01
The Defense Beat
The Pentagon picks eight names. Anthropic is not on the list.
On May 1 the Defense Department finalised classified-network AI deals with OpenAI, Google, Microsoft, Amazon, Oracle, Nvidia, SpaceX, and Reflection AI — the entire frontier roster, minus one. Anthropic is out, with officials citing supply-chain risk after a contract dispute earlier in the year over whether Claude could be deployed for "all lawful purposes," language that would have covered autonomous weapons and mass surveillance. Anthropic refused, sued, and won an injunction. It just lost the procurement.
Read the move on its own terms: the U.S. government's classified-AI estate is now built without the lab whose public posture leans hardest on safety constraints. That is either an indictment of those constraints or an indictment of the buyer. Either way, the bifurcation is real — there is now a defence-flavoured stack and a constraint-flavoured stack, and they no longer share a vendor.
~/agents — claude-code · pid 47
02
Agent loops · proxyFOR YOU
DeepClaude swaps the brain while keeping the body.
$ npx deepclaude — a local proxy on localhost:3200 that pretends to be the Anthropic API and routes Claude Code's agent loop to whatever's behind the curtain. Default backend is DeepSeek V4 Pro at $0.87/M output tokens; OpenRouter sits at the same tier; Fireworks at $3.48/M; the real Anthropic Claude is a $15/M fallback. Swap mid-session with /deepseek, switch back with /anthropic, no restart.
For anyone running long Claude Code loops on personal credit, the math is brutal: the project claims 17× cheaper for the same UX. The catch is the obvious one — DeepSeek is a different model with different failure modes, and "identical UX" is a UI claim, not a capability claim. Worth pinning to a side terminal and trying on a throwaway repo before the bill arrives.
For You
03
Claude · creative pipelines
Claude moves into Photoshop, Blender, Ableton — and the lines between code and craft blur.
Anthropic released nine creative-tool connectors on April 28: Adobe Creative Cloud (50+ tools across Photoshop, Premiere, Express), Blender (natural-language interface to its Python API), Autodesk Fusion, Ableton, Splice, Affinity, SketchUp, Resolume. Available on every Claude plan, immediately. This is the same connector pattern Anthropic shipped for consumer apps the previous week — Spotify, Uber, Instacart — but pointed at production tooling.
The Blender one is the giveaway: it's wired to the Python API, which means Claude can both author scenes and read state. For anyone doing scripted creative work — generative art, motion design, asset pipelines — this is no longer "ask Claude how to do it, paste into the app." It's the app, with a model in the loop. Worth a real session, not a demo.
04
Frontier models
GPT-5.5 ships, and OpenAI starts behaving like a vendor that wants the enterprise contract.
OpenAI is positioning GPT-5.5 as the strongest agentic model it has built — coding, computer use, research, longer agent-like workflows. The sales motion has changed too: aggressive pricing into Office-365-shaped accounts, deep ChatGPT-as-platform plays (the OpenClaw deal — 3.2M users now able to run autonomous agents on GPT-5.4 for $23/month — landed in the same news cycle), and the Pentagon contract from Spread 01.
Anthropic's last public marker is Opus 4.6, with 4.7 already in production for users on the Claude tier. The race for "best coding model" is now decided in dot releases, on benchmarks the labs run themselves, with vendor lock-in arriving via tooling rather than weights.
05
Enterprise · governance
Microsoft Agent 365: a control plane priced at $15 a head.
Microsoft launched Agent 365 on May 1 as a dedicated control plane for enterprise AI agents — governance, identity, and security wrapping for whatever the org has built on Copilot, Azure AI Foundry, or third-party model platforms. The product sits one layer above the agents themselves and is priced like an Office 365 SKU: $15 per user per month, billed alongside the rest of the Microsoft estate.
The bet is that 2026's enterprise AI problem is no longer "get a model in" but "track every agent that's already there." Inventory, policy, audit. It's a governance product dressed as a platform, and it competes less with OpenAI than with the in-house "agent registry" projects half the Fortune 500 has been quietly building since Q4.
06
Essay · craftFOR YOU
The hidden costs of great abstractions.
Joel Gritter's essay — quietly the best 12-minute read of the week — argues that the same abstractions that lower the floor on software development also lower the ceiling on understanding it. Early computing required deep technical knowledge to do anything at all. Today, frameworks and LLMs let anyone ship something functional, and almost nobody ships something good — because discerning good from bad is itself an act of expertise that the abstraction layer is now obscuring.
"Discerning good from bad requires expertise. The inexperienced prospector often mistakes pyrite for gold."
It's not a Luddite take. It's a working programmer's worry that the velocity gains being celebrated this year are being paid for in deferred quality, and that the bill comes due at the senior-engineer review desk. Read it before your next code review.
07
On the calendar
Two weeks out: Gemini 4, Astra, Veo, Android 17 — Google's annual rebuttal.
19
May · Shoreline Amphitheatre · Keynote 10am PT
I/O 2026 lands May 19–20. The leaked betting card: a Gemini 4 reveal sized to OpenAI's GPT-5.5 launch, deeper Project Astra integration (the persistent multimodal assistant), Veo updates for text-to-video, and a fresh push on agentic Gemini across Workspace and Android 17. The interesting question is whether Google can land a story dense enough to pull oxygen back from the OpenAI-Pentagon-Microsoft news cycle, or whether the keynote becomes the "and Google also" footer to a quarter that is no longer about Google.
08
Editor wars · pricingFOR YOU
Cursor's credit-meter, Windsurf's million users, Zed's millisecond.
The 2026 review-up of AI-native editors is not flattering to incumbents. Cursor's June 2025 pivot from request-based to credit-based pricing — Pro at $20/month now buying ~225 Claude requests instead of ~500 — has been compounding through the spring, with users moving down the funnel.
Windsurf has crossed a million active developers and is being treated by the trade press as a peer, not a follower. Free unlimited autocomplete is the wedge; the multi-file coherence is closing the gap.
Zed, written in Rust from scratch, keeps winning the latency benchmarks: sub-second startup, edit latency about 40% below VS Code, GPU-accelerated rendering. For anyone who lives in the editor eight hours a day, the milliseconds compound.
The strategic read: the AI editor market has stratified. Cursor is the polish play and increasingly the premium-tier play. Windsurf is the volume play. Zed is the engineer-engineer's editor. Pick by what you're optimising — token economics, multi-file coherence, or input lag — not by hype.
09
Weird science · quantum control
Oxford pulls off "quadsqueezing" — a fourth-order quantum effect that was theoretical until last week.
Squeezing — the trick that lets you trade noise in one quantum variable for noise in another — has been a staple of optical and atomic physics for forty years, with first- and second-order effects routinely exploited in things like LIGO. Quadsqueezing is the fourth-order version, which in plain language means a much finer knob over the shape of a quantum state's noise distribution. The Oxford team is the first to demonstrate it experimentally.
Why it matters outside the lab: every step in the squeezing-order ladder unlocks a class of measurement and sensing protocols that were previously dominated by classical noise floors. The practical applications — gravitational-wave detectors, biological imaging, quantum-enhanced clocks — won't ship next quarter. But the toolkit just got one tier deeper.
10
Foundations · interpretation
MIT shows the double-slit can be derived from classical "least action" — same answer as Schrödinger.
An MIT team published a result on April 21 showing that quantum problems normally solved with the Schrödinger equation — the double slit, quantum tunnelling — can be reproduced exactly by applying the principle of least action, a classical-physics concept that's been around since Lagrange. Same numbers, different machinery.
The paper isn't a claim that quantum mechanics is secretly classical. It's a claim that the seam between the two regimes is thinner than the textbooks let on, and that the "weirdness" of quantum behaviour can be reframed in terms physicists already had vocabulary for. For software engineers it's a reminder that the abstractions we work over — including the ones we built our careers on — sometimes have shorter paths to the same answer hiding in adjacent disciplines.
Hacker News. The Front Page
Top five from the last 24 hours
01
GameStop makes $55.5B takeover offer for eBay.
111 points · 71 comments · 1.6h ago
The internet's most committed meme stock writes a real cheque. GameStop — sitting on a balance sheet stuffed with the proceeds of every retail mania since 2021 — has made a $55.5 billion offer for eBay. The pitch frames it as a "two-sided commerce platform" play, but the deeper read is the war chest meets a market mood: GameStop's leadership has spent four years insisting they could acquire something serious, and a marketplace business with eBay's scale is the first target large enough to make that claim look credible.
HN discussion →
BBC →
02
Trademark violation: a fake Notepad++ for Mac is making the rounds.
152 points · 49 comments · 1.5h ago
A site at notepad-plus-plus-mac.org is impersonating the project — name, biography of creator Don Ho, the works — to push what it calls an "official" macOS build. There is no official macOS build. Don Ho is asking the community to push back on the misinformation and is pursuing the operator directly. The interesting subtext: classic open-source-trust-erosion attack, with the obvious payload risk if anyone runs the binary. If you've been excited about Notepad++ landing on Mac in your feeds today, that's the impersonator's reach.
HN discussion →
Statement →
03
Debunking the CIA's "magic" heartbeat sensor.
33 points · 28 comments · 11.4h ago
After officials credited a quantum sensor codenamed "Ghost Murmur" with locating a downed pilot in Iran from up to 40 miles away, physicists ran the numbers and found the basic claim incompatible with magnetic-sensing physics. The heart's magnetic field is barely detectable at 10cm; at 1m the signal drops by ~1000×. The likely real story: the airman activated a Combat Survivor Evader Locator beacon, and the heartbeat narrative is either spin or an obfuscation of a more boring intercept method.
HN discussion →
Video →
04
BYOMesh — a multiband LoRa mesh dev board claiming 100× the bandwidth.
390 points · 124 comments · 17.1h ago
Dataparty's BYOMesh is a small-form-factor multiband LoRa interface aimed at MeshTNC and rfparty, designed to give hobbyists access to all common LoRa bands on one board. The 100× bandwidth claim is the headline; the HN comments are unanimous that the number is achieved partly by ignoring FCC duty-cycle and bandwidth limits that Meshtastic and MeshCore also routinely break. Useful, but the regulatory envelope is the asterisk no marketing copy includes.
HN discussion →
Product →
05
Southwest Headquarters Tour.
259 points · 80 comments · 18.1h ago
Katherine Michel's tour of Southwest Airlines' Dallas HQ is the kind of operational-photography post HN occasionally elevates to the front page. The Network Operations Center coordinates 4,000+ daily flights — more than any other airline — with simulators that cost up to $14.2M each and a TechOps facility that maintains the world's largest 737 fleet. The piece reads as an artifact of a vanishing kind of corporate visibility: an outsider walking through every floor, photographing the systems, and leaving with the culture intact in the captions.
HN discussion →
Article →
Architecture in the Wild ● Feature
Cloudflare's internal AI engineering stack — built on the platform they ship.
Ayush Thakur · Scott Roe-Meschke · Rajesh Bhatia · blog.cloudflare.com · April 20, 2026
Cloudflare took the unusual step of publishing the architecture of its own internal AI platform — the one its engineers actually use — and the takeaway is that the entire thing is built on Cloudflare's own publicly available products. Workers, Durable Objects, AI Gateway, Workers AI, Access, Workflows, Sandbox. No proprietary infrastructure, no special path. Eat-your-own-dog-food at enterprise scale: 3,683 active users (93% of R&D), 47.95M AI requests over 30 days, 20.18M routed through AI Gateway, 241 billion tokens processed, 51 billion of those running on Workers AI on open-weight models inside the platform. Merge requests jumped from ~5,600/week to a peak of 10,952.
The architecture splits into three layers and the choices are worth absorbing. The platform layer handles auth (via Cloudflare Access for Zero Trust), routing (AI Gateway), and inference. The single most consequential decision was a centralised proxy Worker enforced from day one — clients never speak to model providers directly, which means policy, rate-limiting, anonymisation, and cost attribution all live in one chokepoint that can be evolved without touching every consumer. The knowledge layer is a service catalog in Backstage plus AGENTS.md files inside each repository, so agents pick up structured context about what the code does and how to operate on it. The enforcement layer wires AI code review into every merge request and checks against an internal Engineering Codex.
Two micro-decisions deserve a callout. First, Code Mode at the portal level: 34+ tool schemas collapsed into two tools, cutting context-window overhead by ~7.5% across the entire fleet. That's the kind of optimisation that only shows up when one team owns the whole pipe. Second, anonymous UUIDs replace user emails before requests hit external providers — privacy gain, tiny cost, only possible because the proxy is centralised. The post is a textbook for senior engineers thinking about how to roll out AI internally without spawning a sprawl of per-team integrations that no one can govern six months in.
"One thing we got right early: routing through a single proxy Worker from day one."
— Ayush Thakur, Cloudflare
If your org is currently shipping the agent-per-team architecture this post argues against, treat it as a roadmap. The proxy-first pattern is portable; the rest of the stack is where the company-specific calls live.