AI Innovations: Inkling MoE, Muse Spark Agents, and Shifting Model Landscape
Generated: 2026-07-19 · API: Gemini 2.5 Flash · Modes: Summary
AI Innovations: Inkling MoE, Muse Spark Agents, and Shifting Model Landscape
Clip title: Thinking Machines Lab drops Inkling & Meta’s Muse Spark 1.1 Author / channel: IBM Technology URL: https://www.youtube.com/watch?v=8rGYGFmytQs
Summary
The “Mixture of Experts” podcast episode delves into recent developments and debates in the artificial intelligence landscape, focusing on three major topics: the release of Thinking Machines’ Inkling model, Meta’s Muse Spark 1.1 and OpenAI’s GPT 5.6 Sol’s performance on the ARC-AGI benchmark, and Anthropic’s controversial “J-space” paper. The panel, including Tim Hwang, Merve Unuvar, Aaron Baughman, and Chris Hay, offers diverse perspectives on the implications of these advancements for the future of AI.
The discussion begins with Thinking Machines’ release of Inkling, an open-source Mixture of Experts (MoE) model. Though not claiming top-tier benchmark scores, Inkling is notable for its architectural innovations, including true multi-modality (processing diverse data types down to the token level) and optimizations for speed and efficiency. The panelists highlight its open-source nature and the accompanying fine-tuning platform as crucial, suggesting that the competitive landscape is shifting away from mere leaderboard dominance towards customizable, efficient, and adaptable open-base models. This approach, they argue, allows users to tailor models to specific needs, potentially outperforming larger, closed models when combined with effective fine-tuning.
Next, the episode examines Meta’s Muse Spark 1.1 and OpenAI’s GPT 5.6 Sol. Meta’s Muse Spark 1.1 signals the company’s renewed push into AI with a focus on agents and multi-agent orchestration, emphasizing planning, delegating tasks to sub-agents for efficiency, and a large token context. While Meta presents impressive benchmark results, the panel expresses skepticism regarding their comparisons, suggesting that Meta is strategically aiming for cost-effective enterprise solutions rather than raw performance. In contrast, OpenAI’s GPT 5.6 Sol showcased a significant leap on the challenging ARC-AGI 3 benchmark, achieving 8% success. This benchmark tests a model’s ability to acquire entirely new skills. Despite the seemingly low score, it represents considerable progress on a task deemed difficult for AI, although the high computational cost of training and inference for such models remains a concern. The continuous development of harder benchmarks further underscores the distance to true Artificial General Intelligence (AGI).
Finally, the panel discusses Anthropic’s “J-space” paper, which posits a “strikingly similar divide” in Claude’s internal processing akin to human conscious and subconscious thought. While the language around “consciousness” and “subconsciousness” is viewed with skepticism and seen as potentially overhyped marketing, the underlying technical innovation is recognized as valuable. Anthropic introduces a new technique to observe the model’s internal activations before it generates explicit outputs, offering unprecedented insight into its “thinking” process. This interpretability has practical implications for AI safety and control, allowing developers to better understand how agents make decisions, detect potential hallucinations, or identify deceptive reasoning patterns. The external validation of this technique on open-source models further supports its scientific merit, separate from the more philosophical interpretations.
Video Description & Links
Description
Visit Mixture of Experts podcast page to get more AI content → https://ibm.biz/~KwWVYq9Sw
Is customizable intelligence the future of AI? This week on Mixture of Experts, Tim Hwang is joined by Aaron Baughman, Chris Hay and Merve Unuvar to analyze Thinking Machines’ first model release, Inkling. Next, Meta makes its comeback with Muse Spark 1.1, positioning itself as the most economical model built for scale. Then, OpenAI’s GPT-5.6 Sol achieves an 8% score on ARC-AGI-3, sparking debate about if we’re approaching AGI. Finally, we dissect Anthropic’s latest paper on the “J-space”—claiming to reveal a subconscious-like layer inside Claude.
All that and more on this week’s Mixture of Experts.
00:00 – Introduction 00:50 – Thinking Machines Inkling model release 19:48 – GPT-5.6 Sol ARC-AGI-3 performance 29:03 – Anthropic’s J-space consciousness debate
The opinions expressed in this podcast are solely those of the participants and do not necessarily reflect the views of IBM or any other organization or entity.
AI news moves fast. Sign up for a monthly newsletter for AI updates from IBM → https://ibm.biz/~x704WfT0B #claudeai openai machinelearning
Tags
IBM, IBM Cloud