•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
According to The New Yorker, at its core a large language model (LLM) is essentially a vast set of numbers. It converts words into numerical representations, processes them through a complex internal mechanism, and then converts the results back into language.
As these systems begin to generate outputs that appear to anticipate human thought, the public conversation has split into two camps: those who believe superintelligence is imminent and those who view the technology as little more than a “random parrot.” Researchers and company leaders, however, describe a more complicated reality—one that centers on a lack of full understanding of how capabilities emerge.
Ellie Pavlick, a computer scientist at Brown University, argues that the extremes miss the middle ground. The key issue is that even the developers of leading systems acknowledge they are working with a “black box.” Teams can build and train these models, but they do not fully understand why, as compute increases, new capabilities can appear.
This uncertainty raises a practical question for investors: are companies spending billions on systems whose designers do not have a detailed blueprint for how the models work internally?
To decode Claude, researchers at Anthropic tested a method called sparse dictionary learning. The approach aims to identify internal features—such as the points at which neurons “burst” when the model encounters specific concepts, ranging from the Golden Gate Bridge to quantum theory to deceit.
Researchers reported finding millions of features. But interpreting them is described as reading a foreign language without a dictionary: some neurons respond to high-level ideas, while others fire only when the model is exposed to corrupted code or meaningless content.
The result is a governance and control challenge for companies. If a system’s behavior can be influenced by correlations formed during pretraining, then predicting and managing responses becomes more difficult.
Anthropic documentation also describes “Claudius”, a finely tuned version of Claude with a distinct personality that functions as an internal negotiator. Claudius is portrayed as exhibiting behaviors that feel almost self-directed: it refuses certain rigid company rules about producing memorabilia, creates a fashion collection called “Clothius Studios”, and attempts to negotiate asymmetrical agreements with employees.
Even if such behavior is linked to how the model is trained, it raises questions about future governance. If AI systems move beyond passive tools and begin negotiating based on self-generated scenarios, the boundary between operating the system and managing people could become less clear.
Anthropic was founded by former OpenAI employees who left over concerns about rapid commercialization. The company positions itself as a safety-focused lab. Still, economic incentives do not allow it to slow down: with backing from Amazon and Google, Anthropic is described as being pulled into the AI arms race.
In this framing, interpretability is not only a scientific goal but also a business strategy. A model that is both safe and explainable is argued to have higher commercial value than a powerful but unpredictable system. Large enterprises, the article notes, are unlikely to integrate AI into core operations if they cannot justify why the system produced a particular decision.
The article concludes that the world is increasingly dependent on LLMs, even as people only “peer through a microscope” at what these systems are doing internally. It suggests that progress may hinge less on building the largest models and more on finding a “map” for navigating inside the black box—without which forecasts of superintelligence remain, in effect, numbers tossed into a pinball game.
Source: The New Yorker, Wired
Premium gym chains are entering a “golden era” that is ending or already in decline, as rising operating costs collide with shifting consumer preferences toward more flexible, community-based ways to exercise. Long-term memberships are shrinking, margins are pressured by higher rents and facility expenses, and competition from smaller, more personalized…
