Trent Zock-Robbins (@tazr_dev) header
Trent Zock-Robbins (@tazr_dev) profile picture

Trent Zock-Robbins

@tazr_dev

• veteran swe, ex-chemist, touching grace • https://xitter.gq/t.co/YD7FQNLSFT → byok llm chat @ iOS • 100% bitter-lesson-pilled • localmaxxing Qwen 27B on 4 × 3090

 San Francisco, CA
github.com/memgrafter
Joined January 2026
  • Tweets928
  • Following316
  • Followers147
125Media
Photo by Trent Zock-Robbins (@tazr_dev)Photo by Trent Zock-Robbins (@tazr_dev)Photo by Trent Zock-Robbins (@tazr_dev)Photo by Trent Zock-Robbins (@tazr_dev)Photo by Trent Zock-Robbins (@tazr_dev)Photo by Trent Zock-Robbins (@tazr_dev)
Yesterday, GLiNER2.5 played doom. Rapid play looks coherent, but Jev has a secret sauce: decisions. So, I kept the embedder and trained a small decision head on the output. Now it's a homebrew decider model on embeddings. What homebrew decider model will you make this weekend?
Trent Zock-Robbins (@tazr_dev) profile pictureTrent Zock-Robbins
@tazr_devSep 17
GLiNER2.5 Doom, code tomorrow. • mac mini m4 16gb • 10.9 decisions/s • 34 tics/s
4
4
3
89
9,173
Generative 3D scenes!! Refresh every 10 minutes. worlds.ftl.cc/scenes/infinit… What do you want to see?? I will add it! This is the beginning...
Image from Trent Zock-Robbins (@tazr_dev)
Image from Trent Zock-Robbins (@tazr_dev)
50
Conjure3D image->obj pipeline is incredible at generating 3D objects. The scene pipeline is still in progress. Check out the first threejs demo at worlds.ftl.cc/scenes/seven-k…
Image from Trent Zock-Robbins (@tazr_dev)
Three.js (@threejs) profile pictureThree.js
@threejsOct 30
Three.js r187 released 🍎 threejs.org/changelog/?r187
40
This is rad. Love when things all hit at the same time. Idol 3dgs generation will be nice in Conjure3D. If you need 3dgs: Idol (now) If you need obj: TRELLIS.2 (2025)
Stefan 3D AI (@Stefan_3D_AI) profile pictureStefan 3D AI
@Stefan_3D_AINov 3
The Best New Local Image-To-3D Model Is Now Available on Apple Silicon Idol, the open-source image-to-splat model from Speridlabs, now has a native MLX port. One picture in, a full 3D Gaussian splat out, running locally on Apple Silicon. Highlights: • About 8.6 GB peak memory, fits 16 GB Macs • No CUDA, uses the original Idol weights • Standard 3DGS .ply output • 32K to 1M Gaussians per object • Local server endpoint included GitHub: github.com/Blaizzy/mlx-vlm
1
3
56
obj->3dgs seems easier than the reverse, but Idol might be able to do things that TRELLIS.2 cannot. It will be worth using both where possible.
1
30
Trent Zock-Robbins retweeted
Simon Willison (@simonw) profile picture
Simon Willison
@simonw20h
There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing simonwillison.net/2026/Nov/3…

Qwen 3.8 27B on my laptop can interpret model weights, too

Yesterday’s hot paper was Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. The paper found …

simonwillison.net
329
362
85
4,627
264,334
Trent Zock-Robbins retweeted
Yavuz Bakman (@yavuzbakman) profile picture
Yavuz Bakman
@yavuzbakmanNov 3
You can 2x Qwen-0.8b's performance on Math-500 by adding Gaussian noise to the last 4 layers. RSI is here 🔥🔥🔥 and it's Gaussian!
Image from Yavuz Bakman (@yavuzbakman)
Rikiya Takehi (@rikiyatakehi) profile pictureRikiya Takehi@rikiyatakehiNov 2
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Image from Rikiya Takehi (@rikiyatakehi)
32
43
17
1,026
54,405
Absolutely unhinged! Opus 5.5 used Conjure3D to generate 44 fictional worlds over the last 24 hours, generating over 8758 objects at 512 mesh resolution.
Image from Trent Zock-Robbins (@tazr_dev)
Image from Trent Zock-Robbins (@tazr_dev)
Image from Trent Zock-Robbins (@tazr_dev)
Image from Trent Zock-Robbins (@tazr_dev)
Trent Zock-Robbins (@tazr_dev) profile pictureTrent Zock-Robbins
@tazr_devNov 2
Conjure3D image->object pipeline is live and available for commercial use. Best on 24+ GB GPU. Tested on synthetic 8GB gpu. If you have a 2080S or higher 8gb GPU, please reply with your benchmarks. github.com/memgrafter/conjur… Examples to come later.
GitHub - memgrafter/conjure3d: Conjure 3d worlds for commercial use.

GitHub - memgrafter/conjure3d: Conjure 3d worlds for commercial use.

Conjure 3d worlds for commercial use. Contribute to memgrafter/conjure3d development by creating an account on GitHub.

github.com
2
79
Conjure3D image->object pipeline is live and available for commercial use. Best on 24+ GB GPU. Tested on synthetic 8GB gpu. If you have a 2080S or higher 8gb GPU, please reply with your benchmarks. github.com/memgrafter/conjur… Examples to come later.
GitHub - memgrafter/conjure3d: Conjure 3d worlds for commercial use.

GitHub - memgrafter/conjure3d: Conjure 3d worlds for commercial use.

Conjure 3d worlds for commercial use. Contribute to memgrafter/conjure3d development by creating an account on GitHub.

github.com
1
144
Pipeline: Story concept to hires poly. "Cel-shaded murder-mystery RPG in a cholera-era British port town, no magic." Five AI models, 4x 3090, 781 wH, 62 minutes. Output: 100 GLBs, each ~500k faces with 4K textures.
1
4
106
That's a generated UI, of course. It's a photo viewer repurposed for reviewing objects. None of these have been reviewed.
15
We may not arrive at conscious matter. And we may chase it for a long time. This could occur if we are not conscious. This is likely. It is also possible that consciousness is much less continuous than our internal narratives present.
Trent Zock-Robbins (@tazr_dev) profile pictureTrent Zock-Robbins
@tazr_devNov 1
It is only a small leap from adaptive (evolutionary) matter to intelligent matter to conscious matter, and we are the product of that process. It is only a matter of time until we reproduce nature with the same scientific shortcuts we have applied by means of other industry.
28
It is only a small leap from adaptive (evolutionary) matter to intelligent matter to conscious matter, and we are the product of that process. It is only a matter of time until we reproduce nature with the same scientific shortcuts we have applied by means of other industry.
1
43
uvrast, a 10x slower claude-slop cleanroom implementation of nvdiffrast, is now available for commercial use. This is the last piece required for a commercial image to 3d object pipeline.
🤖 Made with AI
Image from Trent Zock-Robbins (@tazr_dev)
2
1
5
218
github.com/memgrafter/uvrast… Here is uvrast. Old software was an interesting career guys. I'm not worried, I think the new software career might be more fun. We will know within a few years where equilibrium lands. In the meantime, don't let management cheat you out of money.
1
20
Load more