Attested Enclaves and Asterisked Benchmarks

Google has rebuilt its federated learning (FL) stack around Trusted Execution Environments (TEEs). The aim is for outside auditors, not just Google, to be able to verify the privacy guarantees. Gboard already uses it for English and Japanese next-word prediction, and Google reports stronger privacy, better accuracy, and much faster training.

How it works

The system has four parts:

• Upload: Devices encrypt their training examples locally and upload them. Each upload is bound to an access policy, which lists the TEE computations allowed to process that data. Devices require the policy to be published to Rekor, a public transparency log.
• Key release: A Key Management System, itself a cluster of TEEs running Raft, hands out decryption keys only to TEE workloads whose attested code matches the policy.
• Execution: A "root" TEE runs a Python training loop and sends parallel subtasks to worker TEEs. The distributed logic is written in Federated Language, an open-source orchestration language derived from TensorFlow Federated. Only metrics and differentially private model weights leave the enclaves.
• Recovery: At the end of each round, the program saves its state encrypted under the KMS, so it can recover from root or worker failures without exposing extra data.

The KMS and data-processing binaries can be reproducibly built from the open-source Confidential Federated Compute repo. To keep model architectures and preprocessing proprietary, the Python program can load serialized logic at runtime (sideloading). This keeps verifiability only if all privacy-relevant logic, such as DP noise addition, stays hardcoded in the attested program. Google says side-channel attacks on sideloaded workloads are still an open issue, and that the TEE guarantees are "subject to current-generation TEE limitations."

Ai2 has open-sourced AstaBrief 8B and its training data. The model takes a research question plus retrieved literature excerpts and writes a cited report. It now runs as "Fast mode" in Asta's report generator, next to the Claude-powered "Thinking mode." Across the full pipeline, Fast mode averages 51.1 seconds per report versus 178.5, about 3.5× faster. Ai2 says generation time alone is down by nearly an order of magnitude. Most of the speedup comes from changing the system design. Thinking mode summarizes and clusters snippets, then writes section by section. AstaBrief writes the whole report in one pass, and Ai2 reports no loss in quality from doing so.

The recipe

The base model is Qwen3-8B. Ai2 considered RL, which helped its earlier DR Tulu model, but chose plain SFT + DPO because RL is unstable, expensive and harder to debug. Most of the effort went into the data instead:

• Queries: real Asta/ScholarQA user logs, with bots, very short prompts, non-English and non-scientific requests, and personal information removed. That left 90K queries.
• SFT targets: reports from the multi-step ScholarQA pipeline, backed by Claude 3.5/3.7 Sonnet, o3, o4-mini and GPT-4.1. 47K examples survived filtering.

Cloudflare has released Cloudflare Traces in open beta. It takes the automatic instrumentation that Workers Tracing introduced last year and applies it to the whole HTTP request path. Every supported step a request passes through becomes a span with timing, outcome and attributes: security rules, transform rules, cache decisions, routing, Worker execution and origin handling. You enable it per domain. You don't need to add code, plugins or instrumentation config.

What the spans tell you

The post describes which spans answer which debugging questions:

• Security: custom and managed rule evaluation shows up with its duration and resulting action. Span events name the rule behind a block or challenge.
• Rewrites: the http_request_transform span lists each change, the part of the request it touched and the rule responsible. It also shows where the rewrite happened relative to routing and origin.
• Routing: workers_routing shows whether a route matched, which routing type was used and the pattern that matched.
• Latency: cache, upstream and origin spans nest inside each other. In Cloudflare's example, a cache miss spent 527 ms of a 539 ms request waiting on the origin.

Cloudflare says its own internal traces often contain thousands of spans from dozens of services. The public version is a subset of that.

Sampling, propagation, export

You set a baseline sampling rate per domain, such as 1%. override it for matching traffic and use the same Rules language as the rest of Cloudflare, so you can target hostname, IP, header, path, method or geography. One example: trace 100% of requests from a single customer, or 100% of requests carrying a temporary debug header, while everything else stays at the baseline.

Cloudflare now reports that it is the fastest provider in 74% of the world's 1,000 largest networks, up from 60% in April 2026. That works out to 150 more networks and 38 more countries where it ranks first. The more useful part of the post is how the measurement works and how much the new number depends on a change to it.

How the ranking is computed

Cloudflare takes the top 1,000 networks by estimated user population from APNIC data. For each network it measures connection time, meaning how long the TCP handshake takes from a real user's browser. Providers are ranked by the trimean of those times, a weighted average of the 25th, 50th and 75th percentiles chosen to dampen outliers.

Since 2021 the samples have come from Cloudflare-branded error pages. When a user lands on one, the browser quietly fetches small files from Cloudflare, Amazon CloudFront, Google, Fastly and Akamai and records how long each connection took and whether it completed.

Pi is a coding-agent harness often grouped with OpenClaw, and it now belongs to Earendil. Two Pi releases reached the Hacker News front page on the same day. Latent Space's AINews issue covers both, along with a busy day of model launches.

Pi 1.0 and Pi Durable

Pi 1.0 is a feature release:

• "Codemode," with native support for MCP, Jev and image models
• Extension support for virtual models
• Deferred tool loading
• Cache warming for Anthropic models
• Mid-conversation system messages, so prompts and tools can change based on the transcript so far
• A new TUI theme, with full-screen mode now the default

Pi Durable is the more interesting release architecturally. It ports Pi to TypeScript and moves every stateful component out of the process. Each step is recorded as a checkpointed task. If a process crashes or restarts, agents and subagents resume from their last exact state. Pi Durable runs on any JavaScript runtime (Node, Bun, Cloudflare), uses pluggable storage (memory, SQLite, JSONL) and supports local or remote execution.

Other mechanisms:

• Concurrency: one harness runs several parallel, branching conversations, such as a main channel plus threads, without them blocking each other.
• Durable extensions: installable bundles of system prompts, tools, hooks and durable tasks, for example a multi-step checkout with rollback.
• Background compaction: older messages are summarized to stay within token limits, and the agent keeps working while this happens.