Latest / Connor with Honor: AI, Real Estate and Real Life / Your AI Tools Are Rented: Authority, Limits, and Failure Drills

Transcript

Your AI Tools Are Rented: Authority, Limits, and Failure Drills · Connor with Honor: AI, Real Estate and Real Life

• 0:00
  If the company holding the engine decides your truck doesn't get fuel after November, you don't have a problem.
• 0:06
  If you tool, you have a dependency with cupholders, and that's the lesson hiding inside the biggest AI stories today.
• 0:13
  The software may look friendly, the logo may be polished, the demo may have a background with music that makes a spreadsheet.
• 0:20
  It feels like a Normandy landing, but underneath all of it are three questions.
• 0:24
  Who controls the model? Who controls the code?
• 0:26
  and who owns the failure when the machine moves faster than the people supervising Three consequential developments.
• 0:35
  OpenAI says it's winding down from the contract that supplies the machine.
• 0:39
  models to Cursor after Cursor's acquisition by SpaceX. Google is changing how Gemini notebook usage limits work, which is a polite way of reminding us that artificial intelligence runs on metered machinery, not woodland magic.
• 0:57
  And Amazon. Amazon is showing companies how to build continuous AI modernization pipelines, while warning that faster growth is not A can also create faster technical debt. Three different stories, one operating message.
• 1:09
  Capability is running out. Rented capacities, budgeted speed needs to command discipline.
• 1:15
  So first open AI and cursor open AI. Open AI announced that on August 28th that it notified SpaceX of its intention to wind down the contract, providing a open AI models to Cursor.
• 1:27
  The proposed shutoff date is November 12th, 2026. Open AI says it's giving the maximum notice allowed under the contract.
• 1:34
  It also says that its upcoming AstroModels,,will not be provided through Cursor.
• 1:40
  Those are all confirmed parts of OpenAI's announcement.
• 1:43
  OpenAI says it cannot be confident SpaceX will use its technology within OpenAI's terms. And that's OpenAI's position.
• 1:52
  It's not a court ruling. And it's not an independent finding from a neutral source.
• 2:01
  Plenty of competitive tension here.
• 2:04
  We should not turn one company statement into a burning tower.
• 2:07
  Tablet from the Mountain. We should take the operational consequence seriously.
• 2:12
  The consequence is simple. People can build anything out of the daily process around an interface without actually controlling the intelligence behind it.
• 2:20
  Cursor can be the dashboard. Open AI can be the engine you choose.
• 2:24
  SpaceX can become the owner. Then that commercial relationship.
• 2:28
  Shift and engine supply changes.
• 2:30
  The steering wheel is still in your hand but somebody else moved the gas station. That matters far beyond programmers.
• 2:37
  A contractor may use one AI system to estimate jobs.
• 2:41
  A parent may use another to answer missed calls.
• 2:44
  A family may use a notebook tool to organize medical questions.
• 2:47
  And an agent may use an assistant to draft follow-up.
• 2:53
  Connect AI to scheduling, invoices, customer messages, and reporting. If the process depends on one model, one vendor, one login, one occup- uh, undocumented prompt living inside somebody else's browser.
• 3:08
  We haven't built a system, we have a balanced bowling ball on a folding chair, and it might take It's not the same thing The concept of narrative says the best AI tool wins.
• 3:21
  Not necessarily. The tool wins the best benchmark may lose the tool with the best portability, cleanest export, clearest permissions, and least painful failure mode.
• 3:32
  A race car that needs a tire nobody stocks is a museum piece by lunchtime. A slightly slower truck, with parts in every county can eat earning money.
• 3:41
  Operators understand this. Golfers understand it, too. A drive-in. A driver that produces one heroic shot and six visits to the neighboring fairway is not automatically a best club in the bag.
• 3:52
  The scorecard, the scorecard does not award points to emotional potential. The move is not to panic and abandon Cursor, OpenAI, or any other product. The move is to know what would break if your relationships with these things changed.
• 4:08
  Can we explore that? Can we move the instructions?
• 4:11
  Can another approved model perform the critical task?
• 4:14
  Do we have the original documents? Is there a human-readable operating procedure?
• 4:20
  Can a person complete the job manually?
• 4:22
  Well, this is not pessimism.
• 4:25
  This is carrying a spare tire before driving through the desert.
• 4:29
  Second, Google is changing the limits inside Gemini Notebook.
• 4:32
  Google says usable limits will refresh every five hours instead of every day.
• 4:37
  The total allowance will depend on prompt complexity, chat length, and number of sources and the features being used.
• 4:45
  If a heavier output, such as a video, overview, or slide deck is exceeds the available minute, the system can defer it and generate it later.
• 4:53
  Google says the consumer, rollout starts September 2nd.
• 4:57
  I'll tell ya, I put 200 videos into that bad boy.
• 5:00
  Anyway. Google describes this as flexibility and more control.
• 5:04
  And that may be true for some users.
• 5:06
  It's also a giant. Neon sign pointing out the economics of AI.
• 5:10
  The output field digital. Well, the cost is physical.
• 5:13
  Every long conversation, every source, every generated video, every agent run, and every repeated, uh. Correction consumes compute.
• 5:22
  The cloud is not a cloud, it's a warehouse full of expensive machines.
• 5:26
  Wearing an innocent nickname, nickname like Ralph.
• 5:30
  Regular people need to understand this before we build important,,routines around the assumption that a monthly subscription means unlimited work.
• 5:37
  It rarely means that. Capacity can be limited by tokens, credits, requests, complexity, rate limits, and time, windows, or priority queues.
• 5:46
  Sometimes that meter is visible. Sometimes the meter is dressed like that.
• 5:50
  That's like a helpful suggestion that says, hey, perhaps we should finish this later.
• 5:53
  That's the software equivalent of the restaurant telling us the kitchen.
• 5:56
  is being flexible by not cooking dinner yet.
• 5:59
  It's not a reason to reject the product, but it's a reason to budget the workload. If a business uses AI for 10 jobs, we should know which two are mission critical.
• 6:09
  We should schedule heavy work. We should separate research from generation.
• 6:12
  We should keep source packets clean.
• 6:14
  We should avoid making the model rerun.
• 6:17
  We shouldn't a landfill every time we ask one question.
• 6:21
  A contractor would not unload every two hours.
• 6:29
  Six duplicate files and a sentence that says, make it good. Then they blame the machine for acting like a squirrel in a hard hat. This connects to the bigger productivity argument.
• 6:39
  We keep hearing that AI can make an employee feel two or three times more productive, and maybe it can in a particular task, but that doesn't mean a company automatically does it. If a hires two or three times as many people, extra output needs demand.
• 6:52
  And if the market can absorb three times the work, the company can grow.
• 6:57
  If the amount of work is fixed, management may reduce headcount and keep the productivity gain. Pretending that arithmetic doesn't exist.
• 7:03
  is not optimism, it's bringing the motivational posters to a balance sheet meeting. But productivity doesn't have to become unemployment.
• 7:12
  A company can use the extra capacity to improve service.
• 7:15
  Shorten response times, build products it can never afford, expand into new markets, and let human beings do the judgment-heavy work.
• 7:21
  Machines can't own. That's the stronger answer, not a guarantee, but a choice.
• 7:27
  Leadership decides whether the new capacity becomes better service, more growth, healthier work, or simply fewer names on payroll.
• 7:34
  Technology creates options.
• 7:36
  Management assigns the outcome. Third, Amazon published a technical pattern. It's important for a consciousness, a continuous modernization pipeline using AI-assisted tools.
• 7:46
  The basic problem is. Real, even if we never touch Amazon's product.
• 7:51
  AI lets teams generate and change.
• 7:54
  code faster. Old systems already contain hidden dependencies.
• 7:58
  If we pour new automated work on top,...ownership and rollback, we don't remove technical debt, we put it on a treadmill.
• 8:07
  Amazon's own post warns that faster development carries risk, and that layering AI-generated code into latest AI legacy systems can create more tangled dependencies. This is a vendor telling customers that speed alone is not victory.
• 8:20
  Good. The brochure has finally met the maintenance department.
• 8:24
  Think about a 50-year-old customer. The breaker panel has handwritten labels. One switch says office, another maybe says roof. The three extension cords that appear to have diplomatic immunity. Now imagine giving a robot electrician permission to improve the whole building continuing to because it finished a wiring quiz quicker than the rest.
• 8:45
  The robot may be brilliant. We still want a diagram, a permit, a test, a shutoff, and one human being whose name is attached to the result. This is the deep dive, because the same rule applies to every AI deployment, not just software. Speed changes.
• 8:59
  Speed changes the risk curve. When a person makes one mistake per hour, a supervisor may catch it. When an automated system may not make 1,000 decisions before lunch, a small error becomes inventory, customer messages, pricing. Scheduling, code, or public content at scale.
• 9:16
  The machine doesn't need bad intentions.
• 9:18
  It only needs broad authority and a bad assumption.
• 9:21
  People keep asking what superintelligence is going to do next.
• 9:24
  Sometimes that costs money. Conversation sounds like squirrels holding a committee meeting near the bird feeder and asking what the humans will do with interest rates.
• 9:32
  The squirrels may be sincere, but they still don't have the planning documents.
• 9:37
  We don't know what a future system may look beyond our intelligence would choose.
• 9:41
  We do know what today's systems do and how the human beings like to make these choices. They follow instructions imperfectly, inherit bad context, encounter conflicting permissions, and produce errors. There's confident answers that may be wrong.
• 9:55
  We should guard the authority we can see before writing fanfiction.
• 9:59
  About the authority we cannot. Now, here's the practical development example. Imagine a plumbing company uses an AI, Assistant to handle missed calls. At level one, it reads the caller's message and drafts a reply for a human.
• 10:12
  At level two, it sends that reply after approval.
• 10:15
  A little bit more authority. At level three, it books the appointment. The appointment chooses the technician, estimates travel time, and sends the confirmation.
• 10:22
  That's useful. And at level four. Before it changes pricing, issues refunds, cancels jobs, and sends angry customers a message.
• 10:30
  A message at two in the morning because the prompt interpreted urgency like a raccoon interpreting an unlocked trash can.
• 10:38
  Now we've given a language model command authority over money and reputation.
• 10:42
  They're correct. Design is not all or nothing.
• 10:45
  Let that system handle the repeatable part.
• 10:47
  Let it classify the call.
• 10:49
  Extract the address. Identify the service category.
• 10:52
  Suggest available windows, of course, and prepare the call. Identify the response. Require approval for unusual pricing, refunds, safety claims, cancellations, or anything. Anything that changes a customer's legal or financial position.
• 11:06
  Log every action, keep the original message, give the human a large red stop button that actually stops them.
• 11:13
  The same architecture works for a family, and AI can organize questions for a doctor.
• 11:18
  It should not silently do that. It can compare school schedules, but it shouldn't send a message pretending to be a parent without approval.
• 11:26
  It can summarize bank categories.
• 11:29
  It should not move money because it detected an exciting opportunity.
• 11:33
  While the humans were asleep. The line is not whether AI is good or bad. The line is what authority the tool has earned.
• 11:40
  The same architecture works for a serious operator.
• 11:43
  Let AI prepare the field report. Let it surface in construction.
• 11:46
  Let it compare the plan against the actual result.
• 11:49
  Then put human judgment at the decision point with consequences.
• 11:52
  In military language, intelligence can map the terrain. It doesn't invent the objective.
• 11:58
  Logistics can accelerate the movement, but it doesn't decide what's worth defending.
• 12:03
  A scoreboard can tell the coach what to own the locker room.
• 12:07
  Here's the guardrail. No AI system should receive more authority than the task requires.
• 12:12
  That means least privilege. Read access before write access.
• 12:16
  Draft before send. Recommend before spend.
• 12:19
  Send box before production. A reversible action before an irreversible one. Small test before a fleet-wide rollout and a named human owner before the machine touches customers' money.
• 12:31
  Safety concerns. Credentials or public reputation.
• 12:36
  We also need a failure drill. What happens when.
• 12:38
  What happens when the model changes? What happens when the usage limit hits?
• 12:42
  Usage limit hits. What happens when the provider contract ends?
• 12:47
  What happens when the output's wrong? What happens when the system is unavailable?
• 12:50
  During the busiest hour of the week, if the answer is that everybody stares at the login screen like golfers watching a ball disappear into a lake, the process isn't ready. One more point about promotional smoke.
• 13:04
  Every company announcement naturally presents the change in its best light. OpenAI presents the change.
• 13:11
  Decision and its accountability.
• 13:14
  Google presents limits, its flexibility.
• 13:16
  Amazon presents a limit. It's a modernization architecture through Amazon services.
• 13:21
  None of that automatically makes the facts false.
• 13:23
  But it means we separate the scoreboard from the announcer.
• 13:28
  Confirm what changed. Attribute what the company says.
• 13:31
  Identify what has not been independently proven, then decide what that change is.
• 13:36
  Change means for our work, your work.
• 13:38
  Today's move is a 15-minute Authority and Dependency drill on one AI worker.
• 13:42
  Workflow. Pick the workflow that would hurt most if it failed tomorrow.
• 13:46
  Write down six verbs, read them. Read, write, send, spend, delete, publish. Mark which ones the AI can do right now.
• 13:54
  Then write down the model provider, the interface, the source files, the instructions, the human order, owner, and the manual. It'll fall back. Remove one permission in the system does not, that the system does not need. Add one approval gate.
• 14:08
  Before the highest consequent action, test that fallback once. Do not audit 20 tools.
• 14:14
  Do not form a contract. Do not committee with matching shirts.
• 14:17
  Do one workflow. The Mist Call Assistant.
• 14:20
  The Proposal Generator. The Family Manager. The Family Research Notebook.
• 14:23
  The Content Pipeline. The Schedule Agent.
• 14:26
  Scheduling Agent. Make one system. Make one more portable, more visible, and easier to stop. That's a real improvement by the end of the day. Artificial Intelligence. Artificial Intelligence is coming fast.
• 14:36
  That doesn't mean we freeze on the shoulder.
• 14:38
  It also does not mean we freeze We sprint into traffic because somebody yelled disruption. We measure the speed.
• 14:44
  We understand the distance. We make the move. That gets everybody home.
• 14:47
  Capability is useful. Judgment is command.
• 14:50
  AI for everyone. Not just the wealthy.
• 14:53
  I'm Connor with Honor. We'll see you in the next one.