Thinking Machines Lab drops Inkling & Meta’s Muse Spark 1.1

This episode of Mixture of Experts delves into several significant developments in the AI landscape, starting with Thinking Machines’ Inkling model release, then Meta’s Muse Spark 1.1, followed by OpenAI’s GPT-5.6 Sol performance on ARC-AGI-3, and finally, Anthropic’s controversial paper on the “J-space.”

Thinking Machines’ Inkling Model Release

The podcast begins by discussing Inkling, the first model released by Thinking Machines, a lab founded by former OpenAI CTO Mira Murati. Inkling is a 975B total parameter model, with 41B active parameters, leveraging a Mixture of Experts (MoE) architecture. Unlike many models aiming for top benchmark scores, Inkling is openly acknowledged as not the best from a benchmarking standpoint. Chris Haye praised its open-source nature and innovative architectural choices, particularly its true multimodal capabilities, which process pixels down to token level, and its focus on speed optimizations. Marva Unuvar highlighted its unique fine-tuning platform, Tinker, which allows for automated, closed-loop customization, creating synthetic data and integrating new weights. Aaron Baughman emphasized that Inkling’s core value lies in offering customizable intelligence through open weights, enabling users to adapt and control models for specific needs, rather than solely focusing on a frontier model. The panelists agreed that this approach signals a shift in market competition, moving beyond raw benchmark performance to practical applicability, reproducibility, and architectural novelty.

Meta’s Muse Spark 1.1

Next, the discussion shifts to Meta’s Muse Spark 1.1, released by their Super Intelligence Lab. This model is presented as a strong contender, with Meta claiming benchmark superiority over some proprietary models. Marva noted its focus on agent orchestration, planning, and parallel sub-agents, with a million-token context window and robust coding harnesses, aiming for enterprise use cases due to its cost efficiency and speed. Aaron linked Meta’s AI strategy to its vast consumer platforms (Facebook, WhatsApp, Instagram) and hardware ecosystem, suggesting a goal to integrate intelligence ubiquitously. He also acknowledged Meta’s long-standing contributions to AI (Llama, PyTorch). However, Chris expressed skepticism, calling it a “meh release,” questioning the benchmark comparisons (specifically against older OpenAI models) and lamenting its closed-source nature. He argued that an open-weight release would have generated more community excitement.

OpenAI’s GPT-5.6 Sol ARC-AGI-3 Performance

OpenAI’s GPT-5.6 Sol showed a significant jump from 0% to 8% on the challenging ARC-AGI-3 benchmark, designed to test a model’s ability to learn entirely new skills. This performance reignited the debate around Artificial General Intelligence (AGI).

• Marva was cautious, stating 8% is far from true AGI, especially when compared to a 12-year-old’s 100% capability. She believes AGI will only be evident when models effortlessly handle new, complex tasks out-of-the-box.
• Aaron agreed that Sol is not AGI but noted that this performance indicates the gap between frontier models and AGI is shrinking, potentially turning AGI into an “engineering and scaling problem.” He raised concerns about the immense power consumption required for training (megawatts) and even inference (equivalent to 60 homes for a single job), questioning the cost-effectiveness of current AGI pursuits.
• The panelists debated whether the goalposts for AGI are constantly shifting and if, by earlier definitions, we might already be at an “infant level AGI.” The general consensus was that the definition of AGI remains fluid and crucial to this discussion.

Anthropic’s J-space Consciousness Debate

The final topic centered on Anthropic’s paper introducing the “J-space” in Claude, which the company controversially described as a “subconscious-like layer” where reasoning and processing occur before explicit outputs. Chris viewed the Jacobian technique used by Anthropic as a significant breakthrough in interpretability, allowing researchers to illuminate the “darker space” of internal calculations. While skeptical of the term “consciousness,” he found the observation that Chain of Thought offloads reasoning from the J-space particularly interesting. Aaron largely agreed, calling the consciousness claims “overhype” and “misleading” but praising the paper’s contribution to understanding internal model workings, akin to an fMRI for AI. He highlighted the practical implications for AI safety, such as identifying meaningful features, testing hypotheses, and potentially detecting hallucinations or deceptive reasoning. Marva acknowledged the paper’s controversial nature and potential as marketing but noted external validation from Google DeepMind. She emphasized the practical utility for agent builders, as the J-space offers a “localized, readable, editable internal state”—a crucial “handle” for safety and control by monitoring a model’s internal representations before it acts.

• Tags:
• Agentic AI
• Large Language Models
• AI Safety
• Fine-Tuning
• Enterprise AI
• Multimodal AI
• Interpretability
• AI Benchmarks
• Mixture-of-Experts
• Model Architecture
• Artificial General Intelligence
• Open-Source AI
• ARC-AGI
• Customizable AI
• Model Subconsciousness