Magnets must be respected, but need not be feared. ๐ profile subsumed by Shihan Qu, founder of แ fed bannedแ Zen Magnets. Micromags still @ neoballs.com
Joined July 2009
- Tweets 5,034
- Following 618
- Followers 3,087
- Likes 30,503
469 Photos and videos
Pinned Tweet
Claude Fable helped me with some product design today. Made me a cascading ball design helper that outputs an STL for 3d printing. Used 400k tokens to reach what's shown below.
GPT 5.5 xhigh couldn't come close. Like it never really grasped the concept of the cascading functionality. All sorts of intersecting messes. After 180k tokens across a dozen prompts, I can tell it's not going to get there.
I'm both impressed and frustrated with how far ahead Anthropic is right now. Hoping both OAI and Open Source close the gap soon.
1
2
13
3,663
Alibaba Qwen3.7 slowly fading into irrelevance at the frontier due to proprietary stance.
In it's place we have Minimax M3 and... *checks notes* Rio 3.5 397b, made by the municipal IT company of Rio de Janeiro's city government.
huggingface.co/prefeitura-riโฆ
99
254
2,478
1,155,230
Claude Fable has caught up to humans in simple bench.
Not a single open weights model in the top 20. The only Non-American is Qwen.
2
2
23
3,132
Maybe the reason Ant is so focused on "safety" is because they're growing a Gollum
3
316
Voxel Pagoda Comparison: Gemini 3.1 Pro vs Gemini 3 Flash vs Opus 4.6 vs Kimi k2.5.
See ๐งต for GLM-5 vs GPT-5.3-Codex XHigh vs GLM 5 vs Sonnet 4.6 vs Minimax M2.5 vs Qwen3.5-397B vs Grok 4.2
Gemini 3.1 Pro is very strong, and has a high voxel count. Opus 4.6 seems to have the best taste. At least in these 1 shot comps. K2.5 outputs on par with what previous version of Gemini Pro could. Flash is lazy.
Prompt: "Design and create a very creative, elaborate, and detailed voxel art scene of a pagoda in a beautiful garden with trees, including some cherry blossoms. Make the scene impressive and varied and use colorful voxels. Use whatever libraries to get this done but make sure I can paste it all into a single HTML file and open it in Chrome."
1
8
3,565
Two additions: Minimax M2.7 and Composer 2
1
584
Claude Fable 5 Extra, vs GPT 5.5 xhigh.
Fable on the left is not the highest resolution, but is the best looking voxel pagoda result ever. Diverse vegetation. It's got birds! Voxel count = 43k
GPT 5.5 looking a bit blown out in comparison but quality is still good. Landscape is flat. Voxel count = 16k.
151
Claude Fable 5 Extra, vs GPT 5.5 xhigh.
Fable on the left is not the highest resolution, but is the best looking voxel pagoda result ever. Diverse vegetation. It's got birds! Voxel count = 43k
GPT 5.5 looking a bit blown out in comparison but quality is still good. Landscape is flat. Voxel count = 16k.
Voxel Pagoda Comparison: Gemini 3.1 Pro vs Gemini 3 Flash vs Opus 4.6 vs Kimi k2.5.
See ๐งต for GLM-5 vs GPT-5.3-Codex XHigh vs GLM 5 vs Sonnet 4.6 vs Minimax M2.5 vs Qwen3.5-397B vs Grok 4.2
Gemini 3.1 Pro is very strong, and has a high voxel count. Opus 4.6 seems to have the best taste. At least in these 1 shot comps. K2.5 outputs on par with what previous version of Gemini Pro could. Flash is lazy.
Prompt: "Design and create a very creative, elaborate, and detailed voxel art scene of a pagoda in a beautiful garden with trees, including some cherry blossoms. Make the scene impressive and varied and use colorful voxels. Use whatever libraries to get this done but make sure I can paste it all into a single HTML file and open it in Chrome."
1
2
371
๐ญ๐ฒ๐ป ๐ ๐ฎ๐ด๐ป๐ฒ๐๐ retweeted
Jun 10
When will we see the next open foundation AI model from @Alibaba_Qwen? ๐ค
Really hoping for some larger 3.7 models, all the way up to 397B.
That would make me and a bunch of my friends in the local AI space quite happy and grateful.
Until then, Q-wen?
7
3
27
5,213
Here's the most surprising thing about the Claude Mythos benchmark table... 0% score on a Legal Agent bench by Gemini 3.1???
What's all that broad training data and long context ability doing for you Gemini?
1
1
199
I thought this was a joke. Turned out to be prophecy, but yoinked by Ant
Apr 12
Imagine the alternate reality where we named GPT-5.4-Pro something like Fable.
2
223
Interesting how "this man" has a different voice and tattoos in every clip. But genuinely hard to discern AI video these days
This Man decided the cops arenโt doing enough so now he's launching desserts at speeding cars ๐ญ
5
1,484
I feel similar. Long term Tesla bull. Still holding 4 digit shares, yet undeniable that the Tesla narritive has changed so many damn times.
Used to think they had a 5 year lead on robotaxi at least, but as of right now BYD has an L2 system where the mfg takes liability for all at fault accidents, and xiaomi's fsd competitor drives more like a human in the chaotic streets of urban China than timid FSD does.
Still see Tesla as the dominant player in robotaxi, but if CCP wanted to push their self driving models the same way they've been pushing open weights LLMs, Tesla's moat is thin. Optimus progress also not looking very high tech in the face of Chinese humanoids rn.
Jun 8
As a TSLA investor, I havenโt felt more clueless about whatโs going on/whatโs coming next than I do right now.
FSD is amazing, but the goalpost with Robotaxi keeps getting moved and I have no clue why. No clue when โreasoningโ is going to make parking so good that itโll see spots I donโt see. No clue when half the US population will have Robotaxi access. No clue what the final piece of the puzzle was etc. There were 7 new Robotaxi cities on a list for 1H26 in the Q4 deck, which ended up disappearing from the Q1 deck, and with 3 weeks left in 1H26, only 2 of 7 have hit. No clue why.
Def would be awesome to have something like this. Or at the very least, some new revised roadmap from Tesla AI similar to what we got a year or two ago.
I still have conviction in the long-term story here, but itโs weird how sometimes they choose to communicate with investors, and other times theyโre vague/leave us totally clueless. I canโt tell if theyโre about to go gangbusters or something isnโt going as planned/weโre cooked for a few more years.
1
492
Has to be the cheapest way to run Qwen 35b at a usable speed. Setup looks all cracked out but an old power supply plus $350 for 2x 16gb ~PS5 APUs.
Let some crypto miner's loss be your gain, because these things run low concurrency inference at the same speed as an M3 Pro 36gb macbook that will cost 5x as much. There will be a run on them.
Great find @loktar00
Jun 8
This is actually CRAZY!!! Using llama.cpp RPC
I have 2 BC-250's setup so far, they're able to run Qwen 27b at Q4, and 35b at Q4 as well.
This is without extra CUs unlocked:
Qwen 27b with MTP - 14.5 tk/s
Qwen 35b with MTP - 47 tk/s
For $300 I'm getting these speeds! This is wild!
4
27
4,189
๐ญ๐ฒ๐ป ๐ ๐ฎ๐ด๐ป๐ฒ๐๐ retweeted
Jun 5
Reminder for all young parents:
You only get:
- 1 Summer with your baby
- 3 with your toddler
- 9 with your child
- 5 with your teenager
This time is precious. Donโt rush it.
158
2,501
28,702
926,527
Christ on a bike, is @nvidia going to start gracing some love towards SM120 slackwells and SM121 GB10s?
8
512
I wonder if this counts only Codex App, or also Codex CLI....
Jun 4
Your Codex activity now has a home, and an easier way to share it.
Codex profiles show your activity graph, streaks, lifetime tokens, peak daily tokens, and top features like plugins and /fast mode.
Private by default. Share a card when you want to.
2
493
AA Intelligence scores out for Minimax M3.
-- 54.7 --
Benchmaxxed? I mean, maybe studied for the tests too hard, but kid is clearly smart.
Not like Alibaba is going to give us Qwen3.7 397b anyways. Jury's still out, but I suspect this will be best available on ~384gb vram.
7
1
102
8,240
Oof, 352gb model weights. Guessing just barely out of range for 4x Rtx 6000 pro slackwells
As always, Nemotron 3 Ultra is fully open.
This includes model weights, synthetic data, and post-training recipes. Available now on @huggingface โ nvda.ws/4v1iBhi
1
209
OpenAI taking over @papercliping confirmed
3
218