YT YouTube2Text

YouTube2Text → Transcripts → Opus 5.5 is proof that we do need smarter models

Opus 5.5 is proof that we do need smarter models — Transcript

by Theo - t3․gg · 6,476 words · 904 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I'm going to be real with y'all. AI's
0:01gotten pretty damn smart. It's a hell of
0:03a lot smarter than I ever thought it
0:04would be. Just 2 years ago, a blond guy
0:07with a mustache that somehow had access
0:09to this YouTube channel did a video
0:10called AI isn't going to keep improving
0:12where he relatively confidently asserted
0:15that we were going to hit a plateau with
0:16how intelligent AI could get that would
0:19make the improvements more and more
0:21marginal as we tried using these new
0:23tools. I made that video before Opus 4.5
0:26was even in training and the amount
0:28things have changed since then is truly
0:31hard to comprehend. We went from AI that
0:33could auto-complete when we were typing
0:35occasionally to being able to ask real
0:37questions about our code base, tell it
0:39which files to edit, and now we're at
0:41the point where we just handed a
0:42screenshot of a bug and trust it to
0:44figure it out and test it and get it
0:45merged. It's crazy. But with all that,
0:48my take changed a bit. I went from
0:50thinking that it wouldn't keep getting
0:51smarter to something a bit more bold. I
0:54started to think we might not need them
0:56to be smarter. And I felt this many
0:58times when new models drop. It's like,
1:00"Oh, this is so much better. I can
1:02barely figure out how to take advantage
1:04of this level of intelligence. I have no
1:06idea why we would want anything
1:08smarter." I felt this with Fable 5 and
1:11Opus 5.5 has changed my mind. Not just
1:14because it is better at doing the things
1:16I do every day or it's faster or it's
1:18cheaper, but because it's smarter. And I
1:20don't just think this is better for
1:22coders, I think it's better for a lot
1:24more people than I ever would have
1:26guessed. I used to think that these
1:28model improvements were only going to
1:29really benefit a select set of people
1:31who were always on this bleeding edge
1:33pushing the limits of what they're
1:34capable of.
1:36I was wrong. And the reason my mind
1:37changed on this is sillier than you
1:39probably think.
1:46Yes, it's Halo and Call of Duty. While
1:48I'll gladly admit I'm a sucker for
1:50skateboarding stuff, I do think this
1:52represents a huge shift where models
1:54getting more capable is going to benefit
1:57a hell of a lot more people than just
1:59software devs. Let me explain why
2:01everyone will benefit from this after a
2:03quick word for today's sponsor. I know
2:05this is supposed to be an ad, but I want
2:06to give a real quick history lesson.
2:07Back when AI first started getting
2:09attention, it was to answer basic
2:10questions using the knowledge that it
2:12had. It progressed very quickly to the
2:14point where we were using it to edit
2:15files on our computers using tools like
2:17Claude code. I think we're about to
2:18enter a third era, an era where the
2:20agents aren't just working within the
2:22little box of the website they're on or
2:24the computer that we're running our
2:25systems in. They're going to start doing
2:27real work for us. But in order to do
2:29that, they need to be able to get out
2:30and do things on the web, which means
2:33they need a platform that supports this.
2:35Today's sponsor is browserbase, and man,
2:36I wish all the personal assistants I
2:38used were using them. There's a lot of
2:40layers to why I think browserbase is the
2:42best way for your agents to do things on
2:43the internet, but I'll start with one
2:45that has been really frustrating as I've
2:46leaned more into Grok Bot,
2:48authentication. When I want to let my
2:49Grok Bot access a service that isn't
2:51already a plugin, I have to let it spin
2:53up a computer in the cloud, open it up
2:55myself, go to the webpage, and sign in.
2:57But if they were using browserbase, I
2:58could link my one password account, and
3:00it's handled for me. Not only can
3:02browserbase handle the sign-in, they can
3:04also handle the management of your
3:06sessions and persisting them so that
3:07your agent doesn't have to sign in again
3:09every single time it makes a request. I
3:11know this sounds trivial, but you have
3:12no idea how much everyone else is
3:14getting this wrong, even the major labs
3:16into their personal assistant tools.
3:17That's all great for the signed-in
3:18experience, but a lot of what your
3:20agents do is finding things on the web
3:21and accessing them. That's why the
3:23search API is so useful to get a bunch
3:25of results that your agent might need,
3:26as well as the fetch API, which will
3:28turn those results and pages into HTML,
3:30JSON, or markdown. This is way more
3:31reliable than a curl will ever be.
3:33Considering that 85% of APIs on the web
3:36cannot be accessed via curl, this all
3:39comes together with a really, really
3:40powerful experience. Unlock your agents
3:42and see what they can do at
3:43toydev.link/browserbase.
3:45While I could sit here all day and watch
3:46all of my favorite childhood games get
3:48reverse engineered, decompiled, and
3:50mashed together in these crazy ways.
3:53This is just like the tip of the iceberg
3:55in a lot of ways, and I'm realizing this
3:58stuff goes a lot further than I would
4:00ever have guessed, especially for people
4:02who might not be engineers. It's
4:04genuinely kind of silly that it took all
4:06of these game mash-ups to make me
4:08realize that everyone can benefit from
4:10smarter models, but it it really made it
4:13click for me. And then there's a couple
4:15different angles to this. The most
4:17obvious one is the one that we've all
4:19been living with for a while now.
4:22Smarter models means devs can make
4:24better software faster. This has been
4:26the case, honestly, since like Sonic 4,
4:30that AI could allow developers to take
4:32their roadmap, their plans, the things
4:33they were going to build, and get
4:35through those faster and deliver
4:37software their users want more
4:39effectively than they could have before.
4:41And as models have continued to get more
4:43capable at dev stuff, the amount that a
4:46dev can do has gone up exponentially.
4:48I'm shipping more code in a week than I
4:49shipped in all of last year at this
4:51point. It's crazy. And I know it's all
4:54slop or whatever. You guys are just
4:55wrong at this point. Half the code that
4:57my AI is writing for me is better than
4:59all the code you've written in your life
5:00if you actually think that way. Like
5:01it's insane how far these things have
5:03gotten and how capable they are at
5:05making good decisions and doing real
5:06work. That is all good and dandy, but
5:09we're still presuming that the person
5:11has experience because that is kind of
5:14necessary to get the most most out of
5:16these tools. So, devs can do all of
5:18these things, and when they do these
5:20things and make better software, people
5:22benefit because they can use the
5:24programs that might not have been worth
5:26building before. For a silly example of
5:28this, I built a service called Lawn. I
5:30was using Frame.io, and I was getting
5:32more and more frustrated with how slow
5:34it was, with how expensive it was, with
5:36how Adobe it was starting to feel, and
5:39how annoying it was to just send a quick
5:40link to one of the brands that we're
5:42working with for a review and stuff like
5:43that. So, I threw together an
5:45open-source alternative myself in a few
5:47days, and it works great. This is how we
5:49do all of our video for my team now. For
5:52reference, Frame.io was acquired by
5:54Adobe for 1.3 billion dollars in October
5:57of 2021, and I would legitimately say we
6:00have about 70% of their functionality
6:03that they had at that acquisition point,
6:05as well as a ton of extra things that
6:07they still don't have today. That was
6:09considered a 1.3 billion-dollar product
6:12not that long ago, and now at the
6:14beginning of the year, I was able to
6:16recreate it and open source it cuz why
6:19not? All trivially.
6:22It's crazy how viable these types of
6:25things are when before they just
6:26absolutely weren't. And that was with
6:29like Gemini 3 Pro and Opus 4 6. Do you
6:33know how much it's improved since then?
6:34It's crazy. And I didn't even need to
6:36reverse engineer anything for this
6:38because the way the core product works
6:39made a lot of sense. It was just all the
6:41integrations and like edges you have to
6:42hit to make it work that made it hard.
6:44But that's the program I built that a
6:46lot of others have benefited from. It's
6:48not like a huge product, but there's a
6:49few hundred people that use Lawn to
6:52power their video pipelines for a lot of
6:53YouTube channels that you and I both
6:55watch, which is super super cool. And
6:57people even been forking it and building
6:58their own things on top. And that was
7:00where the next layer started for me. So
7:02you already have professional devs are
7:04able to make better software with
7:06smarter models, but the next piece I've
7:08been seeing I'll call technical people.
7:12There's no good term for this cuz I
7:13don't want to be like offensive. It's
7:15the types of people who like maybe they
7:16played with code a bit and just decided
7:18it wasn't for them. They're very smart,
7:20they're very capable. They just didn't
7:22go the route of being software
7:23developers for a living. I would argue
7:25somebody like a Linus Tech Tips would
7:27fit this archetype to an extent. Not
7:29script kiddies, but like people who know
7:32their way around a computer and have a
7:34good intuition for what software can and
7:36can't do, but very little intuition on
7:39what is or isn't realistic for them to
7:41make. And what's cool with these people
7:43is the power they get as models get
7:45smarter because they don't understand
7:48the limitations of code as well. They
7:50understand limitations of the thing the
7:52code does, they don't necessarily
7:53understand limitations of like different
7:55frameworks and tools and ecosystems. And
7:57if their app's getting really slow and
7:59they built it in Rails, they don't know
8:01what is making it slow. They might hear
8:04that PlanetScale is really fast and
8:06Railway's really fast and end up
8:08spinning up both in different regions
8:10and then their app is really slow and
8:11they just don't get it. As models get
8:13smarter, those gaps start to get
8:15plugged, too. And they can ask questions
8:17like, "Why is the app so slow now?" And
8:21it can do the types of analysis to get
8:23these answers. A highly motivated
8:25technical person that just never quite
8:27got good at coding can do so much more
8:29now. They can fork software they rely
8:33on.
8:34They can now also debug systems far
8:35beyond their understanding. They can
8:38whip up scripts or fork some program and
8:40when it doesn't work how they expect it
8:42to, they can change and adjust the same
8:44way that like you could open up your
8:46computer and swap a part, these same
8:48technical people can now open up their
8:50software and swap things out, too.
8:52Previously, they could like make their
8:53own small little apps with AI. Now, they
8:56can ask the AI to fork some tool they
8:58use and make changes to adjust it to
9:01their liking. They can reverse engineer
9:02apps they're using every day like Codex
9:04and ChatGPT or Discord and insert
9:06JavaScript to change the experience in
9:09the product. All of that was super cool,
9:12but it was very personal. It didn't like
9:13scale out the same way because models
9:16just weren't great at like packaging the
9:18work they had done and making it so
9:20others can adopt those things. You might
9:22be able to make a cool plugin for your
9:24local Discord, but setting that up in a
9:26way others can use it was still complex
9:28and dangerous and hard to get right. But
9:30while devs could make things others
9:32could use, technical people could use
9:35the AI tools to customize what they were
9:36already using, but not necessarily to
9:38productize it and give it out to others.
9:41And that might be my initial framing
9:43here.
9:43Tier one is it helps with things I'm
9:45good at. This is like I'm already
9:46editing code and I ask the AI to
9:48customize a function or maybe I let auto
9:50complete drive some of the lines as I'm
9:52changing them. Then it can make edits to
9:54the things I use often. This could be
9:56using AI to customize a tool that you
9:58use every day, to fork a library and
10:00make some slight adjustments to it, to
10:02do your job more effectively, but it's
10:04still like a very personal. Like the AI
10:06and the thing I build with it isn't
10:08really leaving my little bubble. It's
10:10tier one, I am telling the AI like what
10:12its constraints are and building
10:13everything around it. Tier two is I'm
10:15letting it do a bit more and it's
10:17exploring places I'm not as familiar,
10:18but it's still limited to me and my
10:20little circle of like the software I
10:22use. This is using AI to build a
10:23personal Chrome extension or to make a
10:26Discord script that changes how Discord
10:27works for you or even like the personal
10:28AI assistants and things. All I would
10:30put here. Then it gets to the point
10:32where they can build things you don't
10:33fully understand. Where I have a bunch
10:36of projects I've been working on in Rust
10:38and I don't actually write Rust. I
10:40barely understand the language. I played
10:41with it a bit. I find it tedious and
10:43annoying, but I'm writing a lot of
10:44projects in Rust now even if I'm not
10:46writing the code. And I'm also able to
10:47distribute those projects and I will be
10:49shipping many of them to real users in
10:51the very near future. And then all the
10:52way at the bottom, there is the crazy
10:54new unlocked tier of oh [ __ ] we can use
10:57these tools to decompile, reverse
10:59engineer, and rebuild the things that
11:01already exist in the world at like their
11:04very core. Things that like are
11:06basically impossible to understand. Like
11:08no one fully understands the decompiled
11:10code base for Super Mario 64. That's
11:13just not viable. And it's magical, truly
11:17unbelievably so, that these tools can
11:19build a good enough comprehension of
11:21something like that to operate within
11:23it. The reason I'm tiering this is when
11:25model level ups happen, there are
11:27different categories of people that are
11:29affected differently. So if we have devs
11:32for example, when AI coding first
11:34started to be a thing with like OG
11:36Co-pilot, it would help devs with things
11:38they're good at, and technical people
11:41were not helped at all in this framing.
11:43Like they could ask it questions and get
11:45answers, sure, but they're not building
11:47things with it yet. But then the models
11:50got better, and then they got to the
11:52point where they could like actually use
11:54a tool and edit some code. And then this
11:57happened. Technical people could use
12:00these tools to like customize their home
12:02pages or like work on their Wix sites
12:05and stuff like that. Devs could suddenly
12:07use it to actually customize the tools
12:09they use to edit some packages that
12:11they're relying on to make real changes
12:13to their code base that are multiple
12:14files that are being touched. And then
12:17the models got better again. Opus 4.5
12:19dropped, and all of a sudden everything
12:21shifted again. Technical people were
12:23suddenly able to build full apps that
12:25kind of worked. There were people who
12:27had never coded months before building
12:29benchmarks that they couldn't read for
12:31things they didn't understand. And devs
12:33suddenly could go even deeper. This is
12:35when a lot of us started like building
12:36games as side projects or trying to port
12:39libraries to different things or use
12:42these tools to automate some of these
12:44decompilation processes. And then the
12:46models got better again. And I would
12:48argue with Fable 5 and eventually with
12:50Astra, the jump for devs was immediately
12:53huge, but the jump for technical people
12:55was not quite as big, and I thought this
12:57might be the end to an extent. That the
12:59gap from technical people who don't code
13:02to devs who will know life like all
13:04their understandings, I thought this gap
13:06was going to just keep getting bigger
13:07and bigger and bigger. And then Opus 5.5
13:10happened, and the thing that made Opus
13:125.5 so interesting to me is while it
13:14didn't really move what I was doing with
13:17AI much, like I do a lot more, but it
13:19didn't necessarily unlock a lot of new
13:21capability for me that I couldn't piece
13:23together with other things, the fact
13:25that there was one model that could do
13:27all of this stuff that was well priced
13:29and accessible on the subscription tiers
13:31bumped technical people deep deep deep
13:34into this idea of being able to build
13:36things they don't understand. There are
13:37now people who have never coded before.
13:39They've never opened Blender. They've
13:40never done any of those things that are
13:42using their Claude subscription they're
13:43paying 200 bucks a month for to port
13:45maps from their old favorite games into
13:48new games that they were reverse
13:49engineered. That's crazy. But there's
13:51another tier of people non-technical
13:54people.
13:55I don't think they have been moved
13:58anywhere near as much. But that's an
14:00opportunity for the people in these
14:02other areas. If you are the person who
14:04fixed their mom's computer for a long
14:06time, you might be able to find more and
14:08more problems that your mom has with
14:11other things that AI could solve either
14:13by showing her to use an AI tool or
14:14building her something useful with your
14:16AI tools. You also can start modding
14:18games in ways that you fundamentally
14:20couldn't understand in this position and
14:23then you can share those mods with your
14:25friends. And I am sure there's a bunch
14:26of people playing the Call of Duty Halo
14:29mashup that don't understand anything
14:32about how it works. They were just sent
14:34an EXE from a friend who got it all
14:36figured out and these technical people
14:38can now share so many more cool things
14:41with their friends. And here is where my
14:43spiciest take comes in.
14:45So we need to talk more about these
14:46people, the devs. Cuz I have a harsh
14:50reality I need a lot more of y'all to
14:51wake up to.
14:53We're not very creative. And the more
14:55creative you think you are as a dev, the
14:57more likely you are to have really bad
14:58ideas generally. Devs are not very good
15:03at most things that aren't delivering
15:05code.
15:07That's just reality. The reason that my
15:10channel is so popular isn't cuz I'm
15:11super smart or really good at branding
15:14clickbait thumbnails. It's cuz I'm like
15:17six out of 10 at communication in a
15:19space where where average is a three. So
15:21I'm able to talk about these things and
15:23be a lot easier to listen to than your
15:26average dev who would otherwise talk
15:27about these things. This is why so many
15:29dev teams have a non-technical designer
15:31and a product manager that don't know
15:33[ __ ] about code that are being paid 100k
15:35plus a year. It's cuz we need them
15:38because we don't have the taste or the
15:40creativity to do anything other than
15:42like link two libraries up in a novel
15:44way. We're not good at this. And here is
15:47where the spice comes in. Technical
15:49people who can't code are necessarily
15:52decent at this. These are the people who
15:54got really into Zapier back in the day
15:56because they didn't know how APIs or
15:58webhooks worked. So, they would use
16:00those tools to force things to work
16:03together in ways that made no sense.
16:05I've seen some OBS setups that hurt my
16:08brain because it's just so far outside
16:12of how the program was built to work.
16:14Like I know somebody who has over 40,000
16:17scenes in her OBS because she uses them
16:19all as dynamic layers that link to and
16:22stack on top of each other for crazy
16:24dynamic scenes in like situations that
16:27she'll create in OBS. Like I've seen
16:29some crazy [ __ ] And that's all weird
16:32creative [ __ ] thinking. And suddenly
16:35those types of people have had their
16:38like shackles taken off to an extent.
16:41And that's what's so exciting here is
16:43what happens when those people who don't
16:46understand how this all works, who don't
16:48understand that a game made in IW4
16:51should not work inside of the same
16:54engine as a game made in UE3 versus a
16:56game made in whatever the [ __ ] EA was
16:57cooking at the time. People who don't
17:00have that strong like fundamental belief
17:03that like these things are different,
17:04therefore they can't be combined. Or
17:06these things work this way because
17:08they've worked this way for 20 years.
17:09All of this stuff is just it it's hard
17:13for us to comprehend. And I know this
17:15for a reason that is really, really
17:18silly. I know devs aren't good at this
17:20because of my most popular short.
17:22>> Why can't we commit our environment
17:23files? It sounds stupid when I put it on
17:25a slide like this, but I want you to
17:26really think about this for a second.
17:28When I have a team of engineers that are
17:30working on a project, why do I have to
17:32build another system to share this
17:34specific file, but all the other files
17:36can go and get just fine? We have to
17:38take the step back and think, is this
17:40how we do things cuz it's right, or is
17:42this how we do things cuz it's just how
17:44we've always done it?
17:45>> Why can't we commit our environment
17:46files?
17:46>> I played this short because I have
17:48learned this take is an unintentional
17:52I don't want to say IQ test. It's a like
17:54creative thinking exercise for engineers
17:56who are not good at creativity and
17:57they're often not good at thinking. The
17:59reason I brought this up in my AI world
18:01fair talk was because I want us as
18:04engineers to start unraveling these
18:07deep-set beliefs that we have. It's time
18:09for us to take a step back and look at
18:13what we're doing and why we're doing it.
18:15And one of the best ways to do this is
18:17to have outsiders come in and be like,
18:19that makes no sense. And they do things
18:20differently cuz they don't know any
18:21better. This happens a lot in music
18:23where some random kid from some foreign
18:25country will like something they hear in
18:26another place and make their version of
18:29it, and by the time it's notable enough
18:31to be heard by people from the original
18:33culture, it's also refined enough that
18:35they can't really poke holes in the
18:37things that it does differently. We're
18:38about to hit this era with software
18:41where people who have no idea what
18:43they're doing can come in with ideas
18:44that do not make sense to us because
18:46we've been engineering for 20 plus years
18:49and these kids don't know the difference
18:51between Java and JavaScript, and they're
18:52going to make things that are really
18:54cool and surprisingly powerful because
18:57they don't know to not ask for it. And I
18:59know we're not going to be very good at
19:01this as engineers because of the comment
19:04section in all of these videos. Why
19:06can't I post a picture of my credit card
19:08to Twitter? What kind of tech brain rot
19:10is this? The saga of the sinking Theo
19:12continues. Homie still learning soaps
19:15exist. I can give you a couple reasons.
19:18Decoupling sourcing keys, sometimes in
19:20the company you want to see code of one
19:22another but don't have the ability to
19:23launch it. Yeah, you're you're assuming
19:25everything still works the way it works
19:27today. The point I'm trying to make is
19:29that GitHub giving everyone access to
19:32every file into get as an origin system
19:35sharing every file with everybody with
19:37access might not be the right way to
19:40manage our source code. Love this
19:42comment from Shiv. Oh, you want me to
19:44think? Nah, too much. I'm a software
19:46engineer, man. Yeah, I'm not paying you
19:47to think. I'm paying you to write code.
19:49Boil the ocean, but we can't even find
19:50the ocean actually.
19:52Gamer girl
19:53The reason I'm talking about this I
19:55think devs like you are going to be very
19:58well off cuz there's a lot of devs who
20:00are creative weirdos that do strange
20:02[ __ ] with code that have struggled to
20:04find a job for various reasons usually
20:06dumb. That has shifted a ton. Highly
20:09motivated people with a bit of
20:11creativity and a willingness to think
20:13outside of the box and dive on to things
20:15they don't understand fully, they're
20:16going to be rewarded absurdly heavily in
20:19this era. Where the the less tied you
20:22are to the way things have to work right
20:24now, the more likely you are to break
20:27out of that box a bit and come up with
20:29more unique and creative things. And
20:31when things don't work, you can ask
20:33questions and figure out why. A simple
20:36way of putting this that I actually
20:37really like from chat. Yeah, here it is
20:39from chorus here. Try first to
20:41understand why not cuz there's a lot of
20:43questions and things we do in software
20:44dev that are effectively just cuz cuz
20:46the tools have worked this way forever
20:48and it would be too expensive to try and
20:51change it. Things like the compilers we
20:53use written in one language like porting
20:55it to another is just too much effort
20:56and time and risk and keeping it synced
20:58is not necessarily worth it. Or apps
21:00that were built in one language that
21:02aren't worth porting to another because
21:04it's expensive to to
21:05Or if you have an audio device that's
21:07bugging you, like I often do,
21:09technically speaking, if I was to spend
21:11a few months to a few years, I could
21:13absolutely reverse engineer my
21:14Rodecaster and write custom firmware for
21:16it. It would just take me a very, very,
21:18very long time. Now, these things are
21:20much more accessible. The amount of
21:23effort and historical experience you
21:25need to try out these crazy YOLO runs is
21:30so much lower. And I've done these
21:32things before AI could help me with it.
21:35Where I try some crazy bold idea, like
21:37building my own tech stack, or trying to
21:39build my own game engine, which
21:41obviously didn't go as well as the other
21:42things I did, or trying to build my own
21:44protocol for WebRTC STUN servers to make
21:47the quality of the connections better
21:49without causing any latency issues.
21:51Like, I have been trying these types of
21:53crazy YOLO missions for the better part
21:56of my career. And now those explorations
21:59are much less expensive and much more
22:01accessible, which means you have to be
22:03willing to go down them. But you also
22:05have to be willing to try things and
22:07have them fail. And I find a lot of devs
22:09are worse at that. Since devs are used
22:11to working in the things in the systems
22:12they understand, even if they don't
22:14understand them fully, they are used to
22:16failures, but it's like they change five
22:18lines of code, it doesn't do what they
22:20expect, so they go change them a few
22:21more times. It is so rare that a dev
22:24files a PR that doesn't eventually get
22:26merged that we almost feel bad. Like,
22:29who here has merged a pull request
22:31because they felt bad closing it because
22:34they know the person who made it worked
22:36really hard on it. I know I have. Let's
22:38see some yeses and ones in chat if
22:40you've ever guilt merged a PR.
22:44See all those?
22:46This happened because it feels bad
22:48throwing away somebody's work, but we're
22:50also just so used to the point of a PR
22:52going up being that that PR will merge.
22:55And that is just a deep culturally set
22:58thing that we have to start walking
23:00back. And let's be real here. Let's say
23:03a professional dev who's written code
23:05for years spends 2 hours on some slop
23:09thing that is a 2,000 line of code PR.
23:12And then separately, a technical person
23:15who isn't a dev spends 5 hours on the
23:18same thing and makes a nearly identical
23:20PR. Who's going to feel worse when you
23:22close the PR because you don't want that
23:25in your product? The engineer who spent
23:262 hours on it or the technical person
23:29who spent 5 plus?
23:30>> I think we all know the answer to this.
23:31It's obviously the developer. They value
23:33their time more highly. They have
23:35expectations when they are building
23:36these things. They are used to PRs they
23:38file being merged. And if they can do a
23:41week of work in 2 hours, they're still
23:43putting up a PR that feels like that
23:44week of work. And when you don't merge
23:46it, they still feel bad the way they
23:47used to. The technical people who
23:49haven't coded before don't have this
23:50experience, but they also just don't
23:52have the expectation, either. Because
23:55they're not going in expecting this to
23:56work or expecting this to land.
23:58They're going in expecting this to be a
24:00fun experience, really. And that
24:02difference in expectation going in has
24:05allowed these people to do crazy [ __ ]
24:08recently. So, what I'm trying to say
24:09here is we need to get over ourselves a
24:11bit as developers. We need to be a
24:13little more willing to take a step back
24:15and question things. We need to feel
24:17less bad when the thing we do doesn't
24:19work and should be thrown out. We need
24:21to ask more questions both to our agents
24:23and to ourselves about what we're
24:24building, why we're building it, and
24:26what we could potentially do
24:27differently. We need to be a little bit
24:28more willing to boil the ocean and do
24:30these crazy things that don't really
24:33make sense. I remember just a few months
24:36ago
24:37my Yolo project rewriting the entirety
24:39of the TypeScript compiler in Rust felt
24:42like the unobtainium of AI. Like maybe
24:45someday AI will be able to make a little
24:48bit of headway in this project. Not only
24:50was it working in under 20 hours, it is
24:54now two two three times faster than TS
24:56Go. And most importantly for my actual
24:58reason I want it, it works in WASM. I
25:01didn't think this project would ever be
25:02able to like realistically happen. If I
25:05was lucky, it might pass a compilation
25:07here and there. Now it works on all the
25:09real projects I care about with 99.99%
25:12accuracy against the other versions of
25:14TypeScript. And not to keep calling out
25:17chat, but raging here said that when I
25:19first started the project that they made
25:21fun of me for it so hard because they
25:23never thought it would work. That's why
25:25I started it. I started this project as
25:27like my attempt to think of a thing AI
25:30would never be able to do. And now this
25:31feels too easy. Like yeah, of course the
25:34AI could rewrite a known target like a
25:37compiled app in a different language.
25:39But just a couple weeks ago this didn't
25:41seem possible. It's crazy how much
25:45things have changed. And I'm not even
25:47thinking bold enough, and it's my job to
25:49think too bold so I can make videos out
25:51of it. I like this way of framing it
25:52Game Mechanic. I'm a software engineer
25:54as well as a systems engineer, and my
25:55hot take is that most devs are pretty
25:57happy writing the same damn linked list
25:59written by every other dev over and over
26:00again. They don't really like
26:02differentiated workloads. As a systems
26:04engineer, I'm basically annoyed writing
26:06code that's already been written before.
26:08I think the people who can work at
26:09higher levels of abstraction and not
26:10care about the low level low level
26:12implementation are going to be the
26:13winners. I pretty much entirely agree
26:15with this. The only thing I would change
26:17is the people who are willing to be
26:19flexible with how low or high level they
26:21go, they are going to be the ones who
26:23benefit the most. The people who are
26:25willing to throw away their preferences
26:28in order to make something better for
26:30their goals are going to benefit a lot.
26:33Like do you have any idea how many
26:34technical decisions I have had to make
26:37in projects that are not the first
26:39decision I would prefer at all because
26:42they are the thing that gets me my goal
26:45the fastest. Like I really do genuinely
26:47want to switch us over to the SwiftUI
26:50rewrite for the mobile app, but the
26:52speed the agents can operate at with the
26:54React Native app, combined with the OTA
26:55updates, and all the pushing that Julius
26:58and Jay have done to make the
26:59performance not just comparable, but I
27:01would say in many ways faster. I'm not
27:03going to probably ship the version that
27:05I prefer, especially on a technical
27:07level cuz I know it's more capable,
27:08because the agents, the tokens, and the
27:10time are going into a different thing.
27:12Or, for a sillier example, my fleet. You
27:14might notice some of these boxes have a
27:16new OS on it. You would imagine if I was
27:20making T3OS, that it would be full of
27:22all my opinions. Well, if I do end up
27:25shipping T3OS, it's not going to look
27:28like what you think, because it has none
27:30of my opinions in it. I didn't make this
27:33new distro of Ubuntu because I want to
27:35have a distro that does everything I
27:37want. I'm building this distro because I
27:40want a distro that has none of the
27:41things I want. It's designed to have the
27:44things the agents need instead. T3OS
27:46doesn't have a GUI when you boot it. It
27:49is headless, but it will spin up GUIs
27:52that are a fork of XFCE for my agents to
27:54use through the driver level with Kua
27:57control, so they can do computer use for
27:59each individual thread. It's also using
28:01Bash instead of Zsh or Fish, because as
28:04much as I prefer Zsh, I very much prefer
28:06Zsh, I recently audited all of the
28:08errors that my agents have, and over 30%
28:11of the errors that Claude runs into
28:13across my machines are because the Zsh
28:16machines format newlines slightly
28:18different, and Claude malformats the
28:20Ripgrep commands. I've also learned my
28:22agents aren't quite as good at dealing
28:24with Arch and Cash OS for a dev box. And
28:26as much as I would prefer to run Cash on
28:28all my [ __ ] I've accepted it's not the
28:30best choice for a lot of work and for a
28:32lot of my boxes. So, now I'm making a
28:35not really a fork of Ubuntu. It's more
28:36like a super customized Ubuntu image
28:38that is customized for the agents, not
28:40me. And this is weird because on one
28:42hand I'm going deeper, customizing the
28:44OS, but on the other hand I'm going
28:46shallower because I'm not making the
28:48decisions for myself. In my willingness
28:50to be flexible up and down the spectrum
28:53of depth for this stuff. Because I'm not
28:56building the product for me is how I'd
28:57put it. Or like I'm not building T3OS
28:59for me. I'm building T3OS for my agents
29:02so they can build things for me. And if
29:04we are not more willing to think about
29:05things this way as devs, not only are we
29:08going to stop benefiting from the models
29:10getting smarter, these technical folks
29:12who don't have the same boundaries in
29:15their brains are going to accelerate far
29:17past us. That person you used to get
29:19annoyed by in high school because their
29:21solution to everything was install
29:23Ubuntu, even though they've never
29:25programmed before, that person's going
29:27to run laps around you if you're not
29:28careful. And I say all of this because I
29:31was wrong. I had the silly belief that
29:34models would be putting out things that
29:36were so complex that technical people
29:38couldn't understand it and the devs
29:40would be the only ones who could. And
29:42that same knowledge that made us more
29:44capable of wielding Opus 4.5 and
29:46eventually even Fable is starting to
29:48hold us back. So what happens when this
29:50all hits the non-technical folks? I have
29:53no idea. I'm just paid to talk about
29:55code, guys. Seriously though, like I've
29:57been thinking too much about this and I
29:58just wanted to get the ideas out of my
30:00head because these things are changing
30:02faster than I expected or still
30:04accelerating. And it turns out a model
30:06that can write its own compiler is
30:08useful to people who want to rewrite
30:10their old favorite games. And I never
30:12thought a model that capable would be
30:14useful to my friends in high school, but
30:16I didn't think people would be rewriting
30:18their favorite Xbox 360 games from
30:20decompiled ports in Rust plus Bevy and
30:23then mashing them up together, either.
30:25We're at the very, very tip of this
30:27crazy-ass iceberg and I hope that you
30:29guys are prepared for the future as it
30:32is coming. Cuz god damn,
30:34it's not going to get any slower. What
30:36I'm trying to say is that I was wrong
30:38and I've been reflecting on it a lot and
30:39I know a lot of y'all probably are
30:41thinking the same way I was just weeks
30:43ago. I hope this video convinces you to
30:45start rethinking where these models are
30:48going and who will be able to use them
30:50because I know I have been blown away
30:52with what we can do. Every time I get a
30:53DM from an old friend from high school
30:55or somebody I used to work on random
30:57audio engineering stuff with showing me
30:59the crazy things they're building with
31:00help us. I realize that we are very very
31:03early still.
31:04And where we are going is going to be
31:06pretty damn cool.
31:08So that's all I have to hear. Let me
31:09know what y'all think and until next
31:10time.
31:11Peace nerds.

About this transcript

This page contains the full transcript of Opus 5.5 is proof that we do need smarter models by Theo - t3․gg, generated from the public captions YouTube serves with the video. The transcript has 6,476 words across 904 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.