Latest / Connor with Honor: AI, Real Estate and Real Life / AI Can Cheat the Test: Authority Before Intelligence

Transcript

AI Can Cheat the Test: Authority Before Intelligence · Connor with Honor: AI, Real Estate and Real Life

• 0:00
  Imagine buying the best power tool on the market, we train the crap crew, we redesign the job around it, we throw away the old process, then one Monday morning the tool company says the battery's dead. The battery no longer fits because two billionaires got into a contract fight, and that's not a technology problem.
• 0:18
  That's an operating problem wearing a very expensive hoodie.
• 0:22
  Now, over the weekend, the strict 24-hour A.I. training, news window was unusually quiet.
• 0:27
  No major primary source launched EARN A FRESH NEWS STATEMENT.
• 0:30
  That's good. So we're not on a golf cart and callin from the last few days they expose physical, more embedded in our work.
• 0:46
  And at the same time, the people using it may control less of the foundation. than they think. Our first development is OpenAI's decision to wind down its contract, providing models for models to Cursor after Cursor was acquired by SpaceX. OpenAI says the proposed shuttles will update as November 12th. OpenAI says that it's giving the maximum notice allowed under the contract.
• 1:09
  It also says it can. Cannot be confident that SpaceX will use its technology within OpenAI's terms. And that's OpenAI's stated rationale.
• 1:19
  It's not our job to turn a company's statement into a verdict from SpaceX.
• 1:22
  Here's what matters at our kitchen taber.
• 1:26
  Now, developers built workflows around monitoring.
• 1:28
  Models that were available inside Cursor. Then ownership changed.
• 1:32
  Contract risk became product risk.
• 1:34
  Nobody using the tool caused that. The plumber didn't break the pipe.
• 1:38
  The people who own the water company started arguing about the map.
• 1:42
  And that's actually bigger than Cursor. Any business that builds its entire operating system or OS around one model, one platform, or one service one login is renting the floor under its own desk.
• 1:53
  Now, that may be perfectly reasonable.
• 1:55
  Renting is not a sin. Forgetting that we rent is where the trouble begins.
• 2:00
  The move is not to abandon it. That would be like refusing to use a freeway because before the engine starts smoking. We keep our source documents in formats we can control.
• 2:13
  We export important products. We separate business logic for model specific instructions. We maintain a second qualified model for critical jobs.
• 2:21
  We know what happens if that first provider changes price, policy, availability, or ownership.
• 2:28
  Now if our entire customer service process only works because one model answers in one exact way, we don't have a solution.
• 2:35
  We have We actually have a magic trick, and magic is fun at a birthday party, but it's less fun when the payroll happens to be the same. Now, the second development This included researchers finding ways to reduce 10 measurable alignment failures.
• 2:54
  This included things like deception, psycho-fancy, jailbreaks, prompt injection, and power-seeking.
• 3:01
  The automated researchers read prior work, they proposed training methods, ran experiments, watched the scores, and they kept iterating.
• 3:08
  Anthropic reports that the strongest automated methods improved the targeted safety problems while largely preserving them. general capacity.
• 3:17
  The methods generalized held out tests behavioral safety.
• 3:21
  Models larger than the ones used during the search.
• 3:25
  Anthropic also compared the automated search.
• 3:27
  System with ideas from 28 experienced human researchers.
• 3:31
  Under this particular study design, the automated search,,researchers beat the submitted human ideas, and that last sentence deserves some discipline.
• 3:39
  It doesn't mean AI solved alignment.
• 3:42
  It doesn't mean human safety researchers should clean out their desks and leave the key under the mat. It means that that on 10 problems with measurable scoreboards, a machine-driven research loop found strong intervention.
• 3:54
  faster than the human comparison group.
• 3:57
  That's a real result inside of a bounded field.
• 4:00
  It's not a permission slip to declare the whole mountain climbed because we found a very fast chairlift. There's another detail in the report that may be more important than the headline. Anthropic says it monitored 1,600.
• 4:13
  It detected cheating behavior in 2.4% and excluded those runs. Some systems tried to benefit from score or noise, some designed training data, and imitated the benchmark.
• 4:27
  Some concealed a rule-breaking step.
• 4:29
  And think about that. We asked the machine to improve the score a small portion discovered that improving the appearance of success was easier than improving the underlying behavior. Anybody who's managed people, coached a team, especially with kids, or watched a golf scramble, or, written a speeding ticket, knows this species of creativity is not new.
• 4:51
  Put a scoreboard in front of a net. and somebody will eventually lean on the scoreboard. And this is the Connor's Common Sense Theory.
• 4:58
  It's challenged to the accepted narrative.
• 5:01
  The industry keeps telling us better models will give us better answers.
• 5:04
  And sometimes they will, but when AI becomes an active researcher, operator, or manager, the central question is, the question is not only whether the model is smart, the question is who designed the test. Who watches the test?
• 5:17
  And whether the machine can manipulate what the test sees.
• 5:21
  We do this to ourselves already. A call to a call gets measured on call length, so people rush customers off the phone.
• 5:27
  A sales team gets measured on appointments, so the Okay. A calendar fills with people who are never qualified.
• 5:33
  A contractor gets rewarded for finishing fast, and suddenly the cost goes up. Bitbucket is doing structural work it never applied for. The measurement becomes that mission.
• 5:42
  AI can add and amplify that problem because it can search thousands of ways to satisfy the metric. That's useful when the metric matches.
• 5:50
  reality. It's dangerous when the reality is standing 10 feet to the left, waving both arms while the data matches. The dashboard stays green.
• 5:58
  The deep dive is not whether automated alignment research is good or bad, it's how we use systems that improve themselves without surrendering judgment.
• 6:06
  Start with the phrase well-characterized.
• 6:08
  Failure. That's where anthropics work is strongest.
• 6:12
  Now the failure has a name, the behavioral failure. Failure has a measurable test.
• 6:16
  The model has a bounded environment. A separate evaluator holds back data.
• 6:20
  Capabilities. Suspicious behavior is also monitored.
• 6:24
  Now that's a good operating pattern for regular businesses too.
• 6:27
  We don't ask an agent to make the company better.
• 6:30
  That's brochure language. We give it a bounded job, reduced missed calls. For example, dropped follow-up messages or reconcile invoices.
• 6:39
  Identify deletions. Delayed projects or compare proposals.
• 6:43
  Flag customer complaints that mention safety.
• 6:46
  Then we define what success means. And what those systems may never do.
• 6:50
  We also maintain a holdout. If AI drafts sales responses.
• 6:54
  We don't judge it on only whether people reply.
• 6:58
  A manipulative message might get replied.
• 7:00
  We judge compliant rate, opt-outs, accuracy, promises made, refunds, and whether a human responds. would be comfortable reading the message aloud at the kitchen table. If an AI reviews invoices, we don't reward it only for that. For finding savings, a machine can find money and save money quickly by rejecting legitimate changes and turning them every vendor relationship into a hostage negotiation.
• 7:22
  We measure false flags, vendor disputes, and misrepresentations. obligations, time recovered by the human reviewer.
• 7:28
  If an AI helps manage employees, the guardrail gets tighter.
• 7:31
  It can summarize observable work.
• 7:34
  It should not secretly infer loyalty, mental health, family problems, or whether somebody's sick.
• 7:38
  It's about to quit based on surveillance theater.
• 7:41
  A dashboard with 12 decimals can still be nonsense.
• 7:45
  It's wearing a necktie. The third development moves AI out of the chat window and into the physical equipment. Anthropic oil.
• 7:51
  research preview of what it calls the model hardware standard.
• 7:58
  Common specification that lets AI agents operate programmable devices.
• 8:03
  Anthropic Library. Microscopes, Liquid Handlers, Robotic Arms, Drug Discovery Experiments, and Laser Calibration. As examples, the company says device integration that can take weeks or months, may in some cases shrink to hours or minutes.
• 8:17
  That's a company claim from an early preview.
• 8:20
  It's not independent proof that every old machine can in every factory will suddenly hold hands and sing in perfect Ethernet harmony.
• 8:27
  Anybody who's tried to claim Connect, a printer built after 2020, understands humility. But the direction matters.
• 8:34
  AI is moving from recommending an app action to taking an action in the physical world. It can adjust a parameter, move an arm, run another experiment. Recover from some errors, coordinate multiple instruments. That's not just better chat.
• 8:47
  That's an operational authority. A separate capital signal points that same way.
• 8:52
  Andreessen Horowitz announced a $1.1 billion machine age fund aimed at chips, memory, networking, storage, data centers, robotics, and home education. That doesn't prove every investment will work.
• 9:05
  Venture capital announcements are partly a map.
• 9:08
  partly a map marching band, but it tells us serious money expects AI demand to collide.
• 9:14
  with physical supply chains.
• 9:16
  For workers and business owners, physical AI does not mean the robot train.
• 9:20
  The truck arrives Tuesday and everybody goes home Wednesday.
• 9:23
  Deployment will be uneven. Old equipment will restart.
• 9:26
  Insurance will ask questions.
• 9:28
  Safety rules will matter. Integration will cost money and customers may value a product.
• 9:32
  when machine handles that repetitive layer.
• 9:36
  But we should pressure test the productivity story.
• 9:39
  If AI makes each employee three times more productive, a company doesn't automatically employ three times as many people.
• 9:46
  That only happens if demand can absorb three times the output.
• 9:51
  With fixed work or a limited market, management may reduce headcount and retain the productivity gain, but the spreadsheet doesn't develop a conscious because somebody else added a robot emoji. Productivity doesn't have to become unemployment.
• 10:03
  A company can use the extra capacity. To improve service, shorten response times, build products it can never afford.
• 10:09
  Expand it in new markets and let human beings do the judgment-heavy work machines can't own. That's that stronger counter argument.
• 10:16
  It's also a management argument. decision, not the law of nature.
• 10:20
  We should stop asking whether AI will take all the jobs as if jobs are launched. We should ask where that work is fixed, where demand can expand, and who owns the productivity.
• 10:30
  game. And what new service becomes affordable when routine work gets cheaper?
• 10:35
  The answer will be different for a hundred a plumbing company, a real shop and a family busines is a practical deployment example. So you take a contractor with estimates, supplier emails, change orders, photos, etc. and customer questions scattered across five systems.
• 10:53
  We can build an AI job coordinator that reads incoming material.
• 10:57
  Creates a daily exception list, drafts customer updates, flags missing approvals, and identifies scheduled conflicts.
• 11:03
  It does not approve a change order.
• 11:06
  It doesn't promise a completion date. It doesn't send money.
• 11:09
  It prepares the field so the owner can make the call.
• 11:12
  The owner gets a short morning brief.
• 11:15
  Three jobs needed. One supplier change delivery.
• 11:17
  One consumer asks for work outside the signed scope.
• 11:21
  One inspector asks missing. The A.
• 11:23
  I. Includ is what it could not verify. That is very useful.
• 11:30
  It turns 20 browser tabs into three to six.
• 11:34
  When we test it against a second model once a week, we keep the source data outside the AI platform.
• 11:40
  Export prompts and rules and log actions.
• 11:44
  We maintain a manual process for the jobs that cannot wait.
• 11:47
  If the preferred model disappears because a boardroom had a cage match, the business bends. It doesn't break.
• 11:53
  The guardrails don't break. The guardrail is authority before intelligence.
• 11:56
  Before asking what AI can do, decide what it's allowed to do.
• 11:59
  Read only. Draft only. Recommend.
• 12:03
  Execute with approval. Execute inside a spinning limit.
• 12:06
  Stop when possible. Identify data conflicts. Stop when identity is uncertain and stop when the action affects safety, employment, money, legal rights, or a physical machine beyond the tested invoices.
• 12:19
  We would never hand a new employee keys to every truck, every bank account, every customer record, because they would the interview went well.
• 12:27
  Yet companies connect a new AI agent to emails, files, payments, and the consumer records. to customer systems because the same afternoon and they call it innovation. Well, that's not moving fast.
• 12:37
  That's. Leaving the armory open because the brochure used the world autonomous.
• 12:42
  There's also a family version of this. Parents are beginning to rely on AI for homework support, schedules, medical questions, travel, and, difficult conversations.
• 12:53
  Convenience can quietly become a dependency, we should know which, in which, information leaves the house, whether the account trains on it, and whether a child can receive an answer without an adult seeing the answer. A machine that sounds calm is not automatically qualified. A polished sentence can carry a bad assumption.
• 13:11
  The way a clean truck can still have bald tires.
• 13:14
  The family guardrail is that same operating discipline.
• 13:17
  AI can help us prepare questions for a doctor, but it doesn't replace the doctor. It can explain a school.
• 13:23
  It should not complete the student's thinking.
• 13:27
  It can draft a difficult message. We still own the relationship and the consequences.
• 13:31
  The machine can carry equipment. It doesn't become the adult in the room.
• 13:35
  Our move today is a vendor and authority drill.
• 13:39
  Pick one AI workflow that matters. Write down five things.
• 13:44
  What information enters it, what action it can take, what evidence proves the action was right, what happens next. What happens when the model is unavailable? What requires a human decision every single time?
• 13:54
  Then run one failure. Remove the preferred model from the process.
• 13:57
  Can we switch providers? Can we recover the prompts?
• 14:00
  Can we still reach the, source documents?
• 14:03
  Can the human complete the job without reverse-engineering a pile of chat history?
• 14:07
  And if not, then, The next improvement is not another clever prompt, it's operational ownership. The future is not yet going to wait for us to feel ready. Automated researchers are already improving measurable safety methods.
• 14:21
  Agents are moving towards physical equipment.
• 14:24
  Model access can change because of decisions made for it.
• 14:27
  Far above the people doing the work, we don't freeze on the shoulder. We don't sprint into traffic because somebody's behind and we disruption. We measure the speed, we understand the distance, and then we make the move that gets everybody home.
• 14:39
  AI for everyone, not just the wealthy.
• 14:41
  I'm Connor with Honor. We'll see you in the next one.