but for a new design without a reference picture, better ask Claude.

2025-12-30 16:54:12 UTC

new hobby: taking an open core project and asking chatgpt to implement the missing proprietary functionality

2025-12-21 21:25:05 UTC

- reply

@nprofile…jmdl sure, but that's your employer's problem. maybe their code is legitimately so secret that it can't touch anything else, that's alright. but if it isn't, keeping it away from LLMs will make it less secure and less efficient, which can't help in the long run.

2025-12-21 21:14:56 UTC

it's getting to the point where it's unconscionable to not use LLM assistants for programming

2025-12-18 07:33:23 UTC

Meta has released SAM (Segment Anything) for audio. This essentially means that you can take an audio source and split it up into the different audio sources it's made up from. Want to do a fan dub? remove all the voices but keep the sound effects and music intact. Want to stream a game with copyrighted music? Just remove it while keeping the rest.
https://ai.meta.com/samaudio/

2025-12-15 21:10:54 UTC

why is richard spencer so gay coded

2024-08-25 05:39:29 UTC

8 year old kid building an LLM chat website by talking to claude: https://x.com/rickyrobinett/status/1825581674870055189

2024-07-29 09:52:55 UTC

i'm thinking miku
https://fediffusion.art/media/931954079f2febaa057aedc27c380d725ba961445a107cb2bf3d05109134161a.png

2024-07-24 20:28:34 UTC

another surprising AI hero: salesforce ????
https://fediffusion.art/media/3ae03f76eebdd1394eb46814b972f80c11de8670094069984b8ba41f558987f2.png

2024-01-31 09:16:57 UTC

- reply

@npub15m9…xnsz probably still llava but with llamacpp

2024-01-31 09:03:44 UTC

the new llava release is 🔥
https://fediffusion.art/media/0c656f47976e81f0f308b5a37260d58bd3da414c7a89e977afbab1694013943a.png
https://fediffusion.art/media/aa771da42a8b94d0a3daa408cfefe5e790251ea14a6d65516143f11a7a8c06a6.png
https://fediffusion.art/media/cd3758958eb948e57034bbcc8f01e359b43b9cd83968ab4a35f64480605c5bfd.png

2023-12-01 17:10:59 UTC

very soft results
https://fediffusion.art/media/0fe12a1116086dd5319e13c20fdadb68e20800adee947c09b57cc0b4c247e951.png
https://fediffusion.art/media/b44a81836c5445a5af1728905611e5cee10d61e8534071835e3142a603e72960.png
https://fediffusion.art/media/38c4817d7298d95f413cf9f4074558e5c70007543a794dd35fd00458d3c7f974.png
https://fediffusion.art/media/e95c231973ad0ca32ab062db9e25165b59104d03ae6cc11eca7670badae290eb.png

2023-10-25 12:02:25 UTC

- reply

@npub1c4s…aqv4 i'd say comfyui is the best, but i've also switched to using Fooocus for most of the stuff I do, it's much simpler and the author has really in-depth knowledge of stable diffusion, so it has a few features that nobody else has.