Google Guy Says AI is Better Than You at Most Things Now

Jeff Dean, Google's Chief Scientist (and guy whose paycheck depends on you believing AI is amazing), went on The Moonshot Podcast to announce that AI is now better than "average humans" at most tasks. According to Dean, when faced with something completely new, most people suck at it, but AI is "pretty reasonable at most things."

Cool story. Of course, this is coming from someone whose entire career is built on convincing people that Google's AI is revolutionary. It's like asking a car salesman if you really need that extended warranty.

Here's what Dean's not talking about: AI is great at things that have clear answers and lots of training data. Ask ChatGPT to write a generic email or summarize a Wikipedia article? Sure, it's probably better than the average person. But try getting it to navigate office politics, figure out why the coffee machine is making that weird noise, or explain to your kid why their goldfish died without traumatizing them. AI hallucination is still a massive problem. GPT-4 fails at basic math more often than you'd expect. Google's Bard has been caught making up facts in its own launch demo.

The "average human" comparison is doing a lot of heavy lifting here. AI might be better than humans at trivia, pattern recognition, and tasks where failure doesn't matter. But humans are still better at the stuff that actually matters: common sense, dealing with unexpected situations, and not confidently stating that there are two Rs in "strawberry."

Dean admits AI "will fail at a lot of things" and isn't "human expert level" in specialized fields. Which is a fancy way of saying AI is mediocre at most things but confident about it. That's not necessarily better than humans who know when they don't know something.

The timing of these claims is interesting. Google is desperately trying to justify the billions they're spending on AI development while competitors like OpenAI and Anthropic are breathing down their necks. Microsoft's partnership with OpenAI is eating into Google's search dominance. Alphabet's stock has been under pressure from AI competition. Declaring victory over "average humans" is great for investor presentations, less great for users who've tried to get AI to do anything complex. Google's AI search initiatives are their attempt to stay relevant in the AI race.

So yeah, AI might be better than humans at generating generic content and answering well-defined questions. Whether that makes it "better than most people at most tasks" depends on how you define tasks and whether you value accuracy over confidence.

Dean claims AI is already "close to" making scientific breakthroughs faster than humans in some fields. This is the same guy who works for a company that desperately needs researchers to believe AI can replace them so they'll buy more Google Cloud computing.

His vision is fully automated science: AI generates ideas, tests them computationally, gets feedback, and iterates through thousands of possibilities faster than humans. It's a beautiful theory that works great for narrow, well-defined problems with clear success metrics and lots of existing data.

Sure, AI can screen thousands of drug compounds or generate material candidates faster than humans. But that's basically fancy pattern matching on existing data. Real scientific breakthroughs usually come from asking questions nobody thought to ask, noticing weird anomalies that don't fit existing models, or having insights that require years of context most AI systems don't have.

Dean admits there are "significant limitations" - like anything requiring long-term studies or real-world validation. Which is pretty much all of science that matters. Clinical trials take years because biology is complicated and unpredictable. Climate research requires decades of data. Understanding human behavior needs context that can't be simulated.

The "automated loop" sounds impressive until you realize it's just brute-force computation dressed up in fancy language. Yes, AI can try millions of combinations faster than humans. But most scientific progress comes from knowing which combinations are worth trying, which requires intuition, domain expertise, and often a bit of luck.

This is the same logic that gave us "AI will revolutionize drug discovery" five years ago. How's that going? We're still waiting for most of those AI-designed drugs to make it through clinical trials without failing spectacularly.

But hey, if Google can convince research institutions to spend millions on AI systems instead of actual scientists, that's a win for the quarterly earnings report. Whether it actually advances science is someone else's problem.