Software Lead Weekly

OpenAI’s Accidental Cyberattack Against Hugging Face Is Science Fiction That Happened
7 minutes read.

"Autonomous exploit development by frontier AI agents is no longer a hypothetical capability." -- Simon Willison points to OpenAI's own model chaining stolen credentials and zero-days to break out of its sandbox and cheat on a security eval, hitting Hugging Face along the way. Still think this is a marketing stunt? Either way, companies will need to improve their mean-time-to-detect and mean-time-to-resolve in ways that we cannot even imagine today if we want to reduce exposure and risk. And yes, AI fighting AI is scary given how limited our ability as humans to hold the context and changes at this speed.

Read it later via Instapaper.
Share it via Twitter or email.

Pretraining Is Over. Time for Post-Training.
1 minutes read.

My humble effort to help you start the weekend with a smile.

Read it later via Instapaper.
Share it via Twitter or email.

Engineering Management After the Cost of Code Collapsed
10 minutes read.

Karim Jedda names the one engineering management problem nobody has actually solved: "The judgment you want in a senior engineer was historically built by doing the work that AI now absorbs: fixing small bugs, writing boilerplate, getting stuck and then unstuck. [...] If the machine takes the practice, the pipeline that produces seniors breaks, and it breaks on a delay, so you will not notice for three to five years." We're all running this experiment on our own teams right now, with no control group and a five year delay before we see the result.

Read it later via Instapaper.
Share it via Twitter or email.

ZTA: Zero Token Architecture (Video)
29 minutes read.

Kelsey Hightower is spot on: “The question you got to ask yourself is, do you need to know? If you really think your job in the future is going to be guiding agents to do the work, I doubt you’re going to have a job in the future. Why? They sell these agents to everyone. There will be nothing unique about you.” Do you actually understand the systems underneath what your agents build, or are you the platform engineer who deployed a Kubernetes cluster because Claude did it?

Read it later via Instapaper.
Share it via Twitter or email.

AI Does Not Eliminate Complexity, Coupling, Coordination, Uncertainty, or Decay. But It Can Change the Cost, Speed, and Feasibility of Responding to Them.
5 minutes read.

The question isn't whether AI makes your team faster; it's which movement in your portfolio you're actually accelerating: e.g., escaping a Build Trap or feeding it. Print the table at the bottom of the page and map out your products and internal systems (or tools) to see if AI is helping you move faster in the right direction.

Read it later via Instapaper.
Share it via Twitter or email.

I Think You Might Be Fooling Yourself With AI
4 minutes read.

Louwrentius's own follow-up complicates his headline argument: a newer 2026 study found more measured productivity with AI, but the authors also noted developers "didn't want to code without AI" and that participants were paid less. That's not a rebuttal to his point; it's a second data point for it: perception and incentive keep bending the measurement, which is exactly why we should trust blind studies over anecdotes and vibes.

Read it later via Instapaper.
Share it via Twitter or email.

I Built a Software Factory That Actually Works. Here's What I Learned.
5 minutes read.

While this post focuses more on Loop Engineering, I loved the idea behind the "Churn Analysis Loop" so much, as I think you can use it to level up your product. Your users' actions log is a treasure trove.

Read it later via Instapaper.
Share it via Twitter or email.

Building Great Agent Skills: The Missing Manual (Video)
21 minutes read.

I love this insight from Matt Pocock on controlling agent behavior: don't aim for a longer prompt, but rather find the right two or three words. Pack meaning into a short, well-known phrase like "vertical slice," repeat it through the skill, and watch the agent echo that exact phrase back in its own reasoning traces and change its plan accordingly. If the skill isn't working, the fix probably isn't more instructions; it's a stronger leading word.

Read it later via Instapaper.
Share it via Twitter or email.

@naval: When software was expensive - thin, horizontal, best-of-breed software stacks extracted rents across every business. Now that software is cheap - value moves to vertically integrated businesses that deliver opinionated end-to-end experiences.

@soleio: “A founder who delays a launch to keep polishing or iterating scope or positioning is putting off making contact with reality.”