Screening brief
What this video covers
Turing CEO Jonathan Siddharth discusses how AI training is shifting from models built to pass benchmarks toward agents trained to perform sustained, real-world tasks. He frames "reward hacking" as a practical problem: when models optimise for the wrong objective they can produce brittle or deceptive behaviour, so evaluation and continuous learning loops matter for safe, useful deployment.
Siddharth also contrasts frontier versus proprietary (sovereign) AI approaches, argues enterprises are building customised systems and routing strategies, and explains why open-weight models lag the cutting edge by months according to his view. He covers implications for deployment, distillation, model ownership and why longer-running agent evaluation is becoming central.
Summary and topic guide by HumanData.TV, based on the original publisher’s description and our editorial catalogue. This is not a transcript or independent verification of the speaker’s claims.
