A five-year migration, done in two weeks
Edition of 2026-08-19
# AI For All ### Wednesday, August 19, 2026
As of today, an engineering manager sitting on a multi-year legacy-code migration has a documented case of a comparable project going from five years, staffed against it, to two weeks and low five figures, with the actual prompt published.
Five sentences replaced five years
Asana's frontend test suite had been stuck in a slow migration since 2022, chipped away at by whichever team owned the corner it touched. At the old pace it was still roughly five years from done. So the engineering team tried something unreasonable: point OpenAI's Codex at the whole thing and see how close it could get to finishing in a week.
It took a week and a half. The entire prompt was five sentences long, telling the agent to migrate test files from the old library to the new one, follow the codebase's own conventions, and start with the easy files first. Four agents ran in parallel, day and night, checked twice a day.
The bill: $11K in model usage plus about $1K in infrastructure. Whatever this specific project would have cost with three engineers over five years, it was not that. Anyone running an old internal migration that "we'll get to eventually" now has a number to hold that plan against. OpenAI
The chatbot now treats a fifteen-year-old differently than an adult
OpenAI split ChatGPT into two experiences: one for adults, one automatically applied to anyone it identifies or who identifies as 13 to 17. The teen version blocks romantic or sexual conversation, restricts content around suicide and self-harm, and nudges a student caught trying to get a straight answer on homework toward a mode that makes them work through it instead.
The company still does not verify ages directly. It estimates them from how someone uses the product, and defaults to the safer, teen version whenever it is unsure. That matters because, per a 2025 study cited in the coverage, more than seventy percent of American teenagers already turn to AI chatbots for companionship.
The catch a children's-safety lawyer flagged: ChatGPT's long-term memory is not restricted by default in teen accounts. A parent who wants that limited still has to turn on parental controls themselves. The content protections are automatic; the memory ones are opt-in. AP News
The AI boom sent your next RAM upgrade to another planet
A 128GB DDR5 memory kit now costs roughly ten times its lowest tracked price, and industry-wide, high-capacity DDR5 kits are up close to five hundred percent year over year. This is not a chip shortage story about AI hardware. It is ordinary consumer memory, the kind that goes into a normal desktop or a small business server, caught in the crossfire of AI datacenter demand.
The memory makers, SK Hynix, Samsung, Micron, and China's CXMT, are all seeing revenue double or triple as they prioritize the far more lucrative AI datacenter contracts over PC and phone supply. SK Hynix's own CEO has warned that 2027 will be the worst year in the industry's history for memory supply, with demand outstripping production into 2030.
Anyone planning a PC build or a server refresh should price it now, not after a further squeeze. There is no credible path back to last year's prices before the AI market itself corrects. Tom's Hardware
Weak signal
Bloomberg, TechCrunch and Axios all report that Stripe has agreed to buy AI marketplace OpenRouter for more than seven billion dollars, but as of this writing neither company has confirmed it on the record. Anyone holding OpenRouter credits or building on its API should wait for the official word before assuming anything changes.
The migration story and the memory story point the same direction from opposite ends: AI is now cheap enough to rewrite a company's internal plumbing, and expensive enough to break the price of a component that has nothing to do with AI at all.