Filipe Mendes

A genie offers to increase your average output at work by two or three times. The only condition is that you accept to gradually forget how you do that work. Would you accept it?

This past Saturday, I ran my first marathon and finished faster than my original goals after training for almost a year. Before, I was mostly lifting at the gym and/or playing football once a week. When I started training, I decided to go all out and handle the planning and coaching with LLMs alone, instead of using running apps or personal trainers. I started last year with gemini 3 pro, then moved to codex 5.3, then 5.5, 6.1 sol and more recently grok 4.6. I simply followed the initial plan, logged my workouts, and asked to make adjustments when needed. I also asked LLMs all sorts of questions. From equipment to nutrition, to possible pitfalls, etc. The biggest difference I noticed though, was when I changed from gemini to codex 5.3. It showed the gaps in my plan and where it needed to change. From then on, all other models seemed to agree that the plan was strong. What's cool is that even though I was noticing progress every week, and my numbers were getting better, more recently I was starting to get concerned about not having trained enough to maintain the pace that I needed for the whole race. Turns out I did, and the plan worked all along. I am now convinced that for most people, when provided with ongoing data, frontier LLMs are a good enough solution. Perhaps, even better experience than the more traditional alternatives... (Obviously, be careful with what you do and avoid injuries. In the end, this was an amateur race.)

I really like my setup with grokbots, to coordinate the work and the project, aided by Notion. It really seems like i'm working for a company, with no humans though. The downside is...I now feel that I work for 2 companies at the same time 🤣

My grok 4.6 PRs with reviews from codex and bugbot used to go down on a "bug found -> fix" cycle so often that was becoming unsustainable. I've then moved implementation to Opus 5.5 and cursor's bugbot reviews stopped finding issues. This 5.5 opus medium is a good model.

If you are planning to integrate a tool, library or framework into your codebase, always assume that what AI proposes is out of date.

If you use cursor and are upgrading to the new X subscription, make sure to unify from the cursor website to benefit from 6 months at 50%!

Abstractions are useful in every field.

I didn't believe SWE was ever dead or that AI could do my job better than me. But I do agree this is where we will stand. You start to believe that you can really do anything as long as you know enough to stir the agents. So learning is still needed!

Exactly..."the ironies of automation".

Surely Googlebooks are a great fit for @OmarchyLinux no?

A lazy person before AI, will continue to be a lazy person after AI.

Been using 4.7 High for a few prompts and the output just sounds like Opus 5.0. Too abrupt and direct, difficult to follow and understand.

"What you are willing to write is generally the upper limit of what others are willing to read." This is so true! I've lost track of the number of times people raise PRs with descriptions generated by LLMs, only for me to ask what it's about and why, and have them clarify it in two sentences.

I am sure I can spin off multiple tasks in parallel with dozens of agents and subagents. I have that many ideas and work to do too! I just don't have the capacity to reasonably go through and evaluate all the output that I would get from them.

Im running this experiment where I moved one of my personal projects to be entirely agent driven (developed?), running on cursor cloud agents and Grok, and using pstack. Verification of work by agents really seems to be the game changer but I need more days to evaluate this.

Im starting to hear engineers saying they would struggle to interview now because their skills are starting to get rusty (due to AI usage). I wonder if we will notice higher tenures across the industry because of this.

The challenge with meeting transcription notes is that it hides the participants emotions and body language. People might say X, and get transcribed as such, but think Y.

AI is like coffee.