In this issue
AI stopped looking like a software subscription this week and started looking like a utility bill. On the All-In podcast, Marc Benioff said Salesforce expects to spend $300 million on Anthropic tokens this year, most of it on coding. That single number put a price tag on something a lot of us have been feeling: the meter is running, and it doesn't stop when you close the tab.
Once AI spends land in the same mental bucket as your cloud bill, the questions change. It stops being "is the model smart enough" and becomes "how many round trips does this thing make before it finishes, and what does each one cost me." That reframe runs through today's whole issue, from the engineer's changing job to the bots quietly taking over your traffic.
Topics of the day:
AI spend becomes a metered utility, not a subscription
Netlify's CTO: writing code is no longer the job
ChatGPT admits its old memory was broken
Bots now outnumber humans across the web
Five curated reads on AI economics and operator tools
The Shortlist: Anthropic's pause call, Cursor's Design Mode, McDonald's Archy
AI spend is becoming your infrastructure bill
What's happening: Marc Benioff said on the All-In podcast that Salesforce expects to spend $300 million on Anthropic tokens in 2026, most of it on coding. The figure put a number on what operators have felt for months: AI is no longer a per-seat subscription, it is metered like a cloud bill.
In practice:
Treat agentic tools like metered infrastructure, not SaaS seats, because one request can quietly fan out into dozens of model calls and retries.
Watch round trips, not just the per-token price, since cheaper tokens don't help if your agent burns 40 calls before it finishes a task.
Salesforce measured 30% productivity gains, so the spend buys real output, but someone still has to make that gain legible to whoever signs the check.
Ask vendors how they meter before you commit, since flat "unlimited" plans and metered backends behave very differently once your team leans on them daily.
Bottom line: The new question is how much it costs to run the thing!
Netlify's CTO says writing code is no longer the job
What's happening: At AI Native DevCon in London, Netlify CTO Dana Lawson argued that writing code is no longer the engineer's job. With agents doing the typing, intent expressed in plain language becomes the real programming language, and she expects a billion new apps built by 2029.
In practice:
Lawson's frame is agent experience, AX: design for the agent and the user at once, so the bottleneck moves from typing to deciding.
Steal her line for your own roadmap, you are now the shepherd of production, and the real skill is deciding what not to build.
Treat plain English as your main programming language now, because precision in describing what you want beats raw typing speed.
Watch your non-technical teammates start shipping real tools, since AX removes the CS-degree gate that kept building locked to engineers.
Bottom line: Writing code was never the point, shipping the right thing was. The people who internalize that this year will pull away from the ones still counting lines.
ChatGPT admits its memory was broken, and fixes it
What's happening: OpenAI published unusually candid numbers: its old memory feature had a 41.5% factual recall rate in 2024, meaning it was wrong more than half the time, confidently and without telling you. The fix, called Dreaming V3, rebuilds memory from your history automatically.
In practice:
Check your saved memories again, because recall jumped from 41.5% to 82.8% and preference-following went from 55% to 71%.
Dreaming V3 runs in the background with no commands, synthesizing your chat history instead of waiting for you to save facts manually.
Watch it self-correct now, so once your Singapore trip ends it stops recommending Singapore restaurants and lets stale facts expire.
Plan for free users to get real memory too, since compute costs dropped 5x, so more of your customers now use AI that remembers them.
Bottom line: If you trusted ChatGPT to remember things about your work, half of what it "knew" was wrong until now. Worth re-checking what it still has on file.
Read Later
The AI SaaS squeeze - Tyler Tringas on why AI is compressing software margins, and what that means for anyone selling a tool.
Work with Codex from anywhere - OpenAI's coding agent now follows you from terminal to phone, a preview of always-on AI labor.
OpenAI's Deployment Company - A new OpenAI arm built to help businesses actually build around the models instead of just buying access.
Pit, the AI that runs your business - An agent that wires into operations and tries to run the back office, worth a look if admin is eating your week.
Roughdraft - A markdown-native writing tool built for drafting with AI in the loop, not fighting a chat window.
Bots now outnumber humans online
What's happening: For the first time, automated traffic has crossed 57.2% of all web traffic, according to Cloudflare Radar. AI agents are no longer noise in your analytics, they are becoming their own class of user.
In practice:
Re-read your traffic numbers with this in mind, because a "growing audience" can now be mostly crawlers, scrapers, and agents rather than people.
Expect AI agents to visit your site on a human's behalf, so the page a bot can parse matters as much as the one a person sees.
Tighten your bot rules if you sell or gate content, since scrapers feeding AI models never convert and can quietly inflate your costs.
Watch your ad and SEO metrics, because impressions served to bots burn budget and skew every number downstream.
Bottom line: The open web is now majority machine. If your plan still assumes a human on the other end of every visit, it is already out of date.
The Shortlist
Anthropic called for a coordinated global pause on frontier AI, warning that models are nearing recursive self-improvement, the point where AI builds its own successors with no human in the loop.
Cursor shipped Design Mode, which lets you click an element, draw the change you want, or describe it by voice, and the agent rewrites the underlying code for you.
McDonald's is testing an AI drive-thru called ArchIQ at five US restaurants, where the "Archy" voice assistant has already handled over a million orders with roughly 90% needing no human.
University of Toronto researchers demonstrated a self-spreading AI worm built entirely from free public models that hops device to device and hijacks compute, a preview of the next security headache.
This newsletter is where I (Kwadwo) share products, articles, and links that I find useful and interesting, mostly around AI. I focus on tools and solutions that bring real value to people in everyday jobs, not just tech insiders.
