Notes and news about AI continue to build in my pending folder. Too many to figure out. I’ll start with this one. I saw this news item from Morning Brew, one of my daily news feeds…The AI…it filled the code with bugs.
The amount of bugs popping up in AI-generated code is reaching the loose Sour Patch Kids under a camper’s bunk level. Amazon’s e-commerce senior VP, Dave Treadwell, called an all-hands for engineers at the company yesterday to address the growing frequency of outages, some of which can be traced back to code developed by generative AI, according to the Financial Times.
It continues…
- Last week, Amazon’s store malfunctioned for a few hours, which the company attributed to “a software code deployment.”
- And Amazon’s cloud services unit, AWS, had at least two large outages recently related to AI coding assistants. In December, the company’s cost calculator was down for 13 hours when Kiro, its AI coding tool, tried to change the code, and delete and remake the entire system.
- Though Amazon downplayed the meeting as routine in comments to the FT, the paper reported that Treadwell told employees that senior engineers will now need to sign off on AI-assisted changes made by junior and mid-level engineers.
Solutions?
An expensive solution. Anthropic rolled out a review tool yesterday in Claude Code to (hopefully) catch those vibe-coded mistakes—but with each pull request costing up to $25, it may get pricey fast.
Concurrently with this news, I received a PR request to interview Pramin Pradeep, CEO of BotGauge AI. I receive this sort of thing many times daily. Supposedly, Pradeep wanted to talk about “shadow code” left behind, supposedly maliciously, by AI generated code.
I asked for something in writing. They sent the usual PR thing that mentions shadow code but switches the topic to cybersecurity and then cites an irrelevant “case study”.
However, BotGaugeAI does participate in a market (Claude told me they were 20 out of 128 in that market for what it’s worth) called AI-assisted QA for code.
My research revealed that the basic problem comes from where the LLM AI code was trained. If trained too broadly, it will tend to being in superfluous code. Managers, meanwhile, are discovering that while maybe coders can save time in development using LLMs the task of checking and approving is becoming onerous.
If you’re using LLMs to help code, it would probably pay to check out the companies like BotGaugeAI for automated QA.
Click on the Follow button at the bottom of the page to subscribe to a weekly email update of posts. Click on the mail icon to subscribe to additional email thoughts.
