embracing ai at work

This summer, I had the opportunity to intern at SoFi under the Digital Identity - Identity Core team. It’s my first ‘big tech’ internship experience, and I was constantly inspired and motivated by everyone I worked with. The fast embrace of AI at SoFi was one of my biggest draws to the team—I’m always looking to increase my productivity and learning, and AI makes it possible for me to accomplish so many things I wouldn’t have the technical experience to do on my own. I loved working with a team that shares the new tool they find each day, to be in an environment where people earnestly test new AI features. Instead of reading TLDR AI every day, I could bond with my coworkers over Opus 5 being too verbose, or a new way to organize cmux tabs.

Now on my 10th week at SoFi, I think it’s time to reflect on my relationship with AI since I’ve been here.

Notable insights/lessons learned

1. Increasingly automated processes require increased training for 1) an engineer who can diagnose and resolve errors when critical incidents occur, and 2) an operator/manager who can correct the error and improve prevention measures. Automation brings about an unfortunate irony.The more automated a system becomes, the rarer it is for human involvement to be required. Presumably, the incident in which a human is required is also more critical. I forsee one of the biggest challenges at tech companies in the next few years will be to find the balance between humans/automation when preserving quality.
2. Minimizing the drawbacks of constant context-switching. I’ve run probably 8 terminals at once before, and in a typical day I’ll have 3-4 running at the same time. It’s incredibly draining. For a while, I was confused why I was tired after a few hours of simply prompting AI terminals, but and I felt like my brain was deteorating. Context switching for humans causes declined performance on single-focus tasks, decreases accuracy, and, crucially, most of us underestimate how much multitasking affects us. Unfortunately, it seems to be part of the evolving job so I’m working on mitigating the damage with a centralized dashboard, automated notifications, and maintaining living documentations for each of my AI agents in parallel.
3. Task delegation and management. Wow is it important to delegate AI specific tasks. This summer, I created my first Technical Design Document (we called them TDDs) and Execution Plan, and it was a step I didn’t even know was in the implementation process. We would first include the background for the new design, the purpose, the requirements, the HLD, the LLD, and so on and so forth. My TDD rounded out to 36 pages, which was vaguely AI organized but mostly hand written. AI isn’t great at the high level stuff! It has insane amounts of depth, but in unique breadth and overall system design, it definitely can’t replace engineers yet. This calmed my fears that AI would replace all our jobs in the next couple of years—the space has definitely shifted, but the number of humans needed isn’t decreasing (at least for IC3 up) for any company that’s growing from what I’ve observed. It’s less important now to know all the details of implementation, but engineers instead need experience thinking about edge cases, resiliency, and architectural design.

Bonus: My favorite AI features/tools I’ve discovered this summer!

1. bypass permissions. Most game-changing feature I’ve discovered here. Freed me from clicking “Yes, automatically allow all ___ commands” repeatedly in auto mode. However, I’ve heard auto mode has become much less restrictive in the last month, so I’m looking to play around with that again.
2. cmux. Some days I barely opened my IDE.
3. CLIs > MCPs. Way lower token usage and wayyy faster output.
4. Don’t go over 200k context, unless it’s one task. Cognitive decline of AI models is crazy
5. Open source skills/plugins. It’s been pretty fun playing around with my setup—lately I’ve been enjoying OpenClaw, ccstatusline, and ADHD. SoFi also has a marketplace of (impressively) helpful internal tools for things testing/knowledge management/code review.
6. Having different models evaluate each other’s work. Running code review on Sonnet, GPT Sol, and occasionally Opus for scanning security vulnerabilities and for best practices has yielded different insights for me.
7. git worktrees. It’s super nice to work on different features in the same repository at the same time, like delegating two engineers that need to push changes at the same time from my hackathon experiences.