Learning Library

← Back to Library

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
0:00you know code repair is a long way 0:02behind code authorship from an AI 0:05perspective and I think it's a missed 0:07opportunity at the end of the day the 0:10only thing that really matters is 0:12working code that you can use to get 0:14your job done and right now so much of 0:17the work that I've seen has gone into 0:20crafting coding 0:23experiences that enable someone who is 0:25new to coding to get started and get 0:27into coding fast 0:30so you have things like the repet AI 0:33agent that dropped where it's like hey 0:34let me get started let's do a multi-step 0:36plan and then you can kind of like go 0:38from there and it's often the things 0:40that work are a simple experience right 0:42like it's like let me go through this 0:43multi-step plan and knock up a quick 0:46HTML website for 0:48you but there's been a lot less work put 0:51into the complexities of how you edit 0:54and adjust code reliably with AI and how 0:58AI can fix its mistakes I was using 1:00repet this weekend and it ended up just 1:03sort of giving up and saying please go 1:05talk to a real person at repet because I 1:07can't fix this and it was a mistake that 1:09it had made itself it had tried six 1:11different times to fix it and it had not 1:13thought through systematically despite 1:16my hints what was going on I used the 1:19yellow duck method I like talk to myself 1:22and kind of talk through what the coding 1:23issue was I suggested this is the this 1:26is the actual thing that's happening and 1:28I suggested a plan forward and just 1:30couldn't 1:31engage 1:33and I there are 1:35absolutely opportunities there are 1:37dollars being left on the table here for 1:39an AI that is able to be more 1:41self-critical that is able to actually 1:43think about what has been written 1:45previously and not over index on it and 1:47there's actually really good reasons 1:49from a llm perspective why it's hard for 1:53AIS to correct their mistakes within a 1:55chat but there are also really good 1:57fixes that are on the table if you think 1:58about the interface so the reason why 2:02they fail to correct their mistakes 2:05reliably is because llm chat is context 2:08dependent and so when they make the 2:10initial mistake in the chat they are 2:12predisposed rereading the chat through 2:14the context window to make the same 2:16mistake again and that's why you just 2:17get into this Loop where it says if 2:19you've ever been coding with AI I'm 2:21adding debugging so that this will we'll 2:24figure this out well that's a death Nell 2:26because at the end of the day it adds 2:28the debugging or so it says and it 2:30doesn't actually debug it doesn't do The 2:32Logical thinking process that goes with 2:35debuging and when you think about it 2:39that's the part that matters it's not 2:41the writing the the AI debuging uh log 2:44reads right like you you can get an 2:46audit log of everything that happened in 2:48the code and if you don't do critical 2:51thinking about it it's not going to help 2:53you very much is it so the context 2:55window is super important for 2:58understanding why llms tend to get into 3:00these Loops these spirals but you can 3:02also think as a company about how to not 3:06get hampered by that architecture 3:09limitation for example you could say it 3:13seems like there's multiple responses in 3:15the same vein here maybe I should spin 3:17up a new chat and on the back end give 3:19it a CIS prompt that basically 3:21summarizes the issue and see if the new 3:23chat can help unlock the user here like 3:25that's a really simple one that I've 3:27seen nobody do it would not take any new 3:29technology it's just a slightly 3:31different experience and it would help 3:33us get out of these sort of endless 3:35Loops of debugging and code that don't 3:38go 3:39anywhere and so I guess my challenge is 3:42if you want to make workflows more 3:44useful maybe focus on the workflow 3:47that's already happening with code 3:49authorship and not just on the net new 3:51workflow you want to start look there's 3:53a ton of value that's clearly being 3:55unlocked as cursor and as repet AI come 3:58online and other tools come online help 4:00new people learn to code I love that I 4:02think it's going to be impactful I think 4:04it makes sense from a business 4:05perspective 4:07but it also makes sense to help people 4:12who work in code to edit more 4:14effectively and I know Sonet 3.5 is 4:17working on that with uh AI for 4:20Enterprise and anthropic and that's 4:22great it feels like we have a ways to go 4:25with strong critical thinking edits on 4:29code and I would love to see more 4:31investment in that direction