Today's issue is about the edges — quantum-resistant tunnels, models that see and reason in the same breath, an agent on your face, the price of a TPU second, and a deploy adapter you finally don't have to write.
Quantum-safe tunnels, in production.
Cloudflare just made post-quantum encryption generally available for IPsec, using a hybrid ML-KEM key exchange. Magic WAN and CNI tunnels can flip the toggle today — no client changes, no key-rotation playbook. The first commodity primitive after the NIST PQC standards.
Adapters. Finally a contract.
Next.js 16.2 lands with a stable Adapter API, a public adapter test suite, and a working group spanning Vercel, OpenNext, Cloudflare, Netlify and Sst. If your platform passes the suite, the framework's features pass too. The deploy target stops being a black box.
GLM 5.1 thinks before it acts.
Andrew Ng's roundup leads with Zhipu AI's GLM 5.1, a Chinese reasoning model that drafts a multi-step strategy before executing — closing the gap with frontier US planners on long-horizon agent tasks.
Strategic planning has been the persistent weak spot for open-weight models. GLM 5.1 separates a "plan" pass from an "act" pass, surfacing intermediate reasoning the user can edit. On the OSWorld and AgentBench batteries, it edges past last-quarter's open champions and lands within striking distance of the closed leaders.
The wider issue argues that coding agents accelerate different kinds of software work to wildly different degrees — greenfield CRUD lifts hardest, debugging in big codebases least. A useful corrective to "agents are X% faster" headlines.
See, then reason — in the same breath.
DeepSeek's new paper teaches a model to interleave object detection with language reasoning through "visual primitives" — the model picks where to look, then reasons about what it found, in a single trace. On standard visual benchmarks it edges past GPT-5.4 and Gemini 3 Flash while burning fewer tokens. Code and weights drop with the paper.
| DeepSeek-Vis | 74.2 | 1.0× |
| GPT-5.4 | 72.8 | 1.7× |
| Gemini 3 Flash | 71.4 | 1.4× |
| Qwen3.5-VL | 69.1 | 1.1× |
How a TPU second is priced.
Dwarkesh sits at a blackboard with one of the original TPU architects to walk through the real economics of training and serving — why batch-of-one inference is so brutally expensive, where the speed-versus-price tradeoffs actually live, and what an inference dollar buys you in 2026.
"Once you understand that the matrix-multiply is free and the data-movement is everything, every other architectural decision in a modern accelerator falls out of that one fact." — from the lecture · 1h 40m
An agent on your face.
An open-source Snap Spectacles project lets you place agents in 3D space, drive them with voice, and pair to your desktop coding tools over Bluetooth. Sample apps include an "Agent Center" you can wear around the room. The fastest way today to put a real agent in real space.
Spatial agents
Pin an agent to a corner of the room; it stays there as you walk.
Voice input
On-device speech, no phone tether for the common path.
Bluetooth desktop
Pair to Cursor, Claude Code, your terminal — see logs in your eyeline.
Agent Center
Sample app: a shelf of agents you summon with a glance.
Open source
MIT-licensed; fork the samples, ship your own lens.
Lens Studio
Build with the existing Spectacles toolchain — no new IDE.
$ end-of-issue --date 2026-05-01
> sources fed: cloudflare, nextjs, the-batch, deepseek-ai (github), dwarkesh (youtube), snap-spectacles (github), seeallochnaya (telegram), tochkinadai (telegram).
> rubric: AI tools you could adopt this week · creative software · dev tools & agentic coding · privacy & security · research with a practical kernel.
> six picks today, six distinct sources. quiet day on the model-lab marketing front; loud day for primitives that actually ship.
> tomorrow at 08:00 zürich.
╔══════════════════════════════════╗ ║ ephemeris · 013 · 2026-05-01 ║ ║ vadim.sikora.name/ephemeris/ ║ ╚══════════════════════════════════╝