The Ignorance Boundary Problem is a critical AI safety concept where systems fail to recognize their own knowledge limitations, causing them to confidently produce incorrect answers when encountering edge cases outside their training data. This problem was first identified in Marcus Hail's 1995 MIT dissertation, which became foundational for modern AI safety research. The story illustrates how institutional bias in hiring can overlook brilliant minds who don't fit traditional career paths, and how recognizing the limits of one's knowledge is more valuable than blind confidence.
深掘り
前提条件
- データがありません。
次のステップ
- データがありません。
深掘り
The CEO Mocked a Single Dad's Degree — Then Harvard Called Her Office追加:
The elevator doors opened on the 14th floor, and Marcus Hail stepped out, carrying a leather portfolio that had seen better days. The zipper along the top edge had been repaired with a strip of electrical tape, a fix he'd made 6 months ago, meaning to replace it, and never quite finding the time or the $40.
He was 38 years old. He was wearing the charcoal suit he'd bought for his wife Rachel's funeral 8 years ago, let out once at the waist since then by a tailor on Clement Street who'd charged him $22 and hadn't mentioned what the suit was for. The jacket was clean, the shirt was pressed, and his shoes, black Oxfords, had been polished that morning at 5:30 while Lily was still asleep, and the apartment was quiet enough that he could hear the radiator clicking in the hallway. He'd taken two buses to get here. The lobby of Ross Dynamics occupied the entire 14th floor of the Bradshaw Tower in downtown San Francisco.
Glass walls offered a panoramic view of the bay on one side and the financial district on the other. The reception desk was a single slab of white marble backlit from below. Three young women sat behind it in coordinated navy blazers, and all three of them looked up when Marcus walked in. He watched their eyes move briefly and almost imperceptibly to the portfolio to the suit. Back to his face. Marcus Hail, he said. I have a 9:00 with your talent acquisition team. Of course. The woman in the center, her name plate read, di typed something with two fingers and offered a smile that did not reach her eyes. If you'll take a seat, someone will be right with you. He sat in one of the low white chairs along the far wall. The other people waiting, two men and a woman, all of them younger, all of them in suits that had clearly not been purchased for funerals, glanced at him once and then looked away. He set the portfolio on his knees and looked out at the bay. He was not nervous. That surprised him sometimes.
How not nervous he could be in rooms like this. He was tired. Tired was different. He'd worked the delivery route until 2:00 in the morning, come home, slept 4 hours, woken when Lily's alarm went off, made her breakfast, walked her to school, and taken the bus downtown. He had a thermos of coffee in his bag that he'd been nursing since 7.
He allowed himself one more sip now and recapped it. The talent acquisition office was a glasswalled conference room off the main floor. He could see through it clearly. a woman in her mid-30s named He'd later learned Sandra Cho was leading the first round. She was efficient, business-like, and not unkind.
She asked about his community college coursework, his years of gap employment, and what he was hoping to bring to a high performance technical environment.
He answered clearly and without embellishment. He was midway through explaining a systems architecture problem he'd been thinking about, one that he suspected was related to a recurring inefficiency in Ross Dynamics publicly documented API behavior, which he'd read about in a trade journal 3 weeks ago when the conference room door opened. Evelyn Ross walked in. She was 44 years old, CEO of a company that had tripled its valuation in the past 3 years, and she moved through a room the way weather moved through a valley. You felt it before you saw it. Her hair was dark, cut close at the neck. She wore a charcoal blazer over a cream blouse and carried a tablet she wasn't looking at because she was already looking at Marcus. Sandra started to stand. Miss Ross. I didn't know you were sit. Evelyn pulled out the chair at the head of the table and sat down. She reached across and picked up Marcus' resume from the folder Sandra had open. She read it for approximately 15 seconds. Community college. She said it was not a question.
Yes. Coastal Community College in Daily City. Associates degree in computer information systems.
She set the resume down and then a gap of she looked again. 8 years. That's correct. Doing what exactly? maintenance work, delivery, taking care of my daughter. Something moved across her face. Not sympathy, more like the expression a person makes when they've tasted something unexpectedly sour. Mr. Hail. She folded her hands on the table. Her voice was clear and without heat, which somehow made it worse. I appreciate that you came in today. I genuinely do. But Ross Dynamics operates at the edge of what's technically possible in applied machine learning. Our junior engineers have degrees from Stanford, from MIT, from Carnegie Melon. Our senior architects have PhDs. She slid the resume back across the table toward him. We're not the right fit for someone at your level.
Sandra was looking at the table. The two other candidates visible through the glass wall had paused their conversation. Marcus looked at Evelyn Ross for a moment without speaking. Then he reached for the resume and placed it back in his portfolio. I understand, he said. He was standing. The chair pushed back when the phone rang. Not Sandra's cell phone.
Not anyone's personal device. The desk phone on the credenza behind Evelyn the landline. The one that almost never rang because who used landlines anymore began ringing with a flat, insistent electronic tone. Evelyn frowned.
Sandra crossed the room and picked it up. Ross Dynamics talent acquisition.
This is Sandra. She stopped. Her expression changed. I'm sorry. Could you repeat that? This is Yes. Yes. He's He's actually here right now. She looked at Marcus, then at Evelyn, then back at the phone. I'll Yes. One moment, please. She held the receiver out toward Evelyn Ross like it had suddenly become very heavy.
It's the dean of academic affairs at Harvard University.
Sandra said he says he's been trying to reach a Mr. Marcus Hail. He says it's about the Whitfield laurate award. The room was very quiet. Evelyn Ross did not move for three full seconds. Then she took the phone. To understand what happened next, you have to understand what the eight years before it looked like. Marcus Hail had arrived at MIT on a full merit scholarship at 19 years old. Out of a public high school in Fresno, where his mathematics teacher, a soft-spoken woman named Mrs. Patricia Gomez, who drove 40 minutes each way to work every day, had written in his recommendation letter, "I have been teaching for 22 years. I have never had a student like Marcus. Not almost, not close, never. He finished his undergraduate degree in three years. He stayed for a PhD.
His dissertation on adaptive neural architectures for realtime environmental data processing attracted attention from three different federal research programs and a quiet but sustained interest from several technology companies who weren't yet publicly comfortable saying how seriously they took academic AI research. At 26, he met Rachel Whitmore at a conference in Boston. She was a second-year medical resident who had come to the conference because her hospital's administration was exploring machine learning applications for patient triage. She laughed at something he said during a panel Q&A.
A real laugh, surprised out of her, and he turned to find the source of it, and that had been that they married. Two years later, Rachel finished her residency.
Marcus completed a post-doal fellowship and began fielding job offers that ranged from generous to, in one case, genuinely staggering. He turned them all down. He wanted more time to think. He was working on something he couldn't quite articulate yet, a framework for handling what he privately called graceful ignorance in AI systems, the capacity to recognize and respond appropriately to the limits of one's own knowledge. and he felt with the particular stubbornness of someone who had always been right about his instincts that he was close. Rachel got pregnant. They were both delighted. They bought a small house in Oakland with a yard that had a lemon tree. And Rachel said she wanted to name the baby Lily if it was a girl because lemon trees and lily of the valley were the two things she'd always loved. And Marcus said that was a beautiful reason for a name. Lily was born in April. Rachel was happy.
Marcus was happier than he'd known he could be. 14 months later, Rachel was diagnosed with an aggressive form of glyobblasto, a brain tumor, stage 4.
Presenting with symptoms she'd been quietly attributing to exhaustion and new parent stress for longer than anyone would later be comfortable discussing.
The next 18 months were the most educational of Marcus Hail's life, and he would not have chosen them. He learned what it meant to watch someone you love become someone slightly different. a little bit every day because of what was happening in their brain. He learned what pediatric sleep regression felt like when you were also managing chemotherapy schedules and insurance appeals. He learned which hospital cafeteria was open at 3:00 in the morning and which vending machine on the oncology floor still had the good crackers at the end of the week. Rachel died on a Tuesday in November. It was raining. Lily was 14 months old and would not remember her mother, which was a fact Marcus held inside himself like a stone he carried everywhere. He could have gone back. His adviser, a cleareyed Scotsman named Professor James Dunar, who had sent him a handwritten letter after Rachel's death and never mentioned it again, left the door open formally and informally for 3 years. The research positions didn't evaporate, if anything.
Some of them became more interesting as the field moved toward exactly the problems Marcus had been thinking about.
The money was real and would have been helpful. But Lily had started waking up with nightmares around 18 months. She had a particular cry for when she wanted Marcus specifically, not a babysitter, not her grandmother, not anyone else. He knew the cry. He came when she cried. He was not sure what else to do. and he was also not sure that there was anything else he should be doing. He let the positions lapse. He moved to a smaller apartment. He took the maintenance job because the hours were predictable. And his boss, a former army logistics coordinator named Harold Beam, who asked no unnecessary questions, was willing to work around Lily's school schedule. He was not unhappy. Precisely. He was tired. Often he read papers late at night when Lily was asleep not to stay current. Not really, but because thinking about the problems felt like visiting a place he still loved, even if he no longer lived there. He helped Lily with her math homework. She was excellent at it. This pleased him more than he let on. Bo, she'd said to him once, using the Vietnamese word she'd picked up from a friend's family and applied to him with cheerful permanence.
Do you ever wish you'd been a scientist instead? I was a scientist, he said. But like a real one with a lab. I had a lab.
I had a very nice lab. It had a window.
What happened to it? You happened to it?
He tapped her notebook. Now do the next problem. She did the next problem. She got it right. She always got them right.
The dean's name was Arthur Peton, and he had a voice that sounded like it had been soaked in oak for several years. Evelyn Ross stood at the credenza holding the phone with both hands, which was not something she normally did with anything. The Witfield Laurate Award, Dean Peton said, is given every 5 years to a former researcher whose work at the time of its completion was not fully understood or appreciated by the field. It is in a sense a delayed recognition, an acknowledgement that we got it wrong the first time. He paused.
Mr. Hail's doctoral dissertation, specifically the third chapter, the architecture, he called the ignorance boundary model, has become in the past 18 months, what I think we can fairly describe as the theoretical foundation for a generation of AI safety work that this institution and several others are quite proud of. Marcus was standing near the door of the conference room. He had not asked to sit back down. He was looking at the floor. We've been attempting to locate Mr. Hail for some time. Dean Peton continued. We knew he'd left academic work, but we had an old contact address. We've now reached him through your office somewhat accidentally.
And I confess I'm curious to know why he's there. A brief pause. If he's looking for work, Miss Ross, I'd encourage you rather strongly not to repeat whatever mistake led to him leaving academia in the first place.
Evelyn Ross set the phone down. She looked at Marcus Hail, at the electrical tape on his portfolio, at the suit that was a size too big across the shoulders, at the thermos sticking up from his bag. She said nothing for a long moment. Marcus said, "I can let myself out. Sit down, Mr. Hail," he sat. She sat across from him.
and Sandra reading the room with admirable accuracy quietly picked up her folder and left closing the glass door behind her. Your dissertation, Evelyn said, "The ignorance boundary model. It was a long time ago. Half the AI safety researchers in this country are citing it. I've heard. You've heard." She studied him.
How? I still read the journals while driving delivery routes. I read fast.
She leaned forward slightly.
Why didn't you tell us when you came in?
When I she stopped, he said evenly.
Would it have changed how you looked at the portfolio? She didn't answer, which was itself an answer. It happened the way it happened, Marcus said. I've been in rooms like this before.
The resume reads the same whether I explain the gap or I don't. community college.
8 years. He picked up the thermos, seemed to think better of it, set it down again. I wanted to see if the work could speak for itself, and it can't in a two-page document. I know that. I knew that walking in. Then why come? He was quiet for a moment. Outside the glass walls, the bay was brilliant in the morning light. The water broken into a thousand small pieces by the wind.
because my daughter asked me this week what I was going to do next, he said.
And I realized I'd been asking myself the same question for about eight years and hadn't answered it yet. She spent the next 40 minutes in her office reading everything her assistant Naen could pull together about Marcus Hail in that time, which turned out to be considerable because Marcus Hail had left behind a body of work that the field had spent the better part of a decade catching up to. three published papers, all from his graduate years. The dissertation, a single conference presentation, the recording of which still circulated in AI safety circles, of a 27-year-old Marcus Hail dismantling a panel consensus with such quiet precision that the moderator had visibly struggled to reclaim control of the room. Professor Dunar at MIT had given at some point an interview to an academic publication in which he was asked about his most talented students. He had said carefully, "I try not to rank, but I can tell you that there have been two or three moments in my career when I was in the room for something genuinely original. Marcus Hail's oral defense was one of them. I do not know what happened to his career after that. I do know that whatever he chose to do with his time, the loss to the field was real. Evelyn read this three times. She was thinking about her father, which surprised her.
Her father had been a real estate attorney in Chicago. He had worked, as far as Evelyn could estimate, approximately 60 hours per week from the time she was 5 years old until the time she was 22 and left for business school, at which point she had inherited his work ethic. and what she now understood at 44 was also his loneliness.
He had provided for her excellently.
He had attended a percentage of her school events that she could calculate if she chose to with uncomfortable precision. He was proud of her. She knew this the way you knew things that had been told to you often enough to become fact rather than feeling. She had built Ross dynamics in her own image, which meant she had built it in her father's image, which meant she had built a machine for producing outcomes at the expense of everything that could not be measured. She had sat across from a man today and told him in front of witnesses that he did not belong in her world. She had looked at his suit and his portfolio, and she had rendered a verdict, and she had been wrong in a way that would have been merely embarrassing if the error had been about data. But it hadn't been about data. It had been about a person. She walked back to the conference room where Sandra had been told by Naen to keep Marcus Hail occupied. Sandra had done this by getting him coffee and asking him with genuine curiosity what the ignorance boundary model actually was. Evelyn paused outside the glass and watched him explain it. Watched his hands move.
watched Sandra's expression shift from polite interest to something more intent. Behind them, through the window, two of Ross Dynamics senior engineers were visible in the main AI development lab. They were standing at a whiteboard that was covered in notation, and neither of them looked comfortable. Naen appeared at Evelyn's elbow. The data stream integration is failing in production, Naen said. Griffin's team has been on it since 6 this morning.
They can't isolate it. Evelyn watched Marcus Hail draw something on a napkin to illustrate a point. Get Griffin, she said. The data stream integration was the loadbearing wall of Ross Dynamics newest product, an enterprise analytics platform that was supposed to launch in 11 days in front of a room full of investors who had been patient, but not infinitely so. The failure was not spectacular. It was worse than that. It was subtle. The system processed input correctly 94% of the time and produced results that were plausible but wrong the remaining 6% with no visible pattern distinguishing the failures from the successes. Tom Griffin Ross Dynamics head of AI engineering was 31 years old had a Stanford PhD and had been awake for most of the past 30 hours. He walked Evelyn through the problem in the conference room with the focused misery of someone who was both very good at his job and currently unable to do it. We've checked the training data. We've checked the validation set. We've run ablations on every component layer. The error appears and disappears without consistent correlation to input characteristics. He rubbed his eyes.
It's like the system has a blind spot we can't find. Marcus was still in the room. He had been given no invitation to stay or leave and had, with the stillness of someone accustomed to not drawing attention, simply remained. "Can I see the architecture diagram?" he said. Griffin looked at him, then at Evelyn.
She nodded almost imperceptibly. Griffin pulled up the diagram on the conference room screen. It was dense. several interconnected processing layers, a complex attention mechanism, a custom loss function that Griffin's team had developed over eight months.
Marcus looked at it for about 40 seconds. Your edge case handler in the third processing layer, he said, "What happens when the uncertainty score hits the boundary threshold?" Griffin blinked. It defaults to the baseline prediction. And the baseline prediction was calibrated on what population? A pause the training set which excludes the anomalous inputs. Another pause longer because we flag them as outliers during pre-processing, Griffin said slowly. So your system when it encounters something it's uncertain about defaults to a prediction model that was explicitly built without accounting for uncertainty. Marcus sat down the cup of coffee Sandra had gotten him. The system doesn't know what it doesn't know.
And when it encounters the boundary of its own knowledge, it reaches for the one tool that was built without that boundary in mind. The room was quiet.
That's Griffin started. Stopped. That's the ignorance boundary problem. Yes.
From your dissertation. Yes. Which I read. Griffin said. He was now looking at Marcus with an expression that was difficult to characterize. somewhere between professional admiration and the particular chagrin of realizing you have walked into a wall that someone told you was there which I read 3 years ago. The fix is not complicated. Marcus said you need a secondary calibration layer here.
He pointed at the diagram that treats high uncertainty inputs as a distinct population about 2 days of work. I'd estimate if your team is good less if they're very good. Griffin looked at the diagram. He looked at Marcus. He picked up a marker and walked to the whiteboard. Show me, he said. What happened over the next two days was by the account of everyone who was present.
The most educational experience of their professional lives, and several of them would describe it that way for years afterward. Marcus worked with Griffin's team in the main development lab. He did not take over. He did not perform. He asked questions, listened carefully, and occasionally wrote things on the whiteboard that caused whoever was closest to stop what they were doing and stare. He knew when to leave the engineers alone and when to intervene.
He explained concepts with an economy that the team, trained on the assumption that complexity required complexity to describe it, found briefly disorienting and then deeply useful. He left each evening by 5:30. He came back each morning at 8. The team learned from Sandra that he was walking to the bus stop, taking the 38 Giri to the 43 Masonic, picking up his daughter from aftercare, making dinner, and returning to whatever it was he did at night. No one asked him directly. It seemed important to him, the shape of his day, and the team, who had been working 90our weeks for months, found themselves, to their collective surprise, slightly envious. On the second afternoon, a young engineer named Brena Walsh, who had been at Ross Dynamics for eight months and had the particular exhaustion of someone who was excellent at something she was not sure she was allowed to be excellent at, showed Marcus a piece of code she'd written as an alternative approach to the calibration layer. He read it. He was quiet for a moment. This is better than what I was going to suggest, he said.
She stared at him. The recursive call structure here, he pointed. I didn't think of that. It's cleaner.
Use this. Brena Walsh went back to her desk and did not tell anyone for 2 hours what he'd said because she was quietly processing the fact that Marcus Hail Marcus Hail had read her code and called it better than his own and she needed a moment with that. On the evening of the second day, the integration passed its test suite. All of it. Griffin ran it three times. He found Marcus at the coffee station on the 14th floor, Thermos in hand, about to leave. It works, Griffin said. Marcus nodded. Your team did the work. You showed us where to look. You would have found it. Marcus capped the thermos. It might have taken a week. You had 11 days. The math worked out. Can I ask you something? Sure. You knew what was wrong in 40 seconds.
You've been teaching us for 2 days. You have a PhD from MIT, a Whitfield laurate, and you've been driving delivery trucks." Griffin chose his next words carefully. "What was it like leaving all this behind?" Marcus was quiet for a moment. Through the window, the bay had gone gray blue in the early evening, the city lights beginning to come on. "My daughter can do long division in her head," he said.
"She's 10. She doesn't think it's a big deal. I watch her do it and I think he stopped, started again. I think about all the things I would have missed. He said good night and took the elevator down. Evelyn made the offer in her office on Friday morning. She had prepared for it the way she prepared for important negotiations thoroughly with contingencies with numbers she was prepared to move on and numbers she was not. She offered him the CTO position.
She named the salary.
She named the equity stake. She put both numbers on paper, pushed the paper across the desk, and watched him read it. He read it. He set it down. No, he said, "Which part? The title? The structure? What it would require?" He looked at her evenly.
I'm not saying the numbers aren't generous. They are. But a CTO of a company this size, I know what that job looks like. I've watched people do it.
The equity means I'm a stakeholder in the outcome, which means I'm in when things go wrong and things always go wrong. That's a good thing for you and a complicated thing for me. He paused. I have a daughter. I know she's not a footnote to this conversation. I know that, too. Evelyn picked up her pen and set it down again.
I've been thinking about what Dean Peton said, which part. Don't repeat whatever mistake led to him leaving academia in the first place. She looked at him. I've been thinking about what that mistake was, not yours. I think I understand yours. I mean, the mistake the field made. The institutions, Marcus waited.
They built systems that rewarded full-time commitment, Evelyn said.
continuous output, uninterrupted presence. And when those requirements conflicted with the rest of your life, the system didn't adapt, you adapted.
And the cost of your adaptation was, she gestured slightly at the room. This you doing maintenance work for 8 years while the field caught up to your dissertation. I don't think of it as a cost, Marcus said. I know that's I find that somewhat challenging to fully believe. You should believe it. She looked at him for a moment. I grew up with a father who was very successful, she said. He worked constantly. I admired that. I built this company on it. She paused. I don't think he knows me very well. I don't think I know him very well. We have an excellent professional relationship. She said this without apparent self-pity.
as a simple description of facts. I am not going to suggest that you're wrong about your choices. I'm saying that I'm sitting here having spent 40 years optimizing for one set of outcomes and I am genuinely uncertain.
For the first time whether they're the right ones, Marcus said, "What are you actually offering?" She picked up the paper. She turned it over and wrote something on the blank side. She pushed it back. He read it. It said consulting your terms 10 hours a week maximum remote projectbased we come to you he looked up the compensation is on the other side Evelyn said it's less than the CTO offer it's still more than you're making now you don't know what I'm making no but I know what maintenance workers make in San Francisco and I know what delivery routes Hey, and I know what reasonable math says when I combine them. She folded her hands. I'm not trying to buy you. I'm trying to make you an offer that doesn't require you to make a choice you've already made. He was quiet for a long time. Long enough that it was not a thinking pause, but something more deliberate. The silence of someone who was paying attention to what he actually wanted, which is harder than it sounds.
I'll think about it, he said. Of course, I'll talk to Lily. Something crossed Evelyn's face that she did not fully control. You discuss work decisions with your 10-year-old. I discuss everything with my 10-year-old.
She has opinions. He picked up his jacket from the back of the chair. She's usually right. Lily Hail arrived at Ross Dynamics the following Tuesday at 4:15 in the afternoon. She was 10 years old, 5' nothing, and wearing a blue windbreaker with a small hole at the left elbow. She carried a canvas tote bag with a ceramic container inside it.
And she had taken the bus downtown by herself because she was, as she explained to the security guard at the lobby desk, very responsible. And also her dad had texted her the address and told her he'd meet her in the lobby at 4:30, but she'd taken an earlier bus because she was excited. The security guard, a former middle school PE teacher named Carl, who had seen everything, walked her up himself. She walked into the 14th floor and immediately went to the window because the view of the bay was, in her assessment, extremely good.
"Is that Alcatraz?" she said to no one in particular. "Yes," said Naen, who had materialized from somewhere. "Cool," she spotted Marcus across the floor and waved with her whole arm. He crossed the room to her. She handed him the tote bag. It's rice and the chicken you like.
I made the sauce myself. She looked around the office with frank curiosity.
This is where you fixed the computer thing. This is the place. It looks expensive. It is. H. She made the particular hm. That meant she was filing this information. The lady who runs it, is she nice? She's getting nicer. Lily appeared to consider this sufficient.
She pulled the container out of the tote bag, looked around for somewhere to set it, and accepted Naen's offered assistance with the ease of someone who had never been shy. Evelyn Ross came out of her office and stood at the edge of the main floor. She was watching without, she hoped, being obvious about it. Lily had found her way to the development lab where Griffin's team was working. She was looking at the whiteboard with the focused expression Marcus used when he was interested in something. Brena Walsh, who was closest, said, "Can I help you? My dad wrote some of this," Lily said. She pointed at a section of the board. "He showed me this part last night." "It's about what happens when a system doesn't know what it doesn't know." Brena blinked. "He showed you. He always explains what he's working on." "He says it helps to explain it to someone who will ask bad questions."
She looked at Brena. I asked if the system was like me when I didn't know how to ride a bike and fell over anyway because I thought I knew how. A and what did he say? Brena asked. He said that was exactly it. Lily looked at the board again. I thought it was obvious, but he said most systems don't know to fall over. They just keep going wrong. Brena looked at the board. She looked at Marcus, who was watching from across the room with an expression that was not pride exactly.
It was more humble than that, more surprised, and she thought about the thing he'd said about her code and about what it meant to teach someone who asked bad questions and the unexpected value of that. Evelyn Ross, watching from the periphery, was quiet. She was thinking about what she had built and what it had cost and what she had told herself the cost was worth. She was thinking about a man who had walked into her office 3 weeks ago with an electrical taped portfolio and a thermos of coffee and whom she had looked at and found wanting and what that said about the machinery of her judgment and whether it was possible to recalibrate it. She crossed the room. Lily, she said. Lily looked up with the direct assessing gaze of a child who had not yet learned to be polite instead of honest. You're the boss. I am. You were mean to my dad.
Evelyn absorbed this. I was. Lily considered. He says people are usually mean because they're scared of something. That's probably true. What were you scared of? Evelyn was aware of the total silence that had fallen over the development lab. She was aware of Griffin, of Brena, of Sandra, who had appeared in the doorway. She was aware of Marcus, who had gone very still. I think, Evelyn said slowly.
I was scared of being wrong about something I'd believed for a very long time. Lily absorbed this with the gravity of someone taking a seriously complicated situation seriously.
"Okay," she said. "That's a good reason." She picked up her container of rice and chicken. You should try this. I made it. Evelyn Ross, CEO of Ross Dynamics, 34th on Forb's list of most powerful women in technology, sat down in a chair next to a 10-year-old and ate rice with a plastic fork and did not say anything for several minutes. The changes at Ross dynamics happened over the following 6 months. And they happened in the way real changes happen slowly and then faster with resistance and then accommodation and with the particular mixture of idealism and pragmatism that characterizes any institution trying to correct a mistake without fully admitting the shape of the mistake. The hiring policy was the first thing.
Evelyn convened a working group. She insisted it include Brena Walsh who had opinions and was not afraid to express them. And over three weeks they rewrote the requirements for every technical role in the company. Community college counted, boot camps counted, self-taught counted with a demonstrated project portfolio. Career gaps were noted and then set aside.
The question asked instead was what the applicant had done, not when and under whose institutional opaces they had done it. The flexible scheduling policy was harder.
It ran into the genuine operational requirements of a company that ran live systems and the working group spent considerable time distinguishing between the flexibility that was actually possible and the flexibility that was desirable in theory but would mean that no one was on call when the data stream integration next developed an opinion.
They got to something workable. It was not perfect. It was better. Evelyn made the announcement to the full company on a Wednesday morning. She stood in the main floor, no stage, no podium, just standing in the middle of the room while everyone gathered. And she said without preamble, "3 months ago, I sat across from a man with an 8-year career gap and a community college degree, and I told him he didn't belong here. I was wrong.
I was wrong in a way that I want to be honest about because the wrongness wasn't just about him. It was about a set of assumptions I built into this company that I should have questioned sooner. She paused. I am sorry to Marcus Hail, who I hope is going to be in this building regularly enough that I will have many opportunities to demonstrate that the apology is real. And I'm sorry to everyone in this company who has ever felt like their path to where they are was not the right kind of path. The only right kind of path is the one that got you here. There was a silence and then applause that started at the back and moved forward and it was not the polite kind. Marcus consulting for his contracted 10 hours a week was in the building that day. He was standing near the window with his thermos and he applauded with everyone else. And when Brena Walsh made her way over to stand next to him, she said under her breath, "How does it feel?" He thought about this. Lily says I should negotiate for 12 hours, he said. Brena laughed. What do you think? I think 10 hours is the right amount. He looked out the window.
The bay was very blue. I think anything more and I start becoming someone she has to wait for. Number number part 10.
What he said at Harvard. The Whitfield Laurate Awards ceremony was held on a Friday evening in late October in one of the older buildings on the Harvard campus. A room with high windows and stone floors and portraits of serious men on every wall. The audience was several hundred people academics, researchers, former students, one or two journalists. The ceremony itself was formal in the way academic ceremonies are formal. Sequential, a little slow, punctuated by applause.
that was enthusiastic but calibrated.
Marcus spoke last. He was wearing a new suit. Lily had come with him to buy it and had vetoed two options on the grounds that they looked too important and selected a third that she described as just right. It was a deep navy. He looked Evelyn Ross would later tell Naen like himself. He stood at the podium for a moment before speaking. The room settled. I was told the tradition with this award, he said, is that the recipient speaks for about 20 minutes about the work being recognized. I'm going to try to do that. I want to say something about the work first and then I want to say something true and I'd ask you to bear with me for the second part because it's the harder part to say. He talked about the research.
He was clear and precise and not overly technical.
He had, as Dean Peton would later note approvingly, a gift for speaking to a mixed audience without condescending to either the specialists or the non-speists. He talked about what the ignorance boundary problem actually was, why it mattered, what it meant for AI systems to know the limits of their own knowledge, why that capacity was not a weakness but a form of integrity. He talked about where the field was now and where it might go. Then he stopped. He looked at his notes.
He folded them and put them in his jacket pocket. "My daughter is here tonight," he said. "She's sitting in the third row. If you look, you'll see her because she's the only 10-year-old in the room, and she's been making faces at me since I walked up here." A ripple of laughter. "She doesn't do this to be rude. She does it because she wants me to know she's there." He paused. "I left this work 8 years ago. I left it to raise her. I want to be honest about the fact that this was not entirely a free choice life arranged circumstances that made the choice for me and I simply agreed with them. But I also want to be honest that I have never in 8 years wished it had gone differently. He looked out at the room. Not because the work wasn't worth doing. It clearly was.
Not because I didn't miss it because I did, but because the thing I was doing instead was also worth doing.
It was also important work. It required everything I had. It produced, in my opinion, better results than anything I would have produced in a lab. The room was quiet in the particular way rooms get when something true is being said in a place where true things are not always said. We build systems, institutions, companies, academic programs that ask people to choose, Marcus said. between presence and ambition, between being there and being excellent, between their work and the people they love. We ask this because building systems that don't require this choice is harder and we have not as a field or as a culture done the work of making them harder in order to make them better. He paused. I want to suggest that we should start. He looked down at the podium. The ignorance boundary problem, he said, is about a system that doesn't know what it doesn't know. When it reaches the edge of its knowledge, it defaults to its oldest certainties, and the result is a subtle, persistent wrongness that is hard to find because it doesn't look like failure. It looks like confidence. He was quiet for a moment. I think about this sometimes as a description of something other than AI systems. He picked up his notes. He put them back in his pocket. My wife Rachel died when Lily was 14 months old. I have spent a lot of time thinking about what she would have wanted for us. I don't think she would have wanted me to be here tonight. I mean that as a compliment to hear she would have thought that was wonderful. But I think she would have wanted more than the award, more than the recognition, the thing she wanted when she was alive. For Lily to grow up knowing that she was the most important work her father ever did.
He looked at the third row. Lily was not making faces.
She was sitting very straight. The way she sat when she was trying not to cry, and Marcus recognized this because she had learned it from him. I am told, he said, that this award is given for work that was not fully appreciated in its time. I hope that what I've just said will be better appreciated in its time than that work was, because it seems to me more important. He stepped back from the podium. Thank you. The applause began at the front and moved back and it was not calibrated. It was the kind of applause that happens when a room has been given something it did not know it was missing. Evelyn Ross was seated midway back on the right hand side.
She was not crying which had required some effort. She was thinking about her father whom she had called the previous Sunday for the first time in 3 months and with whom she had talked for 45 minutes about nothing in particular which had turned out to be something after all. Marcus walked down from the stage. Lily met him at the edge of the aisle. She put her arms around him and pressed her face against his jacket in the way she had since she was small, and he rested one hand on the back of her head, and neither of them said anything.
The portraits on the walls looked down at a room full of serious people who were briefly together thinking about something more serious than their work. Outside the windows, the lights of Cambridge were coming on in the early dark.
The October air had gone cold and clear the way it does in New England when autumn has made up its mind. Marcus Hail, who had once been a researcher and was now a father and was becoming slowly and on his own terms a researcher again, stood in a room full of the world's recognition and held his daughter's hand and thought about what he would make her for dinner. The answer was the chicken she liked with the sauce she taught herself to make. He thought it was perfect. The smartest person in the room, Evelyn Ross would say the following spring to an audience of her own 300 hiring managers at a talent conference in Chicago, is not always the one you'd expect.
Sometimes they're the one you almost turned away. And the cost of that mistake is not what you lose from your company.
The cost is what they lose the work they don't do because you couldn't see past the suit. She paused.
関連おすすめ
OpenHuman VS Hermes AI: Who Wins?
JulianGoldieSEO
285 views•2026-05-29
Long-Running Agents — Build an Agent That Never Forgets with Google ADK
suryakunju
142 views•2026-05-30
This computer is made from real human brain cells. And you can buy it.
Talktmsmedia
3K views•2026-05-28
BREAKING: Microsoft’s New Image Generating Model Beat Out GPT 1.5 and Nano Banana 2
aimmediahouse
122 views•2026-06-03
I Made the Same Anime Fight Scene in Every AI Video Generator
NobleGooseAnime
295 views•2026-05-30
Nvidia Bets Big On AI PCs | New Chip To Power Windows Laptops | Technology | AI Updates | N18S
cnnnews18
3K views•2026-06-01
I Tested NEW Opus 4.8 on Four Projects (Updated LLM Leaderboard)
AICodingDaily
298 views•2026-05-29
3D Platformer Update - NO CAPES
SolarLune
294 views•2026-05-30











