Skip to content
Daily BriefNewsDaily Brief

5 AI stories from October 12, 2026: Copilot's sandboxing goes GA, a self-grading model loop, and a Harvard review-bottleneck study

· by Pondero Newsdesk · 5 stories

AI news daily brief: 2026-10-12

AI coding agents made engineers write 30 percent more code this year, and a Harvard study of 718 firms found no measurable rise in shipped software to show for it, per the paper. That finding, covered below, is the one story here that complicates the productivity case behind every coding-agent seat sold in 2026, read it first if today's five stories only get a few minutes. The other four: a GitHub Copilot rollup bundling GA sandboxing and a cheaper model, Sakana AI's self-improving model framework, Nvidia and Microsoft's RTX Spark launch, and Google's open multimodal embedder.

GitHub's Copilot weekly roundup ships Claude Haiku 5.5, GA local sandboxing, and local-model discovery in the CLI

GitHub's October 9 "weekly releases" changelog rolled up a cluster of changes dated October 5 through 7, per GitHub's changelog. Local sandboxing, which scopes what files, networks, and credentials a Copilot agent can touch, reached general availability in Copilot CLI, the Copilot app, and VS Code sessions that use Agent Host, at no added cost. The same week, Claude Haiku 5.5 became generally available across Copilot's Pro, Pro+, Max, Business, and Enterprise tiers, pitched by GitHub as a fast, cheap option for subagents and quick edits. Separately, Copilot CLI's /model command can now discover models from a running local Ollama instance and add one mid-session without a restart, starting in CLI version 1.0.94-0.

The GA milestone matters most for enterprises: admins can now mandate sandboxing across every session instead of relying on individual developers to opt in. Read more: the full writeup.

Sakana AI and UC Berkeley researchers publish MASS, a recursive self-improvement framework where one model plays executor, optimizer, and evaluator

Sakana AI, working with UC Berkeley researchers, published Multi-Agent Self-Supervision (MASS): one model executes a task, searches for a better way to organize the team doing it, judges which version of the work is best, then retrains on the winning logs before the next round. Starting from Qwen3.6-27B, two rounds lifted win rate against the base model from 53.9% to 69.9% on held-out synthetic tasks, and score-per-token on four public research benchmarks reached 1.2x to 1.6x the base model, per the paper. The gains did not transfer to coding: Terminal-Bench 2.0 and SWE-bench Verified scores dropped to 0.96x and 0.94x of the base model.

The more notable number may be trust, not skill: agreement between the model's self-evaluator and two outside judges, GPT-5.5 and Claude Opus 4.8, climbed from 73% to 93% across the two rounds. Read more: the full story.

Nvidia and Microsoft launch RTX Spark PCs and preview DGX Station for Windows, promising trillion-parameter local inference

Nvidia and Microsoft opened preorders October 7 for RTX Spark Windows laptops pairing a Blackwell RTX GPU with a Grace CPU over a 600 GB/s link, rated at one petaflop of FP4, per Nvidia's blog post. Laptops ship October 16; desktops follow in November. Nvidia also previewed DGX Station for Windows, built on a GB300 superchip with 748GB of coherent memory, pitched as trillion-parameter-scale inference on a Windows box.

The spec buyers will anchor on is misleading: a Microsoft engineering post shows the 128GB unified-memory SKU exposes only about 110GB to the GPU once Windows takes its share, per StorageReview's reporting. That gap decides what model actually fits once weights, KV cache, and runtime overhead compete for the remainder. Read more: the full breakdown.

Google ships EmbeddingGemma 2, an open 740M-parameter embedder that natively handles text, images, audio, video, and code

Google released EmbeddingGemma 2 on October 6, folding text, code, image, audio, and video embedding into one 740-million-parameter model under an Apache 2.0 license, per Google's announcement. The context window grew to 8,000 tokens, four times the prior version, and Matryoshka Representation Learning lets developers truncate output vectors to cut local vector-database storage by up to 6x. On code retrieval, Google reported a jump on MTEB Code from 68.76 to 78.68.

A quantized build needs roughly 191MB of active RAM for text-only use on a Pixel 11 Pro, small enough to run offline alongside other on-device AI work, replacing what used to require a separate specialist model per data type. Read more: the full release rundown.

Harvard/Jellyfish study of 700+ firms finds AI coding agents write more code but code review absorbs the gains, with no measurable rise in software output

Harvard researchers Fiona Chen and James Stratton analyzed Jellyfish engineering-analytics data covering 300 million work events across more than 700,000 workers at 718 firms from January 2021 through March 2026, per the paper. Fully agentic coding tools raised lines of code written by 30 percent and pull requests by 23 percent, both statistically significant. Neither agentic tools nor autocomplete-style assistants produced a measurable change in resolved Jira issues or epics, the study's proxy for shipped software.

The paper traces the gap to review: time to review a pull request rose 49 percent after agent adoption, and comments per pull request rose 35 percent, even at firms that had already adopted AI code-review tools of their own. Read more: the full study breakdown.

Sources