today's signal · no scroll
the brief
Speed and control set the tone: DeepMind’s DiffusionGemma leans on diffusion decoding for lower‑latency text, while Simon Willison adds hard numbers. Infra also moved—Cloudflare now routes public hostnames to private origins, OpenAI tied into Oracle Cloud, and Raspberry Pi 5 hit 16GB—as agent tooling matured and unlearning/influence audits sharpened.
DeepMind introduces DiffusionGemma, a diffusion‑based approach that claims up to 4x faster text generation versus standard autoregressive decoders, signaling a new path to lower‑latency LLMs.
DiffusionGemma: 4x faster text generation
New Application Services for Private Origins (closed beta) routes public hostnames to private IPs over IPsec, GRE, CNI, or Cloudflare Mesh—no public IPs or extra connectors.
Route public traffic to private applications with Cloudflare — Application Services for Private Origins is available now in closed beta. Route public hostnames to private IP origins over your existing IPsec, GRE, CNI, or Cloudflare Mesh paths. No public IPs or extra connector software required.
Enterprises can now apply existing Oracle Cloud commitments to OpenAI APIs, consolidating spend and governance while maintaining enterprise security controls for model access.
Access OpenAI models and Codex through your Oracle cloud commitment — Access OpenAI models and Codex through Oracle Cloud, using existing commitments to build and deploy AI with enterprise security and governance.
A 16GB RAM SKU doubles the Pi 5’s memory ceiling, opening headroom for edge inference, databases, and dev workloads previously constrained on 8GB boards.
Raspberry Pi 5 – 16GB RAM — Article URL: https://www.adafruit.com/product/6125?src=raspberrypi Comments URL: https://news.ycombinator.com/item?id=48481857 Points: 179 # Comments: 202
claude‑code v2.1.172 brings sub‑agents that can spawn up to five levels deep, improves Bedrock region detection, adds plugin marketplace search, and fixes a 1M‑context session hang.
v2.1.172 — What's changed Sub-agents can now spawn their own sub-agents (up to 5 levels deep) Amazon Bedrock now reads the AWS region from ~/.aws config files when AWS_REGION isn't set, matching AWS SDK precedence; /status shows where the region came from Added a search bar when browsing a marketplace's plugins in /plugin Added model attribute to the claude_code.lines_of_code.count OTEL metric Fixed sessions using 1M context without usage credits getting permanently stuck — the session now au...
datasette‑agent 0.2a0 lets tools ask users mid‑execution via context.ask_user (yes/no, multiple‑choice, free‑text), tightening supervision loops for safer, more steerable agents.
datasette-agent 0.2a0 — <p><strong>Release:</strong> <a href="https://github.com/datasette/datasette-agent/releases/tag/0.2a0">datasette-agent 0.2a0</a></p> <p>Highlights from the release notes:</p> <blockquote> <ul> <li>Tools can now ask the user questions mid-execution. Tools that declare a <code>context</code> parameter receive a <code>ToolContext</code> object, and <code>await context.ask_user(...)</code> can ask a yes/no, multiple-choice (<code>options=[...]</code>) or free-text (<code>f...
v16.3.0‑canary.48 enables Turbopack’s filesystem cache by default in non‑stable builds, adds a cold‑cache dev indicator, and ships build and CI fixes for early adopters.
v16.3.0-canary.48 — Misc Changes [turbopack] Enable turbopackFileSystemCacheForBuild by default in non-stable releases: #94616 Add a cold cache dev indicator: #94611 instant: build fix-card Copy AI prompt dynamically from title + link: #94637 Stage short-lived 'use cache' entries correctly in dev: #94645 docs: next/root-params references on i18n guide: #94629 [ci] Reduce usage of turborepo cache in build_and_deploy.yml: #94319 [ci] Download fnm directly from GitHub releases and validate check...
Google Research proposes a formal framework and practical audits to verify whether unlearning requests actually remove training influence, helping regulators and developers assess compliance.
New framework for auditing machine unlearning — Algorithms & Theory
LWN’s report details how automated AI agents caused unintended, ecosystem‑wide changes in Fedora and beyond, underscoring the need for guardrails, review, and accountability in agentic tooling.
AI agent runs amok in Fedora and elsewhere — Article URL: https://lwn.net/SubscriberLink/1077035/c7e7c14fbd60fae9/ Comments URL: https://news.ycombinator.com/item?id=48484584 Points: 129 # Comments: 27
OpenAI documents coordinated influence operations using AI to shape U.S. tech narratives—from data centers to tariffs—offering TTPs and indicators relevant to platform defenders.
PRC-linked influence operations are targeting AI debates in the US — A new report from OpenAI details PRC-linked influence operations using AI to target U.S. tech debates, data center narratives, tariffs, and false claims about ChatGPT.
Simon Willison recalls measuring 857 tokens/second on Google’s prior Gemini Diffusion preview and welcomes its return as DiffusionGemma, highlighting real‑world latency gains for developers.
DiffusionGemma — <p><strong><a href="https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/">DiffusionGemma</a></strong></p> Last May Google briefly released an experimental Gemini Diffusion model. I <a href="https://simonwillison.net/2025/May/21/gemini-diffusion/">tried the preview at the time</a> and recorded it running at 857 tokens/second. It was an exciting model, but Google made no further announcements about it.</p> <p>That research ha...