Throw your triage lists at GPT 5.5 and watch them disappear
A developer successfully used GPT 5.5 to solve a complex data migration problem involving millions of rows with unstructured data and edge cases. Previous attempts with other AI tools including Cloud Code and GPT 5.4 had failed to resolve the issue.
Summary
The speaker describes tackling a challenging data migration project involving millions of rows of functionally unstructured, lightly structured data with numerous edge cases. This represented significant complexity for their organization despite the scale not seeming large to others. The developer had previously attempted to solve this problem using various AI coding tools, including Cloud Code (even with Opus) and GPT 5.4, but none were successful. However, after dedicating 6 hours to working with GPT 5.5, they achieved a breakthrough. The error rate in their Sentry monitoring dropped dramatically, indicating the migration problem was effectively resolved. The speaker notes that this type of complex problem was something they had actively avoided in the past because existing AI tools lacked the intelligence to handle it autonomously. The success with GPT 5.5 and what appears to be a coding assistant called 'codeex' has been transformative for their workflow. However, they express concern about the potential costs of using such advanced AI capabilities in production environments due to token pricing.
Key Insights
- The speaker had a data migration problem with millions of rows of functionally unstructured, lightly structured data with tons of edge cases that represented significant complexity
- Cloud Code could not figure out the migration problem even when used with Opus
- GPT 5.4 was unable to solve the data migration problem that the speaker was working on
- After 6 hours of using GPT 5.5, the error rate hit the floor in their Sentry monitoring, indicating the migration problem was resolved
- The speaker states they had truly avoided this kind of complex problem because the AI intelligence was not there to do it autonomously until now
Topics
Transcript
[0:00] So this is an example of a data migration problem with millions of rows which might not sound big to many people but is pretty significant to us in terms of the complexity of the data inside of it with functionally unstructured lightly structured data with tons of edge cases. And I just finally was like GPT 5.5 take me away. I threw cloud code at this. Cloud code could not figure it out even with Opus. I threw GPT 5.4 at it. It could not figure it out. And guess what? 6 hours of GPT 5.5, we saw [0:31] our error rate just hit the floor in our century monitoring. And so I think quality is going to…
Full transcript available for MurmurCast members
Sign Up to AccessMore from How I AI
Cybersecurity Is So Bad I Texted My Parents
The speaker discusses 'Black October,' a period marked by numerous critical cybersecurity vulnerabilities and attacks affecting major internet infrastructure, including PHP, GitLab, SSL issuers, and iOS devices. While acknowledging that AI will eventually improve security, the speaker warns that cybersecurity will worsen significantly over the next 18 months before improving.
What are channels in Claude Duet?
Claude Duet features a channels system inspired by Slack that allows users to organize conversations into separate subthreads within a main chat. This enables more structured collaboration by creating dedicated channels for specific projects, such as a sales channel for sales-related work.
Claude Duet needs multiplayer (please!)
The speaker argues that Claude and similar AI products are currently optimized for single-player use and lack proper multiplayer capabilities. Despite integrations like Claude tags and Slack, the speaker contends that sharing agents with their small team of five people remains problematic, preventing these tools from becoming part of their regular workflow.
Claude Duet is great, but SLOW
The speaker expresses frustration with the slowness of AI model Claude, which impacts their willingness to engage with it. They compare its responsiveness unfavorably against more responsive alternatives, despite acknowledging Claude's capabilities.
Claude Duet should be more annoying
The speaker emphasizes the need for agents like Claude Duet to be more engaging and proactive to avoid being forgotten in the competitive landscape of digital assistants. They critique the current methods of communication, such as morning briefings, as ineffective in attracting user interest.