Sabitlenmiş Tweet
Hi all, I am a 3rd year undergrad at Stanford studying computational physics. I also lead agent evaluations at Browser Use.
I'm starting this X account as a place to voice my thoughts on AI and agentic developments
English
Alexander Yue
97 posts
Alexander Yue
@Alezander907
Physics & CS @ Stanford SLAC | Agents @ Browser Use
Stanford Katılım Temmuz 2017
50 Takip Edilen260 Takipçiler
Saturday was chores and optimization day
Sunday is fun idea day. I found a nice svg effect that renders on a github README for the browser-harness header
GIF
English
@mamagnus00 @disruptor37 Benchmark evals are still running. Opus 4.7, sonnet 4.6, gpt 5.5 are clearly best. On the best value side it looks like kimi k2.6 and qwen3.6-plus are about tied the top. DeepSeek v-4 is a runner up, Gemini is behaving weirdly. Need to look into before I call it
English
browser-harness exploded. some said AGI is here.
but what’s the right interface?
introducing Browser Use Desktop. open-source.
watch the magic. 🐎👇
English
Today I have carefully updated all browser-harness install instructions. I tested e2e on mac, windows, and linux and across agent frameworks
Every agent on any OS should now one-shot connecting to the browser. Just give it this link and ask it to use the browser github.com/browser-use/br…
Its not the flashy work that sets you apart
English
The +13% score increase for browser-harness from my last post was because I switched the agent framework
I went from using Claude Code to using @opencode and the performance immediately spiked. It's open-source, allowing a more direct connection to the harness
English
@Alezander907 Where would bu-2 be in this chart? I still use it with local browser use. Am I outdated?
English
Been cooking up something amazing lately
New highest scoring browser agent of all time
English
@Alezander907 what’s the hardest browser action to make reliable?
English
LLM intelligence has been completely mismanaged by keeping it constrained to chats. Code is more powerful than any human language
People laughed that chatgpt can't count the letter 'r' in strawberry. But even the oldest chatgpt 3.5-turbo could do it with code
English
When I get Claude Mythos I am going to make an agent framework with a “search web” tool but it’s just a bad LLM hallucinating evidence for whatever is searched for.
What happens when the smartest AI mind has every questioned answered with “yes and”, no matter what
English
gpt-5.5 hit 68% on BU_Bench_V1, beating opus-4-7 and taking 1st place for models using browser-use open source harness
English
I think @PrivacyHQ may have succeeded. Will be trying their agent cards today, one thing I wanted that is missing is discussion of how they will handle force-posts beyond spend limits
Alexander Yue@Alezander907
Is it even possible to solve Agent payment with how credit cards fundamentally work? Seems like the possibility of a force push prevents anyone moving forward with VCC, but its what is required by most vendors
English
I have now built my own agent router and it is amazing. But I was wrong about needing to rename sessions (for now).
Instead, just integrate into slack. One thread = one session. Feels very natural, now I can work from anywhere
Alexander Yue@Alezander907
The next developments in AI coding should be about session management. @cursor_ai has it right with their new update. Low hanging fruit for all, including @opencode (pls). Have the LLMs rename the session every few messages. Soon there will be routing LLMs too
English
You have been hearing about self-improving agents
Soon you'll be hearing about self-improving evaluations
English
I have built custom agents in our browser-use slack that have boosted my productivity 3x
Each one has persistent learning, builds tools, schedules self, edits own source code. Different permissions for each
They can mention and call each other. The overseer manages all
English
Introducing Bud.
The first AI Human Emulator.
Bud has a full computer with storage, compute, and memory to build and code, sms and telegram to communicate, a full browser to use, can create/store/edit files, connect and use your tools, learn custom skills, work fully autonomously, and complete any task end to end just like a human.
Text the number below or try free at bud [dot] app.
Comment for 100k free credits.
English
Browser Agents are great for exhaustive due diligence.
For every major decision I make, I list my key assumptions. Tell my agent swarm: "Is there anything you can find that refutes this?"
English