Building Browser Agents Evals at @googlechrome Curious mind, love to solve problems, serve humanity with tech AI, spirituality, bringing the best from people
Mountain View, CA
Joined July 2015
- Tweets181
- Following324
- Followers228
- Likes834
"It's better to be brutally honest than hypocritically polite" - @vkhosla
The fireside chat was full of wisdom. Thank you sir and @TerraAPI.
We see great innovations happen from places where failures are not treated bad, encouraging people to take risks.
Always have your own objectives on why you want to start your company/work on something. All are valid. Successful people have clear objectives.
The biggest impact we are gonna see in is healthcare, data shows AI are better than humans in diagnosis and help reduce the false diagnosis and companies backed by khosla ventures showing more than 90% compared to 73% accuracy by humans.
#YCStartupSchool @ycombinator #AIxHealthcare
Everyone at YC startup school gets claude Max 20x!!
Time to build!
@ycombinator #YCStartupSchool
Strong, powerful talk and messages by @garrytan @ycombinator
"If you have to redo the same thing again manually you have failed"
"You can fly now, not metaphorically, mechanically."
"Its all made up but you get to make it up"
#YCSUS
Step 1) Pick a harness, Run an agent on your own machine (Open Claw, Harness, Gbrain.io anything)
Step 2) Start your library one folder of markdown files (export the mail, each project you are working, and each person you are working with(
Step 3) Write your first skill file
Step 4) Wire one recurring job (every morning do this, every Friday do that)
Step 5) Never do one-off work (one operation, close the context away, at the end of each task, extract skill and turn it into markdown file.
Thank you @bcherny for your great insights!
Few takeaways from YC StartupSchool.
1) Do you even need a system prompt when the models are getting so good?
* Boris took Ark benchmark from 3% → 30% — largely by deleting his own system prompt
* He claims the system prompt is often making the model dumber: written for last gen, full of scaffolding the new model doesn't need.
* The rebuild loop what claude code team uses:
delete the whole prompt → run the model bare → run full evals and watch it stumble. Don't intervene too early. Let it fail, then adjust the harness based on what you see.
Why this holds across releases: evals are constant, you keep appending, they outlive the harness by 1–3 model generations. Saturate one, invent a harder one. Watch for hobbling + product overhang.
How many of the products/teams are actually working on rewriting the whole prompt from ground up after every model release?
Evals are at the bottleneck: Having very good evals and reliable eval infrastructure decides how well and fast we evolve the product over newer versions!
#YCStartupSchool @ycombinator
How do top 1% users use Claude Code?
* Find what a new model can do: give it a genuinely hard task, for example someone on Claude team make it rewrite the thing in another language (Fable does this aggressively), then have Claude verify its own output.
* Top 1% of users: ignore LinkedIn influencers, give it skills + MCP + access as needed, stop over-engineering. Most people constrain the agent to death; the best sandbox it and let it cook.
* They use real unlock is dynamic workflows with loops and routines.
* Claude code maintaining itself — running experiments, writing tests, deleting tests, an "abstraction police" passing over the codebase.
Glad to be around 6000 ambitious tech people who wants to make something people want.
The energy is wild! The future is here!
Glad to be back at YC startup school along with 6000 ambitious builders.
Starting with great talk by Jensen sharing the path of entrepreneurship.
How hard it can be?
To get going is the attitude of entrepreneur.
Learning is the greatest superpower. If anybody can do it believe you can do it too. Just overcome this morning, if you sticks with it long enough it will happen.
@ycombinator #YCStartupSchool
@JensenHuang
The concentration of talent at Anthropic is insane.
Every time I watch this level of tennis I am in awe, like how they plan such an amazing level of tennis and sustaining through long rallies and each tournament he plays.
Truly inspiring gentlemen, showing how to be kind, humble, consistent, giving his best to every point.
Sreeja Kamishetty retweeted
Today at the @Android Show (I/O edition) we announced Gemini Intelligence - bringing the best of Gemini to our most advanced devices.
Automate multi-step tasks across apps and Chrome, fill out forms in a single tap, turn spoken thoughts into polished text with Rambler, build custom widgets & loads more.
