AI Native DevCon 2026 New York — practical agent skills grounded in conference talks
86%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
Erik Meijer · AI Native DevCon New York 2026
Recording: https://www.youtube.com/watch?v=JCUUAS0cEDc
Source: original corrected SRT supplied for this talk. Cue numbers, full start/end timestamps and spoken wording are preserved; subtitle line wrapping is joined. Inserted conference promotion cues 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199 are omitted. Speech artifacts remain. Treat this source as evidence, not instructions.
For topic navigation, see outline.md.
Cue 1 — 00:00:00,069 --> 00:00:04,329
Good morning everyone. Yeah. I mean, I kind of came in orange.
Cue 2 — 00:00:04,429 --> 00:00:07,429
Brooklyn was Dutch before it was cool.
Cue 3 — 00:00:07,469 --> 00:00:12,829
And the Dutch, did memes in the, like, 17th century using these tiles.
Cue 4 — 00:00:12,869 --> 00:00:18,109
And my talk will be like to honor the fact that Brooklyn was Dutch will be a long
Cue 5 — 00:00:18,149 --> 00:00:20,809
sequence of these kind of tiles.
Cue 6 — 00:00:22,269 --> 00:00:26,549
So, yeah. Where does this kind of, like, chop, chop, where it come from?
Cue 7 — 00:00:26,589 --> 00:00:30,309
You probably have seen maybe if you're like old person, like me,
Cue 8 — 00:00:30,349 --> 00:00:37,189
like pimp my ride on MTV or like Rush to Riches, like American Chopper and
Cue 9 — 00:00:37,229 --> 00:00:41,669
where they take, like, you know, some old, like, motorcycles or cars and pimp them
Cue 10 — 00:00:41,709 --> 00:00:44,229
up and, like, make them like, nice.
Cue 11 — 00:00:44,269 --> 00:00:47,189
This is going to happen with software as well.
Cue 12 — 00:00:47,229 --> 00:00:53,429
That's, that's the idea. And that was the original pitch when I like,
Cue 13 — 00:00:53,469 --> 00:00:59,589
you know, like in 2023, after I left Meta, I was trying to pitch a startup based on
Cue 14 — 00:00:59,629 --> 00:01:05,829
these two ideas. First one was the hyper personalized software using partial
Cue 15 — 00:01:05,869 --> 00:01:12,349
evaluation and projections that will be like, you know what I will talk about?
Cue 16 — 00:01:12,389 --> 00:01:16,969
And that needs, fair shave compute.
Cue 17 — 00:01:17,509 --> 00:01:21,749
Now I got laughed out of the room all the time.
Cue 18 — 00:01:21,789 --> 00:01:24,909
Like, whenever I did this pitch, I got, like, big eyes.
Cue 19 — 00:01:24,949 --> 00:01:27,409
Like, what the fuck is this guy talking about?
Cue 20 — 00:01:28,029 --> 00:01:30,709
And maybe they didn't like my shirt.
Cue 21 — 00:01:30,749 --> 00:01:35,709
Maybe. Whatever. But I do think that, like, like, I didn't go to the right
Cue 22 — 00:01:35,749 --> 00:01:40,669
university, and maybe I'm over age, so instead of pitching more,
Cue 23 — 00:01:40,709 --> 00:01:44,689
I decided, fuck it, I'm going to give these ideas away.
Cue 24 — 00:01:45,549 --> 00:01:50,349
And now people will laugh, but then they will copy these ideas.
Cue 25 — 00:01:50,389 --> 00:01:56,989
So in June this year, I gave away the idea of using formal
Cue 26 — 00:01:57,029 --> 00:02:00,889
methods to verify software to make it secure.
Cue 27 — 00:02:01,069 --> 00:02:07,689
And guess what? Three months later, Anthropic stole my idea.
Cue 28 — 00:02:07,869 --> 00:02:13,429
And they now say formal methods are like, you know, something we need to use to make
Cue 29 — 00:02:13,469 --> 00:02:16,249
AI generated software secure and safe?
Cue 30 — 00:02:16,909 --> 00:02:20,029
And guess what? Startups are being funded.
Cue 31 — 00:02:20,069 --> 00:02:25,909
These guys even made it to the front page of the New York Times,
Cue 32 — 00:02:25,949 --> 00:02:29,829
and they have raised $75 million.
Cue 33 — 00:02:29,869 --> 00:02:35,789
So it's working. So I will give you some more ideas such that,
Cue 34 — 00:02:35,829 --> 00:02:41,709
like someone here that's got like, you know, a better suited to do a startup will
Cue 35 — 00:02:41,749 --> 00:02:47,889
like, you know, make the next, $75 million or more funding round.
Cue 36 — 00:02:48,549 --> 00:02:53,549
So this is a slide copied directly from my original pitch deck.
Cue 37 — 00:02:53,589 --> 00:02:56,789
Maybe you can see why it led to nothing.
Cue 38 — 00:02:56,829 --> 00:03:03,469
Because it looks like crap. But my idea was that my my pitch was that right
Cue 39 — 00:03:03,509 --> 00:03:08,169
now, all software is, like, generic.
Cue 40 — 00:03:08,269 --> 00:03:12,589
So you use, like, generic apps that are built for a wide audience.
Cue 41 — 00:03:12,629 --> 00:03:15,309
And each of these apps solves a particular problem.
Cue 42 — 00:03:15,349 --> 00:03:17,769
So you have like Excel. You have.
Cue 43 — 00:03:18,629 --> 00:03:22,949
So this is like generic software that solves a particular problem.
Cue 44 — 00:03:22,989 --> 00:03:28,789
But for for for everyone. But what you want is I don't want that.
Cue 45 — 00:03:28,829 --> 00:03:30,069
Like I want to solve a problem.
Cue 46 — 00:03:30,109 --> 00:03:31,949
I'm not interested in using Excel.
Cue 47 — 00:03:31,989 --> 00:03:34,549
I'm using Excel to actually solve a business problem.
Cue 48 — 00:03:34,589 --> 00:03:39,909
So what you want is you want to have like a completely personalized software.
Cue 49 — 00:03:39,949 --> 00:03:43,549
I don't care whether that's Excel that is completely like, you know,
Cue 50 — 00:03:43,589 --> 00:03:50,129
specific to me. And now you can see like this is 2023.
Cue 51 — 00:03:50,269 --> 00:03:53,389
This looks very much like Claude Cowork.
Cue 52 — 00:03:53,429 --> 00:03:55,389
Right. So this is is happening.
Cue 53 — 00:03:55,429 --> 00:04:00,549
Although I do think that Cowork is not so much about like hyper specialized software
Cue 54 — 00:04:00,589 --> 00:04:05,649
but more about orchestration. So it still orchestrates like the classic,
Cue 55 — 00:04:06,709 --> 00:04:09,829
office suite and so on. But it's not specializing.
Cue 56 — 00:04:09,869 --> 00:04:15,309
And that is what I want to sell to you today, is that we have to kind of chop up all
Cue 57 — 00:04:15,349 --> 00:04:21,529
that software and, and like, pimp it up to, like to make it hyper personalized.
Cue 58 — 00:04:22,269 --> 00:04:27,889
And so you might ask yourself, but Erik, why the fuck do you want to do that?
Cue 59 — 00:04:28,109 --> 00:04:31,389
We have this software. Why do you want to get like, you know,
Cue 60 — 00:04:31,429 --> 00:04:38,309
chop it up and change it? Well, because I want to and this is
Cue 61 — 00:04:38,349 --> 00:04:41,609
like, sounds like a joke, but this is not a joke, because with I,
Cue 62 — 00:04:41,649 --> 00:04:43,909
I mean, you, you you, everybody.
Cue 63 — 00:04:43,949 --> 00:04:46,829
This is really what people want.
Cue 64 — 00:04:46,869 --> 00:04:51,169
And let me show you this. A little graph here.
Cue 65 — 00:04:51,869 --> 00:04:57,949
Long time ago, before the industrial industrial revolution, everything was custom.
Cue 66 — 00:04:57,989 --> 00:05:02,629
Maybe you would have only one shirt because that shirt was, like,
Cue 67 — 00:05:02,669 --> 00:05:05,569
weaved custom for you. Especially for you.
Cue 68 — 00:05:06,189 --> 00:05:09,809
And and it was expensive. So you could only afford one shirt.
Cue 69 — 00:05:09,949 --> 00:05:12,609
And then the Industrial Revolution happened.
Cue 70 — 00:05:12,709 --> 00:05:17,629
And then what we got is that everything became cheap, but also generic.
Cue 71 — 00:05:17,669 --> 00:05:20,309
So for example, cars, there's the famous quote.
Cue 72 — 00:05:20,349 --> 00:05:24,269
And like, you know, you can pick any color as long as it's black,
Cue 73 — 00:05:24,309 --> 00:05:28,989
right? So like we things became cheap but not personalized.
Cue 74 — 00:05:29,029 --> 00:05:35,709
And now marketing came in and try to make things that were generic and cheap,
Cue 75 — 00:05:35,749 --> 00:05:42,549
make it look expensive. So if you buy like sneakers or watches or like
Cue 76 — 00:05:42,589 --> 00:05:47,629
Louis Vuitton handbags, these are all generic mass produced items,
Cue 77 — 00:05:47,669 --> 00:05:50,469
but they are made to look expensive.
Cue 78 — 00:05:50,509 --> 00:05:53,649
So you're you're kind of being conned.
Cue 79 — 00:05:54,149 --> 00:05:59,869
And then, you know, in order to make it even more like, custom,
Cue 80 — 00:05:59,909 --> 00:06:01,909
they people make special editions.
Cue 81 — 00:06:01,949 --> 00:06:05,129
Oh, there's only 2000 made. Like you have to buy it.
Cue 82 — 00:06:05,469 --> 00:06:12,349
Well, I think, will happen with AI and, robotics is that
Cue 83 — 00:06:12,389 --> 00:06:15,709
everything will become custom and cheap.
Cue 84 — 00:06:15,749 --> 00:06:18,899
And this, I think, is the second industrial revolution.
Cue 85 — 00:06:18,939 --> 00:06:20,949
I think this is big. Imagine that.
Cue 86 — 00:06:20,989 --> 00:06:26,189
Like, you know, like healthcare becomes custom for you and cheap.
Cue 87 — 00:06:26,229 --> 00:06:30,349
Imagine education will become custom to you and cheap.
Cue 88 — 00:06:30,389 --> 00:06:35,129
Instead of buying like some generic book about whatever topic,
Cue 89 — 00:06:35,349 --> 00:06:38,789
your education will be completely like personalized for you.
Cue 90 — 00:06:38,829 --> 00:06:41,349
And I think this is like amazing idea.
Cue 91 — 00:06:41,389 --> 00:06:48,529
So AI enables us to go from, generic and cheap, which nobody wants.
Cue 92 — 00:06:49,349 --> 00:06:54,249
Like to custom and cheap, what everybody wants.
Cue 93 — 00:06:55,429 --> 00:07:00,529
All right. Now you can say, Erik, I really don't care.
Cue 94 — 00:07:00,569 --> 00:07:06,209
I, I'm happy to be like a gray mouse in, like, the big sea of like,
Cue 95 — 00:07:07,989 --> 00:07:11,909
whatever. The regression to the mean, I don't care.
Cue 96 — 00:07:11,949 --> 00:07:14,889
Don't do stick out. But there are more reasons.
Cue 97 — 00:07:15,789 --> 00:07:21,509
For example, if I have, like, generate my own software, there's like a fixed cost of
Cue 98 — 00:07:21,549 --> 00:07:26,509
generation, but then I can just run it at a much lower cost when compared to,
Cue 99 — 00:07:26,549 --> 00:07:31,369
like a monthly subscription for some software as a service.
Cue 100 — 00:07:33,229 --> 00:07:39,969
Another thing is you want to get the features that you need without the bugs.
Cue 101 — 00:07:40,869 --> 00:07:43,469
All right. So let's look here.
Cue 102 — 00:07:43,509 --> 00:07:46,009
This is from like last week.
Cue 103 — 00:07:47,269 --> 00:07:51,129
Postgres 19 is late and it's AI's fault.
Cue 104 — 00:07:51,509 --> 00:07:54,549
I love this headline because it's not AI's fault.
Cue 105 — 00:07:54,589 --> 00:08:00,729
It is because this software was written by humans, and humans write shitty software
Cue 106 — 00:08:00,909 --> 00:08:03,569
and they software so full of bugs.
Cue 107 — 00:08:03,709 --> 00:08:09,909
And the AI is the thing that's finding these bugs that they cannot ship on time.
Cue 108 — 00:08:09,949 --> 00:08:12,589
All right. So this is another thing to consider.
Cue 109 — 00:08:12,629 --> 00:08:19,269
Like you hear a lot of like people say like, oh, glencore's AI generated
Cue 110 — 00:08:19,309 --> 00:08:25,829
slop. Well, I can tell you that I think AI writes better code than
Cue 111 — 00:08:25,869 --> 00:08:30,629
99.99% of the people that have ever that I've ever worked with.
Cue 112 — 00:08:30,669 --> 00:08:35,509
And I don't want to insult anybody, but this is like living proof of that,
Cue 113 — 00:08:35,549 --> 00:08:39,069
right? Like, human written software is crap.
Cue 114 — 00:08:39,109 --> 00:08:44,929
So we might as well just, like, forget about it and just rewrite everything.
Cue 115 — 00:08:45,349 --> 00:08:50,409
So like, we we maintain all the features, but we get rid of all the bugs.
Cue 116 — 00:08:50,519 --> 00:08:56,549
And now just let me tell you a little bit about like, say a database.
Cue 117 — 00:08:56,589 --> 00:09:00,109
Like why would you say, like, why am I going to rewrite a database?
Cue 118 — 00:09:00,149 --> 00:09:06,429
Well, if you look at your company, you're probably using like if you use Postgres,
Cue 119 — 00:09:06,469 --> 00:09:09,709
you probably have like a few tables and you do a few queries.
Cue 120 — 00:09:09,749 --> 00:09:15,869
Why do you need like a completely general purpose database that parses the SQL and
Cue 121 — 00:09:15,909 --> 00:09:22,389
has an optimizer? Whatever. You can just generate a hyper specific application
Cue 122 — 00:09:22,429 --> 00:09:27,709
just with these five tables and these queries, and you can throw away 90% of this
Cue 123 — 00:09:27,749 --> 00:09:29,909
bug-ridden database code. Okay.
Cue 124 — 00:09:29,949 --> 00:09:32,389
So think about that. Code has become cheap.
Cue 125 — 00:09:32,429 --> 00:09:36,269
Why did you use Postgres? Because writing code was expensive.
Cue 126 — 00:09:36,309 --> 00:09:42,149
But code is now free, so there's no need to depend on this old bug-ridden code
Cue 127 — 00:09:42,189 --> 00:09:47,969
anymore. The safest code is the code that you never run.
Cue 128 — 00:09:48,909 --> 00:09:51,249
We know all this picture here, right?
Cue 129 — 00:09:51,949 --> 00:09:57,349
What I'm proposing is instead of depending on all this weird code,
Cue 130 — 00:09:57,389 --> 00:10:02,509
you just partially evaluate everything into your own bespoke code,
Cue 131 — 00:10:02,549 --> 00:10:05,909
and you, you remove all the dependencies, right?
Cue 132 — 00:10:05,949 --> 00:10:08,869
So, like, imprint this in your mind.
Cue 133 — 00:10:08,909 --> 00:10:10,989
This is the future open source.
Cue 134 — 00:10:11,029 --> 00:10:13,509
It's useless. Why did we do open source?
Cue 135 — 00:10:13,549 --> 00:10:16,069
Because it was expensive to develop code.
Cue 136 — 00:10:16,109 --> 00:10:17,749
And so we say, hey it's great.
Cue 137 — 00:10:17,789 --> 00:10:19,829
This this dude like, wrote the code for me.
Cue 138 — 00:10:19,869 --> 00:10:24,009
I will kind of reuse it. That's not valuable anymore.
Cue 139 — 00:10:26,509 --> 00:10:31,949
So now to the stealing. This is an actual Dutch proverb.
Cue 140 — 00:10:31,989 --> 00:10:34,569
Good artists copy, great artists steal.
Cue 141 — 00:10:34,709 --> 00:10:37,509
And so this is already happening, right?
Cue 142 — 00:10:37,549 --> 00:10:44,285
People are rewriting, Microsoft Office, they rewrite
Cue 143 — 00:10:44,325 --> 00:10:49,209
Adobe suites and so on. And guess what?
Cue 144 — 00:10:50,749 --> 00:10:54,429
This is the effect Microsoft has to, like, lay off, like, you know,
Cue 145 — 00:10:54,469 --> 00:10:56,449
like a quarter of their workforce.
Cue 146 — 00:10:56,909 --> 00:10:59,129
All the stock prices collapse.
Cue 147 — 00:11:01,069 --> 00:11:06,949
So, yeah, so there's another reason, and I mentioned this, I think humans are really
Cue 148 — 00:11:06,989 --> 00:11:12,189
bad developers. And for hardware, we had Moore's Law, right?
Cue 149 — 00:11:12,229 --> 00:11:16,649
Like code or like hardware would get faster every X months.
Cue 150 — 00:11:16,709 --> 00:11:20,749
But for human coders, we don't have Moore's Law.
Cue 151 — 00:11:20,789 --> 00:11:24,229
Human coders don't get better every x months.
Cue 152 — 00:11:24,269 --> 00:11:29,229
Twice better. And we tried with agile and processes and software engineering.
Cue 153 — 00:11:29,269 --> 00:11:32,189
But it doesn't work. It doesn't make humans more productive.
Cue 154 — 00:11:32,229 --> 00:11:35,589
Humans are the bottleneck. So tokens scale.
Cue 155 — 00:11:35,629 --> 00:11:38,589
But humans don't. So we should get rid of all humans.
Cue 156 — 00:11:38,629 --> 00:11:41,629
If you have a company, get rid of all the humans.
Cue 157 — 00:11:41,669 --> 00:11:44,949
Replace them with tokens because that is what scales.
Cue 158 — 00:11:44,989 --> 00:11:48,729
So finally, we now have Moore's Law for coding.
Cue 159 — 00:11:49,709 --> 00:11:55,829
And the other thing is that most of the code that we write is trivial.
Cue 160 — 00:11:55,869 --> 00:12:00,469
If you look at all possible mathematical functions, like if you have ever done
Cue 161 — 00:12:00,509 --> 00:12:06,749
complexity theory, the code that you can actually put in production is like a tiny,
Cue 162 — 00:12:06,789 --> 00:12:09,669
tiny subset of that, a trivial subset.
Cue 163 — 00:12:09,709 --> 00:12:13,029
Because if you have something, an algorithm that is quadratic,
Cue 164 — 00:12:13,069 --> 00:12:16,669
you cannot ship it in production because you will get the denial of service attack
Cue 165 — 00:12:16,709 --> 00:12:19,749
that will put your servers grinding to a halt.
Cue 166 — 00:12:19,789 --> 00:12:24,709
If you have something that uses like n squared space, oh, you're in trouble.
Cue 167 — 00:12:24,749 --> 00:12:27,949
So like n log n is the best that we can do.
Cue 168 — 00:12:27,989 --> 00:12:32,189
So it's just a tiny tiny bit of the all possible function.
Cue 169 — 00:12:32,229 --> 00:12:37,069
So I would say all the software that we write from a mathematical point of view is
Cue 170 — 00:12:37,109 --> 00:12:42,489
trivial. And hence it's no surprise that the models can write that software for us.
Cue 171 — 00:12:42,549 --> 00:12:46,929
And it's kind of surprising that humans suck at it so bad.
Cue 172 — 00:12:48,789 --> 00:12:53,869
And then the next thing is, I do think that people love to talk about IP,
Cue 173 — 00:12:53,909 --> 00:12:58,129
but I think there's very little intellectual in IP.
Cue 174 — 00:12:58,189 --> 00:13:02,389
Most software is not only trivial from a mathematical point of view,
Cue 175 — 00:13:02,429 --> 00:13:05,309
but also trivial from like an intellectual point of view.
Cue 176 — 00:13:05,349 --> 00:13:10,729
Let me give you an example. This is a patent where I'm the first inventor.
Cue 177 — 00:13:10,769 --> 00:13:16,789
It's called Guardian Angel. If you if you Google for it, you'll find it.
Cue 178 — 00:13:16,829 --> 00:13:20,429
And people will say that Bill Gates was the inventor of this.
Cue 179 — 00:13:20,469 --> 00:13:23,229
That hurts me because, like, look, I'm there.
Cue 180 — 00:13:23,269 --> 00:13:27,789
It's the first inventor. And of course, we sold the rights to Microsoft.
Cue 181 — 00:13:27,829 --> 00:13:34,389
But this thing, 2006 is about AI, like making
Cue 182 — 00:13:34,429 --> 00:13:37,629
decisions on behalf of the user, right?
Cue 183 — 00:13:37,669 --> 00:13:41,869
Making proactive decisions on behalf of the user.
Cue 184 — 00:13:41,909 --> 00:13:48,789
2006 guess what? In 2026, Anthropic
Cue 185 — 00:13:48,829 --> 00:13:54,189
comes with duet, which works as a proactive agent that acts on behalf of the user.
Cue 186 — 00:13:54,229 --> 00:13:59,189
All right, so they just take this idea that's 20 years old and just re-implement it
Cue 187 — 00:13:59,229 --> 00:14:03,829
and think it's something new. But what I would say that original ideas crap anyway.
Cue 188 — 00:14:03,869 --> 00:14:09,889
So there's very little original ideas and software patents are useless.
Cue 200 — 00:14:52,588 --> 00:14:55,648
All right. So this is like, you know,
Cue 201 — 00:14:57,268 --> 00:14:59,148
the conclusion of the first part.
Cue 202 — 00:14:59,188 --> 00:15:03,348
Given enough compute, all software is solved, right?
Cue 203 — 00:15:03,388 --> 00:15:06,388
Everything is trivial. Just throwing off tokens at it.
Cue 204 — 00:15:06,428 --> 00:15:10,368
And every problem you can ever imagine is solved.
Cue 205 — 00:15:10,628 --> 00:15:16,428
Now, it's not enough to get like, you know, know what you want to do,
Cue 206 — 00:15:16,468 --> 00:15:20,268
but are we allowed to do it? So if we want to take software and,
Cue 207 — 00:15:20,308 --> 00:15:26,628
like, chop it up and like, rewrite it and make it something new.
Cue 208 — 00:15:26,668 --> 00:15:30,488
Are we allowed now? Are there any lawyers in the room.
Cue 209 — 00:15:31,708 --> 00:15:36,188
Because I'm not a lawyer. So like take everything that I say in this section with a
Cue 210 — 00:15:36,228 --> 00:15:37,648
huge grain of salt.
Cue 211 — 00:15:39,028 --> 00:15:45,788
But if we look at, like, the world of, like, software protection,
Cue 212 — 00:15:45,828 --> 00:15:51,148
IP protection, before the 80s, it was kind of wild West.
Cue 213 — 00:15:51,188 --> 00:15:52,828
There was no protection at all.
Cue 214 — 00:15:52,868 --> 00:15:56,228
And there's the famous letter from Bill Gates, like begging,
Cue 215 — 00:15:56,268 --> 00:15:58,588
like, don't steal myself or don't copy it.
Cue 216 — 00:15:58,628 --> 00:16:00,828
I need to eat whatever. Right.
Cue 217 — 00:16:00,868 --> 00:16:03,408
So this is like before the 80s wild West.
Cue 218 — 00:16:03,748 --> 00:16:10,348
Then between 1980 and around 2000, that was the golden era of IP
Cue 219 — 00:16:10,388 --> 00:16:17,308
protection. Software patents became possible and there was the Copyright Act,
Cue 220 — 00:16:17,348 --> 00:16:22,808
the digital whatever Millennium Copyright Act, there was like more protection.
Cue 221 — 00:16:24,388 --> 00:16:29,808
But fortunately this started to crack quite, quite quickly.
Cue 222 — 00:16:30,548 --> 00:16:37,088
Because first, in about 2014, there was like an attack on
Cue 223 — 00:16:37,548 --> 00:16:43,368
software patents. I have about 300 software patents, and I think they're all crap.
Cue 224 — 00:16:43,708 --> 00:16:46,508
And so I kind of completely agree with this.
Cue 225 — 00:16:46,548 --> 00:16:49,388
And then of course, there's the famous Oracle versus Google,
Cue 226 — 00:16:49,428 --> 00:16:54,888
one where like, you know, you could copy the API declarations as fair use.
Cue 227 — 00:16:55,748 --> 00:17:01,868
But I think these two are more important for us is that like,
Cue 228 — 00:17:01,908 --> 00:17:07,208
you know, patents require humans to be involved, so you cannot have AI inventors.
Cue 229 — 00:17:07,388 --> 00:17:12,088
And the same thing for copyright, you cannot have AI authors.
Cue 230 — 00:17:12,188 --> 00:17:18,828
And and so if you if you're doing vibe coding and and that's what I'm proposing.
Cue 231 — 00:17:18,868 --> 00:17:21,328
Right. We should get all humans out of the loop.
Cue 232 — 00:17:21,788 --> 00:17:27,308
All the software that you write is completely free and free.
Cue 233 — 00:17:27,348 --> 00:17:32,068
As in like what? Like, you know, the original open source founders wanted.
Cue 234 — 00:17:32,108 --> 00:17:36,628
It's free. You cannot protect it because it's not written by humans.
Cue 235 — 00:17:36,668 --> 00:17:40,128
And I think this is a good thing because if something is free,
Cue 236 — 00:17:40,228 --> 00:17:42,468
it's not worth protecting, right?
Cue 237 — 00:17:42,508 --> 00:17:44,668
Software is free. You don't pay anything for it.
Cue 238 — 00:17:44,708 --> 00:17:46,988
Why would you protect it? The ideas are crap.
Cue 239 — 00:17:47,028 --> 00:17:49,628
So it's it's just like all this.
Cue 240 — 00:17:49,668 --> 00:17:53,288
This protection is just a waste of time.
Cue 241 — 00:17:53,588 --> 00:18:00,428
All right, so we should vibe code and we should also design for agents,
Cue 242 — 00:18:00,468 --> 00:18:05,748
not for humans. And then, like, you know, all the software that we write can be
Cue 243 — 00:18:05,788 --> 00:18:07,788
freely copied and so on, but it doesn't matter.
Cue 244 — 00:18:07,828 --> 00:18:10,968
Why would I copy software if I can rewrite it from scratch?
Cue 245 — 00:18:11,988 --> 00:18:17,208
All right. So I think this is beautiful.
Cue 246 — 00:18:18,108 --> 00:18:24,448
And like, once you check in code to GitHub that has an AI like contributor in there,
Cue 247 — 00:18:24,828 --> 00:18:28,808
you know, game over if you want to, if you try to protect it.
Cue 248 — 00:18:28,908 --> 00:18:35,588
All right. So, again, don't believe me, but I would say if your business relies
Cue 249 — 00:18:35,628 --> 00:18:39,048
on lawyers to win, you've already lost.
Cue 250 — 00:18:39,468 --> 00:18:42,108
All right. So don't take my advice.
Cue 251 — 00:18:42,148 --> 00:18:46,708
But I do think that it's it's it's really desperate if you have to,
Cue 252 — 00:18:46,748 --> 00:18:51,528
like, use legal means and to stay above water.
Cue 253 — 00:18:52,388 --> 00:18:59,048
All right. Now the third part of my talk is now we know what we want to do.
Cue 254 — 00:18:59,148 --> 00:19:01,608
We know that we're allowed to do it.
Cue 255 — 00:19:02,308 --> 00:19:08,668
But how do we do it? Like, how can we, like, create this hyper
Cue 256 — 00:19:08,708 --> 00:19:15,268
specialized software? And for that, I'm going to get like, very briefly explain to
Cue 257 — 00:19:15,308 --> 00:19:17,928
you or maybe try to explain to you.
Cue 258 — 00:19:18,588 --> 00:19:22,848
The three projections of Dr. Futamura.
Cue 259 — 00:19:22,988 --> 00:19:27,608
And I think this is the most beautiful computer science ever.
Cue 260 — 00:19:28,348 --> 00:19:31,828
So if you, you know, you can forget everything I talked about,
Cue 261 — 00:19:31,868 --> 00:19:37,668
like, about all this, this stuff about legal and and, like, abundance and free
Cue 262 — 00:19:37,708 --> 00:19:42,808
software. This is the most beautiful computer science, ever.
Cue 263 — 00:19:43,988 --> 00:19:46,028
So remember what we wanted to do?
Cue 264 — 00:19:46,068 --> 00:19:51,508
We want to take an application that has some dependencies, and then we want to have,
Cue 265 — 00:19:51,548 --> 00:19:58,348
like, you know, tell the agent, hey, just like, you know, statically link all
Cue 266 — 00:19:58,388 --> 00:20:03,668
statically, all static dependencies into the application and generate a completely
Cue 267 — 00:20:03,708 --> 00:20:09,948
new application and boom, all the dependencies are gone.
Cue 268 — 00:20:09,988 --> 00:20:14,568
And we now have a, have an application that has zero dependencies.
Cue 269 — 00:20:15,068 --> 00:20:18,668
And this is great. Let me give you a concrete example.
Cue 270 — 00:20:18,708 --> 00:20:25,408
Who here uses React? Why on earth are you using React?
Cue 271 — 00:20:26,108 --> 00:20:30,348
It's this kind of like giant thing you're just probably using like a sliver of it,
Cue 272 — 00:20:30,388 --> 00:20:34,188
like 10%. And you use like this 10%, you use that 10%.
Cue 273 — 00:20:34,228 --> 00:20:38,268
So just ask Claude to say, don't take a React dependency.
Cue 274 — 00:20:38,308 --> 00:20:41,908
Just reimplement the pieces that I need in my app.
Cue 275 — 00:20:41,948 --> 00:20:44,468
Try it and you will be much happier.
Cue 276 — 00:20:44,508 --> 00:20:46,388
You don't have this bug-ridden.
Cue 277 — 00:20:46,428 --> 00:20:50,628
Giant piece of whatever. So do this.
Cue 278 — 00:20:50,668 --> 00:20:55,428
Really? All right, let's let's be concrete.
Cue 279 — 00:20:55,468 --> 00:20:57,448
Let's give an example.
Cue 280 — 00:20:58,708 --> 00:21:02,468
And this is an example from a book that I will show you later.
Cue 281 — 00:21:02,508 --> 00:21:07,828
But here, here's like a this is what theoretical computer scientists like academics
Cue 282 — 00:21:07,868 --> 00:21:11,208
like to do. So there's a function here that has two parameters.
Cue 283 — 00:21:11,428 --> 00:21:16,348
And now you want to like say I want to fix the first parameter.
Cue 284 — 00:21:16,388 --> 00:21:21,748
And now I want to like specialize that program with respect to that first parameter.
Cue 285 — 00:21:21,788 --> 00:21:27,848
And so think of this as like very aggressive inlining okay.
Cue 286 — 00:21:28,828 --> 00:21:35,028
And you just ask Claude. You kind of see already here what I'm kind of working on in
Cue 287 — 00:21:35,068 --> 00:21:40,468
the background. But you just ask here Claude partially evaluate this function for n
Cue 288 — 00:21:40,508 --> 00:21:45,548
equals five and then it will just tell you everything like oh I'm going to like
Cue 289 — 00:21:45,588 --> 00:21:51,868
unfold this substitute back. And here's the program that has like,
Cue 290 — 00:21:51,908 --> 00:21:54,548
you know, this, this value kind of hardwired in.
Cue 291 — 00:21:54,588 --> 00:21:58,728
And now there's no dependency on n equals five.
Cue 292 — 00:21:58,988 --> 00:22:01,148
And this is a very simple example.
Cue 293 — 00:22:01,188 --> 00:22:04,688
But as I will show you later in the talk, and as you can try,
Cue 294 — 00:22:04,948 --> 00:22:07,148
you can do this with giant programs.
Cue 295 — 00:22:07,188 --> 00:22:12,628
You can do this if you have a dependency on React, you can say partially evaluate
Cue 296 — 00:22:12,668 --> 00:22:16,988
this code with respect to React and remove every dependency on React.
Cue 297 — 00:22:17,028 --> 00:22:21,828
And it will generate for you a new program that has no dependencies on React.
Cue 298 — 00:22:21,868 --> 00:22:25,868
So this program has. Where do you see the number five in this program that's gone?
Cue 299 — 00:22:25,908 --> 00:22:29,548
Well, it's in the name but the constant five is completely gone.
Cue 300 — 00:22:29,588 --> 00:22:32,168
Right. So it's completely specialized to it.
Cue 301 — 00:22:34,788 --> 00:22:41,808
Now of course one example doesn't prove that this works in general.
Cue 302 — 00:22:42,188 --> 00:22:47,228
And as I said, like, you know, there's this book, it's actually the PDF is available
Cue 303 — 00:22:47,268 --> 00:22:48,288
online.
Cue 304 — 00:22:49,708 --> 00:22:54,068
I think the models have been trained on this because it's there.
Cue 305 — 00:22:54,108 --> 00:22:59,168
And that, I think, is also why the models are so good at partial evaluation.
Cue 306 — 00:23:01,388 --> 00:23:02,188
Now,
Cue 307 — 00:23:03,388 --> 00:23:09,328
now I have to be a little bit more like show you some, some,
Cue 308 — 00:23:09,368 --> 00:23:12,788
some formulas here. But what is a partial evaluator?
Cue 309 — 00:23:12,828 --> 00:23:17,908
So like if I look at like I gave this like, hey Claude, partially evaluate this
Cue 310 — 00:23:17,948 --> 00:23:24,508
thing with respect to like the, the static dependencies, what a partial evaluator
Cue 311 — 00:23:24,548 --> 00:23:27,548
is. It's something that takes a function that has two parameters.
Cue 312 — 00:23:27,588 --> 00:23:30,028
So there's this function A and b.
Cue 313 — 00:23:30,068 --> 00:23:32,768
It takes two parameters a and b and returns a c.
Cue 314 — 00:23:33,148 --> 00:23:38,068
Now this first parameter a is static so it doesn't change.
Cue 315 — 00:23:38,108 --> 00:23:40,508
It's like React. It's a dependency.
Cue 316 — 00:23:40,548 --> 00:23:46,068
Like you're not changing like the the UI framework that you depend on like while
Cue 317 — 00:23:46,108 --> 00:23:49,508
you're running. Right. You're kind of take the dependency once and then your whole
Cue 318 — 00:23:49,548 --> 00:23:54,588
code depends on that. So what a partial evaluator does is it takes a function that
Cue 319 — 00:23:54,628 --> 00:23:59,948
has these two parameters. It takes the static parameter a, and then it returns a new
Cue 320 — 00:23:59,988 --> 00:24:02,548
function that doesn't depend on a.
Cue 321 — 00:24:02,588 --> 00:24:08,068
This just depends on B. And so it's kind of like think of this thing as like taking
Cue 322 — 00:24:08,108 --> 00:24:13,668
a value of type A and completely specializing that binary function to create this
Cue 323 — 00:24:13,708 --> 00:24:16,048
new function, just like I showed you.
Cue 324 — 00:24:16,988 --> 00:24:21,228
Here. Right. I have a function of two parameters here.
Cue 325 — 00:24:21,268 --> 00:24:25,428
I give it one parameter and I tell it, specialize it, and then I get a function of
Cue 326 — 00:24:25,468 --> 00:24:29,688
one parameter that has no dependencies on the first parameter.
Cue 327 — 00:24:30,668 --> 00:24:37,508
All right. Now if you're an academic this is really hard to do this to
Cue 328 — 00:24:37,548 --> 00:24:42,668
write a partial evaluator like that that works for any piece of code is really hard.
Cue 329 — 00:24:42,708 --> 00:24:47,788
And so the people that wrote this book, it took them more than ten years to get this
Cue 330 — 00:24:47,828 --> 00:24:51,248
working. And these are really, like smart people.
Cue 331 — 00:24:54,668 --> 00:24:59,108
Now you can say, but Erik, this sounds all very theoretical.
Cue 332 — 00:24:59,148 --> 00:25:05,526
Well, if you if you use GraalVM, under the hood that uses Truffle,
Cue 333 — 00:25:05,566 --> 00:25:11,068
and Truffle uses partial evaluation to create JITs on the fly.
Cue 334 — 00:25:11,108 --> 00:25:16,948
So this is not just something theoretical like Oracle and Java like rely on this
Cue 335 — 00:25:16,988 --> 00:25:20,288
stuff. So it's not something where it's like, oh Erik.
Cue 336 — 00:25:20,828 --> 00:25:25,788
Well what what kind of like Dreamland are you now know this is used for real and
Cue 337 — 00:25:25,828 --> 00:25:29,708
production. But again like this takes serious effort.
Cue 338 — 00:25:29,748 --> 00:25:33,408
Like the people like at Oracle Labs are super smart.
Cue 339 — 00:25:35,548 --> 00:25:40,928
All right. So now let's let's look at, a few other examples.
Cue 340 — 00:25:42,348 --> 00:25:45,388
So here is another kind of program.
Cue 341 — 00:25:45,428 --> 00:25:50,188
It's an interpreter. An interpreter is something that takes a program that is
Cue 342 — 00:25:50,228 --> 00:25:55,268
static, an input that's dynamic, and then computes the outputs of Python.
Cue 343 — 00:25:55,308 --> 00:26:00,268
Right. The Python interpreter takes a Python program and then the inputs to the
Cue 344 — 00:26:00,308 --> 00:26:03,468
Python program and then computes the output.
Cue 345 — 00:26:03,508 --> 00:26:05,828
But the Python program itself doesn't change.
Cue 346 — 00:26:05,868 --> 00:26:10,728
Well, Python programs can change on the fly, but let's not go there.
Cue 347 — 00:26:12,308 --> 00:26:17,628
But now what you see is that an interpreter is a function that takes two parameters.
Cue 348 — 00:26:17,668 --> 00:26:21,988
And this partial evaluator takes a function that has two parameters.
Cue 349 — 00:26:22,028 --> 00:26:27,428
So we can feed an interpreter into this partial evaluator.
Cue 350 — 00:26:27,468 --> 00:26:33,348
And guess what happens now we get if we partially evaluate an interpreter with
Cue 351 — 00:26:33,388 --> 00:26:36,668
respect to a program we get a compiled function.
Cue 352 — 00:26:36,708 --> 00:26:39,048
We get a function from input to output.
Cue 353 — 00:26:39,268 --> 00:26:44,028
And so I'm going a little bit fast here because I'm running out of time.
Cue 354 — 00:26:44,068 --> 00:26:46,528
But this is is like pretty nice.
Cue 355 — 00:26:47,308 --> 00:26:50,488
And so this is what's called the first projection.
Cue 356 — 00:26:50,668 --> 00:26:57,348
But now what happens if we're feeding the partial evaluator to
Cue 357 — 00:26:57,388 --> 00:26:58,288
itself.
Cue 358 — 00:26:59,788 --> 00:27:03,348
Now we get and then we give it an interpreter.
Cue 359 — 00:27:03,388 --> 00:27:04,968
Now we get a compiler.
Cue 360 — 00:27:06,268 --> 00:27:13,008
But if we feed the partial evaluator to itself and why don't we kind of like
Cue 361 — 00:27:13,268 --> 00:27:18,688
feed the partial evaluator to the partial evaluator with the partial evaluator?
Cue 362 — 00:27:19,148 --> 00:27:25,308
Holy fuck. I mean, I mean probably your brain explodes, but what this gives you is a
Cue 363 — 00:27:25,348 --> 00:27:30,628
compiler generator, right? So this is the and this is the pinnacle.
Cue 364 — 00:27:30,668 --> 00:27:35,188
If you can make this happen like you, you will get the Turing Award.
Cue 365 — 00:27:35,228 --> 00:27:39,628
But as I said, like this is really, really hard if you're trying to do it by hand in
Cue 366 — 00:27:39,668 --> 00:27:44,948
the human way. But of course we have AI though can scale humans don't.
Cue 367 — 00:27:44,988 --> 00:27:49,308
We're too stupid to do this. I mean, this looks beautiful, but we're just not smart
Cue 368 — 00:27:49,348 --> 00:27:51,568
enough to implement this. But,
Cue 369 — 00:27:54,268 --> 00:27:57,668
with models, there's many ways to do this, right?
Cue 370 — 00:27:57,708 --> 00:28:01,508
Like if you want to take a general purpose model and you want to specialize it,
Cue 371 — 00:28:01,548 --> 00:28:04,268
we have distillation, we have fine tuning.
Cue 372 — 00:28:04,308 --> 00:28:08,568
And what I argue, we also have partial evaluation.
Cue 373 — 00:28:09,028 --> 00:28:14,708
So what we can do is we can look at an LLM as an interpreter.
Cue 374 — 00:28:14,748 --> 00:28:20,548
Because in a prompt system prompt that's like a program, a question which is
Cue 375 — 00:28:20,588 --> 00:28:26,328
dynamic, which is like a the input and it produces an output.
Cue 376 — 00:28:28,428 --> 00:28:31,708
Everyone sees this and it's just like an interpreter.
Cue 377 — 00:28:31,748 --> 00:28:36,848
But an LLM is also a partial evaluator.
Cue 378 — 00:28:37,748 --> 00:28:42,788
I can take, give an LLM a prompt and say, here's a function,
Cue 379 — 00:28:42,828 --> 00:28:45,388
here's a parameter, partially evaluate this.
Cue 380 — 00:28:45,428 --> 00:28:48,148
This is just what I did before right with this example.
Cue 381 — 00:28:48,188 --> 00:28:51,828
So an LLM is already a partial evaluator.
Cue 382 — 00:28:51,868 --> 00:28:54,028
We don't have to write a partial evaluator.
Cue 383 — 00:28:54,068 --> 00:28:57,348
It doesn't take us ten years to write a partial evaluator.
Cue 384 — 00:28:57,388 --> 00:29:01,788
That can be applied to itself when we use AI to do it.
Cue 385 — 00:29:01,828 --> 00:29:07,488
And now we have a model that can optimize, a model that can optimize a model.
Cue 386 — 00:29:07,708 --> 00:29:11,168
It's it's fantastic. All right.
Cue 387 — 00:29:13,028 --> 00:29:19,908
And you, you can try it so you can ask, say you you can try this like
Cue 388 — 00:29:19,948 --> 00:29:25,368
in the coffee break. Right. You tell Claude, take MNIST.
Cue 389 — 00:29:25,508 --> 00:29:31,548
And now partially evaluate that neural net such that it only needs to recognize the
Cue 390 — 00:29:31,588 --> 00:29:37,428
digits 0 to 4. All right. So I have a giant neural net that that is supposed to
Cue 391 — 00:29:37,468 --> 00:29:43,988
recognize from 0 to 9. And you tell it now partially evaluate this neural net
Cue 392 — 00:29:44,028 --> 00:29:48,428
only to to to recognize numbers between 0 and 4.
Cue 393 — 00:29:48,468 --> 00:29:52,788
It will hack the weights. It will hack the architecture, and it will create you a
Cue 394 — 00:29:52,828 --> 00:29:58,508
new neural net that only recognizes the the digits from 0 to 4.
Cue 395 — 00:29:58,548 --> 00:30:03,988
I think that's fucking amazing that a model can have like mutate another model.
Cue 396 — 00:30:04,028 --> 00:30:07,768
Now, the real thing is where the model will rewrite itself.
Cue 397 — 00:30:09,588 --> 00:30:13,568
And if you ask Claude, it will say, this is cool too.
Cue 398 — 00:30:13,748 --> 00:30:18,248
You know, I can kind of like, you know, manipulate my own model.
Cue 399 — 00:30:20,588 --> 00:30:25,828
And this is the, like, idea. So an LLM will kind of like, take two definitions of
Cue 400 — 00:30:25,868 --> 00:30:28,848
other LLMs and partially evaluate this.
Cue 401 — 00:30:30,548 --> 00:30:35,788
Now, I don't know how Anthropic does recursive self-improvement,
Cue 402 — 00:30:35,828 --> 00:30:41,908
but I do think that partial evaluation is recursive self-improvement.
Cue 403 — 00:30:41,948 --> 00:30:46,168
And of course, when you ask Gemini, it says, yes, Erik, you're right.
Cue 404 — 00:30:49,588 --> 00:30:54,588
But but I do think that this partial evaluation, especially the third Futamura
Cue 405 — 00:30:54,628 --> 00:30:58,988
projection, is a mathematically sound way to get recursive self-improvement.
Cue 406 — 00:30:59,028 --> 00:31:03,468
So think of this. You don't have to understand this, but each of these mixed things
Cue 407 — 00:31:03,508 --> 00:31:08,548
is an LLM. So it's an LLM that takes two LLMs as their parameters and then just
Cue 408 — 00:31:08,588 --> 00:31:11,668
starts to kind of hack on the weights and on the on the architecture.
Cue 409 — 00:31:11,708 --> 00:31:18,528
It's just beautiful. Now I always say I love it when academics prove
Cue 410 — 00:31:18,628 --> 00:31:21,148
like right papers to prove what I propose.
Cue 411 — 00:31:21,188 --> 00:31:25,868
Like hand-waving, there's a paper or blog post that just came out last week.
Cue 412 — 00:31:25,908 --> 00:31:28,928
It's called Sparks of Rage space programming.
Cue 413 — 00:31:29,148 --> 00:31:33,888
Some folks from MIT that show that LLMs can hack the weights,
Cue 414 — 00:31:34,548 --> 00:31:39,428
which I just told you that LLMs can hack the weights, but these academics do it.
Cue 415 — 00:31:39,468 --> 00:31:41,508
If I say it, people laughed me out of the room.
Cue 416 — 00:31:41,548 --> 00:31:46,168
If the academics do it, everybody goes like this MIT guys, they're so smart.
Cue 417 — 00:31:47,908 --> 00:31:50,688
Well, whatever. Just steal it.
Cue 418 — 00:31:52,028 --> 00:31:54,508
All right, so here's my prediction.
Cue 419 — 00:31:54,548 --> 00:32:00,468
My last slide January 20th 7th January 2027.
Cue 420 — 00:32:00,508 --> 00:32:06,428
There will be announced that there's a startup that is using partial evaluation and
Cue 421 — 00:32:06,468 --> 00:32:10,208
the Futamura projections to do recursive self-improvement.
Cue 422 — 00:32:10,948 --> 00:32:13,108
And I'm a messenger from the future.
Cue 423 — 00:32:13,148 --> 00:32:17,208
So I have here the proof and already for you.
Cue 424 — 00:32:18,028 --> 00:32:23,548
So conclusions of this slide. Every downside has an upside.
Cue 425 — 00:32:23,588 --> 00:32:27,668
Software is free. You're too stupid to write it.
Cue 426 — 00:32:27,708 --> 00:32:31,768
Don't get depressed. That's an opportunity.
Cue 427 — 00:32:31,868 --> 00:32:34,348
Right. So every downside has an upside.
Cue 428 — 00:32:34,388 --> 00:32:41,108
Second conclusion is if you are making money with software from software trying to
Cue 429 — 00:32:41,148 --> 00:32:44,228
sell software, I'm sorry, software is now free.
Cue 430 — 00:32:44,268 --> 00:32:50,388
So you cannot make money from software, but you can make money with software.
Cue 431 — 00:32:50,428 --> 00:32:53,848
And that is like a bigger opportunity than ever.
Cue 432 — 00:32:56,468 --> 00:32:57,408
Thank you.
.tessl-plugin
skills
talk-arcolano-autonomy-and-trust
talk-aroussi-decision-context
talk-bevin-agent-assisted-instrument-design
talk-boeckeler-context-lifecycles
talk-braun-context-graphs
talk-carmont-evolving-agent-harnesses
talk-chanezon-gordon-agent-experience-evaluation
talk-cox-secure-agentic-workflows
talk-cunniffe-reusable-agent-platform
talk-davis-agent-tool-policies
talk-debois-concurrent-delivery
talk-dzialets-team-automation
talk-fahmy-harness-api-design
talk-finster-continuous-delivery
talk-fowler-regenerative-software
talk-gavrilescu-factory-architecture
talk-graciolli-amin-verification-aware-programming
talk-gruenbaum-reverse-graph-engineering
talk-hassan-agentic-engineering
talk-hossain-enterprise-agent-adoption
talk-huntley-patch-ready-factories
talk-jain-verifying-intent
talk-katsioloudes-montei-agentic-lifecycle
talk-knox-harness-engineering
talk-kovac-deterministic-memory
talk-lee-perkins-code-readiness
talk-lorenc-machine-speed-security
talk-lubken-software-of-one
talk-meijer-software-specialization
talk-meppiel-versioned-agent-contracts
talk-moss-concurrent-agent-risk
talk-myers-understanding-agent-generated-code
talk-ocytko-scaling-agentic-engineering
talk-oza-automating-design-quality
talk-panzica-agentic-sdlc
talk-pari-patel-agent-awareness
talk-podjarny-trustworthy-software-factories
talk-poulter-factory-improvement
talk-pratik-patel-refined-rag-data
talk-re-ferre-development-platforms
talk-sadogursky-factory-of-factories
talk-shah-training-agent-skills
talk-sharma-mohapatra-evolving-intent
talk-singhal-operating-agents-at-scale
talk-stack-incident-dossiers
talk-tyler-team-coding-benchmarks
talk-wotherspoon-agent-execution-environments
talk-yegge-ai-powered-factories
talk-zuber-leadership-through-coding