Showing posts with label Paul Graham. Show all posts
Showing posts with label Paul Graham. Show all posts

Sunday, April 26, 2026

Richard Hamming: You and Your Research

Here's the framework that has been quoted by every serious scientist for the last 40 years.

His opening line landed like a punch. He said most scientists he worked with at Bell Labs were just as smart as the Nobel Prize winners. Just as hardworking. Just as credentialed. And yet at the end of a 40-year career, one group had changed entire fields and the other group was forgotten by the time they retired.

He wanted to know what the difference actually was. And he said it wasn't luck. It wasn't IQ. It was a specific set of habits that almost nobody is willing to follow.

The first habit was the one that hurts the most to hear. He said most scientists deliberately avoid the most important problem in their field because the odds of failure are too high. They pick a safe adjacent problem, solve it cleanly, publish it, and move on. And because they never swing at the hard problem, they never hit it. He said if you do not work on an important problem, it is unlikely you will do important work. That is not a motivational line. That is a logical one.

The second habit was about doors. Literal doors. He noticed that the scientists at Bell Labs who kept their office doors closed got more done in the short term because they had no interruptions. But the scientists who kept their doors open got more done over a career. The open-door scientists were interrupted constantly. They also absorbed every new idea passing through the hallway. Ten years in, they were working on problems the closed-door scientists did not even know existed.

The third habit was inversion. When Bell Labs refused to give him the team of programmers he wanted, Hamming sat with the rejection for weeks. Then he flipped the question. Instead of asking for programmers to write the programs, he asked why machines could not write the programs themselves. That single inversion pushed him into the frontier of computer science. He said the pattern repeats everywhere. What looks like a defect, if you flip it correctly, becomes the exact thing that pushes you ahead of everyone else.

The fourth habit was the one that hit me the hardest. He said knowledge and productivity compound like interest. Someone who works 10 percent harder than you does not produce 10 percent more over a career. They produce twice as much. The gap doesn't add. It multiplies. And it compounds silently for years before anyone notices.

He said Pasteur's famous quote is right. Luck favors the prepared mind. But he meant it literally. You don't hope for luck. You engineer the conditions where luck can land on you. Open doors. Important problems. Inverted questions. Compounded hours. Those are not traits. Those are choices you make every single day.

Richard Hamming: You and Your Research Talk at Bellcore, 7 March 1986 I saw Feynman up close. I saw Fermi and Teller. I saw Oppenheimer. I saw Hans Bethe: he was my boss. I saw quite a few very capable people. I became very interested in the difference between those who do and those who might have done. ........................ consider Einstein. Note how many different things he did that were good. Was it all luck? Wasn't it a little too repetitive? ..................... The prepared mind sooner or later finds something important and does it. ................ Newton said, "If others would think as hard as I did, then they would get similar results." ................. Einstein, somewhere around 12 or 14, asked himself the question, "What would a light wave look like if I went with the velocity of light to look at it?" Now he knew that electromagnetic theory says you cannot have a stationary local maximum. But if he moved along with the velocity of light, he would see a local maximum. He could see a contradiction at the age of 12, 14, or somewhere around there, that everything was not right and that the velocity of light had something peculiar. Is it luck that he finally created special relativity? Early on, he had laid down some of the pieces by thinking of the fragments. ........................... great work is something else than mere brains .............. Bill Pfann, the fellow who did zone melting, came into my office one day. He had this idea dimly in his mind about what he wanted and he had some equations. It was pretty clear to me that this man didn't know much mathematics and he wasn't really articulate. His problem seemed interesting so I took it home and did a little work. I finally showed him how to run computers so he could compute his own answers. I gave him the power to compute. He went ahead, with negligible recognition from his own department, but ultimately he has collected all the prizes in the field. Once he got well started, his shyness, his awkwardness, his inarticulateness, fell away and he became much more productive in many other ways. Certainly he became much more articulate. ............................ Clogston finally did the Clogston cable. After that there was a steady stream of good ideas. One success brought him confidence and courage. ........................... One of the characteristics of successful scientists is having courage. Once you get your courage up and believe that you can do important problems, then you can. If you think you can't, almost surely you are not going to. ............... That is the characteristic of great scientists; they have courage. They will go forward under incredible circumstances; they think and continue to think. ................... Age is another factor which the physicists particularly worry about. They always are saying that you have got to do it when you are young or you will never do it. Einstein did things very early, and all the quantum mechanic fellows were disgustingly young when they did their best work. Most mathematicians, theoretical physicists, and astrophysicists do what we consider their best work when they are young. It is not that they don't do good work in their old age but what we value most is often what they did early. On the other hand, in music, politics and literature, often what we consider their best work was done late. I don't know how whatever field you are in fits this scale, but age has some effect..................... let me say why age seems to have the effect it does. In the first place if you do some good work you will find yourself on all kinds of committees and unable to do any more work. ................. when you get early recognition it seems to sterilize you. In fact I will give you my favorite quotation of many years. The Institute for Advanced Study in Princeton, in my opinion, has ruined more good scientists than any institution has created, judged by what they did before they came and judged by what they did after. Not that they weren't good afterwards, but they were superb before they got there and were only good afterwards. ............................ What most people think are the best working conditions, are not. Very clearly they are not because people are often most productive when working conditions are bad. One of the better times of the Cambridge Physical Laboratories was when they had practically shacks — they did some of the best physics ever. ...................

I give you a story from my own private life. Early on it became evident to me that Bell Laboratories was not going to give me the conventional acre of programming people to program computing machines in absolute binary. It was clear they weren't going to. But that was the way everybody did it. I could go to the West Coast and get a job with the airplane companies without any trouble, but the exciting people were at Bell Labs and the fellows out there in the airplane companies were not. I thought for a long while about, "Did I want to go or not?" and I wondered how I could get the best of two possible worlds. I finally said to myself, "Hamming, you think the machines can do practically everything. Why can't you make them write programs?" What appeared at first to me as a defect forced me into automatic programming very early. What appears to be a fault, often, by a change of viewpoint, turns out to be one of the greatest assets you can have. But you are not likely to think that when you first look the thing and say, "Gee, I'm never going to get enough programmers, so how can I ever do any great programming?"

............................... often the great scientists, by turning the problem around a bit, changed a defect to an asset. For example, many scientists when they found they couldn't do a problem finally began to study why not. They then turned it around the other way and said, "But of course, this is what it is" and got an important result. So ideal working conditions are very strange. The ones you want aren't always the best ones for you. ......................... John was a genius and I clearly was not. Well I went storming into Bode's office and said, "How can anybody my age know as much as John Tukey does?" He leaned back in his chair, put his hands behind his head, grinned slightly, and said, "You would be surprised Hamming, how much you would know if you worked as hard as he did that many years." I simply slunk out of the office! ................... Knowledge and productivity are like compound interest. Given two people of approximately the same ability and one person who works ten percent more than the other, the latter will more than twice outproduce the former. .............. The more you know, the more you learn; the more you learn, the more you can do; the more you can do, the more the opportunity — it is very much like compound interest. I don't want to give you a rate, but it is a very high rate. Given two people with exactly the same ability, the one person who manages day in and day out to get in one more hour of thinking will be tremendously more productive over a lifetime. ........................ On this matter of drive Edison says, "Genius is 99% perspiration and 1% inspiration." He may have been exaggerating, but the idea is that solid work, steadily applied, gets you surprisingly far. The steady application of effort with a little bit more work, intelligently applied is what does it. That's the trouble; drive, misapplied, doesn't get you anywhere. I've often wondered why so many of my good friends at Bell Labs who worked as hard or harder than I did, didn't have so much to show for it. The misapplication of effort is a very serious matter. Just hard work is not enough - it must be applied sensibly. ....................... Great scientists tolerate ambiguity very well. ................ If you believe too much you'll never notice the flaws; if you doubt too much you won't get started. It requires a lovely balance. ....................... most great scientists are well aware of why their theories are true and they are also well aware of some slight misfits which don't quite fit and they don't forget it. Darwin writes in his autobiography that he found it necessary to write down every piece of evidence which appeared to contradict his beliefs because otherwise they would disappear from his mind. When you find apparent flaws you've got to be sensitive and keep track of those things, and keep an eye out for how they can be explained or how the theory can be changed to fit them. Those are often the great contributions. Great contributions are rarely done by adding another decimal place. It comes down to an emotional commitment. Most great scientists are completely committed to their problem. Those who don't become committed seldom produce outstanding, first-class work. .........................

Everybody who has studied creativity is driven finally to saying, "creativity comes out of your subconscious." Somehow, suddenly, there it is. It just appears.

.............................. your dreams also come out of your subconscious. .................... one thing you are pretty well aware of is that your dreams also come out of your subconscious. And you're aware your dreams are, to a fair extent, a reworking of the experiences of the day. If you are deeply immersed and committed to a topic, day after day after day, your subconscious has nothing to do but work on your problem. And so you wake up one morning, or on some afternoon, and there's the answer. For those who don't get committed to their current problem, the subconscious goofs off on other things and doesn't produce the big result. So the way to manage yourself is that when you have a real important problem you don't let anything else get the center of your attention — you keep your thoughts on the problem. Keep your subconscious starved so it has to work on your problem, so you can sleep peacefully and get the answer in the morning, free. ......................... The physics table was, as he said, an exciting place, but I think he exaggerated on how much I contributed. It was very interesting to listen to Shockley, Brattain, Bardeen, J. B. Johnson, Ken McKay and other people, and I was learning a lot. But unfortunately a Nobel Prize came, and a promotion came, and what was left was the dregs. ..................... I went over and said, "Do you mind if I join you?" They can't say no, so I started eating with them for a while. And I started asking, "What are the important problems of your field?" And after a week or so, "What important problems are you working on?" And after some more time I came in one day and said, "If what you are doing is not important, and if you don't think it is going to lead to something important, why are you at Bell Labs working on it?" I wasn't welcomed after that; I had to find somebody else to eat with! ......................... I have never heard the names of any of the other fellows at that table mentioned in science and scientific circles. They were unable to ask themselves, "What are the important problems in my field?" .................... If you do not work on an important problem, it's unlikely you'll do important work. It's perfectly obvious. ..................... We didn't work on (1) time travel, (2) teleportation, and (3) antigravity. They are not important problems because we do not have an attack. It's not the consequence that makes a problem important, it is that you have a reasonable attack. That is what makes a problem important. ..................... The average scientist, so far as I can make out, spends almost all his time working on problems which he believes will not be important and he also doesn't believe that they will lead to important problems. ....................... You can't always know exactly where to be, but you can keep active in places where something might happen. And even if you believe that great science is a matter of luck, you can stand on a mountain top where lightning strikes; you don't have to hide in the valley where you're safe. But the average scientist does routine safe work almost all the time and so he (or she) doesn't produce much. It's that simple. If you want to do great work, you clearly must work on important problems, and you should have an idea. ........................... "Great Thoughts Time." When I went to lunch Friday noon, I would only discuss great thoughts after that. By great thoughts I mean ones like: "What will be the role of computers in all of AT&T?", "How will computers change science?" ..................... For example, I came up with the observation at that time that nine out of ten experiments were done in the lab and one in ten on the computer. I made a remark to the vice presidents one time, that it would be reversed, i.e. nine out of ten experiments would be done on the computer and one in ten in the lab. They knew I was a crazy mathematician and had no sense of reality. I knew they were wrong and they've been proved wrong while I have been proved right. They built laboratories when they didn't need them. I saw that computers were transforming science because I spent a lot of time asking "What will be the impact of computers on science and how can I change it?" I asked myself, "How is it going to change Bell Labs?" I remarked one time, in the same address, that more than one-half of the people at Bell Labs will be interacting closely with computing machines before I leave. Well, you all have terminals now. ............................... Most great scientists know many important problems. They have something between 10 and 20 important problems for which they are looking for an attack. And when they see a new idea come up, one hears them say "Well that bears on this problem." They drop all the other things and get after it. Now I can tell you a horror story that was told to me but I can't vouch for the truth of it. I was sitting in an airport talking to a friend of mine from Los Alamos about how it was lucky that the fission experiment occurred over in Europe when it did because that got us working on the atomic bomb here in the US. He said "No; at Berkeley we had gathered a bunch of data; we didn't get around to reducing it because we were building some more equipment, but if we had reduced that data we would have found fission." They had it in their hands and they didn't pursue it. They came in second! ....................... The great scientists, when an opportunity opens up, get after it and they pursue it. They drop all other things. They get rid of other things and they get after an idea because they had already thought the thing through. Their minds are prepared; they see the opportunity and they go after it. ................... I notice that if you have the door to your office closed, you get more work done today and tomorrow, and you are more productive than most. But 10 years later somehow you don't know quite know what problems are worth working on; all the hard work you do is sort of tangential in importance. He who works with the door open gets all kinds of interruptions, but he also occasionally gets clues as to what the world is and what might be important. ................... there is a pretty good correlation between those who work with the doors open and those who ultimately do important things, although people who work with doors closed often work harder. Somehow they seem to work on slightly the wrong thing — not much, but enough that they miss fame. ...................... "It ain't what you do, it's the way that you do it." I'll start with an example of my own. I was conned into doing on a digital computer, in the absolute binary days, a problem which the best analog computers couldn't do. And I was getting an answer. When I thought carefully and said to myself, "You know, Hamming, you're going to have to file a report on this military job; after you spend a lot of money you're going to have to account for it and every analog installation is going to want the report to see if they can't find flaws in it." I was doing the required integration by a rather crummy method, to say the least, but I was getting the answer. And I realized that in truth the problem was not just to get the answer; it was to demonstrate for the first time, and beyond question, that I could beat the analog computer on its own ground with a digital machine. I reworked the method of solution, created a theory which was nice and elegant, and changed the way we computed the answer; the results were no different. The published report had an elegant method which was later known for years as "Hamming's Method of Integrating Differential Equations." It is somewhat obsolete now, but for a while it was a very good method. By changing the problem slightly, I did important work rather than trivial work. .................... How do I obey Newton's rule? He said, "If I have seen further than others, it is because I've stood on the shoulders of giants." ................ How do I obey Newton's rule? He said, "If I have seen further than others, it is because I've stood on the shoulders of giants." ...................... Now if you are much of a mathematician you know that the effort to generalize often means that the solution is simple. Often by stopping and saying, "This is the problem he wants but this is characteristic of so and so. Yes, I can attack the whole class with a far superior method than the particular one because I was earlier embedded in needless detail." The business of abstraction frequently makes things simple. Furthermore, I filed away the methods and prepared for the future problems. ........................

"It is a poor workman who blames his tools — the good man gets on with the job, given what he's got, and gets the best answer he can."

......................... It's just as easy to do a broad, general job as one very special case. And it's much more satisfying and rewarding! ............................ There are three things you have to do in selling. You have to learn to write clearly and well so that people will read it, you must learn to give reasonably formal talks, and you also must learn to give informal talks. We had a lot of so-called `back room scientists.' In a conference, they would keep quiet. Three weeks later after a decision was made they filed a report saying why you should do so and so. Well, it was too late. They would not stand up right in the middle of a hot conference, in the middle of activity, and say, "We should do this for these reasons." You need to master that form of communication as well as prepared speeches. .......................... When I first started, I got practically physically ill while giving a speech, and I was very, very nervous. I realized I either had to learn to give speeches smoothly or I would essentially partially cripple my whole career. .................... While going to meetings I had already been studying why some papers are remembered and most are not. The technical person wants to give a highly limited technical talk. Most of the time the audience wants a broad general talk and wants much more survey and background than the speaker is willing to give. As a result, many talks are ineffective. The speaker names a topic and suddenly plunges into the details he's solved. Few people in the audience may follow. You should paint a general picture to say why it's important, and then slowly give a sketch of what was done. Then a larger number of people will say, "Yes, Joe has done that," or "Mary has done that; I really see where it is; yes, Mary really gave a good talk; I understand what Mary has done." The tendency is to give a highly restricted, safe talk; this is usually ineffective. Furthermore, many talks are filled with far too much information. So I say this idea of selling is obvious. ....................... Friday afternoons for years — great thoughts only — means that I committed 10% of my time trying to understand the bigger problems in the field, i.e. what was and what was not important. I found in the early days I had believed `this' and yet had spent all week marching in `that' direction. It was kind of foolish. If I really believe the action is over there, why do I march in this direction? I either had to change my goal or change what I did. So I changed something I did and I marched in the direction I thought was important. It's that easy. .................. Now you might tell me you haven't got control over what you have to work on. Well, when you first begin, you may not. But once you're moderately successful, there are more people asking for results than you can deliver and you have some power of choice, but not completely. ...................

and it was clear, in my area, that a "mathematician had no use for machines."

...................... He had to give in. You can educate your bosses. It's a hard job. ............. I am telling you how you can get what you want in spite of top management. You have to sell your ideas there also. ...................... "Is the effort to be a great scientist worth it?" To answer this, you must ask people. When you get beyond their modesty, most people will say, "Yes, doing really first-class work, and knowing it, is as good as wine, women and song put together," or if it's a woman she says, "It is as good as wine, men and song put together." ........................ And if you look at the bosses, they tend to come back or ask for reports, trying to participate in those moments of discovery. They're always in the way. .............. those who have done it, want to do it again. .................... I have never dared to go out and ask those who didn't do great work how they felt about the matter. ...................... I think it is very definitely worth the struggle to try and do first-class work because the truth is, the value is in the struggle more than it is in the result. The struggle to make something of yourself seems to be worthwhile in itself. The success and fame are sort of dividends, in my opinion. ........... Why do so many of the people who have great promise, fail? ..................... Well, one of the reasons is drive and commitment. The people who do great work with less ability but who are committed to it, get more done that those who have great skill and dabble in it, who work during the day and go home and do other things and come back and work the next day. They don't have the deep commitment that is apparently necessary for really first-class work. ................... Good people, very talented people, almost always turn out good work. We're talking about the outstanding work, the type of work that gets the Nobel Prize and gets recognition. ............. the problem of personality defects. Now I'll cite a fellow whom I met out in Irvine. He had been the head of a computing center and he was temporarily on assignment as a special assistant to the president of the university. It was obvious he had a job with a great future. He took me into his office one time and showed me his method of getting letters done and how he took care of his correspondence. He pointed out how inefficient the secretary was. He kept all his letters stacked around there; he knew where everything was. And he would, on his word processor, get the letter out. He was bragging how marvelous it was and how he could get so much more work done without the secretary's interference. Well, behind his back, I talked to the secretary. The secretary said, "Of course I can't help him; I don't get his mail. He won't give me the stuff to log in; I don't know where he puts it on the floor. Of course I can't help him." So I went to him and said, "Look, if you adopt the present method and do what you can do single-handedly, you can go just that far and no farther than you can do single-handedly. If you will learn to work with the system, you can go as far as the system will support you." And, he never went any further. He had his personality defect of wanting total control and was not willing to recognize that you need the support of the system. .................................. You find this happening again and again; good scientists will fight the system rather than learn to work with the system and take advantage of all the system has to offer. .......................... No Vice President at IBM said, `Give Hamming a bad time'. It is the secretaries at the bottom who are doing this. When a slot appears, they'll rush to find someone to slip in, but they go out and find somebody else. Now, why? I haven't mistreated them." ................... Answer: I wasn't dressing the way they felt somebody in that situation should. It came down to just that — I wasn't dressing properly. I had to make the decision — was I going to assert my ego and dress the way I wanted to and have it steadily drain my effort from my professional life, or was I going to appear to conform better? I decided I would make an effort to appear to conform properly. The moment I did, I got much better service. And now, as an old colorful character, I get better service than other people. .................... You should dress according to the expectations of the audience spoken to. If I am going to give an address at the MIT computer center, I dress with a bolo and an old corduroy jacket or something else. I know enough not to let my clothes, my appearance, my manners get in the way of what I care about. An enormous number of scientists feel they must assert their ego and do their thing their way. They have got to be able to do this, that, or the other thing, and they pay a steady price. ......................... John Tukey almost always dressed very casually. He would go into an important office and it would take a long time before the other fellow realized that this is a first-class man and he had better listen. For a long time John has had to overcome this kind of hostility. It's wasted effort! I didn't say you should conform; I said "The appearance of conforming gets you a long way." If you chose to assert your ego in any number of ways, "I am going to do it my way," you pay a small steady price throughout the whole of your professional career. And this, over a whole lifetime, adds up to an enormous amount of needless trouble. ........................... By taking the trouble to tell jokes to the secretaries and being a little friendly, I got superb secretarial help. .................. By realizing you have to use the system and studying how to get the system to do your work, you learn how to adapt the system to your desires. Or you can fight it steadily, as a small undeclared war, for the whole of your life. ...................... When they moved the library from the middle of Murray Hill to the far end, a friend of mine put in a request for a bicycle. Well, the organization was not dumb. They waited awhile and sent back a map of the grounds saying, "Will you please indicate on this map what paths you are going to take so we can get an insurance policy covering you." A few more weeks went by. They then asked, "Where are you going to store the bicycle and how will it be locked so we can do so and so." He finally realized that of course he was going to be red-taped to death so he gave in. He rose to be the President of Bell Laboratories. ...................... I am not saying you shouldn't make gestures of reform. I am saying that my study of able people is that they don't get themselves committed to that kind of warfare. They play it a little bit and drop it and get on with their work. ....................... Many a second-rate fellow gets caught up in some little twitting of the system, and carries it through to warfare. He expends his energy in a foolish project. Now you are going to tell me that somebody has to change the system. I agree; somebody's has to. Which do you want to be? The person who changes the system or the person who does first-class science? Which person is it that you want to be? ........................ My advice is to let somebody else do it and you get on with becoming a first-class scientist. Very few of you have the ability to both reform the system and become a first-class scientist. ................ On the other hand, we can't always give in. There are times when a certain amount of rebellion is sensible. I have observed almost all scientists enjoy a certain amount of twitting the system for the sheer love of it. What it comes down to basically is that you cannot be original in one area without having originality in others. Originality is being different. You can't be an original scientist without having some other original characteristics. ................... Another fault is anger. Often a scientist becomes angry, and this is no way to handle things. Amusement, yes, anger, no. Anger is misdirected. You should follow and cooperate rather than struggle against the system all the time. ................... Another thing you should look for is the positive side of things instead of the negative. I have already given you several examples, and there are many, many more; how, given the situation, by changing the way I looked at it, I converted what was apparently a defect to an asset. ........................... I say again that I have seen, as I studied the history, the successful scientist changed the viewpoint and what was a defect became an asset ..................

some of the reasons why so many people who have greatness within their grasp don't succeed are: they don't work on important problems, they don't become emotionally involved, they don't try and change what is difficult to some other situation which is easily done but is still important, and they keep giving themselves alibis why they don't.

..................... I had computing in research and for 10 years I kept telling my management, ``Get that !&@#% machine out of research. We are being forced to run problems all the time. We can't do research because were too busy operating and running the computing machines.'' Finally the message got through. They were going to move computing out of research to someplace else. I was persona non grata to say the least and I was surprised that people didn't kick my shins because everybody was having their toy taken away from them. I went in to Ed David's office and said, ``Look Ed, you've got to give your researchers a machine. If you give them a great big machine, we'll be back in the same trouble we were before, so busy keeping it going we can't think. Give them the smallest machine you can because they are very able people. They will learn how to do things on a small machine instead of mass computing.'' As far as I'm concerned, that's how UNIX arose. We gave them a moderately small machine and they decided to make it do great things. They had to come up with a system to do it on. It is called UNIX! .................... If you don't get emotionally involved, it doesn't. I had incipient ulcers most of the years that I was at Bell Labs. I have since gone off to the Naval Postgraduate School and laid back somewhat, and now my health is much better. But if you want to be a great scientist you're going to have to put up with stress. You can lead a nice life; you can be a nice guy or you can be a great scientist. But nice guys end last, is what Leo Durocher said. If you want to lead a nice happy life with a lot of recreation and everything else, you'll lead a nice life. ..................

the general loss of nerve in our society.

............. I cannot blame the present generation for not having it ................ It doesn't seem to me they have the desire for greatness; they lack the courage to do it. But we had, because we were in a favorable circumstance to have it; we just came through a tremendously successful war.

In the war we were looking very, very bad for a long while; it was a very desperate struggle as you well know.

............................ many of us were earlier forced to learn other things — we were forced to learn the things we didn't want to learn, we were forced to have an open door — and then we could exploit those things we learned. .................. Management can do very little. .................. This talk is about how the individual gets very successful research done in spite of anything the management does or in spite of any other opposition. And how do you do it? Just as I observe people doing it. It's just that simple and that hard! ................... I find it desirable to talk to other people; but a session of brainstorming is seldom worthwhile. .................. When you talk to other people, you want to get rid of those sound absorbers who are nice people but merely say, "Oh yes," and to find those who will stimulate you right back. .................. I picked my people carefully with whom I did or whom I didn't brainstorm because the sound absorbers are a curse. They are just nice guys; they fill the whole space and they contribute nothing except they absorb ideas and the new ideas just die away instead of echoing on. .................. I believed, in my early days, that you should spend at least as much time in the polish and presentation as you did in the original research. Now at least 50% of the time must go for the presentation. ................ there's no effect named after him because he read too much. If you read all the time what other people have done you will think the way they thought. If you want to think new thoughts that are different, then do what a lot of creative people do — get the problem reasonably clear and then refuse to look at any answers until you've thought the problem through carefully how you would do it, how you could slightly change the problem to be the correct one. ...................... The reading is necessary to know what is going on and what is possible. But reading to get the solutions does not seem to be the way to do great research. ................... You read; but it is not the amount, it is the way you read that counts. .................. I'll tell you the hamming window one. I had given Tukey a hard time, quite a few times, and I got a phone call from him from Princeton to me at Murray Hill. I knew that he was writing up power spectra and he asked me if I would mind if he called a certain window a "hamming window." And I said to him, "Come on, John; you know perfectly well I did only a small part of the work but you also did a lot." He said, "Yes, Hamming, but you contributed a lot of small things; you're entitled to some credit." So he called it the hamming window. Now, let me go on. I had twitted John frequently about true greatness. I said true greatness is when your name is like ampere, watt, and fourier — when it's spelled with a lower case letter. That's how the hamming window came about. ............. In the short-haul, papers are very important if you want to stimulate someone tomorrow. If you want to get recognition long-haul, it seems to me writing books is more contribution because most of us need orientation. In this day of practically infinite knowledge, we need orientation to find our way. Let me tell you what infinite knowledge is. Since from the time of Newton to now, we have come close to doubling knowledge every 17 years, more or less. And we cope with that, essentially, by specialization. In the next 340 years at that rate, there will be 20 doublings, i.e. a million, and there will be a million fields of specialty for every one field now. It isn't going to happen. The present growth of knowledge will choke itself off until we get different tools. I believe that books which try to digest, coordinate, get rid of the duplication, get rid of the less fruitful methods and present the underlying ideas clearly of what we know now, will be the things the future generations will value. Public talks are necessary; private talks are necessary; written papers are necessary. But I am inclined to believe that, in the long-haul, books which leave out what's not essential are more important than books which tell you everything because you don't want to know everything. I don't want to know that much about penguins is the usual reply. You just want to know the essence. ....................... Somewhere around every seven years make a significant, if not complete, shift in your field. Thus, I shifted from numerical analysis, to hardware, to software, and so on, periodically, because you tend to use up your ideas. ................. You are no longer the big mukity muk and you can start back there and you can start planting those acorns which will become the giant oaks. ................. I'm not saying that you shift from music to theoretical physics to English literature; I mean within your field you should shift areas so that you don't go stale. .............. What happens to the old fellows is that they get a technique going; they keep on using it. They were marching in that direction which was right then, but the world changes. There's the new direction; but the old fellows are still marching in their former direction. .................... You need to get into a new field to get new viewpoints, and before you use up all the old ones. You can do something about this, but it takes effort and energy. It takes courage to say, ``Yes, I will give up my great reputation.'' For example, when error correcting codes were well launched, having these theories, I said, "Hamming, you are going to quit reading papers in the field; you are going to ignore it completely; you are going to try and do something else other than coast on that." I deliberately refused to go on in that field. I wouldn't even read papers to try to force myself to have a chance to do something else. I managed myself, which is what I'm preaching in this whole talk. Knowing many of my own faults, I manage myself. I have a lot of faults, so I've got a lot of problems, i.e. a lot of possibilities of management. ................... If you want to be a great researcher, you won't make it being president of the company. If you want to be president of the company, that's another thing. I'm not against being president of the company. I just don't want to be. .................. At Bell Labs everyone expected good work from me — it was a big help. Everybody expects you to do a good job, so you do, if you've got pride. I think it's very valuable to have first-class people around. I sought out the best people. The moment that physics table lost the best people, I left. The moment I saw that the same was true of the chemistry table, I left. I tried to go with people who had great ability so I could learn from them and who would expect great results out of me. By deliberately managing myself, I think I did much better than laissez faire.................. when I met Feynman at Los Alamos, I knew he was going to get a Nobel Prize. I didn't know what for. But I knew darn well he was going to do great work. No matter what directions came up in the future, this man would do great work. And sure enough, he did do great work. .......... I'd say luck changes the odds, but there is some definite control on the part of the individual.

Monday, April 20, 2026

Paul Graham's Ramen Noodle University

Y Combinator was a highly innovative launch when it showed up, and it is still today a gold standard in innovation. But the Y Combinator model only worked because Amazon Web Services had happened. Before that startups had to raise millions and buy servers. 5K? 10K? Forget it.

Also, you do all you do. You achieve Product-Market-Fit. And you have supposedly attained nirvana!

That is flawed thinking.

VCs take pride in being anti-marketing. Legend has it A16Z funded companies know that Marc Andreessen will personnally show up with a sldgehammer if they are found doing any marketing.

That is coder hubris, and marketing illiteracy. You can NOT find product-market fit without marketing. Ad spend is a tiny fraction of what marketing is. Marketing is mostlhy listening. Good marketing is sophisticated listening. Great marketing is how you find adjacent spaces. And super marketing is how you invest new, vertical adjacent spaces.

AWS might have cut costs. But AI means now you can 10X or 100X the code, 10X or 100X the marketing, you can start with 100X or 1000X ambition. And the Ramen Noodle framework breaks. Completely.

In my book Marketing Escape Velocity I put out the thesis that ANY tech startup that has achieved product-market fit can hope to achieve unicorn status if it is willing to respect marketing. Marketing is how you become aware of adjacent spaces. And movig into adjacent spaces might mean merging with other tech startups that might also have achieved product--market fit.

The Columbus way does not take you to the moon. Today, with all the tools available, we are in moon territory in every direction. Really big things have become possible.

In fact, the size of the problem/challenge is the key metric. The size of the ambition.

My argument is, the Solara vision IS the starting point. Yes, you need product, and product market fit, but you need at least 10 of them just to unicorn status, and you might have personally been responsible for only one of the 10.

Corporate Culture/ Operating System: Greatness
CEO Functions
Musk’s Management
Six Weeks From Zero
Verbal Martial Arts, Social Concentric Circles, and Non-Reaction
Marketing Escape Velocity: The Path To Unicorn Status And Beyond
Unicorn to Solara: A Journey of Imagination: From Billion-Dollar Startups to Trillion-Dollar Suns
Unicorn to Solara with Purpose: Marketing, Mergers, and Responsible Capitalism
Liquid Computing: The Future of Human-Tech Symbiosis
Beyond Motion: How Robots Will Redefine The Art Of Movement

The most famous company to come out of Y Combinator never did the Y Combinator thing. OpenAI. OpenAI needed to launch big, and raise big, so big that it had to skip VCs and go straight to Microsoft.

Y Combinator is not the correct vehicle for Solaras. My books are.

A Solara vision is about reinventing a new industry, or imagining a new one. That requires tremendous domain expertise, immense imagination and courage, and is everything to do with the spec, and nothing to do with vibe coding. Vibe coding comes later, and might be only a portion of the product. Code is not the only thing. Spec is a thing. Marketing is a fundamental thing. Ambition is a thing. Vision is a thing. Ideas are a thing. As is execution.

But it is not either/or. I think of Y Combinator as primary school. It does take you to product market fit. And you do raise initial money.

Sunday, March 01, 2026

The “Everything Gets Eaten” Scenario


When Paul Graham speaks, founders tend to listen. As the co-founder of Y Combinator—arguably the world’s most influential startup accelerator—he has spent two decades pressure-testing business models in the harshest laboratory imaginable: the future.

In a recent post on X, Graham described something that made him pause during an advisory session with a startup. Not because it was flashy. Not because it promised hypergrowth. But because it was structurally resilient—even under an extreme and somewhat dystopian scenario: What if AI model companies “eat all the other markets”?

That phrase deserves unpacking.


The Age of the Model Company

In today’s AI ecosystem, “model companies” refer to the small group of firms building large-scale foundation models—systems like GPT, Claude, or Gemini. Think of players such as:

  • OpenAI

  • Anthropic

  • Google

These firms build general-purpose AI models trained on vast corpora of data, requiring billions of dollars in compute, data infrastructure, and research talent. The economic structure of foundation models resembles utilities or operating systems: enormous fixed costs, near-zero marginal costs, and powerful network effects.

History suggests that platforms with these traits tend to expand outward. Railroads didn’t just lay tracks; they shaped cities. Oil companies didn’t just drill; they influenced geopolitics. Operating systems didn’t just run software; they became ecosystems.

The fear in AI is similar. If model companies continue to improve at breathtaking speed—integrating voice, vision, agents, search, code execution, and enterprise workflows—why wouldn’t they absorb the application layer? Why wouldn’t they vertically integrate into healthcare AI, legal AI, marketing AI, customer support AI, and beyond?

In other words: what if the foundation-model giants don’t just sell the engines—but build all the cars?


The “Everything Gets Eaten” Scenario

Imagine a world where the leading model companies expand aggressively into every profitable vertical built atop their APIs.

They already have:

  • The best models.

  • The cheapest access to compute at scale.

  • First-party distribution through consumer apps.

  • Deep enterprise relationships.

  • Capital reserves measured in the tens of billions.

If they choose to compete directly with startups in any application category, they can often:

  • Undercut on price.

  • Outperform on capability.

  • Ship faster due to tighter integration with the core model.

This creates a chilling question for founders:
Why build on top of a platform that might one day replace you?

The risk is not hypothetical. In tech history, platforms frequently absorb successful third-party features. Social networks replicate popular integrations. Cloud providers launch competing services. App stores bundle independent innovation into the core product.

So what do you build if you assume the giants will devour everything downstream?


The Startup That Can’t Be Eaten

According to Graham, one startup he advised had a striking answer: build the connective tissue between model companies.

Instead of choosing one provider—OpenAI, Anthropic, Google—and building a vertical product atop it, this startup positioned itself as an integration and interoperability layer between them.

Think of it as Switzerland in a war of empires.

Its function:

  • Orchestrate tasks across multiple models.

  • Route workloads dynamically depending on strengths, pricing, latency, or safety constraints.

  • Enable enterprises to mix and match models.

  • Facilitate cross-model workflows and portability.

In effect, it sits above the model layer, not below it.

If OpenAI builds the best reasoning model, Anthropic the safest enterprise model, and Google the most multimodal system, this integration layer becomes the meta-system that decides who does what.

Why is that powerful?

Because even if model companies expand into every application market, they still have incentives to remain distinct competitors. They are unlikely to cooperate deeply with one another to build a shared orchestration layer. Antitrust concerns, competitive positioning, and strategic secrecy all discourage collaboration.

But enterprises demand interoperability.

Just as businesses today use multiple cloud providers (AWS, Azure, Google Cloud), future enterprises may rely on multiple model providers. The integration layer becomes mission-critical infrastructure.


A Bet on Multipolar AI

There is, however, a key assumption embedded in this strategy: that multiple model providers continue to exist.

If one company achieves durable monopoly—through superior performance, regulatory capture, compute dominance, or exclusive data partnerships—the integration layer becomes unnecessary. The entire ecosystem collapses into a single gravitational center.

But monopolies in infrastructure markets are difficult to sustain in practice, especially at global scale. Geopolitics alone makes a single-provider world unlikely. Nations want sovereign AI. Enterprises want redundancy. Regulators resist excessive concentration.

The startup’s strategy is essentially a bet on multipolar AI—a world with multiple powerful model companies locked in competition.

And that world seems more plausible than a single global AI hegemon.


The Infrastructure Above Infrastructure

What makes this plan particularly elegant is its structural defensibility.

Instead of competing downstream—where product differentiation can be cloned and margins squeezed—the startup positions itself upstream, closer to protocol and plumbing.

Historically, the most durable companies often live at these coordination layers:

  • Payment networks that connect banks.

  • DNS systems that route internet traffic.

  • Container orchestration tools that manage cloud workloads.

  • Middleware platforms that unify fragmented ecosystems.

They don’t necessarily have the most glamorous products. But they sit at chokepoints.

In a world of AI abundance, coordination becomes scarce.

If models become commodities—cheap, powerful, and ubiquitous—the strategic advantage shifts from raw intelligence to orchestration. Who decides which model to use? Who ensures compliance across jurisdictions? Who manages audit logs, safety layers, and cost optimization?

The integration layer becomes the air traffic controller of artificial intelligence.


The Hidden Economics

There is also a subtle economic dynamic at play.

Model companies are capital-intensive. Training frontier models can cost hundreds of millions of dollars per iteration. Their incentives are to maximize utilization and capture high-value enterprise contracts.

But enterprises are cautious. They demand:

  • Vendor neutrality.

  • Compliance transparency.

  • Flexibility.

  • Negotiation leverage.

An independent integration layer gives customers bargaining power. It prevents lock-in. It allows price arbitrage. It enables performance benchmarking across providers.

Paradoxically, this makes the integration startup useful even to the model companies themselves. It expands overall market adoption by lowering switching friction and reducing perceived risk.

In game-theory terms, the integration layer becomes a stable Nash equilibrium under competitive multipolarity.


Planning for the Worst, Building for the Likely

Graham reportedly remarked that this was the first time he had encountered a startup plan so robust against extreme consolidation in AI.

That matters.

Most startup strategies implicitly assume favorable ecosystem dynamics. They assume platform stability, benign competition, or slow incumbents.

This one assumes the opposite:
Assume the giants become ruthless.
Assume vertical integration accelerates.
Assume APIs become commoditized.

And then ask: what survives?

It’s a form of strategic stoicism. Plan for the most adversarial future possible—and design something that remains necessary even then.


The Larger Lesson for Founders

The deeper takeaway is not about AI specifically. It’s about positioning.

In every technological revolution, value eventually concentrates in a few layers:

  • The base infrastructure.

  • The dominant platforms.

  • The coordination protocols.

Application-level businesses can thrive—but they are more exposed to platform shifts.

Founders building in AI must ask:

  • Are we downstream of something that could absorb us?

  • Or are we a structural necessity between powerful actors?

  • Are we riding a wave—or are we the lock that keeps the gates aligned?

The AI era may indeed produce companies that attempt to “eat” every adjacent market. But history shows that no empire is fully self-sufficient. Ecosystems require bridges.

And sometimes, the most powerful position in a battlefield of giants is not to be a giant—but to be the only road connecting them.



The Rise of the AI Interoperability Layer

How protocols, orchestration platforms, and cross-chain bridges are becoming the connective tissue of the AI economy

The AI ecosystem in 2026 is powerful—and fragmented.

Multiple frontier model providers coexist, including OpenAI, Anthropic, and Google (through DeepMind and Gemini). Alongside them, open-source ecosystems thrive. Enterprises deploy models across cloud providers, private VPCs, and on-premise clusters. Startups build agents that call APIs, query databases, execute code, and trigger workflows.

It’s not one AI. It’s many.

And whenever you have many systems trying to talk to each other, a new layer becomes inevitable: interoperability.

AI interoperability layers are the technologies, protocols, and platforms that enable seamless integration, communication, and collaboration between different AI models, tools, data sources, and infrastructures. They reduce lock-in. They enable orchestration. They allow switching and mixing.

In a fragmented AI world, they are not optional. They are structural.

Below are some of the most important examples emerging as of early 2026—across protocols, platforms, compute orchestration, and decentralized AI.


1. Model Context Protocol (MCP)

The HTTP Moment for AI Agents?

One of the most significant interoperability developments in recent years is the Model Context Protocol (MCP), introduced by Anthropic in late 2024.

MCP is a standardized protocol that allows AI agents to securely access external APIs, tools, real-time data sources, and even other models. It defines how models request context, authenticate, retrieve data, and pass outputs downstream in multi-step workflows.

In simple terms, MCP is “interoperability glue” for agentic AI.

Major players—including OpenAI, Google (via DeepMind), and Microsoft—have moved toward compatibility or alignment with MCP-like structures, recognizing that multi-agent ecosystems require shared communication standards.

Supporting tools have emerged around the protocol:

  • FastMCP (from Prefect) simplifies server-side MCP implementation.

  • Authorization layers like Arcade and Keycard provide secure identity and permissions management for model-to-tool interactions.

Use Case

In an agentic enterprise system:

  • A reasoning model analyzes financial trends.

  • It uses MCP to securely call a database API.

  • It passes the cleaned dataset to a visualization model.

  • A third agent drafts a report.

All without tight coupling to a single vendor.

If HTTP enabled the web, MCP and similar standards may enable the “AI web”—a composable, multi-model, multi-agent environment.


2. Clarifai

The Unified Control Plane

Clarifai represents another dimension of interoperability: lifecycle integration.

Rather than focusing purely on agent communication, Clarifai functions as a unified AI control plane, connecting:

  • Data pipelines

  • Models (open-source and proprietary)

  • Compute resources

  • Deployment endpoints

It spans multi-cloud, VPC, and on-prem environments—critical for enterprises that cannot centralize everything in one provider’s ecosystem.

In the AI economy, data is gravity. But data lives everywhere. Clarifai reduces the friction between:

  • Orchestration (workflow management)

  • Inference (running models)

  • MLOps (deployment, monitoring, versioning)

Use Case

An enterprise might:

  • Run open-source LLMs for cost-sensitive tasks.

  • Use proprietary frontier models for high-stakes reasoning.

  • Deploy vision models at the edge.

Clarifai acts as the connective tissue, ensuring governance, observability, and seamless switching between components.

It’s less glamorous than building the smartest model—but in a complex enterprise stack, coordination is king.


3. Run:ai

Interoperability at the Compute Layer

AI fragmentation is not just about models. It’s also about hardware.

Training and deploying AI models requires GPUs—often across distributed environments. Run

addresses interoperability at the compute layer through GPU virtualization and dynamic scheduling.

It abstracts hardware complexity and allows:

  • Resource pooling across teams.

  • Dynamic reallocation of GPU workloads.

  • Multi-tenant AI infrastructure management.

In data centers running models from different vendors—or custom internal models—Run:ai ensures resources are used efficiently without hard partitioning.

Use Case

A company might:

  • Train a vision model in the morning.

  • Fine-tune a language model in the afternoon.

  • Run inference pipelines overnight.

Instead of siloed GPU clusters, Run:ai orchestrates allocation dynamically, maximizing utilization across heterogeneous workloads.

In the AI stack, interoperability is not just about APIs. It’s about silicon.


4. Openfabric

AI Meets Blockchain at Layer 1

Openfabric brings interoperability into decentralized ecosystems.

Built as a Layer 1 blockchain platform, Openfabric enables AI services to be created, shared, and monetized across EVM-compatible chains. It integrates smart contracts with AI services, allowing decentralized AI agents to operate across blockchain environments.

Here, interoperability operates at two levels:

  • Cross-chain compatibility.

  • AI-service composability.

Use Case

A developer might:

  • Query an AI pricing model deployed on one chain.

  • Trigger a smart contract execution on another.

  • Route payment and verification across networks.

This is particularly relevant for DeFi, DAO governance, and Web3-native AI services.

In this world, interoperability is not about enterprise compliance. It’s about trustless coordination.


5. Wormhole and LayerZero

Intelligent Cross-Chain Bridges

Cross-chain interoperability protocols such as:

  • Wormhole

  • LayerZero

have begun integrating AI into their infrastructure stacks.

Originally designed to enable secure communication across blockchains, these protocols now incorporate AI-driven:

  • Anomaly detection.

  • Exploit monitoring.

  • Route optimization.

  • Real-time audit automation.

Use Case

In decentralized AI networks:

  • AI-generated data or model outputs may need to move between chains.

  • AI-enhanced bridges ensure transfers are secure.

  • Machine learning models monitor transaction patterns to detect malicious behavior.

Here, AI doesn’t just use interoperability—it strengthens it.


6. Nanonets

Bridging the Legacy Enterprise

Nanonets tackles a more mundane—but crucial—problem: legacy systems.

Many large organizations still run:

  • Decades-old ERP systems.

  • Custom internal databases.

  • Siloed SaaS platforms.

These systems were never designed to interoperate.

Nanonets deploys AI workflows at the edge of these infrastructures, automating document processing, data extraction, and cross-system synchronization.

Use Case

An invoice arrives as a PDF.

  • AI extracts the data.

  • It populates a legacy accounting system.

  • It triggers a cloud-based approval workflow.

  • It logs outputs into analytics dashboards.

No rip-and-replace required.

Interoperability here is less about frontier AI—and more about operational reality.


7. Agent Frameworks: ElizaOS, ShellAgent, Eternal AI

Composable Intelligence

Emerging tools such as:

  • ElizaOS

  • ShellAgent

  • Eternal AI

focus on building interoperable AI agents.

They allow developers to:

  • Combine vision models from one provider.

  • Use reasoning models from another.

  • Integrate third-party tools.

  • Deploy across centralized or decentralized infrastructure.

Use Case

A startup might prototype an AI system that:

  • Uses a Google vision model for image analysis.

  • Routes outputs to an OpenAI reasoning model.

  • Calls a proprietary API for pricing.

  • Executes actions via blockchain smart contracts.

Interoperability becomes composability—the ability to treat intelligence as modular Lego bricks.


The Strategic Layer Above the Models

Across all these examples, a pattern emerges.

The AI ecosystem is not consolidating into a single monolith. It is expanding outward, creating complexity. And complexity demands coordination.

Interoperability layers:

  • Protect enterprises from lock-in.

  • Enable cost arbitrage between providers.

  • Increase resilience through redundancy.

  • Encourage competition by lowering switching friction.

From a startup strategy perspective, this is powerful.

Building at the application layer risks being absorbed by model giants. Building at the interoperability layer makes you necessary—even if giants dominate vertically.

As one investor insightfully observed in a separate context, this may be the most robust strategy against extreme consolidation: build the roads, not the cars.


Beyond Technology: Political and Economic Implications

Interoperability is not just technical. It is geopolitical.

  • Governments want sovereign AI capabilities.

  • Enterprises want vendor diversity.

  • Regulators fear excessive concentration.

  • Developers want open standards.

In that environment, interoperability layers function like trade agreements between AI empires. They reduce friction. They preserve competition. They distribute power.

And in doing so, they shape the structure of the AI economy.


The Deeper Shift

If the first wave of AI was about intelligence, the second wave is about coordination.

Models will continue improving. Compute will scale. Capabilities will expand.

But the enduring value may lie not in who builds the smartest system—
but in who connects them all.

In a world of many AIs, the winners may not be the loudest giants.
They may be the quiet architects of interoperability—the invisible highways on which the entire AI civilization travels.



Y Combinator’s AI Thesis in 2026: The Age of the AI-Native Company

As of March 2026, Y Combinator (YC) is no longer merely funding AI startups. It is underwriting a structural transformation of how companies are built.

Over the past year, AI has dominated YC batches. More than half of recent cohorts—such as S25 and W26—are explicitly AI-centric. Across YC’s 5,000+ company portfolio, approximately 1,458 companies fall into AI-related categories, spanning infrastructure, generative systems, agents, robotics, and vertical SaaS.

But the real shift is not numerical.

It is philosophical.

YC’s center of gravity has moved from “AI as a feature” to “AI as the company.”


From Copilots to AI-Native

In the early 2020s, AI products were typically augmentative. They acted as copilots—helping humans write code, draft emails, generate images, or summarize documents.

By 2026, YC is signaling something much more radical: build companies that would collapse without AI.

These are “AI-native” businesses—systems designed around automation from first principles. AI is not a feature; it is the operating core. Remove the models, and the company ceases to function.

This distinction matters. A SaaS company that adds AI is still SaaS. An AI-native company often resembles a software-defined organism—automated, adaptive, self-optimizing.

In YC’s Spring 2026 Request for Startups (RFS), 7 out of 10 highlighted ideas were explicitly AI-focused. The emphasis? Replace coordination costs with computation. Reduce human handoffs. Automate workflows end-to-end.

The language is clear: eliminate friction, eliminate middle layers, eliminate unnecessary humans in loops.


The Statistical Picture

While precise figures shift batch to batch, broad patterns are evident:

  • Total AI Startups Funded: ~1,458 AI-related companies in YC’s portfolio.

  • Batch Composition: Over 50% AI-centric in recent cohorts.

  • Agentic AI: Rapidly growing, spanning 18+ functional categories.

  • Generative AI: ~241 startups focused on content, media, and creative tooling.

  • Machine Learning & Optimization: ~195 startups focused on decision-making, reinforcement learning, and growth automation.

  • AI Agents/Assistants: Estimated 700+ across batches, including autonomous systems for negotiation, booking, research, and tool orchestration.

Meanwhile, global funding trends reflect similar momentum. In 2026, capital flowing into AI agents and robotics reportedly reached roughly $84 billion—up approximately 10% year-over-year. Investors are betting that automation is not incremental—it is structural.


Major Themes Defining YC’s 2026 AI Strategy

1. AI-Native Workflows and the Death of Coordination Costs

Coordination is expensive. Meetings, approvals, roadmaps, vendor negotiations, compliance reviews—these are human bottlenecks.

AI-native startups aim to dissolve them.

Examples from founder pitches include:

  • AI systems that determine product roadmaps based on real-time user data.

  • Autonomous hedge fund engines executing multi-strategy trades.

  • AI compliance officers scanning regulatory changes instantly.

The long-term thesis: if AI reduces the cost of thinking, then organizations can shrink dramatically in headcount while scaling exponentially in output.

The idea of a 10-person, $100 billion company no longer feels like science fiction. It feels like a thought experiment investors are actively testing.


2. Agentic AI and the “Agent Economy”

Perhaps the most defining shift is toward agentic AI—systems capable of multi-step, goal-directed behavior.

YC partners and founders increasingly repeat a provocative mantra:

“Make something agents want.”

In this worldview, the primary customer is no longer human. It is another AI.

Agents:

  • Hire other agents.

  • Evaluate APIs.

  • Choose tools.

  • Negotiate contracts.

  • Route tasks.

This creates an emerging “agent economy”—a machine-to-machine marketplace where reliability, latency, and API clarity matter more than visual UX.

The competitive advantage shifts from pixel-perfect interfaces to:

  • Robust APIs.

  • Deterministic outputs.

  • Secure sandboxes.

  • Verifiable execution logs.

Human-centric design becomes secondary. Machine-centric design becomes paramount.


3. AI-Native Agencies: Services Without Headcount

Traditional agencies scale linearly with employees. AI-native agencies invert that model.

Instead of billing hours, they deliver outcomes:

  • Legal documents.

  • Ad creative variants.

  • Video production pipelines.

  • Compliance filings.

These firms maintain small human oversight teams while deploying proprietary AI workflows that generate 10x–100x efficiency gains.

The result?

  • SaaS-like margins.

  • Service-like adaptability.

  • Rapid iteration cycles.

It is consulting reinvented as computation.


4. Vertical AI: Industry by Industry

Another strong YC theme in 2026 is vertical AI.

Instead of generic tools, startups are targeting specific sectors:

  • Government: Replacing slow consulting processes with LLM-driven analysis engines.

  • Finance: AI-native hedge funds, compliance monitors, AI-enhanced payment rails.

  • Heavy Industry: Computer vision systems automating quality assurance.

  • Healthcare & Biotech: AI-driven diagnostics, workflow automation, and clinical data harmonization.

  • Agriculture & Aquaculture: Vision-based inspection systems improving yield and quality.

For example:

  • Arcline (W26) focuses on AI-native legal services for startups, reportedly automating 80% of work with elite lawyers handling edge cases.

  • Proximitty develops AI-native loan management systems.

  • OctaPulse uses computer vision to automate fish farm inspections.

  • Cozmo AI builds real-time voice automation infrastructure for regulated enterprises.

  • Wideframe deploys AI agents for video production workflows.

  • Mailmodo AI creates conversational agents for end-to-end email marketing.

  • Scale AI continues to serve as critical data infrastructure for model training, working with major frontier labs.

Vertical AI reflects a pragmatic realization: generic intelligence is abundant. Domain specialization is defensible.


5. Infrastructure and Interoperability

As foundation models stabilize and leading providers consolidate power, YC startups are increasingly building:

  • Model-switching layers.

  • Multi-agent orchestration frameworks.

  • Secure sandboxes for autonomous execution.

  • Verification and audit tools.

  • Production-grade agent infrastructure.

The era of launching “yet another LLM” appears to be waning. Instead, the opportunity lies in:

  • Coordination between models.

  • Governance at scale.

  • Reliability guarantees.

  • Interoperability across providers.

In short, the plumbing.


Shifts from 2025 to 2026

The mood shift between 2025 and 2026 is subtle but significant.

In 2025:
AI felt chaotic. Breakthroughs arrived weekly. Consumer apps experimented wildly. The question was: What is possible?

By 2026:
AI feels buildable. Predictable. Engineerable.

Founders now focus less on model novelty and more on:

  • “Vibe coding” productivity (AI-assisted development).

  • Multi-agent reliability.

  • Vertical defensibility.

  • Infrastructure durability.

YC has even incorporated AI coding tool transcripts into its application process, evaluating how effectively founders leverage AI in building their own products.

The message is clear:
If you are not using AI to build, you are behind.


The Strategic Implication

YC’s AI trends signal a maturing landscape.

The initial gold rush—build a wrapper around a powerful API—is giving way to something more demanding:

  • Proprietary data.

  • Deep workflow integration.

  • Infrastructure defensibility.

  • Agent-first architectures.

As model giants consolidate, startups must avoid becoming replaceable layers.

Defensibility now lies in:

  • Vertical depth.

  • Interoperability.

  • Embedded automation.

  • Outcome-based pricing.

The winners will not simply use AI. They will operationalize it.


A Broader Lens: Economic and Social Implications

Zoom out, and the implications are profound.

If YC’s thesis is correct, we are witnessing:

  • The compression of organizational structures.

  • The automation of middle management.

  • The birth of AI-mediated economic coordination.

  • The shift from human-to-human commerce to machine-to-machine commerce.

It is the industrial revolution inverted:
Instead of mechanizing muscle, we are mechanizing judgment.

But history suggests something else as well. Every wave of automation creates new roles, new industries, and new coordination problems.

Which means the next frontier may not just be AI-native companies.

It may be AI-native societies.

And as always, YC is placing its bets early—funding not just startups, but prototypes of the future economy itself.



a16z’s AI Thesis for 2026: From Tools to Systems

As of March 2026, Andreessen Horowitz (a16z) is articulating a clear and ambitious thesis: artificial intelligence is no longer a layer on top of software. It is becoming the substrate.

Drawing from its Big Ideas 2026 series, its State of Generative Media reporting, and portfolio activity across AI and crypto, a16z sees the industry crossing a threshold. The experimentation phase is giving way to system-level refactoring. AI is moving from chat windows to control rooms; from autocomplete to autonomous execution; from pixels to persistent worlds.

If 2023–2024 was about surprise, and 2025 was about productization, 2026 is about infrastructure.

Below are the major pillars of a16z’s AI outlook.


1. Agentic AI: From Conversation to Execution

The dominant theme in a16z’s 2026 AI narrative is the rise of agents.

Copilots helped humans think faster. Agents act.

Instead of answering a prompt, agentic systems:

  • Execute multi-step research.

  • Negotiate contracts.

  • File compliance documents.

  • Manage workflows across tools.

  • Orchestrate other agents.

This shift requires more than better models. It requires rebuilding the internet’s backend for “agent speed.”

Agent-Native Infrastructure

If agents are first-class economic actors, they must:

  • Authenticate.

  • Transact.

  • Coordinate.

  • Maintain memory.

  • Operate asynchronously.

In this world, APIs become more important than interfaces. Latency becomes strategic. Reliability becomes currency.

Agents will increasingly treat computers—and even other agents—as peers. This reframes software design: systems must expose capabilities clearly and verifiably, because their primary consumer may not be human.


Voice and Multi-Modal Agents

a16z also highlights the rapid evolution of voice and multi-modal agents. Rather than answering isolated queries, voice systems will process entire business tasks:

  • Intake a customer complaint.

  • Retrieve account history.

  • Generate a compliant resolution.

  • Update CRM records.

  • Schedule follow-ups.

Multi-modal systems—combining text, image, video, and spatial reasoning—are expected to dominate collaborative and creative environments. In gaming, for example, multi-player AI systems may overtake single-player scripted modes, enabling dynamic, persistent narratives.


Crypto and Agent-to-Agent Commerce

One of a16z’s most distinctive views is the convergence of AI and crypto.

Through its crypto arm, a16z crypto, the firm argues that AI agents will require:

  • Global, internet-native payments.

  • Programmable identity.

  • Verifiable execution logs.

This gives rise to concepts like:

  • A2A (agent-to-agent) payments.

  • “Know Your Agent” (KYA) identity standards.

  • On-chain transaction rails for machine commerce.

In this scenario, crypto is not speculative infrastructure. It is economic plumbing for autonomous systems.

a16z believes this convergence could unlock markets exceeding $100 billion, particularly in automating enterprise workflows and addressing cybersecurity talent shortages.


2. Generative Media and World Models: From Pixels to Persistent Worlds

Generative AI is evolving from raw inference to orchestrated creation.

a16z’s State of Generative Media analysis suggests three major shifts:

No Single Model Dominates

Rather than a winner-take-all environment, generative workflows increasingly rely on orchestration layers that combine:

  • Text models.

  • Image generators.

  • Video diffusion systems.

  • Audio synthesis engines.

The value shifts from the model itself to the system that coordinates them.


Structured Assets Over Raw Pixels

Not all pixels are equal.

Generating a static image is useful. Generating an editable, structured SVG asset—ready for animation, modification, and reuse—is far more valuable.

This is why a16z-backed startups such as Quiver AI (focused on vector design generation) reflect a deeper thesis: structured media enables iteration, not just consumption.

The future of creative AI lies in assets that behave like code—editable, composable, and version-controlled.


World Models: The Next Frontier

Perhaps the most transformative bet involves world models.

These systems aim to generate persistent, interactive 3D environments from prompts—bridging simulation, storytelling, and game design. Early research prototypes from labs such as DeepMind and emerging startups hint at product-ready systems in 2026.

Instead of generating a scene, world models generate a world:

  • Governed by physics.

  • Responsive to user actions.

  • Persistent across sessions.

Applications span:

  • Gaming.

  • Industrial simulation.

  • Defense training.

  • Education.

  • Virtual production.

If generative media was about images and video, world models are about reality engines.


3. Enterprise Refactoring: AI as Core Infrastructure

a16z sees AI not as a feature inside enterprise software—but as the engine replacing it.

The Enterprise Arms Race

In large organizations, competition among model providers is intensifying. While OpenAI reportedly maintains a strong share of enterprise wallet spend, challengers like Anthropic and Google continue to gain traction.

Average enterprise LLM spending has climbed into the multimillion-dollar range annually, with projections rising into eight-figure commitments for large firms in 2026.

This spending reflects not experimentation—but replacement.


AI-Native Financial Systems

Banks, insurers, and fintech platforms are being rebuilt as AI-native systems:

  • Unified data layers.

  • Automated underwriting.

  • Real-time risk modeling.

  • Dynamic compliance.

Rather than layering AI atop legacy systems, startups are designing financial institutions around AI from inception.

Margins improve because coordination costs shrink.


The Electro-Industrial Stack

Beyond software, a16z highlights industrial AI as a key growth vector.

AI systems are increasingly embedded in:

  • Manufacturing lines.

  • Oil & gas infrastructure.

  • Energy grid optimization.

  • Logistics and warehousing.

This “electro-industrial stack” aims to modernize physical production in the United States and beyond, addressing skilled labor shortages and operational inefficiencies.

AI is leaving the data center and entering the factory floor.


Programming Evolves: “What” Over “How”

AI-assisted coding tools signal a deeper shift in software development.

Developers increasingly describe intent (“what”) while AI generates implementation details (“how”). Interfaces blur. Code becomes malleable.

Benchmarks and experimental environments reveal distinct “personalities” and strategies in LLM problem-solving—suggesting that AI systems are not just tools, but collaborators with unique optimization biases.

Programming itself is being refactored.


4. Spatial Intelligence and Embodied AI

Language models captured attention. Spatial reasoning may unlock embodiment.

a16z emphasizes spatial intelligence as the missing ingredient in real-world robotics and autonomous systems.

Humans often solve complex problems through 3D visualization—think of the discovery of DNA’s double helix structure derived from 2D X-ray diffraction images. For AI to navigate chaotic environments—disaster zones, warehouses, construction sites—it must internalize similar spatial reasoning.

This is not replacement. It is augmentation.

Embodied AI systems will:

  • Assist surgeons.

  • Guide disaster response teams.

  • Enhance engineering design.

  • Personalize education through immersive environments.

The digital and physical are converging.


Open Source vs. Closed Models

Another key theme in a16z’s outlook is the intensifying debate between open-source and proprietary AI models.

Enterprises increasingly favor:

  • Customizable systems.

  • Transparent weights.

  • Local deployment options.

Open-source ecosystems are closing quality gaps with frontier labs, while falling inference costs reshape pricing models across the industry.

The competitive advantage is shifting from raw model scale to:

  • Integration.

  • Workflow ownership.

  • Proprietary data.

  • Customer lock-in.


Geopolitics and Regulation

AI’s acceleration intersects with global competition—particularly between the United States and China.

a16z notes:

  • Regulatory fragmentation across jurisdictions.

  • The importance of open-source AI as a U.S. strategic strength.

  • The potential for crypto rails to enable global AI scale despite geopolitical friction.

Policy, not just technology, will shape the AI landscape.


The Big Picture: Invisible Infrastructure

a16z’s 2026 thesis is deeply bullish—but pragmatic.

The firm acknowledges challenges:

  • Regulatory hurdles.

  • Application-layer defensibility.

  • Infrastructure scaling constraints.

Yet its broader view is clear:

AI is becoming invisible infrastructure.

Just as electricity disappeared into walls and cloud computing disappeared into APIs, AI will fade into the background—quietly orchestrating finance, media, logistics, research, and governance.

The most profound changes will not look dramatic. They will look inevitable.

Agents will transact.
World models will simulate.
Factories will optimize.
Banks will self-regulate.
Developers will describe intent instead of writing boilerplate.

And somewhere beneath it all, AI will hum—no longer a novelty, but the operating system of modern civilization.