☕️ Meta doubled the efficiency of its ads AI model

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

📈 Meta doubled the efficiency of its ads AI model

🔌 Cloudflare Workers now support TCP

⚡ Vercel eve builds Next.js for AI agents

🌐 Kubernetes Gateway API adds stable TCP and UDP routing

🔑 NuGet API keys now expire in 30 days

Plus: 🎁 13 other news you might like, 🧰 6 tools, and 📚 5 papers.

📈 Meta doubled the efficiency of its ads AI model LINK

Meta rebuilt its Generative Ads Recommendation Model (GEM) training stack to double end-to-end efficiency to 20-25% Model FLOPs Utilization while scaling training FLOPs 4x over 12 months across several thousand latest-generation GPUs.

Compute gains came from custom recommendation kernels-Jagged Flash Attention drops up-to-50% padding waste, BlockAttention cuts long self-attention from O(L²) to O(L), and MXFP8 attention plus MLP add low-precision speed without regressing CTR/CVR quality.

Scaling used topology-aware 5D parallelism with SM-free collectives, pairing 2D FSDP plus Expert Parallelism for dense parameters and Fully Sharded 2D Model Parallelism for trillions of sparse ones, mapped to Meta's multi-tiered network to hide communication.

🔌 Cloudflare Workers now support TCP LINK

Cloudflare Workers can now accept inbound TCP connections through a new `connect(socket)` handler, letting you route raw sockets between Workers, Durable Objects, and Containers to run gRPC and other TCP-based protocols on Cloudflare.

You can forward a socket from a Worker to a gRPC server running in a Container written in any language, enabling full-duplex bidirectional streaming across Cloudflare's 330+ locations for low-latency voice AI and colocated inference workloads.

Without a container, Workers can serve unary and server-streaming gRPC APIs using `@connectrpc/connect`, with Cloudflare automatically translating between gRPC and gRPC-web so existing clients and backends need no changes; it's launching in private beta.

⚡ Vercel eve builds Next.js for AI agents LINK

Vercel has open-sourced eve, an in-house framework that builds AI agents using a filesystem-first approach modeled on Next.js, organizing behaviors, tools, and subagents into files and folders rather than boilerplate registration code.

Eve ships with durable execution that checkpoints steps and resumes after human approval, runs agents in hardware-isolated microVMs on the server and Docker locally, and records every model call and tool invocation as OpenTelemetry traces.

Requiring Node.js 24 or higher, eve defines tools as TypeScript files auto-discovered by filename with Zod-validated inputs, loads Markdown skills into context only when relevant, and deploys the same agent across GitHub, Slack, Salesforce, or Notion.

🌐 Kubernetes Gateway API adds stable TCP and UDP routing LINK

Gateway API v1.6.0, released June 30th, graduates TCPRoute and UDPRoute to Standard in the v1 API, giving raw L4 TCP and UDP workloads like databases, DNS, and VoIP a portable way to plug into a Gateway.

TCPRoute and UDPRoute route traffic by port and protocol alone, with no L7 awareness: a route's parentRefs attaches to a Gateway listener, and omitting sectionName binds it to every matching listener instead of one.

New experimental resources now live in a separate gateway.networking.x-k8s.io group with an X prefix, such as XBackend and XMesh, dropping the prefix on graduation to Standard rather than relying on version strings to signal stability.

🔑 NuGet API keys now expire in 30 days LINK

Microsoft is capping new NuGet.org API keys at a 30-day lifetime starting August 17, 2026, down from the previous 365-day option, as part of an effort to tighten supply chain security across the ecosystem.

Keys created before August 17 will expire on November 1, 2026, and Microsoft recommends migrating to NuGet Trusted Publishing, which uses OpenID Connect to issue temporary keys and already supports GitHub Actions and GitLab.

Publishers sticking with API keys must update automated workflows for shorter rotations and narrowly scope permissions, a change Microsoft justified by citing the NX NPM breach where stolen credentials triggered 6,000 malicious activations in 36 minutes.

Other news you might like

• How to implement JWT authentication in NestJSLINK

• Azure and Community Guidelines on Choosing Between a Skill or a Sub-AgentLINK

• Gap Decorations Are Now Available, Here’s What’s NewLINK

• Alibaba’s AI coded for 16 days straight and every commit is on GitHubLINK

• How Stripe Built Kai on Deep Agents in 1 WeekLINK

• Valve publicly releases Lepton and FEX compatibility tools to power Steam Frame gamingLINK

• How to build a cloud software factory - computer use verificationLINK

• Smaller, faster, safer: running Kimi and GLM at scaleLINK

• Alibaba shares rally after unveiling its 'most powerful' AI model as U.S.-China competition heats upLINK

• Your agent needs a computer, not a container, introducing @cloudflare/computerLINK

• Azure Private Link for Elastic Cloud Serverless is now generally availableLINK

• Core Ubuntu package to be turned into a Snap; Deb ditchedLINK

🧰 Trending tools

Prefactor: an evaluation layer that scores agent runs in real time, catching quality regressions and drift before they impact production customers.LINK

SKI: lets you voice-code with Claude Code, Codex and other agents, hearing spoken responses back so you build hands-free at thinking speed.LINK

Humalike x Hermes: gives AI agents turn-taking, timing, and memory APIs so they know when to speak, wait, or interrupt naturally.LINK

Openbase: helps developers pick reliable open-source packages by comparing popularity, activity, and reliability metrics alongside real user reviews.LINK

Pushary: sends AI agent approvals and questions to your phone's lock screen, letting Claude Code, Cursor, or Codex keep working while you're away.LINK

Claude Code usage tracking by LangWatch: monitors token usage, costs, cache efficiency, and tool calls across coding sessions with full terminal replay, free for individuals via npx.LINK

📚 Trending papers & reports

Cross-benchmark skill training shows that teaching an AI agent on 363 unrelated tasks still boosted its performance on five separate outside tests by roughly 3 to 10 percentage points, even on tasks like software engineering it never practiced, suggesting it learned genuinely transferable problem-solving habits rather than just memorizing test patterns.LINK

Code authorship detection nails identifying programmers from coding-contest submissions, hitting 92.6% accuracy for 10 authors, but collapses to near-zero on real classroom coursework, undermining its use for catching student cheating.LINK

Android app mapping combines code analysis with smart automated exploration to test apps, covering ~18% more of an app using 345 fewer interactions than standard random testing, speeding up quality checks before release.LINK

Android app testing maps an app's internal screen logic in detail, generating tests that cover ~16% more of the app's code while taking ~84% less time than existing tools.LINK

App testing tool maps an app's full menu structure so its automated tester skips random guessing, reaching complete screen coverage roughly 10x faster than today's random-tap testing tools.LINK

See you tomorrow for a new dose of ☕️ Devshot!