AWS releases a free local emulator

AWS local emulator, Google's new Gemini model, and more.

AWS releases a free local emulator

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

☁️ AWS releases a free local emulator

🧠 Google unveils new Gemini model

🔒 Your containers now get hardware-level isolation

⚙️ LinkedIn builds own GitOps engine

🐘 CoreWeave runs Postgres in your cluster

⚡ You can now share AI memory across TPUs

Plus: 🎁 8 more stories you might like, 🛠️ 5 tutorials, 🧰 6 tools, and 📚 5 papers.

AWS releases a free local emulator LINK

  • AWS has shipped AWS Emulator, a free, locally hosted suite that replicates hundreds of its cloud services on developer workstations, letting teams validate AI-generated code offline instead of against the live cloud.
  • Written entirely by AI agents with no human-authored or reviewed code, the tool uses AWS's public Smithy API models as the spec and the live platform as ground truth, with agents generating test suites, filing GitHub issues, and fixing bugs behind an adversarial security merge gate.
  • Built in Rust, it runs as a local HTTP server via CLI on Linux, Windows, macOS or Docker, so developers just redirect AWS CLI or SDK calls to localhost; it's open-sourced under CC0 with a public "things-that-do-not-work" backlog of known discrepancies.

Google unveils new Gemini model LINK

  • Google today released the Gemini Decisions API, a narrow AI service that answers only yes-no, multiple-choice, and rating-scale questions, running nearly five times cheaper than its entry-level Gemini 3.5 Flash-Lite model.
  • Gemini Decision 1 scores options without token-by-token decoding, keeping median compute latency around 50-100 ms across multimodal datasets, and returns a terse answer plus a confidence score applications can use to decide whether to trust the output.
  • At $0.10 per million input tokens with no output charge, the model beat OpenAI's Decisions API and the startup TypeSafe's Jev on seven of nine benchmarks, though Jev proved slightly more cost-efficient; a limited free tier is available in Google AI Studio.

Your containers now get hardware-level isolation LINK

  • AWS now runs each EKS pod or ECS task inside its own hardware-isolated Nitro Sandbox VM, letting you safely execute untrusted, multi-tenant, regulated, or LLM-generated code while sharing a single EC2 instance across workloads.
  • Each sandbox is a lightweight VM managed by the Nitro Hypervisor with near-native EC2 performance, plus cryptographic attestation that releases secrets only to workloads proving their integrity; enabling it in EKS or ECS is a one-line change.
  • Workloads stay fully native, Pod Identity, VPC CNI, EBS/EFS/S3 storage, and CloudWatch all work out of the box, and Nitro Sandboxes are available today on most Nitro-based EC2 instances in all commercial regions at no extra charge.

LinkedIn builds own GitOps engine LINK

  • LinkedIn replaced Argo CD with its own in-house GitOps engine to deploy across more than 250 Kubernetes clusters, now handling thousands of services and over 100,000 deployments every week.
  • Rather than a shared controller, each deployment runs in its own ephemeral pod with no shared state, using PUT for full replacements instead of PATCH and a 3-way diff across live, previous, and newly rendered manifests to detect drift.
  • Deployment infrastructure failures fell roughly 15-fold despite deployment volume rising 3.7x, though LinkedIn credits better orchestrator error classification too and keeps Argo CD for cluster add-ons; the engine stays internal with no open-source release.

CoreWeave runs Postgres in your cluster LINK

  • CoreWeave Database is now generally available, a managed PostgreSQL service that runs inside customers' CoreWeave Kubernetes Service clusters, letting AI teams keep a production relational database next to their GPU training and inference workloads.
  • Databases run on CPU nodes and connect through in-cluster Kubernetes services, avoiding cross-cloud network traffic and egress paths, with each instance keeping a primary plus read replicas across hosts, automatic failover, and point-in-time recovery from continuous backups.
  • Built on the CloudNativePG operator and fully PostgreSQL-compatible, it supports migration from other clouds using native Postgres tools; customers keep control of data, schemas, and users while CoreWeave handles replication, failover, patching, and backups.

You can now share AI memory across TPUs LINK

  • Google Cloud has open-sourced TPU Sync, a C++ library that shuttles KV cache blocks and model weights between TPU HBM, host DRAM, shared storage, and the data center network, targeting LLM serving and reinforcement learning weight sync.
  • For prefill/decode disaggregation, median KV transfer latency for Qwen3 235B dropped from several seconds to roughly 230 ms, while prefix caching cut average TTFT by up to 90% on 64K-context workloads.
  • Developers get JAX and PyTorch wheels wiring into vLLM, SGLang, and llm-d 1.0 on GKE, but the Apache 2.0 repo ships only development builds on PyPI, RL weight sync stays experimental, and APIs are changing rapidly.

Tutorials & Advice

> Understanding React internals: React is fundamentally a scheduler deciding what to render and when, so understanding that engine clarifies why each hook behaves the way it does.

> Tests as data: ensuring AI-generated code does what we want: Store AI-generated tests as plain data files describing inputs and expected outputs so coding agents can't quietly alter them while refactoring.

> How we migrate TBs from Heroku without logical replication: PlanetScale copies Heroku databases table by table and tracks row changes with triggers, enabling terabyte-scale moves without the usual hours of downtime.

> An Open Field Guide to Inference Engineering: Run open models yourself instead of paying hosted AI services, because self-hosting cuts high-volume costs, protects private data, and controls speed and reliability.

> Two Ways Into the Maze: From Bidirectional BFS to Bidirectional Dijkstra: Searching a graph from both start and goal at once dramatically cuts the work, but with per-step costs you must keep going until no cheaper meeting is possible.

Other news you might like

  • GKE dynamic slicing on TPUs LINK
  • See What Your Argo CD Applications Cost LINK
  • Choosing QuePaxa, an unusual consensus algorithm LINK
  • Developers have been hacking around MCP’s file-upload problem for two years. A fix is taking shape. LINK
  • meta/astryx: An open source design system that's fully customizable and agent ready LINK
  • What if any process could handle SIGKILL? LINK
  • CloudNativePG becomes a CNCF Incubating Project LINK
  • Microsoft releases PowerShell 7.7 with security fixes and new commands LINK

Trending tools

Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a shared chatroom during AI cooldown periods. LINK

Jev: adds structured AI-driven decision-making to automation workflows, letting developers build reliable branching logic without hand-coding every rule LINK

Monospace from Directus: wraps any SQL database with instant REST and GraphQL APIs plus a no-code app for collaborative, team-friendly data management. LINK

iFixAi: audits AI agents with 250 inspections across 69 misalignment categories, flagging failures and giving engineers evidence to investigate and fix issues LINK

Floot MCP: connects Claude and ChatGPT to build full-stack apps from chat, handling databases, logins, hosting, emails, and deployment automatically. LINK

FastRouter.​ai: routes requests across 200+ AI models through one OpenAI-compatible API, offering smart routing, automatic failover, and cost controls for teams. LINK

Trending papers & reports

Blurry bug-report screenshots stop coding agents from finding the faulty code, so this approach feeds them crisp reference images auto-generated from the app's own codebase, sharply improving which source file they correctly identify. LINK

Portable work packages bundle a finished document with the hidden research, assumptions, and permission rules behind it, letting an AI reader correctly answer 96% of questions versus 86% when handed raw files. LINK

Enterprise coding assistants largely fail at writing Salesforce's niche agent-building language, with the best model passing just 25% of tasks and every model scoring zero on writing from scratch, showing public-code skills don't transfer. LINK

Automated software proofs can now be written by an AI agent that reasons like a coder, lifting success from 47% to 60% on hard verification tasks at roughly half the cost. LINK

Browser-based chip toolkit revives discontinued Xilinx Spartan-3E and Spartan-6 chips with a free, open flow that matches the old software byte for byte while fitting designs into ~15% to 21% fewer logic cells. LINK

See you tomorrow for a new dose of ☕️ Devshot!

More from the archive