For 15 years I've built engineering teams off roughly the same playbook. This spring, our team doubled its output with no extra headcount — and I realized the playbook itself had to change. So I got hands-on. Between meetings (I do still have a day job as CEO) I shipped 100 PRs to production, to feel where the old way breaks. The short version: the engineer's job is shifting from delivering a product to delivering the system that delivers the product. I wrote up what we changed: how to move from local to global optimizations, what happens to quality when agents write almost all the code, and whether different types of work now need different processes. swarmia.com/blog/the-new-way…

Here's a typical flow of signing into Deel: 1) Login with Google on desktop 2) Google wants my physical 2FA device 3) Deel wants another 2FA from mobile app 4) Mobile app updates and loses login info 5) Login with Google, enter password 6) Google wants my physical 2FA device on mobile 7) Use Authenticator to log in on mobile 7) The auth notification has disappeared 8) Send it again and tap it on mobile, still doesn't work 9) Use Authenticator to log in on desktop 10) "Your login session has timed out, try again" 11) Go to step 1

If your bottleneck is not writing code, it will be something else. Reviewing code, releasing changes, training your own customer-facing teams, or ultimately your customers’ tolerance for change.

Measuring AI coding tool productivity hits the same hurdles we've faced in developer productivity for a decade. Well-meaning and easy-to-measure metrics turn out to be misleading. When someone says that AI writes 90% of the code, it may sound like you can now reduce your engineering headcount by that much. But that's not what it means. In reality it doesn't matter who wrote the code. Someone had to understand the context, drive the process, and check the work. AI likely saved you some time from writing the code, but the AI didn't do it by itself. Focusing too much on AI-generated code might also cause some cultural issues. "I didn't write this code, it was AI." You should be clear that the developer and the team still owns all code they ship, no matter how the code was written.

We've all seen reports that Claude 3.7 is... eager to generate code. I just asked it to initialize a new repository with a specific name. It used the repository name to figure out what the app was supposed to do and then implemented it.

MacBook Air M1 has to be the best computer I ever owned. It lasted five years as my main machine and originally replaced two Intel iMacs that I had at home and at the office, and an Intel MacBook Pro that I used for travel. I hope the new M4 Air is worth the upgrade finally.

We're working with some Fortune 500 companies and companies from more regulated industries. One of the common questions is: can you continuously deploy software while maintaining quality and security? The answer is clearly yes: - We've never had a security incident. Even external security audits haven't found anything major. - We have tens of thousands of automated test cases covering our product. - Our uptime has always been at least 99.97% annually. Last year we were at 99.992%. - We typically deploy 30 changes to production on a given weekday. Our desire to move fast is not a reason to ignore reliability. Quite the opposite: it's the great infrastructure and foundation that allows us to move quickly.

Join Swarmia Circle on March 27 to hear seasoned engineering and product leaders, including @Geek_Manager, @johncutlefish, and Nathen Harvey, discuss the challenges and opportunities of leading and scaling engineering organizations. Sign up today: swarmia.com/circle/

The best part of joining the Forbes Technology Council is that all of the AI SDRs just pick the latest company from my Linkedin profile. Very easy to filter out spam.

1/ Can Large Language Models (LLMs) truly reason? Or are they just sophisticated pattern matchers? In our latest preprint, we explore this key question through a large-scale study of both open-source like Llama, Phi, Gemma, and Mistral and leading closed models, including the recent OpenAI GPT-4o and o1-series. arxiv.org/pdf/2410.05229 Work done with @i_mirzadeh, @KeivanAlizadeh2, Hooman Shahrokhi, Samy Bengio, @OncelTuzel. #LLM #Reasoning #Mathematics #AGI #Research #Apple

Before your company rebrands as .ai, remember it's also a British Overseas Territory. every.to/p/the-disappearance…

I should have gotten a burner phone for the NYC apartment search. Some of these management companies won't stop contacting me, even two months after I told them I found another place.

"When is it ready?" Well, it depends... You can now forecast cross-team initiatives in @SwarmiaHQ. swarmia.com/changelog/2024-0…

The AI meeting notes in @Gong_io were suboptimal when they first launched. Guess what? Now they're actually on par with human-written notes.