Amazon Scales Back Most of Its In-House AI Models in Major Strategy Shift
According to a report from Business Insider, Amazon is undertaking a radical overhaul of its AI strategy, moving to phase out most of its in-house flagship AI model series, "Nova." The company has invested heavily in developing its own foundation models, but intensifying competition and profitability challenges appear to have forced a change of course.
Going forward, Amazon is expected to continue developing some of its own models while shifting to a strategy centered on partnerships with external providers such as Anthropic. The company appears to be focusing on expanding Bedrock, its model delivery platform that leverages its strengths as a cloud provider — a shift from "building" AI to "aggregating and offering" it. The success of this platform strategy will directly determine its competitiveness in the cloud market in the years ahead.
DSpark Speculative Decoding Merged into llama.cpp — Local Inference Gets Even Faster
llama.cpp GitHub PR #25173 has been merged, integrating the "DSpark" speculative decoding method developed by DeepSeek. Members of Reddit's LocalLLaMA community have already begun sharing test results, reporting significant improvements in token generation speed (tg).
DSpark-compatible models such as DeepSeek-V4-Pro-DSpark and Bonsai-27B-antidoom-1bit-DSpark have been published on Hugging Face. Speculative decoding speeds up inference by predicting tokens with a small draft model and verifying them with the larger model, and it promises practical benefits for running LLMs in local environments.
New Sandbox Tools for AI Agents — Microsoft's Quicksand and Hotcell
Two open-source projects for building sandboxed environments to run AI agents safely were released almost simultaneously.
Microsoft has released "Quicksand," a tool that can sandbox AI agents on Windows without requiring Docker or WSL, dramatically lowering the barrier to adoption.
Meanwhile, "Hotcell," released by independent developer Sina Meraji, lets you create, pause, and manage sandboxes on any device — laptops, Linux VMs, or bare metal. It achieves isolation using Docker, Apple VZ, or Firecracker, and rather than injecting API keys directly into the sandbox, it issues per-sandbox tokens. Developer experience is well considered too: you can launch multiple sandboxes from AI coding tools like Claude Code with a single line.
While the two tools take different approaches, both address the demand for building secure, easy-to-set-up execution environments for AI agents. As autonomous agent execution becomes widespread, sandboxing technology is growing in importance as a cornerstone of the infrastructure stack.
AI Revenue Is Growing Fast — but Still Can't Keep Up with Expectations, The Economist Analyzes
The Economist argues that while AI companies' revenues are growing rapidly, they still fall short of investor and market expectations. AI-related sales continue to grow at a historically unprecedented pace, but massive infrastructure investments and competitive pressures leave the path to monetization uncertain.
This analysis offers one perspective on the debate over the sustainability of the AI boom. How the market digests the reality that technological progress and business success are not necessarily the same thing will be worth watching.
Engineering Standards in the Age of LLMs — "Use Your Brain"
An article by Paolo Galeone titled "Use Your Brain: Engineering Standards in the Age of LLM" is drawing attention on Hacker News. He argues that even as LLM-based coding tools become ubiquitous, the fundamental principles of software engineering — testing, architecture design, code review, and domain knowledge — remain as important as ever.
The message — that engineers are responsible for actively guaranteeing quality rather than accepting AI-generated code as-is — gives many developers pause for thought. LLMs are powerful tools, but they are no substitute for engineering judgment.