20 | uc berkeley cs | @CalBlockchain | ai research, building stuff | warriors fan
California
Joined May 2020
- Tweets92
- Following94
- Followers137
- Likes687
amazing read. i'd add that people think they build "tokyo"-like codebases via SDD but the issue is people rarely have the situational awareness to understand the mutations they make to their codebases. that's why i find the best teams have their llm wiki almost like a bible that live and die by.
Satwik Panigrahi retweeted
Today, we’re sharing a new state of the art for computer use.
Our system holds the two highest verified scores on OSWorld, the standard benchmark for AI agents that operate a computer like a person: 83.6% using Claude Opus 4.7 and 81.5% using Claude Sonnet 4.6. The human baseline is 72.4%.
🧵 1/7
Satwik Panigrahi retweeted
Excited to announce @tryramp as the official finance partner of Blockchain at Berkeley.
From our consulting team to every B@B Xcelerator cohort, Ramp is now the default for corporate cards, bill pay, and stablecoin accounts. One platform for fiat and on-chain — the way it should've always been.
Get started → ramp.com/partners/blockchain…
Wrote this for @CalBlockchain when I first joined. Super cool protocol called @3janexyz
OpenAI scaling its Trusted Access for Cyber with GPT-5.4-Cyber is obviously a response to Project Glasswing. However, what I find more interesting is the irony of it all. OpenAI now champions (kinda) democratized to specialized tools. Meanwhile Anthropic, who has always had an open ethos, opts for a more gated frontier capabilities through closed programs.
All the latest architectural "unhobblings" on the agent side keep pointing to the same truth: context bloat is the enemy. And while most commentary points to the benchmarks, I see something more fundamental at play.
What results from this is a self-perpetuating RL system that continuously learns its internal model, adapts to new environments with minimal overhead, and scales intelligence without shear token usage.
The software engineering landscape is no longer "how fast can we code this?" The new constraints will be:
1. How fast can we decide WHAT to build? (vision, strategy, priorities)
2. How do we integrate these systems into messy legacy stacks without everything breaking?
I just listened to @gokulr talk about this — PMs are now hands-on prototypers that decide the evals for non-deterministic code. The new moat is therefore judgement + outcomes + integration.
A pattern I've been seeing lately is that most research boil down to clever normalization tweaks, which from first principles, makes sense. Recently ViT-5 achieve massive benchmark advancements off of replacing LayerNorm for RMSNorm + LayerScale.
Similarly GDPO decoupled reward normalization which led to sharper preference optimization and stable training. Nothing flashy, yet leads to compounding returns. With more stable training, a lot of these vision tasks will lead to stronger transfer to world simulaation.
Crypto is slowly realizing the path to ubiquity is through chain abstraction; turning it into an invisible universal backend that powers everyday txns w/o users ever touching a wallet or knowing details.
PMF is there, just look at @blackbird.
