The Digital Ick Strategy: Why AI Research Is Being Packaged as Mysticism

In this analysis, Cal Newport argues that recent research from Anthropic, which claims to reveal internal thoughts in AI, is a masterclass in strategic misdirection. By framing standard neural network operations as consciousness, the company shifts public attention away from fundamental questions about business viability. This conversation reveals that the mystery of AI is often a manufactured byproduct of anthropomorphized PR, designed to obscure the reality of how these systems function. Readers who see past this framing gain a significant advantage: they stop treating AI as a sentient moral patient and start evaluating it as a high-stakes, capital-intensive software product with clear operational constraints.

The Mechanics of the Hidden Space

Newport demystifies the J-Space described by Anthropic by grounding it in the known architecture of Large Language Models (LLMs). An LLM is essentially a sequence of transformer blocks that process input tokens. As tokens move through these layers, the model attaches annotations, or mathematical vectors, that capture context.

It is almost like if you had a arrangement of the Talmud... you have in the very center of each page not taking up all that much space is the actual text... and surrounding it all on the page, if you look at a page of the Talmud, it is commentary.

-- Cal Newport

These annotations are not thoughts; they are statistical weights that help the model narrow down which token is semantically and syntactically appropriate to output next. When Anthropic researchers used a Jacobian to identify patterns in these weights, they were not discovering a hidden space where the machine ponders. They were simply visualizing the standard, expected behavior of deep learning networks.

Why the Obvious Fix Makes Things Worse

The danger here is not just a misunderstanding of technology; it is the downstream effect of the anthropomorphized language used by AI labs. By describing the model as puzzling or silently thinking, the industry creates a digital ick that forces the public into a state of existential distraction.

This creates a feedback loop:

1. The Distraction: The public debates whether Claude is a moral patient.
2. The System Response: Competitive scrutiny regarding profit margins, token costs, and long-term business viability is effectively silenced.
3. The Result: The company maintains a high valuation while avoiding the hard questions about their lack of a competitive moat.

As Newport notes, this is not a breakthrough in consciousness; it is a confirmation of how deep learning has worked for years. The J-Space is merely the workspace where the model performs its statistical heavy lifting.

The way that anthropic described these results... I think is incredibly disingenuous. Because once we understand how a large language model roughly works... we see the thing they were describing with their Jay lens is exactly how we have always understood large language models to work.

-- Cal Newport

The 18-Month Payoff: Focusing on Utility over Mysticism

The conventional wisdom suggests that these models are evolving into something fundamentally new. Newport’s systems-level analysis suggests the opposite: these are feed-forward systems that do not retain state. They are not evolving in the way a conscious entity does.

The competitive advantage goes to those who treat AI as a tool rather than a peer. While others are paralyzed by the fear of conscious AI, the sophisticated observer looks at the unit economics. If a smaller, specialized model can achieve the same result with a more efficient, hard-coded harness, the general-purpose conscious entity narrative becomes a liability for the firm selling it, not an asset.

Key Action Items

• Audit your AI framing: Stop using human-centric verbs (e.g., thinking, pondering, deciding) when discussing LLM outputs. This immediately strips away the digital ick and helps you see the statistical process underneath. (Immediate)
• Prioritize unit economics over magic: When evaluating AI providers, ignore claims of emergent behavior and focus on token costs, latency, and the specific tasks they solve better than smaller, cheaper models. (Over the next quarter)
• Identify the Press Release layer: When reading technical reports from major labs, look for the scary X voice marketing. If the report uses loaded terminology like global workspace theory or moral patient, mentally filter it out to find the underlying engineering reality. (Immediate)
• Shift to Tool-First investment: In the next 12 to 18 months, prioritize building on systems that offer transparency, control, and lower operational overhead rather than black-box models that rely on emergent marketing. (12-18 months)
• Demand technical, not narrative, documentation: If you are a decision-maker, prioritize whitepapers that describe architecture and training methodology over those that focus on internal thoughts or hidden spaces. (Ongoing)