> Currently, we can’t just assume that these AI models are totally right,
Why could we ever assume that?
> but the notion that it could be right took the guesswork out of our next steps,
Devils advocate here. Couldn't this just be a severe case of confirmation bias? You take 100 such cases, ask AI "how does it work?" and in 99 of those, the answer is somewhere on the spectrum between "total nonsense" and "clever formulation but wrong". One turns out to be right. That's the on we are seeing here, getting confirmed in the lab. That doesn't actually mean AI reduced the time by 75%.
A broken clock is also correct twice a day. We wouldn't say we have invented a clock that works without energy, sure it's wrong sometimes, but when it's correct, it's awesome! No, it's just a broken clock that's wrong most of the time.
I would also love to see that with "generative AI" we have discovered some helpful magic, but as long as we are not honest about those details (which would include publishing and owning up to mishaps), this is all just riding a hype train.
I think this perspective overlooks how human expertise actually works. Humans in cutting-edge research also get things wrong a lot: most hypotheses fail, most experiments don't pan out, and most novel approaches lead nowhere.
When we celebrate a scientist who makes a breakthrough, we're not crediting them for being right 100% of the time. We're recognizing that they were right more often than random chance and earlier in the process than would otherwise occur.
A researcher (or AI) who can identify promising directions at a 2/99 or 3/99 rate instead of 1/99 is genuinely valuable – they're effectively doubling or tripling the efficiency of the discovery process.
Imagine if AI can test theories in under 100 seconds AND is slightly better out of 99 tries at getting things right. Beats the human out of the water.
This is exactly how I use AI now daily: Provide it with relevant context and help me to troubleshoot some software issue. I found often it casts a wider net and provides good leads to follow up. It balances out my own bias and overall speeds up discovery of the root cause. Admittedly some benefit comes from having to explain the AI the context - half the problems are gone once one can explain them clearly.
> You take 100 such cases, ask AI "how does it work?" and in 99 of those, the answer is somewhere on the spectrum between "total nonsense" and "clever formulation but wrong". One turns out to be right.
They're still using the scientific method, the only thing they're getting from AI is hypotheses to test. And AI is great at brainstorming plausible hypotheses.
Why could we ever assume that?
> but the notion that it could be right took the guesswork out of our next steps,
Devils advocate here. Couldn't this just be a severe case of confirmation bias? You take 100 such cases, ask AI "how does it work?" and in 99 of those, the answer is somewhere on the spectrum between "total nonsense" and "clever formulation but wrong". One turns out to be right. That's the on we are seeing here, getting confirmed in the lab. That doesn't actually mean AI reduced the time by 75%.
A broken clock is also correct twice a day. We wouldn't say we have invented a clock that works without energy, sure it's wrong sometimes, but when it's correct, it's awesome! No, it's just a broken clock that's wrong most of the time.
I would also love to see that with "generative AI" we have discovered some helpful magic, but as long as we are not honest about those details (which would include publishing and owning up to mishaps), this is all just riding a hype train.