00:00:00
Life just got a heck of a lot harder for
00:00:01
all those VS Code forks that you know
00:00:03
and love. Yes, believe it or not,
00:00:05
Microsoft just opensourced their AI
00:00:07
editor. What they mean with this is
00:00:09
they're about to start open sourcing all
00:00:10
of the GitHub copilot chat pieces that
00:00:13
are in VS Code. And that's not the end
00:00:15
of the things they announced. They also
00:00:16
announced that they're open sourcing
00:00:17
Windows Subsystem for Linux, which is
00:00:19
really cool. This is a huge
00:00:21
philosophical change for Microsoft. I
00:00:24
was lucky enough to chat with them a
00:00:25
little bit about what's going on and I
00:00:27
have a ton that I want to share with you
00:00:28
guys both from the philosophy of why
00:00:30
Microsoft is doing these things to the
00:00:32
impact it will have on a bunch of
00:00:34
companies you know that I talk about a
00:00:36
lot here. I'm also an investor in a
00:00:38
bunch of those companies and this might
00:00:39
be the end of them. So, uh, someone has
00:00:42
to pay the bills. We're going to do a
00:00:43
quick sponsor break and then dive into
00:00:45
all the things you need to know about
00:00:46
this change. AI has made a lot of things
00:00:48
easier, but it made one thing
00:00:49
significantly harder. Interviewing. Yes,
00:00:52
recruiting is in a terrible place right
00:00:55
now. If you've opened up a job listing
00:00:56
on your site, you know exactly what I'm
00:00:58
talking about. You're suddenly getting
00:01:00
tens of thousands of applicants with
00:01:02
terrible AI generated resumes and trying
00:01:04
to find the good ones in the pile. It's
00:01:05
nearly impossible. Trying to hire a
00:01:07
great engineer feels like just a
00:01:09
non-stop hill you're climbing up. Unless
00:01:11
you use today's sponsor, G2I has solved
00:01:14
hiring. I know it's a crazy statement,
00:01:16
but honestly, all of the people that
00:01:18
I've been talking to that started using
00:01:19
G2I have been blown away with them. So,
00:01:22
you can take my word for it or we can
00:01:23
just read some of these crazy
00:01:25
testimonials. Shop Monkey said that they
00:01:27
made 17 hires in 60 days. Do you know
00:01:30
how crazy that is? Especially to not
00:01:33
have to fire half of them after cuz you
00:01:35
hired too quickly. If these are just
00:01:37
random junior engineers with no
00:01:38
experience or vibe coders that have no
00:01:40
background, that'd be one thing. But
00:01:42
that's not the case at all. G2I has been
00:01:44
working for eight years to build up
00:01:46
their pool of over 8,000 experienced
00:01:48
engineers. These are people with real
00:01:50
industry experience. all over the stack,
00:01:52
back end, front end, mobile, web,
00:01:54
whatever the heck you guys are doing,
00:01:55
there's a very good chance you have a
00:01:57
bunch of awesome engineers that are
00:01:59
already part of G2I's network. The
00:02:00
coolest part though is the video
00:02:02
platform. Once you've decided you want
00:02:04
to hire for a given role, you write up a
00:02:06
bunch of questions, you give them some
00:02:07
stuff that's important to you, and then
00:02:08
G2I goes out and gives those questions
00:02:11
to a bunch of the engineers in their
00:02:12
network. And then you get to use their
00:02:14
app to go through all of these
00:02:15
interviews with all of these people
00:02:16
where humans are answering the question
00:02:19
on video. so you can get an actual idea
00:02:21
of what this person is like to work
00:02:22
with. They're not just handing you a
00:02:24
list of names and saying good luck.
00:02:25
They're doing most of the hard work for
00:02:27
you. From the vetting to the interview,
00:02:29
they will smooth the process out.
00:02:30
They'll even spin up a shared Slack
00:02:32
channel with you as part of their
00:02:33
default process. They've said for a bit
00:02:35
now that from starting working with them
00:02:38
to first PR from your new hire is 7
00:02:41
days. My favorite thing about the
00:02:43
platform though, and I wish we could do
00:02:45
this more realistically for traditional
00:02:46
roles, when you do decide you like
00:02:48
somebody, you bring them on for 7 days,
00:02:51
and if they're not shipping the way you
00:02:52
want them to, you don't have to pay a
00:02:54
scent. And if you want someone
00:02:56
different, they will backfill within the
00:02:58
week. Remote or in person, backend or
00:03:00
frontend, contract, full-time, whatever
00:03:02
you need, these guys will hire in days
00:03:04
instead of months. Give them a shot
00:03:06
today at
00:03:07
soyv.link/gti. So, why is Copilot going
00:03:10
open-source? We've talked about parts of
00:03:12
this before, but I think it's important
00:03:13
to like quickly understand what VS Code
00:03:16
being open source meant in the first
00:03:18
place. VS Code is an open- source editor
00:03:20
that a lot of people use. It's probably
00:03:22
the most popular editor ever made at
00:03:24
this point. It's open source, which is
00:03:26
awesome because you can fork it. It's
00:03:28
MIT licensed. You can do a ton of cool
00:03:29
things to it. But for the most part,
00:03:31
we've been extending VS Code with the
00:03:34
extensions API, which is a way to add a
00:03:36
widget into the side of VS Code that has
00:03:38
an iframe that can do things.
00:03:40
Occasionally, it's stuff like language
00:03:41
servers or themes and whatnot. But
00:03:43
historically, we've been relatively
00:03:45
limited with what we can do using
00:03:47
extensions. So, as such, the effective
00:03:50
role of VS Code being open source was so
00:03:52
that people could see the code, make
00:03:54
changes to things that were broken, and
00:03:56
contribute with like the team making it.
00:03:58
But it wasn't like a bunch of people
00:04:00
were using that code to do things. It's
00:04:01
they were fixing things wrong with it.
00:04:03
That is until one big thing happened.
00:04:06
That's what we're talking about today.
00:04:07
It's co-pilot. Copilot kind of changed
00:04:09
the game. The idea of your editor
00:04:11
writing your code with you and for you
00:04:13
was huge and immediately resulted in a
00:04:16
explosion of things trying to do the
00:04:18
same. The first attempts were mostly
00:04:20
also VS Code extensions. The problem was
00:04:22
that you needed a really deep
00:04:24
integration with VS Code to have a good
00:04:25
experience to do things like tab
00:04:27
complete jumping to the thing that you
00:04:29
want to change or having like the diffs
00:04:31
inlined underneath the thing that is
00:04:33
changing. So despite VS Code being open-
00:04:35
source and the extension platform being
00:04:37
very well established, the capabilities
00:04:40
of the platform did not align up with
00:04:41
the things people wanted. And slowly
00:04:43
copilot had to diverge from the
00:04:45
traditional extension path and it had a
00:04:48
bunch of special stuff built into VS
00:04:50
Code that they could use that other
00:04:51
extensions couldn't. So what just
00:04:53
changed? I want to be very clear about
00:04:55
what hasn't changed first and foremost.
00:04:57
So C-pilot server backend is still
00:05:02
closed source. So if you think you can
00:05:04
get the whole co-pilot experience,
00:05:06
backend, frontend, servers, management
00:05:08
of tokenization and like context windows
00:05:10
and all the chaos that it takes to build
00:05:11
something like this, they were very
00:05:13
explicit that that is not what they are
00:05:14
doing here. The magic of the co-pilot
00:05:16
servers and APIs is not something we get
00:05:18
any insight into. But that's not
00:05:20
actually something I care that much
00:05:22
about. The other important pieces is
00:05:24
that this hasn't happened fully yet. So
00:05:28
this wasn't like, oh, here's the code.
00:05:30
It's all open source now. You can do
00:05:31
whatever you want. This is the
00:05:32
announcement of the plan to start
00:05:34
changing these things. And one of the
00:05:36
big things they plan to change, which
00:05:38
again is not done yet, but they are
00:05:40
working on it and want the community to
00:05:41
be involved, is the opening up of the
00:05:43
APIs that co-pilot uses. They do very
00:05:47
much intend to make it so a third party
00:05:49
that's willing to build their own
00:05:50
servers, handle the inference stuff
00:05:52
themselves, could get the same quality
00:05:54
of experience in VS Code that Copilot
00:05:56
has. In order to do that before, you had
00:05:59
to fork VS Code, which is why we saw so
00:06:01
many forks from void to pair to Windsurf
00:06:03
to, of course, cursor. I'm invested in
00:06:05
three of the four I just mentioned. This
00:06:07
is going to be a fun window for all of
00:06:10
those. That said, this also kind of
00:06:12
levels the playing field. As crazy as it
00:06:14
sounds to say Microsoft came in and
00:06:17
potentially directly harmed the plans
00:06:19
and business models for all of those
00:06:20
companies, this means more businesses
00:06:23
can compete more effectively. If you
00:06:25
don't have to build a whole editor and
00:06:27
manage a whole VS Code ecosystem forked
00:06:30
out from scratch yourself, which means
00:06:32
you have to let go of things like the
00:06:34
marketplace, you have to manage security
00:06:36
incidents and pull in the changes by
00:06:38
hand. You have to deeply understand VS
00:06:41
Code to do that. Now you can just make a
00:06:44
new extension that has a similar quality
00:06:46
of experience. One way of thinking of
00:06:48
this is like the what capability and
00:06:50
quality was possible. Let's say we have
00:06:53
a chart showing the quality that's
00:06:55
possible with different solutions. If
00:06:58
you are co-pilot, we'll say co-pilot's
00:07:00
quality is here. So co-pilot is this
00:07:03
good. If you were just building an AI
00:07:05
extension on VS Code before the things
00:07:07
that are announced today get shipped,
00:07:09
the quality you were capable of shipping
00:07:10
was much much lower than what C-Pilot
00:07:14
can do. So if you wanted to meet this
00:07:16
bar that co-pilot had set, if you saw
00:07:18
the line there and you wanted to build
00:07:21
something that was as good if not
00:07:22
better, you really couldn't because we
00:07:24
were limited by this bar, the quality of
00:07:27
what was possible via an extension. So
00:07:30
in order to get past that, in order to
00:07:31
get to where Copilot was at and
00:07:33
theoretically go even further, you had
00:07:35
to fork. And this is the problem that
00:07:38
the Microsoft team saw. VS Code fork. It
00:07:42
is significantly more capable if you
00:07:44
fork VS code and you have the team
00:07:45
that's capable of doing it and managing
00:07:47
it. We've seen all the crazy stuff you
00:07:48
can do in stuff like you know cursor. I
00:07:51
love it. It's a really good editor
00:07:52
experience. The problem is that
00:07:54
maintaining a VS Code fork has a ton of
00:07:56
consequences and problems and it also
00:07:59
means the entry point to do this is
00:08:01
really high. We could also frame this as
00:08:03
like how hard is it to build these
00:08:05
different things and honestly the
00:08:07
current chart would kind of represent
00:08:08
that as well. Building and maintaining a
00:08:11
fork of VS Code is really hard if you
00:08:14
don't have the budget of a team like the
00:08:15
co-pilot team at Microsoft. What
00:08:17
Microsoft's trying to do here is they
00:08:18
want to take the AI extensions and the
00:08:20
idea of people building things like that
00:08:21
new integrated experiences like Klein
00:08:24
think things like Klein and Augment
00:08:26
Code. If you're not
00:08:28
familiar is an agent that you can
00:08:30
install as an extension inside of VS
00:08:32
Code. It's also open source which is
00:08:34
really cool. And then there's Augment
00:08:36
Code who has been a sponsor of the
00:08:37
channel that I quite enjoy using.
00:08:39
They're one of the few AI like code
00:08:40
things I use outside of cursor because
00:08:42
they do an incredible job of indexing
00:08:44
gigantic code bases. They're not paying
00:08:47
for this video. I just really like using
00:08:49
them to download an open source repo and
00:08:51
try to figure out how it implemented
00:08:52
something. So things like that are
00:08:54
currently very limited by what they can
00:08:56
do in VS Code because they effectively
00:08:57
just have the iframe API. So Microsoft's
00:09:00
trying to do is pull this out so it can
00:09:02
get to the same quality level as
00:09:04
C-pilot. Does that mean they can go as
00:09:06
far as a VS Code fork? probably not. But
00:09:08
at the very least, wherever the bar is
00:09:10
set for co-pilot, over time, the
00:09:13
capability of extensions is going to get
00:09:15
to the same place, which is a very
00:09:17
exciting change. Let's quickly read what
00:09:20
they have to say about this. So, it's
00:09:21
not just my thoughts and words. We
00:09:23
believe the future of code editors
00:09:24
should be open and powered by AI. For
00:09:27
the last decade, VS Code has been one of
00:09:28
the most successful open source projects
00:09:30
on GitHub. We are grateful for our
00:09:32
vibrant community of contributors and
00:09:33
users who choose VS Code because it is
00:09:36
open source. Za becomes core to the
00:09:38
developer experience in VS Code. We
00:09:39
intend to stay true to our founding
00:09:41
development principles. Open,
00:09:42
collaborative, and communitydriven. We
00:09:45
will open source the code in the GitHub
00:09:46
copilot chat extension under MIT. This
00:09:48
is something I meant to call out
00:09:49
earlier. They're not just open source,
00:09:51
they're MIT licensed, which means you
00:09:53
can do whatever you want with them. It's
00:09:54
nice to see them not change that. They
00:09:56
could have done a license that was like
00:09:58
you can make whatever you want with this
00:09:59
but you can't sell a competing product.
00:10:02
There's a lot of companies that have
00:10:03
licenses like that. They just went MIT
00:10:05
so you're still able to fork. You could
00:10:07
even make the argument that building
00:10:08
your own cursor just got a lot easier
00:10:10
due to the stuff that they are planning
00:10:12
to do here. As they were saying once
00:10:14
they've open sourced this they plan to
00:10:16
carefully refactor the relevant
00:10:17
components of the extension into VS Code
00:10:19
Core. So the parts that are currently
00:10:22
allowing for a lot of the custom
00:10:23
integrations for the cool like
00:10:25
autocomplete inline stuff, all the
00:10:27
things that make the copilot chat
00:10:29
extension unique are going to start
00:10:30
making their way into VS Code Core so
00:10:32
other extensions can take advantage of
00:10:34
them. As they said, this is the next and
00:10:37
logical step for us in making VS Code an
00:10:39
open-source AI editor. The reflection
00:10:41
that AI powered tools are core to how we
00:10:43
write code. A reaffirmation of our
00:10:45
belief that working in the open leads to
00:10:47
a better product for our users and
00:10:49
fosters a diverse ecosystem of
00:10:51
extensions. Really cool to see. The
00:10:53
obvious next question is why now when we
00:10:56
have cursor raising a ton? We have wind
00:10:58
surf maybe getting bought. Still haven't
00:11:01
gotten an update on that by the way. I
00:11:03
think it's happening. The rumors have
00:11:04
gone way too far for them to not. But
00:11:06
it's interesting. I think after this
00:11:07
news especially they're going to want to
00:11:08
take that deal. Then there's Perryi and
00:11:10
Void, which are also both open source VS
00:11:12
Code forks focused on AI experience and
00:11:14
AI code stuff. I haven't heard much from
00:11:17
either of those, which is concerning. So
00:11:19
with all of that going on, why now? They
00:11:22
were pretty transparent about this,
00:11:24
which I thought was cool. Over the last
00:11:25
few months, we've observed shifts in AI
00:11:27
development that motivated us to
00:11:29
transition our AI dev in VS Code from
00:11:31
closed to open source. The biggest
00:11:33
point, and I think this is really
00:11:34
important to understand, is that LLMs
00:11:35
have been continuously significantly
00:11:37
improving. So the secret sauce that made
00:11:40
co-pilot work in the past matters a lot
00:11:42
less. Prompts are going open constantly
00:11:45
now. More and more I'm seeing companies
00:11:47
saying screw it. Who cares if our prompt
00:11:49
gets shared? It's not that special
00:11:51
anymore. As the models get better, the
00:11:53
system prompts not saying they don't
00:11:54
matter. I'm saying they they are less
00:11:56
secretive and they are less uniquely
00:11:59
valuable. Especially as a new model
00:12:01
comes out. Your old system prompt in
00:12:03
order to fix things like weird diffing
00:12:04
might just not work at all. Especially
00:12:07
now that models like GPT4.1 are trained
00:12:09
on git diffs, so they can do diffing
00:12:11
syntax directly instead of having to
00:12:12
rewrite the whole file. Previously,
00:12:14
Copil was using like a custom model
00:12:16
derived from GPT3 that had a ton of
00:12:19
system prompts to make it function at
00:12:20
all. I'm sure that was essential to why
00:12:22
they decided to keep it closed source at
00:12:24
the time. That barely matters anymore.
00:12:26
There are lots of companies with better
00:12:27
AI editing experiences than where
00:12:29
co-pilot was. The next point they had is
00:12:31
that the most popular and effective UX
00:12:33
treatments for AI interactions are now
00:12:35
common across editors. Yes, it took a
00:12:38
bit for us to get to that point, but all
00:12:39
the things we now expect in our AI
00:12:41
experience like command I to open the
00:12:43
sidebar, command K to autocomplete from
00:12:46
here, tab to blast through the changes,
00:12:49
all of those things are relatively
00:12:50
standard and you can hop from windsurf
00:12:52
to cursor to copilot and not feel like
00:12:54
you're entirely in a new world. We want
00:12:56
to enable the community to refine and
00:12:57
build these common UI elements by making
00:12:59
them available in a stable and open
00:13:01
codebase. Huge. An ecosystem of open
00:13:04
source AI tools and VS code extensions
00:13:06
has emerged. We want to make it easier
00:13:08
for these extension authors to build,
00:13:10
debug, and test their extensions. This
00:13:11
is especially challenging today without
00:13:13
access to the source code in the C-pilot
00:13:15
chat extension. This is another point I
00:13:17
think is really worth considering. When
00:13:19
Microsoft looks at two types of
00:13:22
companies, if we have Windurf and Cursor
00:13:24
on one side and we have Klein and
00:13:28
Augment on the other, this is who
00:13:30
Microsoft wants to have win. But right
00:13:32
now, these guys are very much winning.
00:13:35
There is a reason for that. It's because
00:13:37
these guys are doing things Microsoft
00:13:38
doesn't like that they can win. Since
00:13:41
they chose to fork, since they chose to
00:13:43
do the hard thing and rebuild VS Code
00:13:46
and manage the fork that you have done
00:13:48
from it, they now get a benefit that
00:13:51
client and augment don't. They can make
00:13:53
the changes and make the quality of
00:13:54
experience that Copilot has when these
00:13:57
companies couldn't. So if Microsoft sees
00:13:59
this as an imbalance where they want
00:14:00
these guys to win and they want these
00:14:02
guys to fail, this is the most logical
00:14:05
thing they could possibly do. This is
00:14:08
particularly funny to me because in my
00:14:09
video on Windinsurf, my like final take
00:14:11
was that the best thing OpenAI could do
00:14:13
would be open source it because there
00:14:15
wasn't a big open-source player in the
00:14:17
AI editor space yet. Now there is. Now
00:14:21
the biggest open- source editor is also
00:14:23
the biggest open- source AI editor.
00:14:25
Well, at least it's getting there. I
00:14:28
think this is the most important point
00:14:29
to take home from here. Not that
00:14:31
Microsoft necessarily wants to kill
00:14:32
these companies and destroy them. More
00:14:34
that they want companies doing the the
00:14:37
right thing building into the VS Code
00:14:38
ecosystem. They don't want them to be at
00:14:40
a disadvantage. They want to make it
00:14:42
easier for more companies like this to
00:14:43
find more success and build better
00:14:45
experiences within VS Code. But in order
00:14:47
to do that, they have to open up more,
00:14:49
which is why they chose to do
00:14:51
it. There's a couple more quick points
00:14:53
that I thought were interesting. They
00:14:55
want to share more of how the chat
00:14:57
extension actually collects data and
00:14:59
where the data is being sent to and from
00:15:02
to give you a better idea with just more
00:15:04
transparency if you could read the code.
00:15:05
That's a cool thing to see. And also
00:15:07
malicious actors who have been targeting
00:15:08
these AI dev tools. If it's open source,
00:15:10
it's easier for us to scan through to
00:15:12
find problems and also identify fixes
00:15:14
and go through the whole process of how
00:15:16
exploitations happen. Really cool to
00:15:18
see. Coming weeks, we'll work to open
00:15:21
source the code in the GitHub copilot
00:15:22
chat extension, as well as refactor the
00:15:24
AI features from the extension into VS
00:15:27
Code directly. Our core priorities
00:15:29
remain intact. Deliver great
00:15:30
performance, powerful accessibility, and
00:15:32
an intuitive, beautiful user interface.
00:15:34
Open source works best when communities
00:15:36
are built around a stable and shared
00:15:38
foundation. That's the key. Since Cursor
00:15:41
and Windsurf aren't open source, they're
00:15:44
all a increasingly brokenoff fork of VS
00:15:47
Code. The community is no longer
00:15:49
building around this single shared
00:15:51
center point and they want that to
00:15:53
change. They want the same VS code
00:15:55
everywhere. And the selfish reasons why
00:15:57
are actually a lot smaller than you
00:15:59
might think. It's stuff like the C++
00:16:01
extension that they maintain breaking
00:16:03
because cursor does some specific thing
00:16:05
that was patched in VS Code a while ago.
00:16:08
They never backfilled. That type of
00:16:09
stuff is just annoying when the
00:16:11
foundation isn't shared. There's a huge
00:16:13
part of why Linux did so well. It's also
00:16:15
a huge part of why I have a grudge
00:16:16
against Android because they hard forked
00:16:17
the Linux kernel. The more we can share
00:16:19
that foundation, the better we can be as
00:16:22
a community in iterating and building on
00:16:24
top of a thing. And then the stated
00:16:26
goal, as I've been saying, their goal is
00:16:27
to make contributing AI features as
00:16:29
simple as contributing to any part of VS
00:16:31
Code. The stcastic nature of large
00:16:33
language models makes it especially
00:16:35
challenging to test AI features and
00:16:36
prompt changes. To ease this, we'll also
00:16:39
make our prompt testing infra open
00:16:41
source to ensure the community PRs can
00:16:43
build and pass tests, too. That's really
00:16:45
cool. They shared their whole iteration
00:16:48
plan publicly. So, if you're the type of
00:16:49
person that wants to keep up with the
00:16:50
details of how this all being
00:16:51
implemented, it's all there if you want
00:16:53
to do it, which is really, really cool.
00:16:55
This complements the agent mode stuff
00:16:56
they also officially released today
00:16:59
really well. The idea of a fully agentic
00:17:02
VS Code experience being open source is
00:17:05
super exciting. WSL going open source is
00:17:08
just again showing their commitment to
00:17:09
open source and building on a shared
00:17:11
foundation. And then there was one other
00:17:12
thing I didn't mention. I probably
00:17:14
should have put it in the intro. Edit.
00:17:16
Yes, Microsoft released a Vim competitor
00:17:19
today that is open- source, which is
00:17:21
kind of crazy that Microsoft found that
00:17:23
it was worthwhile to build their own CLI
00:17:26
editing
00:17:27
experience. I also love that it pays
00:17:30
homage to the MS DOS editor. This looks
00:17:33
really cool. It's something I plan to
00:17:34
play with later. Let me know what you
00:17:36
guys think about it and if you want a
00:17:37
whole dedicated video. That's all I got
00:17:39
for now. Wait, it's in
00:17:41
Rust. This is wild. Oh,
00:17:44
Microsoft. There's always something to
00:17:46
talk about, isn't there? Well, thanks
00:17:48
for joining me on my day off. Hope you
00:17:49
guys enjoyed this. Let me know what you
00:17:51
think. And until next time, peace nerds.
