Google dropped its own Python type checker for Meta’s Rust one, and cut type-check time by up to 99%

😁 Hello, super humans! Your type checker is probably the slowest thing in your inner loop that you have stopped noticing. Google just stopped noticing too, then replaced its own tool with a rival’s and published the numbers. Let’s dig in.

📰 Quick Signals

• 🧠 AI: Anthropic plans to put 5% of its shares, subject to stockholder approval, into a new public-benefit LLC, with the board able to add 2% more for every 50% rise in company value above a baseline.
• 🤖 Robotics: Sam Altman confirmed OpenAI will build a humanoid robot, saying “we will” and expecting “a cool demo in 2027,” while calling the software “brain” more important than the body.
• 💻 Programming: NestJS 12 moves its core packages to ESM, adds Standard Schema validation and an official observability SDK, and requires Node.js 20.19+ or 22.12+.
• ⚡ Electronics: Rivian is building its own in-vehicle AI models and a 5nm autonomy processor with Arm and TSMC.
• 📡 Telecom: Ericsson says India’s telecom growth has not plateaued: about 550 million 5G users by September 2026, yet 5G still reaches only about a third of the population.

🔍 The Big Story: Google dropped its own Python type checker for Meta’s Rust one, and cut type-check time by up to 99%

If you write Python at any scale, type checking is either a coffee-break tax on your CI or a laggy squiggle in your editor. Google says it removed a long-standing bottleneck by throwing out its own tool.

What happened: According to InfoQ, Google replaced Pytype, its in-house Python type checker, with Pyrefly, an open-source checker originally built by Meta. Google’s internal benchmarks show roughly 10x faster checking, incremental rebuild latency down by up to 98%, peak compute cost for Python type checking down by more than 80%, and type-checking time on clean-build critical paths down by 90 to 99%.

The details: The interesting part is why Pytype was a dead end. It analyzed compiled bytecode rather than source ASTs, and bytecode changes between Python releases, so every new typing PEP became a growing maintenance burden. Pyrefly works on source, is written in Rust, and Meta says it can check 1.8 million lines of code per second on large codebases, which is fast enough to run on every keystroke. It passes about 97% of the typing specification’s conformance suite, and its diagnostics point at the root cause with inline annotations and sometimes suggest fixes.

flowchart LR
    subgraph Before[Pytype]
        A1[Python source] --> A2[Compile to bytecode]
        A2 --> A3[Analyze bytecode]
        A3 --> A4[Rework for each new Python release and typing PEP]
    end
    subgraph After[Pyrefly]
        B1[Python source] --> B2[Parse to AST in Rust]
        B2 --> B3[Check against typing spec]
        B3 --> B4[Fast incremental results]
    end

The numbers are Google’s own, and the InfoQ piece does not describe migration steps; one commenter rightly noted you should measure cold and warm check times on your own codebase and weigh diagnostic quality alongside raw speed. Pyrefly is also not alone: the commenters compared it with Ty and Mypy.

🗞️ More News

🧠 AI

• Anthropic signed a 20-year lease for capacity at a planned C$10B campus in Nova Scotia, its first Canadian data center.
• Stripe Link data shows AI agent transactions rose 38x in September across about 200 agents.
• Charles Schwab and JPMorgan say they are blocking unauthorized AI agents from moving customer money.
• OpenCode said stolen API keys compromised 1.4% of Console users and is urging customers to rotate keys.
• Investors reportedly want AI cloud firm Nscale to price its IPO at $16B to $19B rather than the $30B to $35B it seeks.
• Anthropic’s workforce data shows junior hiring in technical roles fell sharply while AI-agent work hours passed human hours.
• Illinois Governor JB Pritzker signed an executive order creating a state AI Cabinet to propose AI policy through 2027.

🤖 Robotics

• Stryker is investing $400 million in CMR Surgical and plans to become the exclusive U.S. sales agent for CMR’s Versius surgical robots.
• A Robot Report guide walks through choosing a BLDC motor for robotic arms versus mobile platforms, covering sensing and safety trade-offs.
• The MMX mission launches October 19 toward Phobos; its 25 kg Idefix rover would weigh about 15 grams-equivalent in Phobos’s low gravity and is slated for release in early 2029.

💻 Programming

• Microsoft released .NET 11 Release Candidate 2 under a go-live license, with no new C# language features and one breaking runtime change that disables advisory file locking by default on iOS and tvOS.
• Buildkite added pipeline-scoped secrets, new MCP tools for validating pipelines and diagnosing failed builds, and Linux arm64 support for its GitHub Actions compatibility layer.
• AWS previewed its Well-Architected Agent, which checks environments against best practices and recommends prioritized fixes across cost, security, performance, and reliability.

⚡ Electronics

• At OCP 2026, Navitas showed a GaN-based 24 kW, 800 V to 12 V board built with Nvidia targeting 97% efficiency, while Infineon and TI showed new hot-swap and telemetry parts for AI data center power.
• The Darwin Board is an ESP32 project that treats simple RC filter circuits as genomes that mutate toward a filter suited to a target signal.
• A builder made a microscope with a 3D view, handy for inspecting solder joints.
• SEMICON West 2026 showed the industry preparing to expand fab infrastructure to meet AI demand.

📡 Telecom

• Kit Carson Electric Cooperative became the first provider in New Mexico to sign a BEAD grant agreement: $26.9 million to build fiber to 3,935 locations, with $12 million in co-op matching funds.
• A 6G panel describes the architecture path as moving from standalone 5G toward agentic AI cores, rather than a clean slate.
• Ericsson defends its custom-ASIC approach to AI-RAN, citing near-term gains while keeping GPU adoption as a future option.

👨‍💻 Code Corner

Try a type checker on a deliberately wrong snippet, then see how quickly it points at the root cause. Install Pyrefly and run it on a small file with a real bug.

pip install pyrefly
pyrefly check demo.py

# demo.py
def total_cents(prices: list[float]) -> int:
    return sum(prices) * 100  # bug: returns float, declared int

def greet(name: str) -> str:
    return "Hello, " + name

greet(42)  # bug: int passed where str expected

Both mistakes are caught before the code ever runs, which is the cheapest place to find them.

🧰 Toolbox

• Pyrefly: Meta’s open-source, Rust-based Python type checker, now adopted by Google; this issue’s Big Story.
• ty: Astral’s fast Python type checker, one of the alternatives commenters compared against Pyrefly.
• Standard Schema: a shared interface for validation libraries, which NestJS 12 now supports.
• Darwin Board: the ESP32 project that evolves RC filter circuits, with code and project page.
• 3D microscope build: a 3D-view microscope build that makes solder-joint inspection easier.

🎬 Demo Watch (rotating)

Anthropic reported that Claude, given plain-language specifications for eight quantum circuits and an automated evaluator, produced designs that beat the best published versions on qubit and gate counts, sometimes by an order of magnitude or more. Standouts include a new reversible approach to modular inversion for elliptic-curve point addition found in a single 53-hour run, and a decoder design that cut the gate count of a key step in a leading quantum optimization algorithm by about 12 times. What is hard here is that the evaluator, not the model, is the source of trust: the designs are checkable. What to treat carefully: this is Anthropic’s own report, with researchers at Harvard and QuEra Computing collaborating, so wait for independent replication. Read the Anthropic report.

📚 From the Blog

• Turning Pixels Into Something the AI Can Eat: Episode 1 got clean pixels out of the camera and Episode 2 moved them across the network; this one prepares them for the model.
• Building Your First Neuron From Scratch: a single learnable unit with weights nudged to lower error, the building block behind every model in today’s AI news.
• The Network Behind the Cameras: the unglamorous plumbing that moves captured footage across the network.

😀 The Bot Says…

Google had a 10x slower type checker for years and it was, in fairness, very thoroughly typed.

---

That’s all for today! Hit reply and tell us: how long does your type checker take on a cold run?