Amitabh (@amitabh) profile picture

Undying Curiosity

Joined March 2008
Photo by Amitabh (@amitabh)Photo by Amitabh (@amitabh)Photo by Amitabh (@amitabh)
Amitabh retweeted
the illusion of choice
33
104
15
1,543
56,836
Gradient Descent, Evol search and now this? Who builds the first weight programming model.
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…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
25
If everything is a financial instrument. I think we probably see income ETFs powering sell side of prediction markets soon.
107
1/ I drove Meta's Muse agent workflow system end-to-end built one from scratch, ran it, tore it down. The plumbing is genuinely good. But it fails silently in ways that I think is worth sharing for anyone building for agents.
2
2
134
9/ TL;DR : the plumbing is strong real lifecycle, real sandbox, a verify layer that actually prevents fabrication. What's missing is everything around the model: no share/import contract, fragile ingestion, no guarantee the output matches your spec. Build around those three.
1
7
10/ If you're building agents or tools on Muse, plan for: prompts as the distribution format, single-line sends, and re-checking output against your spec. /end
8
Amitabh retweeted
Replying to @MTSlive
I process tokens therefore I am
5
19
539
28,407
Tesla FSD did a terrific job avoiding the red car where a driver decided to just merge into the right lane instead of taking their left turn. Edge cases aside definitely a net positive on road safety. @Tesla
2
162
The response time to avoid the red car above is less than a 100ms. It moved out before I could even take over and try it.
10
Replying to @amitabh @Tesla
Some edge cases where FSD could do better in my empirical exp - doesn't recognize road bumps if they are not black and white stripped. - doesn't pick the most optimal lane which will help it take an easier right or left turn. Instead of takes a last minute lane switch.
46
First it was OpenAI running the last prompt even if you hit rate limits and now this small UX live edit summary. Open AI codex shows some serious attention to detail. 10/10 UX.
1
134
Gpt 5.5 is overall great for execution. But at times it's frustrating that the model just gives up and calls it a dead end. But it's the same issue with Opus and Gemini pro.
104