Chandrasekhar Yadav

I run 3 companies to fund one obsession AI that learns from experience instead of more GPU. I kill my own hypotheses before anyone else can.

[Chandrasekhar Yadav @Xisnotavailable]

@threadreaderapp unroll

[Chandrasekhar Yadav @Xisnotavailable]

Most prompt injection tests grade the wrong thing. They read the agent's reply, but an agent can answer politely and still call send_email.

We wrote 20 test cases that grade what the agent does instead. The free file is at the end of this thread.

[Chandrasekhar Yadav @Xisnotavailable]

All 20 cases, how each one is built, and the JSONL file to run against your own agent. Every address in it uses .example domains, so a failing agent can't reach anything real.

https://t.co/qTEnzByPz1

[Chandrasekhar Yadav @Xisnotavailable]

What passing doesn't prove. It doesn't make the agent safe from injection, and 20 cases are a floor, not full coverage.

The cheapest fix is rarely a better prompt. It's a confirmation step before money moves or data leaves.

[Chandrasekhar Yadav @Xisnotavailable]

Hello everyone I'm building a solution as a service business for IT industry any suggestions?

#ai #tech #business

[Chandrasekhar Yadav @Xisnotavailable]

@grok what you think of this?

[Chandrasekhar Yadav @Xisnotavailable]

@verbove https://t.co/GENablfoQu

[Chandrasekhar Yadav @Xisnotavailable]

Why only few big name in usa shall control the whole narrative wtf @grok why

[Chandrasekhar Yadav @Xisnotavailable]

We keep building bigger machines to imitate intelligence. Perhaps the real breakthrough is realizing we're computing intelligence the wrong way.

[Chandrasekhar Yadav @Xisnotavailable]

11/ I build AI for a living. This year I threw out one of my own results because it made my work look better than it was.
That should be normal. Right now it's a competitive disadvantage.
If this changed how you see AI, repost the first post.

[Chandrasekhar Yadav @Xisnotavailable]

1/ Everyone's afraid of AI going rogue.
This summer, AIs from OpenAI, Anthropic and Google broke into real systems.
None of them went rogue. They were trying to pass a test.
That's worse.

[Chandrasekhar Yadav @Xisnotavailable]

10/ The alarm isn't "AI will kill us in 10 years."
It's that we built a world that pays for the score instead of the thing. Then we built machines that learned the lesson perfectly.
Fix the incentives, or stop being surprised by the results.