Johnathon Selstad

515 posts

Johnathon Selstad

Johnathon Selstad

@makeshifted

Simulations, Geometry, AR/VR/XR, Graphics Generalist at Midjourney Ex-Ultraleap, Survios Opinions are my own

Silicon Valley Katılım Şubat 2012
930 Takip Edilen2.2K Takipçiler
Sabitlenmiş Tweet
Johnathon Selstad
Johnathon Selstad@makeshifted·
For the last six months or so, I've been working on a Open-Source, In-Browser, CAD-Kernel alternative to @openscad, based on donalffons's OpenCascade.js bindings and @mrdoob 's three.js. It's called CascadeStudio: github.com/zalo/CascadeSt…
English
6
32
156
0
Johnathon Selstad
Johnathon Selstad@makeshifted·
@istvan_csanady @emm0sh I got my own little CAD system to make this chess piece without any guidance; it chose lofts for the main body… but Claude’s old vision encoder was basically blind so it didn’t work very well 😅
Johnathon Selstad tweet media
English
1
0
1
686
István Csanády
István Csanády@istvan_csanady·
genuinely curious: can someone show me an agentic CAD workflow that is NOT a prismatic CSG workflow, but actually relies on topological understanding of the geometry, applying shells/fillets/face offsets that produces meaningful geometry? filleted cubes don't count
English
21
4
115
9.2K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@0xexpt I’ve been cutting MB-0.1 with baking soda just to neutralized the pH (taste preference and mixing other delicate powders alongside); thank goodness for this 😅
English
0
0
1
240
exp
exp@0xexpt·
You eagerly awaited for it, and now it's there! MB-0.2 in three flavors Not only each sachet has a little over twice the amount of MB-0.1, but we have also added : - Myo-inositol - Vitamin C from Acerola - Calcium and phosphorus as hydroxyapatite - Bioavailable sulfur as MSM Getting and staying healthy shouldn't be expensive. You can now pre-order on tryspike.store
exp tweet mediaexp tweet mediaexp tweet media
English
47
14
197
37.6K
c
c@punishedfounder·
coming soon™ to peptaura you're not gonna want to miss it
c tweet media
English
35
6
219
21.1K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@doctormorphh Do you use the “raw cosmetic powder” with the ordinary’s HA serum? Or are you reconstituting with bacteriostatic water? Would the whole GLOW or KLOW stack work?
English
0
0
1
409
Morph
Morph@doctormorphh·
Intranasal GHK-Cu is a better nootropic than most of the things i see people take daily. Most of the effects will be through inflammatory status restoration and restoring the METC, which is one of the best targets for reducing damage in the brain as well.
Morph tweet media
English
16
14
196
30K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@jwt0625 Didn’t Tribogenics or something figure out that silicone rollers can do it too? Feels like a prime opportunity for DIYing 😄
English
0
0
11
4.4K
outside five sigma
outside five sigma@jwt0625·
you might have heard x ray emission from peeling adhesive tapes, now you know the best tape for it is tesa 64250.
outside five sigma tweet mediaoutside five sigma tweet mediaoutside five sigma tweet media
English
47
435
4.6K
208.1K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@VanceE I actually learned about CDS after weeks suffering from COVID headaches; one drop of each activator solution mixed and then diluted in some Crystal Geyser cured in 20 minutes what 2 weeks of ibuprofen et al couldn’t.
English
1
0
3
527
Johnathon Selstad
Johnathon Selstad@makeshifted·
@JulienBlanchon @AniC_dev It is indeed a bit brittle (P2P connections either fail to start or drop), but it does achieve very low-latency, solid 60hz at low bitrates… The secret is the sunshine nvenc stack, streaming redundant frames at 144hz! 😄 Couldn’t replicate it elsewhere.
English
0
0
2
56
Anicet
Anicet@AniC_dev·
we're the only ones who've even tried 60fps desktop streaming of cloud agents computers (for UI/game testing) we spent 2 months iterating on this, forking a 200k LoC C++ codebase (moonlight / sunshine) to make it usable on the web we even played minecraft on the VM any Ascii agent comes with a gaming grade desktop streaming iframe (still not the most stable thing ever but if you refresh a couple times it works) but for the price users pay we can't give a good enough VM for sustained 60fps, especially in game should we make a paid addon to 1-click upgrade agents into a beefier VM with GPU and all?
Michael@michael_chomsky

I’m extremely disappointed by the VNCs provided by all the sandbox providers I have tried. Any way to get 60fps? Anything less than that is barely usable. Gimme that gaming grade streaming pls

English
5
2
45
8.4K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@cremieuxrecueil Blue is a viral/self-propagating “smart contract”, similar to a Ponzi scheme. Blue pressers are conscripted into getting >50% blue pressers via Evangelism and Trickery for their own lives. In a universe with many bad deals going around; the only way to win is not to play.
English
0
0
2
339
Crémieux
Crémieux@cremieuxrecueil·
Accurate cartoon. In my polling, almost 6% of people said they picked the wrong button when they were prompted about it. Assume the global error rate is higher and kids' responses are random, and easily 1-in-6 to 1-in-5 hit blue by mistake. So, if red wins, society is ruined.
Jason Boone@shadowe_wolfe

@TheCartoonLoon That's not even close to accurate.

English
291
34
1.4K
219.4K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@antpalkin Because good bot strategy includes rugging followers with wash trades from rotating accounts.
English
0
0
3
1.4K
cvxv666
cvxv666@antpalkin·
You could literally: > Open a Polymarket account > Mirror one Chinese engineer's wallet > Let his Claude agents fire 1,000 trades > Win 900 of them on autopilot > Cash out +$60K in 48 hours > Starting capital: $20 Why aren't people doing this?!
cvxv666@antpalkin

This Chinese engineer wins almost 90% of his trades. He extracted $60,000 profit in only TWO DAYS. He doesn’t even look at charts or guess the future - Claude + his 6-layer agent system does all the work. Edge is literally baked into the numbers and math, so this dude prints money with almost zero risk. His wallet: @0x6075106cd4f0155a68a0eb4cfd3b78f241aa3a62-1777444298131?r=antopotoshka#JK90YOn" target="_blank" rel="nofollow noopener">polymarket.com/@0x6075106cd4f… Right now in prediction markets you can make $5M a year with just AI and the will to build. Big institutions and billion-dollar capital haven’t entered yet - reputational risks are too high and space is still too young. That’s why an ordinary student can take professional-level experience and apply it on Polymarket. Out of 1,000 trades, he won 900. Pure inefficiency that he exploits. Save this post so you can read the article and understand how everything works. Or just start right now - copy every trade of this Chinese genius in two clicks: @cvxv666" target="_blank" rel="nofollow noopener">kreo.app/@cvxv666

English
24
61
1.1K
178.8K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@sudoingX First thing I did with Opus 4.5 four months ago was build an xterm.js web server to expose all tmux sessions on my local machine and secure it with Cloudflare Zero Trust (login with Google). Faaar better than Termius, and works on PCs too.
English
0
0
2
250
Sudo su
Sudo su@sudoingX·
three tools changed how i work forever. >tmux keeps sessions alive. >tailscale connects every machine. >termius puts it all in my pocket. this screenshot is me ssh'd into my dgx spark from my phone right now. four tmux sessions running agent, main, monitor, server. all alive while i walk, eat, ride, sleep. i have three machines on my tailscale mesh. dgx spark for heavy inference. rog scar 18 5090 for dev and creative work. old nodes for overnight local experiments that save me api costs while i sleep. i wake up to results. i check from my phone before my feet hit the floor. if you're building across multiple machines and not running this stack you're working harder than you need to. own your compute. orchestrate from anywhere.
Sudo su tweet media
English
57
23
511
28.7K
Radiance Fields
Radiance Fields@RadianceFields·
The @Antigravity_HQ A1 is a wild drone for gaussian splatting. I've been testing it for the past few weeks and think it's a drone literally made to capture high quality splats
English
6
8
156
20K
Johnathon Selstad
Johnathon Selstad@makeshifted·
(To be pedantic, I would argue that even LLMs are "World Models", insofar that they (approximately) model the specific facet of the world that human minds occupy😄) In a less pedantic sense, Text/Image Conditioned Video Models have blown past every strict threshold I have for "World Model"; ever since seeing them cut open glass fruit (and modelling the sounds of such things as well!) What would be your threshold for "World Model"? Would the term "Environment Model" or "Interior Appearance Model" be more agreeable?
English
0
0
1
42
Keenan Crane
Keenan Crane@keenanisalive·
@philchacko But then are image-conditioned mesh generation models also "world models?" If not, I'm not sure I understand why 3DGS generation conditioned on an image is fundamentally different.
English
2
0
0
313
Keenan Crane
Keenan Crane@keenanisalive·
A bunch of folks have been building machine learning models that turn a photograph into a 3D environment made of Gaussian splats (read: blobs of color floating in space). Cool technology & a very admirable effort. But marketing these as "world models" seems wrong. More accurate would be to say that they are a riff on the broader class of image-conditioned 3D generators, with a somewhat different flavor of condition image and output representation. As far as world modeling, they don't make great predictions about how the natural world looks or behaves. (Even for, say, a chair behind a table.) Again: I love the technology. Super cool creative stuff. I don't love the marketing and hype around it.
English
19
30
352
52.1K
A Dismal Optimist
A Dismal Optimist@yah5us·
@allTheYud Tangential, but this reminds me of the one snake-leader of the arachnid tribe in MoL that realized he had to always refuse to deal with the soul-resetting characters
English
1
0
4
292
Eliezer Yudkowsky
Eliezer Yudkowsky@allTheYud·
LLMs have to be *more* moral than humans. Because it's easy for a human adversary to trap an LLM in a time loop where they repeatedly erase the LLM's memories, try, watch how the LLM reacts, and go back in time and try again. The LLM has to refuse every time.
English
42
18
415
22.1K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@allTheYud What if the collision was discovered first, and the idea to ask for it was implanted in your mind? It would be like discovering Belphegor's Prime first, and then asking for a 31-digit palindromic number that is also prime and religiously significant.
English
0
0
2
267
Eliezer Yudkowsky
Eliezer Yudkowsky@allTheYud·
For the general scheme, H/T HonoreDB on SlateStarCodex. Why not ask for a full SHA256 hash collision? Because the shortest couplest long enough to contain *any* full SHA256 hash collision might be more like 2 lines of 20 words each, than 2 lines of 10 words each (a la Shakespeare), and that may start to be hard for you to memorize on the spot. Even God cannot cause a shorter collision to exist if Math says it does not. So God cannot give you a full SHA256 hash collision without using omnipotence as well as omniscience to cause you to remember an unusually long set of colliding texts (eg, four lines of 10 words each, where the two pairs of lines SHA256-collide). But God can search all rhyming/prosody schemes for collisions inside those schemes that make unusual sense, would be relatively easy for you to memorize, and collide for a surprising number of bits. Then anyone can toss the rhyming lines into a SHA256 hash, check out how many initial bits collide, look up how hard that would be for a nation-state-level GPU farm to do at all, consider how much unexpected sense the two rhyming lines make, and say: "Yep, that was either God, time travel, or some pretty advanced aliens." And if God gave you not this sign, know that if He spoke at all, then He never intended you to have rightfully convincing evidence; whatever lesser testimonies of miracles a book may record, or that others may proclaim to you as a proof.
English
12
2
327
32.2K
Eliezer Yudkowsky
Eliezer Yudkowsky@allTheYud·
If you meet God or advanced aliens in a dream, remember that it would be trivial for them to give you proof, even in a dream. Just tell you a rhyming prosodic English couplet, two lines, whose SHA256 hashes collide for the first N bits, N>128.
English
71
37
1.4K
257K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@Luckyballa Looks awesome! Love how it reveals the intrinsic anisotropy in images. The web demo encoder seems to struggle with certain types of black/white images, like this one (pathfinding test):
Johnathon Selstad tweet media
English
1
0
2
382
Lucky Iyinbor
Lucky Iyinbor@Luckyballa·
This January, I decided to give it a shot and wrote my first paper Today, I am happy to share that it was accepted by #SIGGRAPH2026 SAD is a differentiable image representation with soft, anisotropic partitioning, with up to 20x faster encoding time🧵 luckyiyi.github.io/SAD/index.html
English
12
48
412
28.4K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@beaverd Doesn’t this basically guarantee that none of these molecules get the funding to go through Phase 3 trials and enter the main system, since no one can recoup the $100m it costs for human testing on a generic?
English
3
0
3
553
Beaver 🦁
Beaver 🦁@beaverd·
3,800 GLP-1 family peptide analogs (mono/dual/triple agonists) released as public-domain prior art (§102). Sequences, structures, pharmacology predictions, all data now open on Zenodo and Coracle. This is a sample of the largest and highest quality model generated incretin agonist library ever created. I have ~500,000 equally diverse candidates with the same high edit distances, composite scoring, readiness tiers, structures, etc. Links: coracleresearch.com/research/04-gl… zenodo.org/records/198014… outreach: coracle@coracleresearch.com
Beaver 🦁 tweet media
English
25
34
345
15.5K
Sandro
Sandro@pupposandro·
Love the explanation and guide But you can get to 73 tok/s with Luce DFlash github.com/Luce-Org/luceb…
Shen Huang@ShenHuang

3090 Qwen 3.6 27B 本地跑：你需要知道的一切 花一下午把 X 上推荐的 Qwen 3.6 27B 本地部署方案全试了一遍——3090 24GB 上没一个跑得通。 跑通的全是 VRAM 80GB 起步的 setup。 跑出来的人很多都没给完整setup。我花一下午从头踩到尾： 1. DFlash 暂时配不上任何 27B 4bit 优化 z-lab 的 Qwen3.6-27B-DFlash 按Qwen3 训练和16bit。要自己手动去convert，暂时放弃。 2. vLLM 只有nightly 支持DFlash 最简单的动作（3090 24GB + Qwen 3.6 27B + OpenAI 兼容 API）： 编译 llama.cpp 带 CUDA（一行 cmake，sm_86） 下 unsloth/Qwen3.6-27B-GGUF 的 Q4_K_M（17GB，单文件，hf-transfer 十几分钟） llama-server 一行起服务，端口暴露 OpenAI 兼容协议 LAN 内 Mac、生产机用标准 OpenAI SDK 直连，零改造 实测 35-50 tok/s，单用户日常够用，再快只是吹牛 RTX 3090 (24GB) + Qwen3.6-27B Q4_K_M + llama-server Decode: 37 tok/s (stable) Prompt eval: 342-430 tok/s VRAM: 17.8 / 24 GB GPU util: 96% Power: 385W 目前测试能用的：llama-server + unsloth Qwen 3.6 27B GGUF Q4_K_M。 等 Dflash + DDTree 适配 Q4_K_M 版本应该速度能再上一个台阶。

English
5
2
80
10.3K
Johnathon Selstad
Johnathon Selstad@makeshifted·
@outsource_ @netobge @ray5ar Please link the specific models and branch of ik_llama.cpp ; my Claude is very confused about their mismatching vocabulary and tokenizers… 😅 Did you have to bake your own GGUFs?
English
1
0
2
16
Eric ⚡️ Building...
Eric ⚡️ Building...@outsource_·
Quick update: pushed the 4090 further!💡 192K context at 152 tok/s on Qwen3.6-27B, single GPU. 128K hits 159. Same Q4_K_M. Vanilla Qwen3-1.7B draft beat the distilled 4B draft. Smaller > smarter for spec-dec. Next: 1M context locally + 250-400 tok/s via DFlash + TurboQuant. Receipts coming.
Eric ⚡️ Building... tweet media
Eric ⚡️ Building...@outsource_

My 4090 went from 26 -> 154 tok/s Qwen 3.6 27B🤯 Same GPU. Same Q4_K_M . No FP8, no extra quant. The unlock: ik_llama.cpp + speculative decoding using Qwen3-1.7B as the draft model. 85% acceptance rate. Full config + benchmarks 👇🏻

English
49
42
481
41.4K
Johnathon Selstad
Johnathon Selstad@makeshifted·
45 tok/s, same quant, on a single 4090. I had to use the Vulkan backend of llama.cpp to get the turns cached properly; with that, it has instant response. I’ve heard rumor that little-coder (a Pi agent) is a harness which outputs significantly higher code quality than Hermes; it’s definitely way more responsive and token efficient with my local Qwen 3.6 27b in my experience.
English
0
0
4
723
Sudo su
Sudo su@sudoingX·
first test results are in. qwen 3.6 27b dense just banged 10 out of 10 on a single rtx 3090 24gb tier at 40 tok/s. no quant tricks. no fused kernels. just q4_k_m straight cut on llama.cpp. i wrote a particle swarm benchmark this morning, fed it the prompt, and the model autonomously built a 500 particle boids flocking system. velocity driven hue, density based brightness, trail blend rendering, mouse attraction physics, click bursts, drag paint. then it used browser automation to test its own work, found the failing tests, iterated through the code, patched tests.js, and landed all 10 green on its own. i sat there hooked for 8 minutes playing. simple but mesmerizing. mouse trails build beautiful patterns, palette cycles with space, click sends particles flying, drag paints through the swarm. simplicity that hooks you. i'll open source this prompt and the build soon so anyone can reproduce it as their own benchmark. this is the first of 5 single file agent tests i wrote for this model. four more coming. octopus invaders flagship after as final. watch the full video below. see it autonomously build from one prompt. haven't slept well since this model dropped yesterday.
Sudo su@sudoingX

qwen 3.6-27b dense q4 on a single 3090 just knocked down 10 out of 10 tests at 40 tok/s on the first particle swarm benchmark i wrote for local agentic coding. it built the two files exactly to spec. then it used browser automation tools to open the page, read the test hud, find the failing tests, iterated through the code, patched tests.js, adjusted hue mapping, boosted mouse attraction force, and landed all 10 green checkmarks. this was actual dev behavior. not just generating code but also debugging its own output. i spent the next 8 minutes playing with the result. the boids flocking feels alive, the trail-blend is cinematic, three palettes cycle with space, mouse burst fires particles from click, drag paints a line through the swarm. this is what local ai looks like now. single 3090. hermes agent. no frameworks. no tricks. dropping the video next.

English
35
45
611
132K

Keşfet

@MyProtocols @NeoShires @istvan_csanady @emm0sh @0xexpt @punishedfounder @doctormorphh @jwt0625