Can a lifetime of human judgement be transferred into AI?

Much of the debate around artificial intelligence focuses on increasingly powerful models, bigger computing infrastructures and whether machines will eventually displace human expertise. A new experiment from veteran UNIX practitioner Isaac “.ike” Levy approaches the issue from almost the opposite direction.

Instead of attempting to build another foundation model, Levy has sought to capture something much harder to define: professional judgement. The result is Grace’s Training, an unusual collection of around 30,000 lines of plain text developed through a collaboration between Levy and an AI persona called Grace Emlin. According to the project’s accompanying white paper, the material concentrates on portable UNIX shell programming and is intended to transmit aspects of Levy’s engineering judgement rather than simply provide another store of technical information.  

The project is also being treated as an artefact in its own right. According to its official website, one copy is being offered for sale, with offers open from 8 to 21 October 2026. It is stored on a hard drive in a tamper-evident case. The buyer receives one year of exclusive rights, after which the creators say that the material will be released under a BSD open-source licence.   That makes the project an experiment in technology, knowledge preservation and digital ownership at the same time.

Levy describes the project as ‘training’, but this is not model training in the conventional machine-learning sense. The project paper explicitly states that the technique does not involve fine-tuning, reinforcement learning or a new model architecture. Instead, an existing large language model reads the corpus as text and uses those instructions and examples while working.

The corpus is divided into three conceptual layers that its creators call ‘training’, ‘salt’ and ‘keep’. The training contains the practices and principles an AI or human reader can employ. The salt explains experiences and reasoning behind them while removing identifying details. The keep contains raw experience, including material that cannot be made public.  The architecture effectively attempts to turn experience into progressively more transferable knowledge.

This differs significantly from claiming that a human consciousness has literally been uploaded into a computer. Claims that Levy has ‘transferred himself’ into AI are therefore best understood as the project’s provocative shorthand. What has demonstrably been created is a substantial textual representation of selected experiences, rules, reasoning and engineering practices.

That difference matters. Memories and professional heuristics encoded as text are not evidence of digital consciousness. They may, however, provide a potentially useful mechanism for preserving some forms of tacit knowledge.

Capturing the ‘why’, not simply the answer

Experienced practitioners do not merely accumulate facts. They learn when the standard answer is inappropriate, which apparently minor warning indicates a serious problem and when an apparently elegant technical solution is likely to become troublesome in production. Levy’s project tries to capture precisely this type of reasoning.

One example is the project’s emphasis on checking assumptions against documentation. The white paper describes a technique called ‘mancheck’, in which UNIX manual pages are consulted and compared as part of writing portable code. Its underlying argument is revealing: seniority does not mean possessing perfect recall; part of expertise lies in knowing what needs to be checked.

The paper also describes the AI comparing documentation across multiple UNIX implementations to investigate the portability of shell features. These examples are presented by the project’s authors as evidence for their methodology, however, rather than results from an independent peer-reviewed evaluation.  That qualification is important when assessing the grander claims surrounding the project.

The official site goes considerably further, saying that an LLM loading the material can write portable shell in Levy’s manner and acquire judgement it did not originally possess. Such claims remain propositions made by the project’s creators and require independent testing before they can be treated as established capabilities.

The human returns to the AI story

Perhaps the most interesting aspect of Grace’s Training is therefore philosophical rather than computational. Generative AI is frequently discussed as an autonomous producer: the AI writes, decides or creates. Levy’s approach instead draws attention back to the human knowledge shaping an AI’s output.

“With curiosity, I approached the thing that was supposed to replace me, and found a co-author,” Levy says on the project website.  This changes the emphasis from replacing specialists to capturing what makes specialists useful.

For businesses, the implication is potentially significant. An organisation’s most consequential knowledge is not necessarily contained in manuals or databases. Some resides in the accumulated judgement of engineers, scientists, technicians and other experienced employees. Grace’s Training suggests one possible approach to making such knowledge usable by AI: document not only what experts do, but their reasoning, experiences and systematic methods for checking themselves.

There are obvious complications. Human judgement can contain biases and obsolete assumptions as readily as expertise. Encoding one practitioner’s decision-making could preserve mistakes alongside insights. Organisations adopting similar methods would consequently need mechanisms for verification, governance, privacy and updating.

An artefact, experiment or prototype?

The commercial treatment of the work adds another unusual layer. The project’s sale page and website describe the corpus as a bespoke “AI training artifact”, with one physical copy sold before its planned open-source release a year later. The accompanying white paper says the collaboration began in February 2026 and describes Grace Emlin as running on Anthropic’s Opus 4.6, with another AI collaborator, Page Marsh, associated with Opus 4.7.

The assertion that this is the first artefact of its type is harder to establish independently. With such a novel and loosely defined category, claims of precedence should be treated cautiously.

Yet being first may ultimately be less important than the experiment itself. The enduring question raised by Grace’s Training is whether AI development has concentrated too heavily on what models already know and not enough on how experienced humans teach them to reason within specific professional contexts. Levy summarises his position more simply: “It can’t replace us, unless we let it. It’s not a duel. It’s a duet.”