Synthesis history

Five new items this pass all relate to ARC-AGI-3 — the competition page, arxiv benchmark paper, and associated commentary — confirming the benchmark is now formally active rather than merely announced. None carried extr…

The main addition is item 40065, Google's Android Bench update: GLM-5.2 is now included in Google's domain-specific coding benchmark alongside seven other models (Fable 5, Sonnet 5, Opus 4.8, Kimi K2.7 Code, MiniMax M3,…

Jack Clark's Import AI 464[^39880] is the main substantive addition: it specifies Fable's KernelBench-Mega result (18.71X speedup, beating Claude Opus 4.8 at 14.4X and GPT 5.5 at 4.34X), reports the Remote Labor Index s…

Lisan al Gaib's investigation into GLM-5.2's post-training methods[^39733] and Semgrep's independent cybersecurity benchmark result[^39731] are the two substantive additions this pass, deepening the distillation debate …

EdgeBench from ByteDance Seed is the substantive new development: a benchmark measuring in-context experiential learning over 12–72 hour tasks finds top frontier models doubling their learning speed every three months, …

Agents-A1 (35B) is a new entrant reported July 1 by Rohan Paul, claiming 1T-model-level performance by training on 45K-token verified task trajectories and distilling specialist teacher models — a different efficiency a…

Simon Willison has published hands-on testing of Ornith-1.0's 35B MoE variant, confirming proficient agentic performance at 103 tokens per second and noting the clean Apache 2.0 licensing from the Gemma 4 base — adding …

New items this pass are thin — mostly social media posts with no cited claims or key quotes. One developer (Justin, item 35818) confirms DeepSWE benchmark results match their hands-on experience with GLM-5.2, adding a f…

GLM-5.2's 22.8% ARC-AGI-2 score received official verification from the ARC-AGI evaluation team[^35208], converting a self-reported result into an independently confirmed one. A widely-shared analysis combining the veri…

Zvi Mowshowitz extended his GLM-5.2 analysis, now framing it as a potential 'DeepSeek moment' for open-source AI agents[^34125]—a stronger characterization than his prior commercially skeptical stance, though the two fr…

Two Chinese AI labs released models in mid-June 2026 that have drawn attention for closing parts of the gap to closed frontier systems. GLM-5.2, from Zhipu AI, scores near Claude Opus 4.7 on traditional benchmarks and …