Open Models Go Frontier, AI Learns to Improve Itself, and Cyberwar Gets a Scaling Law

Two competing forces are defining the AI moment right now: an accelerating race to push frontier-grade capability into open weights, and a growing recognition that the recursive loop — AI improving AI — may have already begun turning.

Open Source Hits the Multimodal Frontier

Google's Gemma 4 landed this cycle as a genuinely multimodal open model claiming "frontier intelligence on device" — a phrase that would have been marketing nonsense 18 months ago but now roughly describes where the open-weights tier actually sits. IBM's Granite 4.0 3B Vision offers a compact counterpoint: enterprise document understanding in a 3B-parameter package. The pattern is consistent — vision and multimodal capability, once the exclusive domain of GPT-4V and Claude 3 Opus, is now table stakes for any serious open model family.

A Million Tokens That Agents Can Actually Use

DeepSeek-V4 ships with a one-million-token context window, but the more interesting move is the framing: this isn't positioned as a parlor trick for fitting large codebases, but as infrastructure for agents that need sustained working memory across long task sequences. If that claim holds under real workloads, it changes how agentic pipelines get designed — fewer chunking hacks, more trust in the model's own attention.

AI Is Starting to Build Itself

Import AI's issue 455 — "AI systems are about to start building themselves" — isn't hyperbole. It lays out concrete evidence of AI-assisted research pipelines where models generate and evaluate their own training improvements. Issue 454 covered the parallel track: automating alignment research itself. The compounding implication is uncomfortable: if alignment research can be automated, the safety work needed to make recursive self-improvement safe must be completed before the recursion starts in earnest.

Cyberwar Gets a Scaling Law

Import AI 452 introduced something worth sitting with: scaling laws applied to cyberwar. If offensive capability scales with compute the way model performance does, the economics of state-level attacks change fundamentally. Issue 457 followed with the more visceral version — AI Stuxnet: what adaptive, self-modifying cyber-physical attacks look like with a capable model in the loop. The Hugging Face cybersecurity and openness post takes the other side: open models help defenders too, and restricting them doesn't neutralize well-resourced adversaries who won't be bound by export controls anyway.

Benchmarks Are Breaking Down

The Open ASR Leaderboard added what it calls "benchmaxxer repellant" — private evaluation sets specifically designed to catch models that overfit public benchmarks. IBM Research launched the Open Agent Leaderboard with similar integrity goals for agentic tasks, and published a companion VAKRA analysis documenting exactly how and where agents fail. The underlying problem: as models train on or near benchmark distributions, leaderboard positions stop correlating with real-world usefulness. Private held-out sets and failure-mode analysis are the current best answer — but they're also a temporary one.

The thread connecting all of this is compounding velocity: open models absorbing frontier techniques faster, agents gaining longer working memory, AI research pipelines becoming more self-referential, and attack surfaces scaling with compute. Watch whether evaluation infrastructure can keep pace with any of it.