Collection of ideas, thoughts and rambles that come out of my head every day
Lately I found myself naturally gravitating toward Claude Sonnet 5 with Medium reasoning. It just felt “better” somehow, but I had no objective measure to back up that gut feeling. So I did what any curious technologist would do: I ran some tests 😉
I find it fascinating that we are throwing massive LLM models at every problem these days. Claude Opus works incredibly well and we use it for just about everything. But that is not the future.
Remember that 10,000 hour rule from Outliers? The idea is simple: deliberate practice makes experts.
I think we are about to see a massive shift in how companies approach software, and SaaS vendors are going to feel the pain.
Teladoc and the pandemic made completely normal to discuss your health concerns over a video meeting.
An elegant solution for keeping OpenClaw from leaking credentials could be to put a proxy server between the bot and the outside world.
Until recently, startups had a genuine speed advantage. Big corporations were tangled in bureaucracy, committees, and endless approval chains. A small team could ship innovative products in weeks while enterprise companies took months (or years). But Large Language Models changed everything.
LLMs have fundamentally shifted the calculus on whether to buy or build internal tools.
LLMs are incredible for writing code faster. I have seen productivity gains that feel almost magical. Features that used to take days now come together in hours. That said, nobody is quite sure yet if that 10x speed boost comes with hidden costs. Does it also mean 1/10th the stability? After all, nobody actually wrote or deeply reviewed that code. Are we looking at 10x the bugs (and will AI patch…
I have been thinking about how LLMs are trained on code, and honestly, the training data is probably a mess. Most code out there is not great (let’s be real), and I doubt the requirements used to generate that code were stellar either. The training sets for programming-focused LLMs are almost certainly imperfect. Yet it is amazing that models like Codex or Claude work as well as they do.