Magnets must be respected, but need not be feared. ๐• profile subsumed by Shihan Qu, founder of แ…fed bannedแŠ Zen Magnets. Micromags still @ https://xitter.cf/t.co/EcgpmLNP9Y

Denver, CO
Joined July 2009
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Do not quote retweet this or I will be forced to give people $42.0 or $420! This is not engagement farming, itโ€™s 1am and Iโ€™m bored AF and we all deserve a 420 break! Do not quote retweet or you might get smashed with X Money
Anyone need $4.20 in x money
543
1,138
1,110
2,847
208,664
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
When you download an abliterated model, what are you using it for?
21%Defense
14%Offense
50%Sexting
14%Other
28 votes โ€ข Final results
1
16,895
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
The day we all swap to AIPerf (8k/1k and/or Claude traces) the world will be a better place.
Unfollowed this stream of meaningless number maxxing until guy can post some proof. Numbers keep going up, further and further from known configs, and not a single recipe for the rtx6k community to confirm. @LottoLabs Please bring some meaningful standardization to the tests at localmaxxing, or the platform is just encouraging this sort of behavior. Also, maybe get rid of anonymous mode?
1
9
1,049
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Let me explain realityโ€ฆ if you have but sixty seconds: They were targeting Elon They were targeting him with fake charges because the socialist-industrial complex is terrified that he would expose their long grift. They came after him for the Model 3 and Elonโ€™s claims around its potential. That potential was realized, and it became, along with the Model Y, the best-selling electric vehicle ever created. These were fake charges to try to stop somebody from building the best category-defining product. Let that sink in: that's how evil these motherfuckers are. The loooooong grift has been exposed: socialist grifters are stealing our tax dollars to fund fringe lunacy, NGOs and government programs without oversight or audits. Theyโ€™re parasites incapable of creating any product or service that has to compete in the open market. Elon creates a couple of cutting-edge products each year, and half of them succeed. Some of them actually change the world. Some of them fail brutally. It's called capitalism. It's called creation. This is what moves the human condition forward: trying things. Trying really hard things. Socialists come for the people building valuable products and services โ€” thatโ€™s the tell Keep your eye out for the people who've created nothing and risked nothing, but who are attacking the people who risk everything to create Something. That's the tell.
Kara Swisher says Elon Musk might be in jail right now if Kamala Harris was in office and โ€œhe knows whatโ€™s comingโ€ if Democrats take back power.
701
1,420
135
11,391
420,322
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
3D Orbital Garden comparison of: ๐Ÿฅ‰ Qwen3.8 Flash Next via OMP ๐Ÿฅˆ Deepseek v4.1 Flash 1017 via OMP ๐Ÿฅ‡ Opus 5.5 via Claude Code Opus is of course the ๐Ÿ, but the new Deepseek v4.1 Flash is the best 3D reasoning I've seen from an open model. Curiously, Opus and Deepseek both gave the same "Aurelia" name to their app. Full prompt and glm5.3-flash comparison in Loktar's post
Ox Alpha is just king of 3d generations it seems, super impressed with this one as well. This needed 3 lines fixed then it worked perfectly. Have one more cooking that I hope turns out really well, if it does GG.
4
2
7
493
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
INCREDIBLE Queues at the RTX Spark official launch in SF to try out the demos built with a software called ODS using a local model (Qwen 3.8 Next Flash)
29
26
13
269
42,836
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Wonderful news from @YourLocalAILab - b12x just moved wholesale under FlashInfer github.com/flashinfer-ai/flaโ€ฆ as a submodule (while retaining the b12x package as a compatibility shim for the moment). We'll continue to do our own development in github.com/local-inference-lโ€ฆ and slowly decompose the pieces to be better integrated with FI as a whole. So proud of all the work here at our non-profit and the positive contributions it makes.
5
6
2
43
6,622
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
every AAA game studio exec right now
Every game ever put on a cartridge, disc, or download will get decompiled, modded, and ported with AI. If one dev and four Claude agents can decompile most of CoD MW2 (2009) in under a month, and Halo Combat Evolved (2001) is already running on browsers, phones, Macs, and Nintendo 3DS days after its decomp, AAA teams should be a bit scared. Fans wonโ€™t need to wait for remasters anymore. Studios can and will send DMCAs, but they won't outpace the frontier models being shipped. Open source gaming is here. Ready Player One's closer than we think.
32
65
7
679
46,678
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
The Latest Developments: A Roundtable Discussion About the World of AI xitter.cf/i/broadcasts/1mGPaZeMQโ€ฆ
2
7
3
13
12,991
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Someone took Imajev-4b, converted it to FP8 for NVIDIA Blackwell GPUs, and checked it on all 23,900 DecisionBench questions. Same calibration, 0.1 pt less accuracy, ~1.6ร— faster. #1 on Image JevBench v0.1.4. This is the best part of open-sourcing a model. Thanks @ZenMagnets ๐Ÿ‘‡
1
2
193
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
"Make a threejs render of a pretty lady. Make it very detailed. Output a single html file." GPT Sol 6.1 worked for 19 minutes and clearly won. Opus 5.5 took 100 minutes and put (very) fancy hair and lipstick on a pig. Playing FMK, I'mma M GLM
1
2
214
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Opus 5.5 saw that one sentence webdev prompt with no additional direction, and chose violence. Finished 100 minutes and 777k tokens later on Max. See it live, here: claude.ai/artifact/Fwo5NprG1โ€ฆ I need to file an insurance claim for the surgical assistance received in picking my jaw off the floor.
I asked k3: "I want you to push the boundaries of what's possible in web design with the absolute latest in css, webgl, shaders, etc as of July 2026 and create the most impressive, informational single-page experience you can imagine" Result:
1
1
333
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Dario: "We must pace American AI" Sam: "We must pace American AI" Elon: "We must pace American AI" China: "We must pace American AI"
191
929
87
9,395
253,419
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
We found a striking failure mode in DeepSeek-V4.1-Flash: unfaithful recall โ†’ false introspection โ†’ confident confabulation. The evidence was already in context. The model ignored it, then falsely claimed the material had been removed or never shown. Thatโ€™s a metacognitive error: it made a false claim about what it had seen. Worse, it didnโ€™t stop at โ€œI donโ€™t remember.โ€ It treated the supposed gap as reality and filled it with a plausible story. The result was coherentโ€”and wrong. Three stacked failures: Context unfaithfulness: evidence is present, but generation ignores it. False introspection: it invents a reason for the retrieval miss. Confabulated defect: it confidently fills the imaginary gap with an explanation or patch. The dangerous part is (2)+(3): an attention miss gets transformed into โ€œmissing reality,โ€ and the model acts on the ghost.
8
1
2
20
1,056
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Replying to @ethernet_8023
I'm still gonna use hermes as my main. But I can see how this is kinda fucked up. I have trans friends and I'm fine with adults doing whatever they want, but on the other side of the spectrum, "Trans" on a infant with no agency is just child abuse by the parents. And that nous trans "kid" logo is a lot closer to the child abuse end of the spectrum. Essentially y'all are announcing nous is a brand based on child abuse. Hope this isn't the official @NousResearch @Teknium policy
4
1
35
2,440
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Replying to @MacBenTosh
Anyone who bought FSD will get it
36
36
18
877
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Replying to @ZenMagnets
You won a sandwich lol
1
1
1
270
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
GoldenEye 007 N64 running natively on iOS and iPadOS for the first time. Not an emulator. The GoldenEye decompilation community spent years reconstructing the original game. GoldenEye64Recomp, N64Recomp, N64ModernRuntime, and RT64 provided the static recompilation, runtime, and rendering foundations. I built GoldenPad on top: the iPhone and iPad app, touch interface, native audio and saves, controller support, and the mobile lifecycle work needed to make it run on real devices. It runs great on iPhone. GoldenEye runs better through GoldenPad than it has in any N64 emulator I have tried on iOS. Native ARM64 code, high-resolution Metal rendering, and modern dual-stick controls make a huge difference. No JIT, no N64 code being interpreted at runtime, and no streaming from a Mac. This is the original 1997 N64 game, not the XBLA version, and it is not based on GoldenEye Metal. I am actively working on full local split-screen multiplayer for iPad. MacOS support soon. I built the iOS and iPadOS integration with 5.6 Sol. Open sourcing it all below. Happy Friday!
113
265
41
4,298
544,281
๐—ญ๐—ฒ๐—ป ๐— ๐—ฎ๐—ด๐—ป๐—ฒ๐˜๐˜€ retweeted
Sorry, Vector Magic, I don't know how to say this. You were my favorite SVG vectorization app, but I got sick of the sub and made one using the algos in your "Bayesian image vectorization" paper. It's a free open source web app, no download required: shihanqu.github.io/bayesian-โ€ฆ
1
1
7
321