This week: Anthropic went to extraordinary lengths to turn millions of books into AI training data. We look at what that reveals about the value of institutional knowledge, and how AI actually uses information it was never trained on.

AI in Action

Anthropic’s very expensive reading habit [AI Training/Data]

Anthropic bought millions of physical books, cut off their bindings, scanned every page, and discarded the paper copies as part of an effort to build a massive digital library for Claude. The operation, known as Project Panama, began in 2024 and sat alongside a legally riskier effort in which Anthropic had downloaded millions of books from pirate libraries. A federal judge later ruled that converting purchased print books into internal digital copies qualified as fair use under the circumstances, while the pirated library created separate copyright liability. For an industry usually measured in GPUs and data centers, this was a very old-fashioned way to acquire an input.

Books offer something the open web often does not: long stretches of edited, coherent human thought. A technical book can compress years of expertise into one source, while a novel or history can carry context across hundreds of pages. For a model, that makes books useful material for learning how ideas develop, how information connects, and how language holds together over longer spans. Project Panama suggests Anthropic considered that kind of material valuable enough to build a physical acquisition pipeline around it.

The same logic becomes more interesting once companies start applying AI to their own internal knowledge. Take two insurers using the same frontier model. One connects it to policy manuals and public information. The other adds twenty years of claims, underwriting decisions, unusual exceptions, expert judgment, and what happened afterward. The second insurer has a record of its own judgment, including which calls worked and which ones did not. Companies have been accumulating this kind of institutional memory for years through customer files, research archives, compliance systems, project records, and internal notes. AI can make some of that history useful again by bringing past decisions and outcomes into future work.

A competitor can sign up for the same model tomorrow. Recreating twenty years of decisions, mistakes, exceptions, and outcomes is harder. The challenge is making that experience usable without turning every forgotten folder into training material or locking it inside one provider. A SharePoint archive stuffed with “final_v7_reallyfinal” files will not suddenly become strategic. A well-kept record of how your company actually makes decisions might. Anthropic spent millions building a library its models could learn from. Plenty of companies already have one of their own. They just never thought of it that way.

Bits of Brilliance

How does AI learn your business? [Enterprise AI/Data]

Most people assume an AI has to be trained on company information before it can use it. A lot of enterprise AI works differently. The model arrives with general knowledge, then gets access to company information when a task calls for it, whether that is a policy, customer record, contract, or previous case. Training shapes what the model learned beforehand. Context is the information available while it works. This is especially useful inside companies, where policies change, customer records get updated, and access permissions can disappear overnight.

A new employee works in roughly the same way. They bring knowledge from outside the company, then rely on internal systems and records to do the job properly. RAG, short for retrieval-augmented generation, is one way AI systems can search approved sources and bring relevant information into the task. An insurance AI might start with a customer’s policy and claims history, then look up a previous case when something unusual appears. Retrieving information sounds straightforward until the wrong policy version, customer record, or supposedly similar case gets pulled in. The model can reason carefully and still reach the wrong answer because it started with the wrong evidence.

As these systems take on longer tasks, the information in front of them can keep changing. People decide which systems are available, which sources count as authoritative, what stays off-limits, and where human review is required. Within those boundaries, the AI can increasingly decide what it needs to retrieve next. This gives two companies using the same foundation model plenty of room to build very different systems. One may have years of useful knowledge buried across disconnected folders; another may have made that knowledge easy to find, properly permissioned, and reliable enough to use in real work.

The model name only tells you so much about an enterprise AI system. What can it see? Where does it look when it needs something? Which sources does it trust? Who decided those rules? A capable model can still make a bad call when the wrong evidence lands in front of it. Companies have spent decades trying to get the right information to the right person at the right time. Increasingly, they will have to do the same for machines.

_{Byte-Sized Intelligence is a personal newsletter created for educational and informational purposes only. The content reflects the personal views of the author and does not represent the opinions of any employer or affiliated organization. This publication does not offer financial, investment, legal, or professional advice. Any references to tools, technologies, or companies are for illustrative purposes only and do not constitute endorsements. Readers should independently verify any information before acting on it. All AI-generated content or tool usage should be approached critically. Always apply human judgment and discretion when using or interpreting AI outputs.}