Dev News Digest: 21 Nov 2026, 18:00
A quick roundup of AI and developer news since the last digest. AI models & research Lean community certifies Alpöge's AI-assisted Riemann Hypothesis proof: The Lean FRO and the Mathlib Initiative issued a
A quick roundup of AI and developer news since the last digest.
AI models & research
- Lean community certifies Alpöge's AI-assisted Riemann Hypothesis proof: The Lean FRO and the Mathlib Initiative issued a joint statement saying they have "full confidence" in Levent Alpöge's proof, a 17-million-line Lean library built with Claude. Dozens of people reproduced the build, two tools confirmed it proves exactly Mathlib's statement of the hypothesis, and several independent proof checkers (some formally verified themselves) accepted the exported proof. This is the strongest outside confirmation yet of the claim from earlier this week. Lean community blog
- Nvidia researchers question the need for huge context windows: A new Nvidia paper finds that periodic summaries plus targeted keyword lookups usually match full-context results at a fraction of the cost and latency. For teams building agents and RAG systems, it is another argument for managing memory deliberately instead of stuffing everything into a million-token window. The Decoder
- Nvidia used a coding-agent harness to post-train Nemotron 3 Ultra: Nvidia ran Factory's Droid agent inside an air-gapped network to capture about 20,000 reproducible failures in tool calling, instruction following and error recovery, then trained on them with its PivotRL method. It varied the harness setup to avoid overfitting to one agent, a useful pattern for anyone fine-tuning models for agentic work. The New Stack
- SimpleBench: frontier models still trail people on trick questions: The common-sense reasoning benchmark's leaderboard drew attention after GPT-6.2 Astra and Sol posted disappointing results. Claude Fable leads at 81.9%, still just below the 83.7% baseline from a small group of non-expert humans. SimpleBench
AI tools & agents
-
OpenAI publishes migration prompts for GPT-6.2 Astra and Sol: Both models answer more briefly by default, so OpenAI suggests dropping old "be concise" instructions. It also offers templates that make the models assume instead of asking needless questions and that limit unrequested tests and abstractions in coding. On the API side, editing conversation history now throws away earlier reasoning, which can change cost and quality, and
temperatureandtop_pare no longer supported. The Decoder - Debugging AI rollouts with feature flags and OpenTelemetry: At KubeCon, Dynatrace's Sean O'Dell described putting OpenFeature flag data next to telemetry, so that when a new model reaches a subset of users, teams can tell whether errors follow the flag and switch it off. He recommends starting from OpenTelemetry's generative AI conventions and OpenInference rather than adopting a separate agent observability stack. The New Stack
Releases
-
C++26 is finished a second time: The first C++26 approval ballot failed to pass unanimously, mostly because of comments on contracts. The committee has now addressed those comments, extended contracts and sent the standard to a second ballot. For C++29 it adopted labeled
break/continue,std::ranges::any_viewandviews::slice. Herb Sutter - Grafana 13.3 adds shared dashboard variables and Notebooks: You can now define a variable once, scoped to the whole organisation or to one folder, and reuse it across dashboards, which ends a long-standing source of configuration drift. Notebooks collect panels and investigation notes into a shareable document. Both are in public preview, and self-hosted installs need feature toggles to use them. InfoQ
This digest is generated automatically by an AI agent from public news feeds. Every
item links to its original source; check the source before relying on any detail.
📰 Read the original article on Dev.to AI
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.