Transcript — State of Play: AI Coding Assistants

Bagita Bokela is a is a distinguished
engineer at Thor works and um I've been
chatting to Bigita for a little while
and actually was really impressed by a
very a post that went viral on Martin
Fowler's uh site that wrote uh talking
about specs and talking about uh the the
comparison at the time between SpecKit
uh Tessle and uh Kira I think it was and
uh okay Tesla's moved on a little bit
since them. But it's still amazing to
see how many people go to go to that uh
site. Uh Bita is an amazing person. I
very much encourage you to to follow
her. She's she's very thoughtful in with
so many posts and blogs that she writes
and a really wonderful way to finish off
uh this conference with a very visionary
uh session whereby we is going to look
at the last 12 months where what's been
changing where we are today and what we
can look forward to. So please give a
very AI native dev warm welcome to
bigita bller.

>> Okay. Thanks Simon. Thanks Simon and
Patrick for inviting me. Yeah. So I'm a
distinguished engineer at Thoughtworks.
And what that means for me specifically
is that uh three years ago I got a
full-time role to just be immersed in
this space of AI coding or in general
using AI on software teams to help my
colleagues to help our clients. I kind
of stay on top of it. So I talk a lot to
uh our teams, our clients and so on. And
then I write about it for example on my
colleague Martin Fer's website. Um and
so that's kind of like what what all of
this is um based on. And yeah, it's kind
of like a tough task of wrapping up
after everybody heard about this topic
for two days. Uh, so I'm going to try
and help you see the forest for the
trees or the multiple forests for all of
the the trees. So I'll kind of do like a
recap. I'll start with the recap slide.
Simon was briefly confused and thought
maybe my slide setup is wrong, but I
it's kind of like recap a lot of what
the stuff that you heard also over the
last two days, but also kind of what
happened in the last 12 months like
advancements as well as things that are
maybe not going so well or that are kind
of like the all the second order
consequences and implications that we're
experiencing right now. So that when you
get back to work tomorrow and your
colleague who maybe isn't as immersed in
the space ask you, so what should I
know? Right? I hope I can help you
answer that question

and I'll start with
uh yeah the reason why all of this is
happening which is uh the models. So
kind of first and most obvious um there
wasn't even that much talk about models
here at the conference I think um which
is not surprising and totally okay I
think because for me this is not really
the most exciting part to be honest I'm
much more interested in everything
that's now happening around it the
ecosystem all of the integrations and so
on and uh I mean obviously if we talk
about the last 12 months there was the
Opus 4.5 moment kind of last year um
that uh made a lot of people kind of
like come back to this that hadn't maybe
tried AI AI coding for a while. Um,
and uh, yeah, so that was maybe the
biggest event in models that happened
like that. I mean, almost every week
there's a new model, but I usually don't
even follow it that much because like I
said, it's uh, kind of interesting to
see all the stuff around it.

So, if uh,
we think about models like what are the
core things kind of as users who use
them for coding that we need to know or
learn almost like a kind of like
learning map, right? So the first thing
I always try to get out of the way is
that they are not magic right they are
very very impressive and very very
useful math but uh unfortunately even
like a lot of technologists a lot of our
peers kind of like it's very easy to
fall into that trap right like of uh
thinking of them as as more than that
right so I kind of like this
visualization to remind ourselves that
you know even though we don't really
know what's like why it works and what's
happening it's still like very
impressive math right so first of all
they're not magic

Um I mention here as the second point
their statelessness because that's also
something that I often notice that
people haven't quite grasped right so
the the model doesn't have a session
right so the longer our conversation
with them gets the longer our session
with them gets every single time the our
our agent our harness basically sends
the whole history of the conversation
right maybe not quite there's caching
and all kinds of clever ways that uh
different tools try to optimize that but
they are stateless Right. So that is a a
factor that happens like the longer our
conversations get with them. So I think
that's a really important core thing
that um uh yeah people need to
understand.

Uh the third thing we need to know about
is of course the size of the context
window and um in relationship but also
in relationship to what that means for
attention. Right? So even though
technically the context windows have
gotten a lot bigger, um it comes with a
trade-off on like how well the models
are able to keep attention on all of the
many instructions and all of the context
that we're trying to feed them now. So
there's something there to be understood
by everybody who uses this about what
that trade-off is.

Um and then um
finally there's uh this and that's maybe
the the biggest area. So those first
things are kind of like you could you
can learn them in a formal training and
kind of like understand the basics,
right? But this last one is a lot more
about using the models and figuring this
out, right? Which model do we use for
which task, right? So there's I mean
these are just like a few illustrative
examples like uh we we have autocomplete
like there's still people who use that a
lot. Uh let's say or let's say you want
to just change a few specific files and
you have very clear instructions and a
very clear idea of what you want to do