---
title: "AI Career Advice — Andrew Ng and Lawrence Moroney at Stanford"
author: "Andrew Ng & Lawrence Moroney · Stanford CS230"
description: "It really feels like the best time ever to be building with AI and to build a career in AI. A few months ago, on social media and in traditional media, there were questions about whether AI progress..."
source: "https://adhenawer.net/posts/original/ai-career-advice-andrew-ng-lawrence-moroney-stanford.html"
lang: en
---

TL;DR · key points

1. It really feels like the best time ever to be building with AI and to build a career in AI.
 2. One study that most influenced my thinking was done by the organization METR.
 3. The reasons I think this is a golden age to be building come down to two themes: more powerful and faster.

## THE GOLDEN AGE OF BUILDING WITH AI 0:31

It really feels like the best time ever to be building with AI and to build a career in AI. A few months ago, on social media and in traditional media, there were questions about whether AI progress was slowing down. People were asking, is GPT-5 really that good? I think it actually is. But part of the reason that question was even raised is that if a benchmark caps at one hundred percent, rapid progress eventually runs into the ceiling — you simply cannot score above one hundred percent.

One study that most influenced my thinking was done by the organization METR. They measured, over time, how complex a task AI can do, as measured by how long the same task would take a human. A few years ago, GPT-2 could do tasks that took a human a couple of seconds. Then four seconds, eight seconds, a minute, two minutes, four minutes, and so on. The study estimates that the length of tasks AI can do is doubling every seven months. For coding, the doubling time is even shorter — maybe around seventy days. So code that used to take me ten minutes to write, then twenty, then forty — AI can do more and more of that.

The reasons I think this is a golden age to be building come down to two themes: more powerful and faster. All of you in this room can now write software that is more powerful than anyone on the planet could have built a year ago, by using AI building blocks: large language models, RAG, agentic workflows, voice AI, and of course deep learning. It turns out that most frontier LLMs have a decent basic understanding of deep learning, so if you ask one to implement a transformer network, it is actually not bad at helping you build software quickly using those building blocks.

## STAY ON THE FRONTIER OF YOUR TOOLS 3:32

With AI coding, the speed at which you can get software written is faster than ever. I personally find it important to stay on the frontier of tools, because the tools for AI coding change extremely quickly. Several months ago, my personal number one favorite tool became Claude Code, moving on from some earlier generations. Since the release of GPT-5, OpenAI Codex has made tremendous progress. And Gemini 3 was released this morning, which seems to be another huge leap forward.

If you ask me every three months what my favorite coding tool is, the answer probably changes every six months, and quite possibly every three. Being half a generation behind on these tools means being, frankly, quite a bit less productive. Everyone says AI is moving fast, but AI coding tools are one sector where the pace really is tremendous. Staying at the latest generation of tools, rather than half a generation behind, makes you more productive.

## THE PRODUCT MANAGEMENT BOTTLENECK 4:56

With our ability to build more powerful software faster, the advice I now give much more strongly than a year or two ago is simple: go build stuff. Take classes from Stanford, take online courses, and use every opportunity to build things and show them to others.

But there is a weird implication of all this, which is the product management bottleneck. When it becomes increasingly easy to go from a clearly written software spec to working code, the bottleneck shifts to deciding what to build — to writing that clear spec in the first place. When I build software, I go through a loop: write some code, show it to users, get feedback, revise my view of what users want, and iterate. I think of that as product management work.

Because AI coding has made building software much cheaper and faster, it ironically shifts the bottleneck to deciding what to build. In Silicon Valley, people have long talked about an engineer-to-PM ratio — four to one, seven to one, eight to one. But because engineering is speeding up and product management is not speeding up as much, I am seeing that ratio trend downwards, sometimes even to one to one. On some teams I work with, the proposed headcount is one PM per engineer — unlike anything in traditional Silicon Valley.

## ENGINEERS WHO SHAPE PRODUCTS MOVE FASTEST 7:26

The other pattern I am seeing is that engineers who can also shape products move incredibly fast. You take the engineer and the PM and collapse them into a single person. Some engineers genuinely do not enjoy talking to users or the empathetic side of product work, and that is completely fine. But the subset of engineers who learn to talk to users, get feedback, and develop deep empathy so they can decide what to build are some of the fastest moving people I see in Silicon Valley today.

Earlier in my career, I once tried to convince a bunch of engineers to do more product work, and I ended up making good engineers feel bad for not being good product managers. That was a mistake I regretted for years. Part of me worries I am about to repeat it. But the fact is that when you can write code and talk to users to shape what to build, you move much faster, because you are not waiting for someone else to take the product to customers. You write code, you have a gut feeling about what to do next, and you iterate. That velocity of execution is much higher.

## THE PEOPLE YOU SURROUND YOURSELF WITH 9:16

One of the strongest predictors of your speed of learning and level of success is the people you surround yourself with. We are all social creatures; we all learn from those around us. There are sociology studies showing that if your five closest friends are smokers, your odds of being a smoker are much higher. I do not know of studies showing that if your five or ten closest friends are hardworking, determined people trying to make the world better with AI, you are more likely to do the same — but it is almost certainly true.

At Stanford, there is a fantastic student body, a fantastic faculty, and what I consider a unique connective tissue. Many of the people working at the cutting-edge AI labs were former students of Stanford faculty. That rich connective tissue means we often find out about things that are not widely known, through relationships and friendships. When some company does something, one of my friends in the faculty will call someone at the company and ask, “Hey, that's weird, does this really work?” While you are at Stanford, form those friendships and take advantage of that network.

When it comes to applying for jobs, the single most important factor in your career success is the people you work with day to day. I will tell a story I have told in previous classes. A Stanford student I knew, clearly a high flyer, got a job offer from a company with a hot AI brand. This company refused to tell him which team he would join — sign on the dotted line first, they said, then we will figure it out. He signed, and was assigned to the backend Java payment processing system. Nothing against Java backend payment processing, but he was an AI student who did not get matched to an AI project. He was frustrated and left the company after about a year. The unfortunate thing is I told that story in CS230 years ago, and a couple of years later another student in the class went through the exact same experience with the same company.

If a company refuses to tell you what team you will be on, that should raise a serious question. Instead of chasing the hottest logo, find a team with hardworking, knowledgeable, smart people trying to do good with AI. You learn from the people you deal with day to day, not from the logo on the door.

## GO BUILD THINGS — RESPONSIBLY 15:18

My top piece of advice is that it has become much easier than ever to build powerful software quickly. Be responsible — do not build software that hurts others. But there are so many things you can build. The number of ideas in the world is much greater than the number of people with the skill to build them. Finding a job has gotten tougher for fresh college graduates, and at the same time many teams cannot find enough skilled people. A lot of projects in the world will simply not exist if you do not build them.

You do not need to wait for permission. You do not need to wait for someone else to do it first. The cost of a failure is lower than ever — you waste a weekend but learn something. So long as you are responsible, go for it. Try things. Build a lot. That will be the single most important thing for your career.

## ABOUT WORKING HARD 16:28

I am going to say something that is considered impolite in some circles: I am going to encourage you to work hard. Some people do not like this because some people are in a phase of life where they cannot work hard — a new baby, an injury, a disability, valid personal reasons. We should respect and support them. But every single one of my PhD students who became very successful worked incredibly hard. The two a.m. hyperparameter tuning — I have been there, done that. Still doing it some days.

If you are fortunate enough to be in a position in life where you can work really hard, there are so many opportunities right now. If you get excited spending evenings and weekends coding, building, and gathering user feedback, leaning in will increase your odds of being really successful. Between watching some dumb TV show and spending a weekend finding your agentic flow to try something, I will choose the latter almost every time — unless I am watching a show with my kids.

## THE TWO-SIDED INTERVIEW 19:04

I want to reinforce something Andrew said about choosing the people you work with. It is crucial — but the flip side matters just as much. When a company interviews you, they are also choosing you, and good companies want to choose people they actually want to work with.

I have been mentoring young people heavily over the last eighteen months. One young man I worked with was a super lead coder, great experience, well educated. He got laid off from his job in medical software in April. That same month his girlfriend broke up with him, and a couple of weeks later his dog died. Not a good place. When I sat down with him, he had a spreadsheet of three hundred jobs he was tracking. He had been deep in interview loops at Meta, Microsoft, Amazon, and other major companies. Every time, he thought the loop went well. Every time, a day later, the recruiter called to say no.

So I ran some mock interviews and gave him a really tough one — obscure corner cases in his code, picking holes, applying pressure. And I saw it. The advice in recruiting pamphlets often says things like, “You'll have an opportunity to share an opinion — stand your ground, have a backbone, do not bend.” His interpretation of that advice was to be very, very tough. When I challenged him, he became hostile. From an interviewing manager's perspective, he was the clichéd ten-x engineer, but nobody would want him on their team because of that attitude. We fine-tuned it. The good news is that the next company he interviewed with — one where teamwork is highly valued — hired him, and he doubled his salary. So yes, stand your ground. But do not be a jerk while you do it.

## A BRIEF HISTORY OF THE AI JOB MARKET 25:37

I have been working in tech for more decades than ChatGPT thinks there are R's in “strawberry.” I first got involved in AI back in 1992, worked in it briefly just before the AI winter, and got pulled right back in in 2015 when Google launched TensorFlow. What does the career landscape look like today?

The market is not ideal. Look at the news — massive tech layoffs, slowing hiring, entry-level positions that feel scarce, fierce competition. But if you approach it the right way, it is not necessarily a bad thing.

Here is how we got here. In 2021 and 2022, the global pandemic caused a massive industrial slowdown. Companies pivoted toward things that drove revenue directly. At Google, TensorFlow was open source and did not directly drive revenue, so we scaled back. Every company scaled hiring back. Then 2022 and 2023 arrived — we came out of the pandemic with a huge logjam of hiring that had not happened, and AI exploded onto the scene. Every company had to hire like crazy. And when every company hires like crazy, most companies overhire. You had bidding wars, talent grabs, and people with “AI” on their resume getting jobs they were not qualified for. Then 2024 and 2025 became the great wake-up. Companies realized they had overhired, they had unqualified people in roles they could not do, and a big correction started. Today, companies are much more cautious about AI hiring. Opportunity is still there — but you have to approach it strategically.

## THREE PILLARS OF SUCCESS 28:49

You can no longer just put “AI” on your resume and get overhired. Today, you not only have to have the mindset of these three pillars, you have to be able to show it — and thanks to the ability to prompt products into existence, showing has never been easier.

Pillar one: understanding in depth. Academically — machine learning, model architectures, reading papers and understanding them, and knowing how to put that work to use. And also having your finger on the pulse of real trends, where the signal-to-noise ratio favors signal.

Pillar two: business focus. Hard work is a nebulous phrase, so think of it as “you are what you measure.” The trend of 996 — nine a.m. to nine p.m., six days a week — is not a metric of hard work. It is a metric of time spent. I measure hard work by output. People ask me how I have time to write so many books. The answer is baseball. Games take three and a half to four hours. In baseball season, I write while the Mariners or Dodgers are on in the background. The real advice: do not let your output be for the job you have — let it be for the job you want. When I finally got hired at Google (after interviewing twice and failing twice), I wrote a Java cloud application for predicting stock prices and put the code on my resume. My entire interview loop was about my code. It put the power in my hands.

Pillar three: a bias toward delivery. Ideas are cheap. Execution is everything. I have interviewed many people with fluffy ideas and no way to ground them, and many with half-baked ideas that were brilliantly grounded. Guess who gets the job.

## THE FOUR REALITIES OF WORKING IN AI TODAY 36:18

Two or three years ago, if you could build an image classifier you were golden — six-figure salaries, big stock grants. That is not the case anymore. Today the word is “production.” Everything is biased toward production. The bottom line is the bottom line. If you come in with that mindset, that is one of the keys to open the door.

Reality one: business focus is non-negotiable. For most of the last decade, many large companies, particularly in Silicon Valley, focused on developing their people above everything, including bringing their whole selves to work — which led to a lot of in-company activism. There was nothing wrong with activism itself. But overindexing on activism led to real problems — the famous incident where activists broke into a Google Cloud leader's office over a country the company did business with, had a sit-in, and used the bathroom all over his desk. That kind of event has caused the pendulum to swing back hard. Business focus is non-negotiable now.

Reality two: risk mitigation is part of the job. In an AI interview environment, I would argue this is the single most valuable skill. Come in with the mindset of “you are doing a business transformation from heuristic computing to intelligent computing — here are the risks, and here is how you mitigate them.”

Reality three: responsibility is evolving — from fluffy “make the AI work for everyone” to a harder-nosed “make sure the AI works, make sure it drives the business, and then make sure it works for everyone.” The Gemini image generation fiasco a couple of years ago is instructive. I tested it with the same prompt — a young woman in a cornfield wearing a summer dress — across ethnicities. Young Asian, Indian, Black, Latina — all beautiful, adherent results. Caucasian? “While I understand your request, I am unable to generate images of people, as this could potentially lead to harmful stereotypes and biases.” White? Same refusal. A very poorly implemented safety filter. The hack I used to break it: I asked for an Irish woman. It returned images of women — all with red hair. Ireland has the highest proportion of redheads in the world, about eight percent. Associating an entire ethnicity with a hair color is itself problematic. A very narrow view of what is responsible ended up damaging the reputation of the model and the company. Responsible AI has moved out of fluffy social framing into harder-line things tied to protecting the business.

Reality four: learning from mistakes is constant. You will make them. The people around you will make them too. Give them grace.

## TECHNICAL DEBT AND "VIBE CODING" 49:05

The meme is out there that generated code makes engineers less useful. Peel it apart and it is not true. The more skilled you are as an engineer, the better you become at prompting code. And the metric I like to use is technical debt.

Think about debt the way you normally would. Borrow half a million dollars for a house on a thirty-year mortgage, and you pay back about a million with interest. That is probably good debt — the house appreciates, you are not paying rent, you get thirty years of use. An impulse purchase on a high-interest credit card is bad debt — two hundred dollars of shoes for which you end up paying five hundred, getting nowhere near five hundred dollars of value.

Every time you build software, you take on debt — bugs, support, new requirements, documentation. The only way to avoid debt is to do nothing. So when you prompt code into existence, ask: is this worth the technical debt it creates? How do you take on good technical debt?

Clear objectives met. Do not just fire up ChatGPT and start spinning code. I am building a small startup in the movie-making space using code generation almost exclusively. What I have done is build, test, throw away, build again, test, throw away — and each time my requirements sharpen. Code is cheap now; engineered code is not.

Business value delivered. I have seen people spend hours vibe-coding a cool website on Replit, and the only honest answer is, “So what? How does this help the business?” That is the bad-debt version.

Human understanding. The worst technical debt you can take on is shipping code nobody else understands. Then you quit for a better job, and the company is stuck. Variable names, clear algorithms, and documentation are how you avoid that.

Watch out for classic bad debt: solutions looking for problems, spaghetti code from prompt after prompt, and authority over merit — the VP who bought a Replit subscription and now it is your problem to fix what they built.

## NAVIGATING THE HYPE CYCLE 56:29

Hype is one of the strongest forces in the universe, particularly in AI and crypto, the two fields I work in. The currency of social media is engagement, not accuracy. Even LinkedIn is overwhelmed with influencers using Gemini or GPT to write engagement-farming posts. If you can filter signal from noise, and then encourage others toward the signal and not the noise, that makes you immensely valuable in a one-to-one environment like a job interview.

Last year, a European company asked me to help them implement an agent. The correct first question to ask a client in that situation is not “what is an agent to you?” It is not even “what do you want to do?” It is “why?” I spoke with the CEO and peeled it apart. Eventually what he really wanted was to make his salespeople more efficient. Nowhere in that sentence does the word “AI” or “agent” appear. As a trusted advisor, my job was to make his salespeople more efficient, not to shill AI or agents.

I talked to the top salespeople and asked what they hated most about their jobs. They were spending about eighty percent of their time researching companies and contacts, and only twenty percent selling. Salespeople are paid largely on commission, so this was crushing. We set a goal of making them twenty percent more efficient — and only then did we introduce the ideas of AI and agentic AI.

## WHAT AGENTIC AI ACTUALLY LOOKS LIKE 1:02:43

An agent really comes down to four steps. First, understand intent. Large language models are great at understanding. Feed the LLM the task — “I want to meet Bob Smith and sell widgets to Bob Smith, here is what I know about him” — and let it capture the intent clearly. Second, plan. Declare the tools available — browsing the web, searching, and so on — and let the LLM break the work into steps. Third, use the tools to get a result. Fourth, reflect on that result against the original intent. If it does not match, loop back.

We built a pilot for that company's salespeople and they saved between ten and fifteen percent of their wasted time. The unintended consequences were beautiful: they made more sales per week, they earned more money, and their jobs became less miserable. McKinsey recently reported that about eighty-five percent of AI projects at companies fail — largely because they are not well scoped and people jump on the hype bandwagon without reasoning through the problem.

Other hype examples to peel apart: “software engineering is dead,” “Hollywood is dead,” “AGI by year-end.” A year ago at an event in Saudi Arabia, I sat beside a CEO of a generative AI company who told a table full of media executives that by this time next year they would be generating ninety-minute films from a single prompt. Two years later we still cannot — and even if we could, what prompt could tell the whole story of a movie?

## BECOMING A TRUSTED ADVISOR 1:06:46

How do you navigate a world drowning in hype? Look at trends and evaluate them objectively. Identify the genuine opportunities. Ignore fashionable distractions and the people leaning into them. Then lean into the skill of explaining technical reality to leadership.

One piece of counterintuitive coaching I once received: whenever something feels magical, figure out how to make it as mundane as possible. When you can make it mundane, you can ground it, and then you can explain it in detail. Take text-to-video. The mundane explanation is that a model trained on video predicts a sequence of slightly different frames, given a matching prompt. Suddenly it is not magic. And once it is not magic, domain experts — not only the technical side — can come up and do brilliant things with it.

Filter actively. Go deep on the fundamentals. And keep your finger on the pulse, which is the hard part, because that means wading into the cesspits of social media to understand what the people you advise are seeing.

## BUBBLES COME — BUILD FOR LIFE AFTER THEM 1:20:45

The overall landscape is ripe with opportunity. But there are risks. Think about Titanic. Right before the famous “iceberg right ahead” moment, there is a scene in the crow's nest where the spotters are freezing and complaining, and then it cuts away to the crew realizing the binoculars were left behind in port. The whole idea is that arrogance about moving forward made them fail to equip their people to look out for risk. That is a good metaphor for the AI industry today.

The B word — “bubble” — is in the news. There will be a bubble. Look at the dot-com bubble of the early 2000s. It burst. But the people who did dot-com right did not just survive, they thrived — Amazon, Google. Pets.com, famous for Super Bowl ads and “if you build it they will come,” evaporated. The anatomy is always the same: hype on top, massive VC investment, unrealistic valuations, me-too products, and a small kernel of real value at the bottom. I am already seeing VC investment drying up for anything with “AI” on the tin, and startups getting scaled-back rounds.

Focus on the fundamentals. Build real solutions. Understand the business side. And most of all, diversify your skills. Do not be a one-trick pony who only knows one API or framework — the industry moves on, and you get left behind.

## THE COMING BIFURCATION — BIG AI AND SMALL AI 1:24:24

Over the next five years I see a bifurcation. On one side is big AI — the Geminis, Claudes, and OpenAIs driving toward AGI, getting bigger and bigger. On the other side is small AI — open-weight, self-hostable models. I prefer “self-hostable” or “open weights” to “open source.” A recent Y Combinator article reported that around eighty percent of its companies were using small models, with Chinese models doing especially well. Rather than big versus small, think of it as models hosted by someone else on your behalf versus models you can host yourself.

The self-hostable side is currently underserved. The skill I see developers needing most on that side over the next two to three years is fine-tuning — taking an open model and fine-tuning it for downstream tasks. One concrete example: I work a lot with Hollywood studios, and IP protection there is extreme. Studios will not share movie details with GPT or Gemini, because that is sharing IP with a third party. But they could enormously benefit from analyzing movie synopses — what works, what does not, why a movie hit, why another flopped, what time of year helps. Enter self-hosted small models. Today's 7B model is as capable as yesterday's 50B; next year's 7B will match today's 300B. The same logic holds for law firms, medical offices, and anywhere privacy matters.

## SMALL AI AND EMBEDDED INTELLIGENCE EVERYWHERE 1:24:29

Small AI is the next big thing because we are moving into a world of AI everywhere, all at once. The traditional view is that compute platforms are CPU plus GPU. But in mobile there is massive innovation around SME — Scalable Matrix Extensions — which brings AI workloads onto the CPU. The front runners are Chinese phone vendors Vivo and Oppo, who have released phones with SME-enabled chips. The magic is that you do not need a separate chip drawing extra power and taking extra space, and the CPU can run AI workloads at low power.

Take Alipay. They wanted photo search — “places I ate sushi” — without three problems: privacy (sharing photos with a back end), latency (upload, process, download), and the cost of building and operating the cloud service. Moving all of it onto the device solves all three. Apple has invested heavily in scalable matrix extensions. When they talk about neural cores in the A-series and M-series chips, that is the same idea. Embedded intelligence is no longer sci-fi.

## AGENTIC INTENT — A PERSONAL DEMO 1:24:58

I like to call the hidden part of artificial intelligence “artificial understanding.” When you can use models to understand things on your behalf, and craft new things from that understanding, you develop superpowers. Quick demo. I took a picture of my son taking a slap shot in a practice arena — no people in the stands. I thought I was good at prompting, so I wrote a nice prompt asking for a scoring goal. The result: the AI, having no idea that the top right corner was actually a storage area for garbage, painted in a full crowd. My son shot wide. Everyone cheered. He had two sticks. They forgot his name.

I did not go through an agentic workflow for that. Contrast that with a short movie clip I am working on with a startup. The workflow: feed the scene (“the girl is on the bench, upset, and he wants to comfort her”), the full story, and the constraints (eight-second shots, clear dialogue) to an LLM as step one — understand intent. The LLM ended up writing a prompt far more descriptive than I ever would have, with understanding of shot composition, angles, and emotion. Step two: the tool (Veo), with its idiosyncrasies — Veo is bad at high-action scenes but good at slow camera pulls. Step three: actually generate. A Veo generation costs two to three dollars for about four videos, so the planning at step one paid for itself many times over in credits saved. It rarely took more than two or three tries to get something good.

## WHAT HAS SURPRISED ME THE MOST 1:32:39

What has surprised me most in AI, and probably should not have, is how much hype took over. I honestly thought people in important decision-making roles would see the signal better than they did. And the desire for immediate profits over long-term gains surprised me too. After Andrew and I taught the TensorFlow Coursera specializations, Google launched a professional certificate — a rigorous exam that, once passed, helped people find work. It cost Google about one hundred thousand dollars a year to run. The goodwill was immense.

One young man from Syria earned one of the first TensorFlow certificates there, moved to Germany during the civil war, landed a job at a major German firm, and was able to move his family out of the war zone. There were countless stories like that. But the program cost slightly more than expected to run, so Google cancelled it. A shame, given the goodwill.

The other recurring surprise is the people who become successful with AI that you would never expect. A good friend of mine, a former pro ice hockey player who dropped out of school at thirteen to focus on skating, now runs a nonprofit ice rink. He has to present quarterly to a board. He was spending around one hundred and fifty thousand dollars a year on consultants to assemble reports from spreadsheets, PDFs, and compressor-room instruments. He asked me about ChatGPT over a beer. It turned out he had already tried it and it worked brilliantly. The report that used to cost one hundred and fifty thousand dollars in consulting now takes him two hours. That money now goes to underprivileged kids — hockey equipment, lessons. He calls himself the dumbest person alive. I told him, “Congratulations, you are now a developer.”

## HELPING NON-EXPERTS SEE THROUGH THE HYPE 1:38:38

For engineers like us, it is easier to spot the signal in the noise. For non-technical people, that is our opportunity to be trusted advisors — to help them understand the reward mechanism in the hype machine (engagement, not accuracy) and to peel apart what they have tried, what worked, and why. The average person is intelligent, even if not expert in a specific domain. Key in on that intelligence. Help them foster it. Navigate them through the difficult parts and let them shine where they are already strong.

## AI FOR SCIENTIFIC RESEARCH 1:40:15

Is AI a good idea for scientific research? My gut is: almost always. Use the most powerful tools you have available, but always double-check your results against grounded reality. I did my undergraduate in physics and was relatively successful in the lab mostly because I automated things through a computer that other people did with pen and paper.

One quick story from Wales — the poorest country in Western Europe, where I did my undergraduate. I went back to lecture there and met a researcher doing brain cancer work with computer imagery. I asked what his biggest blocker was. His answer: access to a GPU. His department had one GPU shared among ten researchers — half a day each. His slot was Tuesday afternoon. He spent the rest of the week preparing so that Tuesday afternoon's training run would succeed, or he would lose another week. I showed him Google Colab and his brain melted. Free GPU, in the cloud, running a notebook on my phone. Widening access changed everything for his research.

## CAN AI BE A FORCE FOR EQUALITY — OR INEQUALITY? 1:43:44

Yes. It can be both, and it can be neither. Any tool can be used for any means. The important thing is to educate and inspire people to use things for the right ends. There is only so much governance can do, and sometimes governance causes more problems than it solves. I try to live my life assuming good intent but preparing for bad intent. The bad examples I showed earlier were good intent used mistakenly, not bad intent. AI itself has no choice. It is how people use it.