Is AI Safer, or Just Misleading? A Fast Q&A
No slides tonight, just questions fired straight at me and straight answers back. Twelve rounds on the model that got banned by a school district the same morning its own maker called it critical risk.
What happened on September 3rd?
A new model graded critical for cyber capability the same day. LA Unified banned it for every single service.
Critical? Is the press being dramatic? So the thing is actually dangerous?
That is the manufacturer's own system card. They wrote the threshold, then met it, and said so. The company publicized that the model can find unknown security flaws on its own. They chose to print that sentence. Read it.
And the district panicked. Surely the board panicked.
While the district reviews instructional use and safeguards, that is a pause. That is not a verdict. The committee chair says it surprised the board. Make of that what you will.
Is the model safer than the last one?
Yes, genuinely. 3.4 percent malign output against 18.8 percent. More than five times better.
So 3.4 percent is fine?
Three out of every 100 real tasks go somewhere you do not want. Hire a person with that record and see how it feels.
They claim 99.7 percent on prompt injection. That sounds excellent.
21 failures in 10,000. If it reads 10,000 emails this year, that is 21 chances a stranger gives it an email it acts on. It is excellent. It also makes the world less safe than the number alone suggests. A percentage is useless without a denominator, and nobody hands you the denominator.
What's the two clocks idea?
The lab's clock is moving. Institutions move in school years. Nobody is holding both watches.
Who wins that race? What about the UN Red Lines, will they make it?
Whoever moves first, so far that has never once been the slow clock. 200-plus signatories, 10 Nobel winners, deadline of end 2026. That's 114 days from this morning. A year in, no treaty. I would not build my Monday around it.
Are the AI layoffs real? So AI took those jobs?
One month is definitely not a trend. Maybe, or the company needed to cut anyway, or the people got cut to help pay for the AI. Only one of those three ever shows up in the actual memo.
Which AI model should a six-person shop buy?
None yet. You don't have a model problem, you have a leak, and nobody's watching the inbox on a Saturday.
When does an agent actually make sense?
Where a real judgment call happens, then supervise it for 30 days, because 3.4 percent is what a brand new hire actually looks like too.
My AI says my idea is great. What's one move this week?
It says that to everybody. That is a product feature, not an assessment. Write your AI policy on one page. A district with 370,000 students just did it. You haven't.
Common questions
Is the newest AI model actually safer than the last one?
Yes, genuinely, by the company's own numbers. Malign output dropped to 3.4 percent against 18.8 percent, more than five times better. But 3.4 percent of real tasks still going somewhere you do not want is a record no human hire would survive on.
What is the two clocks problem in AI?
The lab's clock moves in weeks. Institutions move in school years. Nobody is holding both watches at once, and so far whoever moves first has never been the slow clock.
Are the AI layoffs real?
A single month of cuts is not a trend. When a company blames AI for layoffs, there are three possible truths in the memo: the software actually does the work now, the company needed to cut anyway, or the people got cut to help pay for the AI. Usually only one of the three shows up in the announcement.
Which AI model should a small business buy?
None yet, if the real problem has not been diagnosed. Most small shops do not have a model problem, they have a leak, and nobody is watching the inbox on a Saturday. Fix the leak first.
When does an AI agent actually make sense for a business?
Where a real judgment call happens, and only after the owner supervises it for 30 days, because a 3.4 percent error rate is what a brand new hire actually looks like too.
Connor T. MacIvor · CalDRE #01238257 · Sync Brokerage, Inc. · DRE #02031490