pepmach

@pepsi_machine

building coding agents @kirodotdev

San Francisco, CA

github.com/pepmach

Joined May 2020

365 Following

33 Followers

pepmach
@pepsi_machine
Nov 14
If this is true it'd be an interesting Silicon Valley episode
山景城小路
@Zen_with_AI
Nov 13
微软 365 的 SRE 写事故报告，写出了灾难片的感觉。一位微软员工发帖自嘲：Azure Front Door 的服务发现功能“开箱即用就是灾难”，一开就把所有流量打到没准备好的区域，全球用户断网，连自家 M365 服务一起陪葬。 Reddit 的评论：“我真喜欢看他们写这个，就像看灾难片。”
pepmach
@pepsi_machine
Nov 14
Read the Stand with Jensen letter and consider adding your name. #standwithjensen
Stand with Jensen | An open letter to the NVIDIA Board
From standwithjensen.com
pepmach
@pepsi_machine
Nov 13
is claude down?
1
pepmach
@pepsi_machine
Nov 13
Whoever wrote this piece doesn’t even bother to conceal their bias, hatred, and the latent desire to make himself known to the power behind the coordinated attack on Jensen. What a terrible way to ruin your reputation and trust as a biographer.
Pat McGovern
@pw_mcgovern
Nov 13
from jensen's biographer (@stephenwitt) in today's NYT
pepmach
@pepsi_machine
Nov 2
did he really just say “read & write weights”
Rikiya Takehi
@rikiyatakehi
Nov 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:
Readers added context
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.…

pepmach

pepmach’s posts