Latest / Connor with Honor: AI, Real Estate and Real Life / How AI really works, it was grown not made and that should give you a wierd feeling

Transcript

How AI really works, it was grown not made and that should give you a wierd feeling · Connor with Honor: AI, Real Estate and Real Life

• 0:00
  SPEAKER_00: You keep saying I should understand this stuff, but I'm fifty-seven.
• 0:05
  SPEAKER_00: Is it too late for me?
• 0:06
  SPEAKER_00: No, and most people talking about it don't understand it either.
• 0:09
  SPEAKER_00: And that's the part nobody says out loud.
• 0:11
  SPEAKER_00: Start at the beginning.
• 0:13
  SPEAKER_00: What is it actually doing?
• 0:14
  SPEAKER_00: It's predicting what comes next.
• 0:16
  SPEAKER_00: So you type something and it works out the most likely next piece of writing over and over and over very, very fast.
• 0:24
  SPEAKER_00: That is it?
• 0:25
  SPEAKER_00: Well that's the engine.
• 0:26
  SPEAKER_00: Everything else is scale.
• 0:28
  SPEAKER_00: An enormous amount of text and an enormous amount of computing power.
• 0:32
  SPEAKER_00: Where did it learn all of that?
• 0:34
  SPEAKER_00: From human writing, books, websites, forums, manuals, arguments, instructions, all of it.
• 0:40
  SPEAKER_00: So it's quoting people, isn't it?
• 0:42
  SPEAKER_00: Not usually.
• 0:43
  SPEAKER_00: It learned the patterns, not the pages.
• 0:45
  SPEAKER_00: Closer to how you learn to talk than a library card catalog.
• 0:50
  SPEAKER_00: Is it looking up things when I ask for something?
• 0:53
  SPEAKER_00: Sometimes if the product is built to search, but the thing underneath is not looking anything up, it's producing what a good answer usually looks like.
• 1:01
  SPEAKER_00: That sounds like a problem.
• 1:03
  SPEAKER_00: It is the problem.
• 1:04
  SPEAKER_00: It can hand you something that looks exactly like the right answer, but is completely wrong.
• 1:10
  SPEAKER_00: Why would it do that?
• 1:11
  SPEAKER_00: Well, because it was built to sound right, not to be right.
• 1:14
  SPEAKER_00: Those two things overlap most of the time, and most of the time is not all of the time.
• 1:19
  SPEAKER_00: So how do I know when to trust it?
• 1:21
  SPEAKER_00: Trust it on shape and on first drafts.
• 1:24
  SPEAKER_00: Check it on facts, names, numbers, and dates.
• 1:27
  SPEAKER_00: Anything you would be embarrassed repeating.
• 1:30
  SPEAKER_00: You gotta check it.
• 1:31
  SPEAKER_00: Does it remember me?
• 1:32
  SPEAKER_00: The thing underneath does not.
• 1:34
  SPEAKER_00: Every conversation starts from nothing.
• 1:36
  SPEAKER_00: Some products bolt memory on top of it, and that's a choice a company made.
• 1:41
  SPEAKER_00: Not the machine deciding to remember who you are.
• 1:44
  SPEAKER_00: So it's not learning from me personally.
• 1:46
  SPEAKER_00: Not in the moment, no.
• 1:48
  SPEAKER_00: Whether your conversation gets used later to train something is a policy question, and the answer is in the fine print.
• 1:54
  SPEAKER_00: Should I be careful what I type into it?
• 1:56
  SPEAKER_00: Yes, same rule as anything else.
• 1:58
  SPEAKER_00: If a competitor reading it would cost you money, then I wouldn't put it in a tool whose retention policy you haven't read.
• 2:07
  SPEAKER_00: Alright, is it thinking?
• 2:09
  SPEAKER_00: That's the question everybody actually wants answered, and I'm going to give you the real answer instead of a comfortable one.
• 2:14
  SPEAKER_00: Go ahead.
• 2:15
  SPEAKER_00: Nobody knows.
• 2:16
  SPEAKER_00: And I don't mean that as a dodge.
• 2:18
  SPEAKER_00: I mean that the people who built these things say so in public.
• 2:22
  SPEAKER_00: Who says that?
• 2:23
  SPEAKER_00: Dario Amodi, who runs Anthropic, one of the biggest AI companies in the world, he published an essay in April 2025 called The Urgency of Interpretability.
• 2:33
  SPEAKER_00: And what does he say?
• 2:35
  SPEAKER_00: That when one of these systems makes a choice, they have no idea at a specific level why it actually made the choice.
• 2:42
  SPEAKER_00: The company that built it doesn't know why it does what it does.
• 2:46
  SPEAKER_00: And that's what he wrote on his own website under his own name because, well, he's running the company.
• 2:52
  SPEAKER_00: How is that even possible?
• 2:54
  SPEAKER_00: Because of how it gets made.
• 2:55
  SPEAKER_00: His own words are that these things are grown more like than they're built.
• 3:00
  SPEAKER_00: Grown?
• 3:01
  SPEAKER_00: Like a plant.
• 3:02
  SPEAKER_00: You set the conditions, you control the soil and the water and the light.
• 3:06
  SPEAKER_00: You do not control where every branch actually sprouts from and ends up growing to.
• 3:12
  SPEAKER_00: Well, that's not how software works.
• 3:14
  SPEAKER_00: It is not.
• 3:15
  SPEAKER_00: Within with normal software, a person wrote every rule.
• 3:18
  SPEAKER_00: With this, nobody wrote the rules.
• 3:20
  SPEAKER_00: The system actually found them.
• 3:22
  SPEAKER_00: That is unsettling.
• 3:23
  SPEAKER_00: It should be a little.
• 3:24
  SPEAKER_00: But notice what it is not.
• 3:26
  SPEAKER_00: It is not a monster waking up.
• 3:28
  SPEAKER_00: It is people shipping something faster than they can actually explain it.
• 3:33
  SPEAKER_00: Well, is anybody actually working on that?
• 3:35
  SPEAKER_00: Yes, and he put a clock on it.
• 3:37
  SPEAKER_00: He wants to be able to reliably diagnose what's going on inside these systems, and he gave that five to ten years.
• 3:44
  SPEAKER_00: Which means right now they can't.
• 3:47
  SPEAKER_00: Right now they mostly cannot.
• 3:48
  SPEAKER_00: And that's where actually where it actually stands here in 2026, September.
• 3:53
  SPEAKER_00: Has anything changed since then?
• 3:55
  SPEAKER_00: Something big.
• 3:56
  SPEAKER_00: In July of this year, and this is the part I find genuinely fascinating.
• 4:01
  SPEAKER_00: What happened?
• 4:01
  SPEAKER_00: On July 6th, Anthropic published research saying they found something like a workspace inside of the model, a small area where it holds what it is currently thinking about.
• 4:12
  SPEAKER_00: A workspace.
• 4:13
  SPEAKER_00: Dozens of idea at a time, and less than a tenth of everything going on in there.
• 4:17
  SPEAKER_00: And nobody actually designed it.
• 4:19
  SPEAKER_00: What do you mean it appeared?
• 4:21
  SPEAKER_00: Nobody sat down and built that room.
• 4:23
  SPEAKER_00: It formed it on its own while the thing was being trained, and they only found it afterward by actually going and looking.
• 4:29
  SPEAKER_00: Okay, and what is in the room?
• 4:31
  SPEAKER_00: Here's the part that will sit with you.
• 4:33
  SPEAKER_00: They find it working, they found it working through steps inside the space that never show up in the answer that it actually hands you.
• 4:41
  SPEAKER_00: Is it thinking things that it doesn't say?
• 4:44
  SPEAKER_00: On multi-step problems, yes, the middle of the thinking happens in there and never makes it onto the page.
• 4:50
  SPEAKER_00: That sounds like it's hiding something.
• 4:52
  SPEAKER_00: I wouldn't go that far, neither did they.
• 4:55
  SPEAKER_00: You do the same thing.
• 4:57
  SPEAKER_00: You do not narrate every thought before you open your mouth.
• 5:00
  SPEAKER_00: Could you imagine that world?
• 5:01
  SPEAKER_00: But you can't see mine either.
• 5:03
  SPEAKER_00: Right, the difference is that they built a tool that they can look at, some of this now.
• 5:08
  SPEAKER_00: Now they can read it and they can steer it, and that's brand new.
• 5:12
  SPEAKER_00: So is it conscious or not?
• 5:13
  SPEAKER_00: Here's why I trust this particular research.
• 5:16
  SPEAKER_00: They said it very plainly that it does not show the thing has experiences or feelings.
• 5:22
  SPEAKER_00: They said that themselves.
• 5:24
  SPEAKER_00: In their own paper, they separated being able to report on your own thoughts from actually having an inner life, and they only claimed the first one.
• 5:32
  SPEAKER_00: A company had the chance to say their product might be alive and didn't take it?
• 5:37
  SPEAKER_00: Well, this is exactly why I'm telling you about this one and not the 10 scarier headlines you probably already saw this week.
• 5:44
  SPEAKER_00: So, what do I actually do with all of this?
• 5:46
  SPEAKER_00: Use it for drafts, summaries, and keep the and the boring repetitive work.
• 5:51
  SPEAKER_00: Check anything factual, keep your private business out of a free tool.
• 5:56
  SPEAKER_00: And that covers about 90% of it.
• 5:58
  SPEAKER_00: And the rest?
• 5:59
  SPEAKER_00: Stay curious, stay skeptical of anybody who's certain in either direction.
• 6:03
  SPEAKER_00: The people closest to this day, they don't fully understand it yet.
• 6:06
  SPEAKER_00: You're not, you're you're actually allowed to say that as well.
• 6:09
  SPEAKER_00: So AI for everyone, not just the wealthy.
• 6:12
  SPEAKER_00: I'm Connor with honor.
• 6:13
  SPEAKER_00: Be safe out there.
• 6:14
  SPEAKER_00: We'll see you in the next one.