Relative to your interests, Friday

I was hoping it would do things like carry around amphorae to refill my wine between each sip and plates of succulent grapes, and then help me stumble home after the bacchanal.

Instead, it’s chainsaws.

Enterprise AI: ROI, Cost, and Reorgs

• How to make better enterprise AI ROI metrics and dashboards – Encode what you should next in AI ROI metrics, don’t leave the decision up to post-metrics vibes. “Make the decision before making the metrics. Translate the decision into language that people use before packaging it up into quantifiable metrics. For example, ‘Should we scale this AI feature from 10% to 100%?’ or ‘Should we fund a newer v2 model training next quarter?’ or ‘Should we replace manual workflow X with automation?’ Clear decision language is needed to form clear images of success. In other words, you have to be able to say your decision out loud before you can imagine what success will look like. This is your best protection from getting fooled by a nice Metric Theater with pretty dashboards.”
• more marketing context, more better AI for marketing – AI just makes good things gooder, and bad things badder: “The real work starts before evaluating tools or issuing an RFP. Marketers need defined personas, clear messaging frameworks, content standards, governance processes, reusable content, and alignment on measuring success. AI performs only as well as the context it’s given. Organizations that establish the richest context, the clearest governance, and the sharpest understanding of what makes their brand distinctive have an advantage, regardless of which solution they choose.”
• Reorganizing the dev teams at Visa to take advantage of AI – When you speed up the SDLC process, you should collapse silo’ed teams into one team to prevent human bottlenecks based on handoffs. Example of that at Visa.
• Tokenomics – how AT&T burns through a trillion+ tokens a month (and rising), but still manages to drive down enterprise costs through ‘moneyballing’ – Over a trillion tokens a month and still growing double digits, but the per-token cost keeps falling thanks to open source models and picking the right model for the job.
• 🤖 Target’s Real AI Moat: Agent Architecture, Autonomy, and Governance – The edge isn’t the models, it’s the architecture, governance, and operations built around them, starting with deciding whether a problem needs an agent at all.

How People Actually Use AI

• How people use AI in 2026: TV, homework, the DMV – Google looked at 15 million Gemini conversations and found people mostly asking about TV, homework, and government paperwork. Work is 13.5% of it, and almost none of that is trying to automate anything. Also: what the study leaves out, which is most of enterprise AI.
• AI moral panic – AI amplifies your current state, moral panic edition.
• 🤖 Google Quit the AI Race to Bet on World Models Over Recursive Self-Improvement – Google didn’t lose the race so much as leave it: DeepMind thinks coding agents building better coding agents is a dead end, and is betting on world models instead.

Platform Engineering

• Not my mistake, now my problem – Whatever your developers build above your infrastructure is a platform, and they’ll soon flit away from it, and now for you it’s: not my mistake, and now my problem.
• 🤖 You Already Have a Developer Platform, and It Is Someone Else’s Mistake – You already have platforms, you just didn’t choose them, and a decade of Cloud Foundry ops-to-developer ratios says one deliberate platform costs a few more people, not a new department.
• 🤖 Shift-Left Gave Developers the Work Without the Capability – Platform engineering worked so well that the platform is now the constraint: built for one persona, before agents and GPU scheduling. Push security and cost down into it as defaults.
• 🤖 The Platform’s Next User Is Not a Person, and VMs Are Back on the Golden Path – The internal developer platform is becoming an agent developer platform with dozens of machine consumers per human, and the VMs that platform engineering 1.0 shoved off the golden path are being deliberately brought back.
• 🤖 Kubernetes Adoption Is Settled; Kubernetes Complexity Is the Bottleneck – Kubernetes won the orchestration argument and promptly became the problem. The scarce skill is now hiding it rather than running it.
• 🤖 Cognitive Load Does Not Disappear at Scale, It Moves Into the Org Chart – Big companies “solve” developer cognitive load by hiring a layer of specialists to absorb it, which feels better for the developer and is worse for everyone else.
• mainframes still run money – COBOL is still behind 40% of online banking, 80% of in-person credit card transactions, and 95% of ATM transactions.

Code, Specs, and Security

• Using frontier models to find security problems in Spring Framework – a trickle turned into a tidal wave – Community vulnerability reports went from 55 in March to 112 in April, and internal frontier-model scanning added another 370 on top. The historical baseline was about seven.
• New spec maks MCP more enterprise-y, scalable, and otherwise performant – The big change is dropping the metadata handshake, which makes MCP apps easier to scale when traffic spikes.
• Why Java May Finally Get a Standard JSON API – Thirty years in, the JDK may finally ship JSON parsing in the box instead of leaving everyone to pick a third-party library.

The Headless Lifestyle

• UX is dead, what’s next? – Headless lifestyle: “I want something from a service now, I hope the service has a CLI so I can send an agent, and I never see the interface at all. If the service doesn’t have a CLI, and it’s something I want to use regularly, then maybe I can build my own.”
• MacPaw releases CleanMyMac CLI for cleaning developer and AI environments on macOS – The headless lifestyle, tidy your laptop edition.
• Which AI actually reads your site? Two months of LLM traffic, measured – A slight rise in AI SEO trick usage, but just stick with HTML for now. Also a prompt to tell the AIs how to do this. There’s effectively no llms.txt usage right now.

Writing, Prompts, and Style

• How I tell AI to write in my style, other people’s styles – and how to strip out shitty AI talk and LinkedIn language – An agent skill to remove AI writing tells, LinkedIn style, and write like Coté.
• A system prompt to get AI to stop pretending to be human – A drop-in system prompt that strips the anthropomorphic tics and gets the machine to describe itself as what it actually is.
• Prompt to rank paintings, but explain how – “How would you explain the differences between a ‘pretty good’ work by Mondrian, and an excellent work, referring to the abstract paintings here.”

Wastebook

• Make the Cheapest Game You Can – “Find a collection of images you did not request and could not control. Study it. Build a playable game from what is already there. It forces you to stop treating art as decoration added after the important work is finished. It teaches you to recognize visual themes, write within limitations, cut material that does not serve the project, and find ideas outside your normal habits.”
• If you like what most people consider “bad music,” finding an endless stream of AI generated music is a gift.
• “Eine Henny.”

Podcasts

• HuggingFace make me gazpacho – Software Defined Talk #583 – How normal people use AI, Alphabet’s Anthropic-fueled earnings, and skills vs. agents. Plus, catching AI-cheating students with a hidden prompt.
• Talking About Modernization – Software Defined Talk #582 – Bun’s move to Rust, OpenAI hunting for revenue, and Stripe’s bid for PayPal. Plus, Coté’s Dad wisdom.

Logoff

I’m trying out new newsletter formats. As ever, I want to publish on my blog more (and more things there), but for those of who read (and like) email versions, you know, I need something. Anything?