Latest / Connor with Honor: AI, Real Estate and Real Life / 3 Out of Every 100 Tasks Go Where You Don't Want. Is AI Safer?

Transcript

3 Out of Every 100 Tasks Go Where You Don't Want. Is AI Safer? · Connor with Honor: AI, Real Estate and Real Life

• 0:05
  SPEAKER_00: What happened on September third?
• 0:06
  SPEAKER_00: Modeled grades critical for cyber capability, same day.
• 0:10
  SPEAKER_00: LA Unified bandit for every single stud.
• 0:15
  SPEAKER_00: Critical is the press being dramatic.
• 0:20
  SPEAKER_00: So the thing is dangerous.
• 0:22
  SPEAKER_00: Student.
• 0:23
  SPEAKER_00: Well, that's the manufacturer's own system card.
• 0:25
  SPEAKER_00: They wrote the threshold, then said that they met it.
• 0:31
  SPEAKER_00: The company publicized that it can find unknown security flaws on its own.
• 0:35
  SPEAKER_00: They chose to actually print that.
• 0:39
  SPEAKER_00: Read their sentence.
• 0:40
  SPEAKER_00: And the district panicked.
• 0:43
  SPEAKER_00: Surely the board while the district reviews instructional use and safeguards, that is a pause.
• 0:49
  SPEAKER_00: That's not a verdict.
• 0:55
  SPEAKER_00: The committee chair says it surprised the board.
• 0:58
  SPEAKER_00: Make of that what you will.
• 1:02
  SPEAKER_00: Is the model safer than the last one?
• 1:04
  SPEAKER_00: Yes, genuinely 3.4% maligned outcomes against 18.8.
• 1:09
  SPEAKER_00: More than five times better.
• 1:17
  SPEAKER_00: So 3.4% is fine.
• 1:19
  SPEAKER_00: Three out of every 100 real tasks go somewhere you do not want.
• 1:23
  SPEAKER_00: Hire a person with that record and see how it feels.
• 1:26
  SPEAKER_00: They claim 99.79% on prompt injection.
• 1:32
  SPEAKER_00: That sounds excellent.
• 1:35
  SPEAKER_00: 21 failures in 10,000.
• 1:38
  SPEAKER_00: If it reads 10,000 emails this year, that's 21 chances a stranger gives it orders.
• 1:46
  SPEAKER_00: It is excellent.
• 1:47
  SPEAKER_00: It also makes the world safe, useless, without a denominator, and nobody hands you the denominator.
• 1:54
  SPEAKER_00: What's the two clocks idea?
• 2:01
  SPEAKER_00: Labs move in weeks, institutions move in school years.
• 2:05
  SPEAKER_00: Nobody is holding both watches.
• 2:07
  SPEAKER_00: Who wins that race?
• 2:11
  SPEAKER_00: What about the UN red lines?
• 2:15
  SPEAKER_00: Will they make it?
• 2:16
  SPEAKER_00: Whoever moves first, so far that has never once been the slow clock.
• 2:24
  SPEAKER_00: 200 plus signatories, 10 Nobel winners, deadline of end 2026.
• 2:30
  SPEAKER_00: That's 114 days from this morning.
• 2:33
  SPEAKER_00: A year in no treaty.
• 2:35
  SPEAKER_00: I wouldn't build my Monday around it.
• 2:37
  SPEAKER_00: Are the AI layoffs real?
• 2:42
  SPEAKER_00: So AI took those jobs.
• 2:44
  SPEAKER_00: Information lost 23,000 jobs in August.
• 2:47
  SPEAKER_00: Construction added 22,000.
• 2:50
  SPEAKER_00: One month is definitely not a trend.
• 2:55
  SPEAKER_00: Maybe, or they needed a cut anyway, or the people got cut to pay for the AI.
• 3:00
  SPEAKER_00: Only one of the three shows up in the memo.
• 3:03
  SPEAKER_00: Which model should a six-person shop buy?
• 3:07
  SPEAKER_00: None yet.
• 3:08
  SPEAKER_00: You don't have a model problem, you have a leak, and nobody's watching the inbox on a Saturday.
• 3:14
  SPEAKER_00: When does an agent make sense?
• 3:17
  SPEAKER_00: Where a real judgment call happens, then supervise it for 30 days because 3.4% is what a new hire actually looks like.
• 3:33
  SPEAKER_00: One move this week.
• 3:34
  SPEAKER_00: This is a product feature, not an assessment.
• 3:41
  SPEAKER_00: A district with 370,000 students just did it.
• 3:45
  SPEAKER_00: You have