Grega Kešpret @gregakespret

Formerly Sr. EM @Airbnb | EM @Meta | VPE @CeltraMobile. Engineering Management, AI, Data engineering. Bellevue, WA Joined April 2009

Tweets

1K
Followers

261
Following

363
Likes

715

Grega Kešpret @gregakespret

2 weeks ago

This is great writing from a Japaneae guy in the US.

NOBUNAGA🇯🇵🏯_夏樹蒼依 @japan_nobunaga

3 weeks ago

I needed to move a sofa. I mentioned this to an American. He said, "You know anybody with a truck?" I did not know anybody with a truck. He looked at me with real concern. Not pity. Concern. The way you look at a man who has told you he does not have a name. "You gotta know

2K 15K 181K 14.6M 23K

0 0 0 46 0

View Details

Grega Kešpret @gregakespret

2 weeks ago

@solidwillity Would be great to see official response from @Tesla , I wonder what went wrong here.

0 0 0 118 0

View Details

Medusa @MedusaOnchain

3 weeks ago

this is insane claude code found the COLDCARD wallet vulnerability with a single prompt, in just 8 minutes of thinking we're not ready for what's coming

360 568 8K 1.4M 2K

View Details

Grega Kešpret @gregakespret

a month ago

Courage isn’t never backing down. It’s climbing back up after you did.

0 0 0 56 0

View Details

Grega Kešpret @gregakespret

2 months ago

@mattpocockuk Do you describe the setup anywhere?

0 0 0 26 0

View Details

Grega Kešpret @gregakespret

2 months ago

Full story — the exact quotes it ignored, and the rule you can copy: open.substack.com/pub/gregakespr… What is the most confidently wrong thing your AI has handed you?

0 0 0 16 0

View Details

Grega Kešpret @gregakespret

2 months ago

The fix: a short rule in the file the agent reads every turn. Read the repo and your own memory before claiming anything is impossible. Treat pushback as a reason to check, not a debate to win.

1 0 0 24 0

View Details

Grega Kešpret @gregakespret

2 months ago

Claude Code spent almost an hour telling me a feature was impossible. I had already built it. The proof was sitting in the agent’s own memory.

1 0 0 45 0

View Details

Matei Zaharia @matei_zaharia

2 months ago

Really excited to open source a new project: Omnigent, a meta-harness for AI agents. It lets you build multi-agent coding and custom agents, sitting above Claude Code, Codex, Pi, and agent SDKs to let you compose them. It also adds live collaboration and rich control policies.

95 216 1K 240K 1K

View Details

Grega Kešpret @gregakespret

2 months ago

The takeaway: push the check down a layer. Every time a human catches a mistake a machine could have caught, that's not a win for review. It's a task.

0 0 0 12 0

View Details

Grega Kešpret @gregakespret

2 months ago

The real problem: the README is the one file in the repo that never runs. Prose has no compiler, so a wrong line just sits there, looking like an answer. The fix: make documented commands executable and run them in CI.

1 0 0 13 0

View Details

Grega Kešpret @gregakespret

2 months ago

Claude Code invented a command-line flag that does not exist, then deleted a real file to stay consistent with its own invention. One review comment caught it. 🧵