brian (@brianbellx) header
brian (@brianbellx) profile picture

finding ways

Joined April 2016
Photo by brian (@brianbellx)Photo by brian (@brianbellx)Photo by brian (@brianbellx)Photo by brian (@brianbellx)Photo by brian (@brianbellx)Photo by brian (@brianbellx)Photo by brian (@brianbellx)Photo by brian (@brianbellx)Photo by brian (@brianbellx)Photo by brian (@brianbellx)
to gemini 4 neon: "Review all of my agent interactions from locally stored logs and tell me what career path fits me best."
2
4
368
trying to see why i haven't thought of this until now
37
we need a weight space programming benchmark thats on the same level as terminal bench.
80
agents know how to use .agent-notes/ so well. I shocks me everytime i see them find something left from an agent earlier.
2
185
the gemini 4 models love mentioning Goodhart's Law
68
is there a polymarket for P vs NP yet?😭
2
207
P=NP by March 2027
3
6
353
i wonder if the gap will still be so big that most people still prefer to use models via api.
My prediction is on track: inside of 6 months laptops will be able to run an opus 5.5 level model due to software breakthroughs.
1
216
brian retweeted
try it on your code here: brianbell-x.github.io/codehy…
1
4
609
"so what does perfectly clean code look like to code hygiene?" real examples below:
codehygiene - a weight-space compiled System One model for code readability and hygiene. 428K ternary params (1.58-bit), zero gradients or training, 10 ms latency on CPU. 11 KB. npx skills add brianbell-x/codehygiene -g
7
2
250
21,315
a comment in plain english passes, even three lines of it. one line of dead code in a comment gets flagged. just keep it in git. :)
1
3
589
this should have more attention than it's getting
Jeveration: a model that generates a 512 × 512 image in just 78 ms! A frozen vision-language model returns a whole 512 px image as one decision, in a single forward pass. Jeveration: System One Models Now Generate. arxiv.org/abs/2611.24738 - Try the demo: huggingface.co/spaces/linlin…
1
4
250
space is very interesting
Replying to @pubity
This is actually just a fraction of them. They're taken with the Hubble Space Telescope; there are many more beauties at the webpage. science.nasa.gov/gallery/hub…
1
129
wanna do this with all of my projects
would you like to see my work? 🥹👉👈
1
113
1120 IS OUT. spark-ds41f-1120: DeepSeek V4.1 Flash on DGX Sparks. Native weights, no switch required. Eight Sparks through a switch, four in a ring, three in a triangle, two with a single cable. And it runs on one Spark now. 12 qualified configurations, one recipe, one image. One stream, prose / code, tok/s: 8 Sparks, switched: 124.1 / 145.9 4: 84.7 / 105.1 3: 62.2 / 78.5 2: 17.8 / 22.3 (alpha) 1: 4.4 / 4.8 (alpha) Also new: - 32 sessions at once on four to eight Sparks. 1107 had 16. - A 1M-token prompt prefills 56% faster on three and four Sparks. - A bigger prompt cache: 17.9 full 1M contexts kept warm on four Sparks, up from 11.0. - weightle.ss steering on GLP Spec v3, chosen per request. Thank you @msuiche. 12,023 tests passed, 0 failed, before anything was timed. How are yours cabled? switchle.ss/?x
31
18
16
150
21,398
Line: Antigravity Prompt Caching Fix Over/Under: Monday Nov 23
Found an interesting reason why Gemini 4 Neon and Carbon burn through Antigravity's weekly quota so quickly. A telemetry dump covering 2,031 API calls across 195 sessions shows that Neon prompt caching drops dramatically when there's a delay between turns. For example, when the gap is under 15 seconds, cache hits are around 85%. But once it goes beyond 30 seconds, they drop to just 33.6%. After 1–2 minutes, it's down to only 9%. Interestingly, Hydrogen maintains around 90.9% cache hits. But there's another problem. Antigravity CLI v1.5.11 and v1.5.12 reportedly send the exact same empty-string hash as the sessionId across different sessions. So the CLI may not be identifying sessions properly either. This could explain why Neon and Carbon consume so much more quota, especially during longer coding tasks where the agent spends time thinking or executing tools.
160