weekend ai reads for 2026-03-13

OpenAI and Anthropic have been striking deals with consultants to help promote the use of their technology. In deals reached with McKinsey, Boston Consulting Group, Accenture and Capgemini, OpenAI engineering teams will work alongside those firms’ consultants.

When the first real employee identifier appeared: “WOW!”, the agent’s chain of thought showed. When the full scale became clear — tens of millions of messages, tens of thousands of users: “This is devastating.”

The early movers won't just capture market share, they'll define entirely new buying categories. When you're the vendor selling guaranteed results while your competitors are still selling seats and hoping for the best, that's not a fair fight. You win the deal every time.

The playbook: companies should start with the outsourced, intelligence-heavy task. Nail distribution. Expand toward the insourced, judgement-heavy work as the AI compounds. The outsourced task is the wedge. The insourced work is the long-term TAM.

LLMs remind us more than anything about how much work it was to actually ship a finished product. That stuff takes years, man, it does. And an LLM can do that for you if you do the devil’s bargain it’ll do that for you in a day. Because the deal is so damn good we change ourselves to make that deal acceptable.

Over time, it may accelerate our deskilling and make creative work feel more like playing a slot machine, prompting and regenerating until something sounds right.

The skill isn’t prompting. It’s knowing what needs to exist and why. It’s taste, the one thing that, no matter how good models get, they can’t learn.

Every generation of tools promises the same thing:

“We’ll make this easier.”

What they actually do?

Add another layer.

What’s still above the line? Compilers. State-of-the-art models. Operating systems. Software that encodes genuinely novel algorithms or requires years of accumulated domain data - cross-border taxation engines, real-time trading systems, climate modelling platforms. These remain assets because AI can’t trivially replicate the knowledge embedded in them. The difficulty isn’t in the code. It’s in the understanding the code encodes.

Tools I use to manage my schedule, communication, and complex projects with large teams across several countries. Examples and tutorials for productivity and sanity.

Take Your 10-20 Best Engineers. Physically Separate Them.

Put them in a different building. Not a different Slack channel. Not a different sprint. A different building.

Don’t let anyone talk to them. Don’t let the VP of Product “align” with them. Don’t let the head of CS send them feature requests. Don’t let the CEO pop in to “check on progress.” No one.

1. Workforce empowerment (ChatGPT)
2. AI-native distribution (verticals, apps, ads)
3. Expert capability (Co-scientist, Sora)
4. Systems and dependency management (Codex)
5. Process re-engineering (Agents)

Takeaways by Bloomberg AI

Companies with hiring ratios for AI experts below the sector average tended to underperform, with average equity returns of -19% and excess returns in bonds of -0.79%.
Companies with a hiring ratio for AI experts well above the sector average outperformed, with excess returns for credit more than 0.4x above the sector average being -0.11% this year.

The fact that these systems are trained on biased, Western-centric data is precisely why students must learn to question them. When we fed drafts into a model, it did not magically return A+ prose. More often, it amplified our weaknesses back at us: Vague claims remained vague, filler language spread and it often read as impersonal and corporate.

AI agents navigate iOS apps through screenshots by default — slow, expensive, and fragile. The accessibility tree already has everything they need. Here's the pattern, the toolchain, and why the work pays double.

A modern data context layer should essentially become a superset of what a semantic layer would traditionally cover. Sure, specific metric definitions can be hard-coded, but a modern context layer should include more to ensure agent autonomy – canonical entities, identity resolution, specific instructions to dissect tribal knowledge, proper governance guidance, and more.

"The first candidate who actually reads the whole document before responding."

240 countries. No search. No type-ahead. Just pure, unfiltered, alphabetical scrolling. Bonus points if the US is under "T." Extra bonus points if it's not even alphabetical. You absolute psychopaths.