Newton: Google's AI models are trained to 'teach to the test'
Casey Newton · Bold Predictions For Tech and AI in 2025 — With Casey Newton · Jan 5, 2025 · at 25:57
Casey Newton gives his theory on why Google's Gemini models lead public LLM benchmark leaderboards despite underwhelming product execution.
“Well, I have a conspiracy theory, which is you know kind of not great to be sharing on a podcast, but I'm gonna do it anyway, which is, I think Google is teaching to the test. You know what I mean? Yes. Like, I think that they know what the benchmarks are, and they train the models to be really good at the benchmarks, and then they train the models, and they come out and they say, look, it does really well on the benchmarks, and then you just try to use it for other stuff, And it kind of sucks.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →