GLM-5.2 Passes Vibe Check; Z.ai Forecasts Open Fable by December
Latent Space39 days agoBenchmarkZhipu AI's GLM-5.2 has passed broad informal community vibe checks, drawing favorable comparisons to GPT-class models and signaling a meaningful quality leap for open-weights AI. Z.ai, the company behind GLM, is additionally forecasting release of an open frontier-tier model — dubbed Open Fable — by December 2026. Together, these developments suggest open models are genuinely competing at the frontier rather than perpetually trailing closed proprietary systems.
Frontier Post-Training Recipe Review with Finbarr Timbers
Interconnects (Nathan L.)42 days agoCommentaryIn the 18th installment of his interview series, Interconnects author Nathan Lambert speaks with Finbarr Timbers about the post-training techniques used at frontier AI labs. The conversation examines the methodologies — including supervised fine-tuning, reinforcement learning from human feedback, and preference optimization — that shape model behavior after pretraining. The discussion offers a practitioner's perspective on the evolving landscape of alignment and capability tuning at scale.
Inside Anthropic's Weekend Fight Over Claude Mythos 5 and Fable 5 Export Controls★ 76
The Verge AI42 days agoRegulationAnthropic received a US export control directive at 5:21 PM Friday ordering immediate suspension of access to its newest frontier models, Mythos 5 and Fable 5. Rather than comply, the company spent the weekend contesting the Trump administration's order. The dispute signals a significant escalation in US government willingness to treat frontier AI models as controlled technology subject to export restrictions.
U.S. Government Orders Anthropic to Disable Claude Fable 5 and Mythos 5★ 78
TechCrunch AI45 days agoRegulationTechCrunch reports that the U.S. government ordered Anthropic to immediately disable Claude Fable 5 and Claude Mythos 5 worldwide, citing national security concerns. Anthropic says the order appears tied to a claimed narrow jailbreak of Fable 5, but argues the cited capability is already common in other public models. The move highlights a potential backlash against Anthropic’s safety-first messaging around especially powerful AI systems.
Initial Impressions of Claude Fable 5★ 71
Simon Willison's Weblog48 days agoCommentaryAnthropic released Claude Fable 5 and Claude Mythos 5 simultaneously; Fable 5 matches Mythos 5 in capability but adds strict safety classifiers, with new API fallback mechanisms for rejected requests. Both models offer 1M token context, 128K max output, January 2026 knowledge cutoff, priced at $10/$50 per million tokens — double Opus 4.x. Simon's knowledge-breadth test shows Fable 5 substantially outperforms Opus 4.8, listing dozens of his open-source projects with approximate dates from memory alone.
Anthropic says these topics are too dangerous to let its Fable 5 model talk about
Ars Technica AI48 days agoEthicsAnthropic has announced that its latest frontier model, Fable 5, enforces hard refusals on topics deemed too dangerous, specifically cybersecurity, biology, and chemistry. The move reflects the company's ongoing effort to balance capability with safety as models grow more powerful. For developers and researchers in these fields, the restrictions may limit practical usability in legitimate professional contexts.
Mistral AI partners with NVIDIA to accelerate open frontier models★ 74
Mistral AI News50 days agoBusinessMistral AI announced it is a founding member of the NVIDIA Nemotron Coalition, a global initiative for open frontier foundation models. The partnership combines Mistral AI’s model architecture, training techniques, multimodal capabilities, and enterprise fine-tuning tools with NVIDIA compute, development tools, and synthetic data pipelines. The coalition’s first initiative is a DGX Cloud-trained base model that will support the upcoming NVIDIA Nemotron 4 family and be open-sourced for specialization.
Reality: The Final Eval — Lukas Petersson and Axel Backlund of Andon Labs
Latent Space53 days agoBenchmarkLatent Space talks with Lukas Petersson and Axel Backlund of Andon Labs, the authors behind VendingBench. The episode focuses on evaluating Claude models across a range from Haiku to Mythos. It also discusses how they build frontier evals from scratch, with an emphasis on creating benchmarks that remain useful and meaningful over time.
Google DeepMind 與大韓民國宣布建立合作關係，利用前沿 AI 模型加速科學突破★ 75
Google DeepMind Blog92 days agoBusinessGoogle DeepMind officially announced on April 27, 2026, a strategic partnership with the government of the Republic of Korea (South Korea). The core objective…
Google DeepMind 強化其「前沿安全框架」(Frontier Safety Framework)，以應對先進 AI 模型的嚴重風險★ 75
Google DeepMind Blog277 days agoReleaseGoogle DeepMind has recently announced the strengthening of its Frontier Safety Framework (FSF) — a systematic mechanism designed to proactively identify…