"AI Medley" is an AI tech sharing column by "Salty Talk," casual chats covering recent tools, papers, and practical advances. Today, we look at two major developments: an academic consortium developing the Ataraxos AI to defeat the top human player in the classic board game Stratego, and Google's announcement of its internal Gemini 4 Argon model along with its production engineering applications.
1. Academic Consortium Develops Ataraxos AI to Defeat Top Human Stratego Player
Following chess, Go, and poker, the classic strategy game Stratego had long remained unconquered by machines; even DeepMind, despite its vast budget, failed to build a system capable of consistently defeating elite human players. Now, a joint research team from Carnegie Mellon, MIT, New York University, and Stanford University has developed Ataraxos AI. Using just 16 GPUs and a few thousand dollars in compute costs, Ataraxos defeated Pim Niemeijer—widely recognized as the greatest Stratego player in history—with a record of 15 wins, 1 loss, and 4 ties. In this imperfect-information game where each player commands 40 pieces of varying ranks revealed only during combat, NYU researcher and co-author Eugene Vinitsky noted: "Stratego has a very unique feature that it has a large amount of hidden information that unfolds over a very long time horizon" [Source: arstechnica.com](https://arstechnica.com/science/2026/10/ai-finally-beat-the-best-stratego-player-in-history-and-did-it-on-a-budget/).
2. Google Announces Gemini 4 Argon Model, Kept Internal for Now
Google promised Gemini 3.5 Pro back in June, and after steadily shipping smaller Flash models, the company is aiming to reclaim the frontier with Gemini 4 Argon. Google claims this new AI delivers state-of-the-art performance across coding, knowledge work, and cybersecurity, though it remains unavailable to the public. While external developers must wait, Google reports that its internal engineers are already using it extensively. Powered by "fleet telemetry data," Argon reportedly helped Google save 300 TiB of memory across its data centers. Meanwhile, its autonomous agents have been driving the internal migration of C/C++ codebases to Rust, spanning thousands of lines in core re2 and libgav1 libraries and over 800,000 lines of code in the Fuchsia OS Zircon kernel. On the DeepSWE v1.1 software engineering benchmark, Gemini 4 Argon scored 77.9%, outperforming GPT-6 Astra, Fable 5.1, and Opus 5.5. Google promises comparable gains on long-horizon tasks and highlighted top-tier scores on the Vals Index economic analysis benchmark [Source: arstechnica.com](https://arstechnica.com/google/2026/09/google-announces-gemini-4-argon-ai-model-but-you-cant-use-it-yet/).
From an academic consortium training Ataraxos on 16 GPUs to beat Stratego's best human player, to Google leveraging Gemini 4 Argon to migrate the Zircon kernel and critical libraries to Rust, cutting-edge research and real-world engineering continue to push forward in their respective domains. Whether tackling hidden information in imperfect-information games or proving capabilities on software engineering benchmarks and massive system migrations, empirical data and practical results remain our most reliable measures of technical progress.
> Note: 2 items were excluded due to topic scope boundaries: b039f703ebf7885f (politician statements and public health policy controversies) and 3615960163537c76 (government-led industry regulation policies and public governance issues).
---
*Text compiled with AI assistance; audio generated with synthetic AI voice.*
