Model Strategies

• Thinking Machines, a lab linked to former OpenAI CTO Mira Murati, released "Inkling" with 975B total parameters and 41B active parameters, prioritizing a fully open-weight, natively multimodal architecture over SOTA status.1:07
• The company's "Tinker API" supports closed-loop fine-tuning, allowing users to generate synthetic data and customize the model locally; this orchestration is presented as a primary differentiator against closed-model providers.4:45
• Meta's Muse Spark 1.1 is positioned as an agent-friendly, cost-efficient engine featuring a 1 million token context window, targeting enterprise-scale orchestration and delegation rather than pure intelligence jumps.11:19

Progress & Interpretability

• OpenAI's GPT 5.6 "Sol(s)" achieved an 8% score on the ARC AGI 3 benchmark, rising from ~0% in earlier versions; speakers describe this as meaningful progress but remain wary of labeling such results as AGI.19:59
• Anthropic’s recent research uses Jacobian space methods to map internal activations, offering a way to inspect model reasoning in real-time, which Google DeepMind researcher Neil Nanda has reportedly replicated on open-source models.29:08
• Panelists caution that while this interpretability work provides a handle for detecting deception or hallucinations, the associated marketing framing around "consciousness" or "subconsciousness" lacks technical evidence.30:05