AI Code Repair Still Lagging
Key Points
- Code repair lags far behind code generation in AI tools, leaving a missed opportunity to deliver reliably working code that users actually need.
- Current AI coding experiences focus on getting beginners started quickly (e.g., multi‑step plan agents) while offering little robust support for editing, adjusting, and fixing code errors.
- LLMs struggle to self‑correct because they reread the same flawed context window, causing loops where the model repeats earlier mistakes instead of performing genuine logical debugging.
- There’s a market and technical opportunity to build AI systems that are more self‑critical, can reason about prior code, and overcome context‑window limitations with better interfaces and debugging workflows.
Full Transcript
# AI Code Repair Still Lagging **Source:** [https://www.youtube.com/watch?v=E3Mry80TEgs](https://www.youtube.com/watch?v=E3Mry80TEgs) **Duration:** 00:04:34 ## Summary - Code repair lags far behind code generation in AI tools, leaving a missed opportunity to deliver reliably working code that users actually need. - Current AI coding experiences focus on getting beginners started quickly (e.g., multi‑step plan agents) while offering little robust support for editing, adjusting, and fixing code errors. - LLMs struggle to self‑correct because they reread the same flawed context window, causing loops where the model repeats earlier mistakes instead of performing genuine logical debugging. - There’s a market and technical opportunity to build AI systems that are more self‑critical, can reason about prior code, and overcome context‑window limitations with better interfaces and debugging workflows. ## Sections - [00:00:00](https://www.youtube.com/watch?v=E3Mry80TEgs&t=0s) **AI Code Repair Lagging Behind Generation** - The speaker argues that while AI tools quickly help novices write code, they still fall short at reliably editing and fixing code, as illustrated by a recent failure of an AI coding assistant, highlighting a missed commercial opportunity. ## Full Transcript
you know code repair is a long way
behind code authorship from an AI
perspective and I think it's a missed
opportunity at the end of the day the
only thing that really matters is
working code that you can use to get
your job done and right now so much of
the work that I've seen has gone into
crafting coding
experiences that enable someone who is
new to coding to get started and get
into coding fast
so you have things like the repet AI
agent that dropped where it's like hey
let me get started let's do a multi-step
plan and then you can kind of like go
from there and it's often the things
that work are a simple experience right
like it's like let me go through this
multi-step plan and knock up a quick
HTML website for
you but there's been a lot less work put
into the complexities of how you edit
and adjust code reliably with AI and how
AI can fix its mistakes I was using
repet this weekend and it ended up just
sort of giving up and saying please go
talk to a real person at repet because I
can't fix this and it was a mistake that
it had made itself it had tried six
different times to fix it and it had not
thought through systematically despite
my hints what was going on I used the
yellow duck method I like talk to myself
and kind of talk through what the coding
issue was I suggested this is the this
is the actual thing that's happening and
I suggested a plan forward and just
couldn't
engage
and I there are
absolutely opportunities there are
dollars being left on the table here for
an AI that is able to be more
self-critical that is able to actually
think about what has been written
previously and not over index on it and
there's actually really good reasons
from a llm perspective why it's hard for
AIS to correct their mistakes within a
chat but there are also really good
fixes that are on the table if you think
about the interface so the reason why
they fail to correct their mistakes
reliably is because llm chat is context
dependent and so when they make the
initial mistake in the chat they are
predisposed rereading the chat through
the context window to make the same
mistake again and that's why you just
get into this Loop where it says if
you've ever been coding with AI I'm
adding debugging so that this will we'll
figure this out well that's a death Nell
because at the end of the day it adds
the debugging or so it says and it
doesn't actually debug it doesn't do The
Logical thinking process that goes with
debuging and when you think about it
that's the part that matters it's not
the writing the the AI debuging uh log
reads right like you you can get an
audit log of everything that happened in
the code and if you don't do critical
thinking about it it's not going to help
you very much is it so the context
window is super important for
understanding why llms tend to get into
these Loops these spirals but you can
also think as a company about how to not
get hampered by that architecture
limitation for example you could say it
seems like there's multiple responses in
the same vein here maybe I should spin
up a new chat and on the back end give
it a CIS prompt that basically
summarizes the issue and see if the new
chat can help unlock the user here like
that's a really simple one that I've
seen nobody do it would not take any new
technology it's just a slightly
different experience and it would help
us get out of these sort of endless
Loops of debugging and code that don't
go
anywhere and so I guess my challenge is
if you want to make workflows more
useful maybe focus on the workflow
that's already happening with code
authorship and not just on the net new
workflow you want to start look there's
a ton of value that's clearly being
unlocked as cursor and as repet AI come
online and other tools come online help
new people learn to code I love that I
think it's going to be impactful I think
it makes sense from a business
perspective
but it also makes sense to help people
who work in code to edit more
effectively and I know Sonet 3.5 is
working on that with uh AI for
Enterprise and anthropic and that's
great it feels like we have a ways to go
with strong critical thinking edits on
code and I would love to see more
investment in that direction