Wed · 29 Apr 26 29·iv
A frontier issue. The model labs publish their receipts; the network reports its first drone-strike outages; an LLM trained only on text from before 1930 quietly asks what we have left to teach it.
AI infra · System cards
5.5, on the record.
OpenAI shipped GPT-5.5's system card alongside the model: red-team transcripts, deployment guardrails, and the residual-risk envelope OpenAI is willing to ship behind. Read it the way you'd read a quarterly filing — the news isn't always the headline number.
- 38pages
- Card length
- 11eval suites
- Capability + safety
- Med.CBRN
- Risk classification
Network · Geopolitics
Drones hit clouds. The map redraws.
Q1 2026 broke the previous record for internet disruption events. Nationwide shutdowns in Uganda and Iran, conflict-driven outages, and the first drone strikes targeting cloud infrastructure — Cloudflare's quarterly review reads like a logbook of the internet getting more political.
Three patterns recur across the quarter. First, the increase in deliberate, government-directed shutdowns, often timed to elections, exams, or protests; second, conflict-driven outages whose geography matches active battle lines; third, a new category — direct kinetic strikes on data-center sites in contested regions, raising the cost of "the cloud is somewhere else."
The Radar team's data shows visibility into individual countries' connectivity collapsing for hours to days at a time, with submarine cable damage layered on top. The fix is not technical alone: cross-border peering, multi-region failover, and contracts that tolerate sovereign-level outages all become procurement conversations.
For anyone running a service with users outside their home market, the practical question is no longer which region to pin to but how cheaply you can survive losing one. The full report includes per-country traffic charts, ASN-level outage maps, and a roll-up of the named incidents that defined the quarter.
UgandaNationwide · 5 days
IranRecurring · multiple events
UkraineKinetic · regional clouds
CubaPower-driven · 36h
Dev tools · Agent runtime
The agent has its own machine.
Sprites are Fly's bet that every coding agent should run inside a per-task VM, not on your laptop. Kurt Mackey walks through the loop end-to-end: spin one up from the CLI, ssh in, type claude, and throw the box away when the task is done.
fly sprites create — pick a region, pick a base image
fly sprites ssh — drop into the VM, no setup
claude — the agent is now somewhere it can break things
fly sprites destroy — the side effects evaporate
Growth · Architecture
Microfrontends as a growth loop.
Cursor unified a sprawl of marketing properties into one Next.js shell, then wrapped every page in feature flags for surgical experiments. The case study reads like a playbook for any product team whose landing pages outgrew the engineering team that owns them.
The marketing site is now a place engineers are willing to ship to — because changing it doesn't require a re-deploy of the app.
Cursor split the surface area: a small, slow-moving root, and many fast-moving zones owned by whoever ships them. Microfrontends keep the bundle costs honest; flags replace branch-heavy A/B with in-product toggles.
The interesting part is the org chart that fell out. Growth experiments now happen in the same repo as the app, in days instead of sprints, with the marketing team unblocked from engineering review on copy-only changes.
Engineering · iOS
Swift enums that age well.
Part II of Sentry's tour through their metrics-API rewrite: replacing dynamic types with protocols, future-proofing Swift enums against new cases, and the small ergonomic choices that decide whether an SDK survives three OS releases.
// before: stringly-typed metric, easy to break across SDK versions func capture(_ name: String, value: Double) // after: protocol-driven, exhaustively switchable, additive-safe protocol Metric { var name: StaticString { get } } enum PerfMetric: Metric { case launchTime(seconds: Double) case scrollFrameDrop(count: Int) // new cases land without breaking call sites }
Protocols over dynamic types. Enums with associated values over stringly-typed metadata.
Research · Public domain
An LLM that has never heard of the iPhone.
Talkie-LM trained a 13-billion-parameter model exclusively on text published before 1930 — every Dickens, Conan Doyle, and Bradstreet line, no Wikipedia, no Reddit, no benchmarks. The result is a curiosity and a question: what does today's AI think when it has never met today?
A demonstration model from the Talkie-LM lab. Trained as a probe: every parameter shaped by writing that predates radio, antibiotics, and the modern news cycle. Useful as a baseline for studying what frontier-model capabilities are downstream of scale versus recency of data.
Analysis · Frontier models
What 5.5 actually buys you.
Zvi Mowshowitz reads the GPT-5.5 system card and the first week of reactions: where the benchmarks miss, which capability jumps actually land in product, and which workflows just changed under your feet without anyone announcing it.
The system card mostly told us what we expected — which is itself the news. — Zvi Mowshowitz
That's today.
Seven picks for Wednesday, 29 April 2026. Today's sources: OpenAI, Cloudflare, Fly.io, Vercel, Sentry, Hugging Face via Telegram, and The Zvi on Substack. Picked against the rubric — tools you could adopt this week, dev infra worth knowing, and the occasional idea worth slowing down for.
The Batch · DeepLearning.AI
The Zvi · Substack
Import AI · Jack Clark
OpenAI
Anthropic Engineering & News
GitHub Engineering
Sentry Engineering
Vercel Blog
PostHog Blog
Cloudflare Blog
Fly.io Blog
Next.js Blog
@seeallochnaya · Telegram
@denissexy · Telegram
@rvnikita_blog · Telegram
@TochkiNadAI · Telegram
@ProductsAndStartups · Telegram
Jeff Bullas