GPT-5.6 Sol tops ARC-AGI-3, Mercor $20B, Anthropic hires Bernanke

OpenAI’s GPT‑5.6 Sol broke the ARC‑AGI‑3 leaderboard, becoming the first model to win a public game (87% success on FT09). It excels at reading unfamiliar scenes and replanning, outperforming peers by correctly orienting itself before acting. The result signals a shift toward LLMs that can navigate new environments before generating code.

Flex-Forcing lets a single video diffusion model switch between bidirectional and autoregressive generation by chunking time and denoising steps. This yields up to faster inference and better long‑range coherence, closing the gap between quality and speed that has limited video generation. Researchers can now adapt generation to hardware budgets without retraining.

The Gradient essay argues that coupling vision, audio and text won’t yield human‑level AGI because true intelligence must be grounded in a physical world model and sensor‑motor interaction. Scaling large multimodal models merely stitches heuristics together, while embodiment‑first approaches could produce the reasoning, planning and coordination AGI needs.

A character‑level transformer trained on Nietzsche reveals a single attention head that behaves like a copying mechanism, identified via eigenvalue analysis of its OV circuit. The author’s initial guesses about the head handling quotation or parenthesis grammar were disproved, highlighting how mechanistic interpretability can overturn its own hypotheses.

Mercor is reportedly negotiating a new round that would value the AI‑training startup at $20 billion, double its October valuation. The deal would underscore continued investor appetite for specialized data pipelines that power models at OpenAI, Anthropic and Meta, and could accelerate the race for industry‑specific AI expertise.

Anthropic added former Fed Chair Ben Bernanke to its Long-Term Benefit Trust, the independent board that vets the company’s AI agenda. Bernanke’s macro‑economic expertise is meant to steer Anthropic’s impact on workforces and economies, tightening accountability as AI scales. The move signals a push for institutional safeguards around powerful models.

The New York Times and Daily News say OpenAI lied about its ability to search training data and concealed a 78 million‑record internal log database while deleting billions of ChatGPT conversations. The plaintiffs ask a judge to sanction the firm and block its heavily redacted 20 million‑log sample from evidence. If true, the move could force stricter discovery rules for AI developers.

Fake media detectors lose the arms race; the post argues shifting to default distrust of audiovisual content unless cryptographically signed multimodal provenance can be verified, raising spoofing costs dramatically. This provenance‑first approach, paired with platform policy and legal liability for untagged generated media, could restore trust in visual evidence.

The post argues that tracking AI capability progress, exemplified by the METR time‑horizon graph, is a weak lever for reducing existential risk. It shows how such metrics primarily feed investors and hype, not safety decisions, and calls for a broader re‑weighting of AI‑safety work across organizations.