"These [LLM] systems actually don't understand what it takes to write secure code" – Gary Marcus #BHUSA #LivePost
There hasn't been a fundamentally new invention or discovery in AI since 2017: everyone is following more or less the same formula. – Gary Marcus #BHUSA #LivePost
Neural networks have had all the funding, but symbolic AI systems (like GPS!) have a ton of value. They're not in competition, they're complementary – Gary Marcus #BHUSA #LivePost
Musk figured out that you can make wild AI promises and get money whether those promises are accurate or not. He said every year we're on the cusp of ubiquitous driverless cars, but we're actually not all that close (even if we'll get there eventually) – Gary Marcus #BHUSA #LivePost
OpenAI is a great example: selling stock at a $500B valuation with no clear path to making any kind of profit. It's a form of performance art – Gary Marcus #BHUSA #LivePost
The world favors bold predictions over accurate predictions; which means you have to be skeptical. It's easy to make an "80% demo" of just about anything, but that last 20% is often very hard. – Gary Marcus #BHUSA #LivePost
example: anybody can put an agent demo together, then in the real world it deletes your database. There's so many edge cases in any interesting problem that getting to competence is a 10-year problem rather than a 6-month problem – Gary Marcus #BHUSA #LivePost
I had to raise my probability of doom because it would take a person with arbitrarily large amounts of money being involved in AI in a sloppy manner... and then I realized that guy is here and doing that. But I'm still only at 3%, because humans are resilient – Gary Marcus #BHUSA #LivePost
"Coding agents are the best agents... [but] coding agents have huge security problems" – Gary Marcus #BHUSA #LivePost
AI has a sim-to-real problem; it's easy to make something that works in a lab or a simulation, but the real world is a lot messier. – Gary Marcus #BHUSA #LivePost
You win a virtual Darwin award if you give an AI agent access to your life. (Context: people losing their entire wallet trying to do crypto trading with AutoGPT) – Gary Marcus #BHUSA #LivePost
The best analog is autocorrect -- that's really all these systems do. And that has value, but just letting it do stuff for you is nuts. – Gary Marcus #BHUSA #LivePost
You get answers that sound like they're based on solid reasoning, yet aren't. But you get people who base important decisions and even values on it. It's undermining critical thinking skills. – Gary Marcus #BHUSA #LivePost
I'm really worried that we're headed toward a world where we don't have entry-level jobs that train younger people to be experienced and senior people. We're seeing a tiny effect already (which media overstates), but we're on that path. – Gary Marcus #BHUSA #LivePost
There's an illusion of understanding created when people look at a diagram or see a video explanation or the like. But the understanding isn't genuinely there and breaks down when challenged – Gary Marcus #BHUSA #LivePost
We don't have enough clinical data to know what do do about "ChatGPT psychosis", but we do have enough clinical data to know that it's real, and that it has ended people's lives – Gary Marcus #BHUSA #LivePost
We do have a huge problem with loneliness. Social media has been a huge part of that. AI systems do have the potential to help with this, but they also have a large potential for harm. And a lot of it comes from opportunity costs (using chatbots for companionship means not developing social skills and building real relationships). May provide short-term comfort, but there's serious concern about long-term harms. – Gary Marcus #BHUSA #LivePost