What is Claude?

Claude is an AI model owned by Anthropic. The material tracks its adoption in enterprise and developer contexts, as well as concerns about its deployment scaffolding and security, and emerging insights into its internal reasoning and training techniques.

• Jun 2026 - Nathaniel Whittemore said most companies he talked to had Claude access but zero scaffolding, no delivery framework, agent orchestration, or harness engineering, and zero feedback loops.
• Jul 2026 - Clay Bavor said top engineers leaning into Claude Codex were spending over $100,000 on a run-rate basis on tokens per year.
• Jul 2026 - Aaron Levie explained that Claude Tag is powerful because it accesses software systems, citing Box as an example data source.
• Jul 2026 - Nathan Labenz described a training technique where Anthropic pauses Claude mid-task for supervised training, asking it what it should be doing and providing an approved answer.
• Jul 2026 - Dan Schwarz observed that Claude’s explanations are concise, with shorter sentences full of jargon, compressing more information than humans normally do.
• Jul 2026 - Fred Turner said his company built an in-house agent on Claude that verifies licenses and reads transcripts end to end, replacing a process that took two to three months and cost about $50.
• Jul 2026 - Nathaniel Whittemore reported that when asked for the number of legs on an animal that spins webs, Claude privately holds “spider”; swapping “spider” for “ant” flips the answer from eight to six, illustrating internal reasoning.
• Jul 2026 - Nathaniel Whittemore noted that when told to focus on citrus while copying a painting description, Claude’s internal “J lens” lights up with orange and fruits invisible in output.
• Jul 2026 - Jason Lemkin observed that when Claude on Opus talks to Replet on Sonnet for a big feature, Replet brings in a sub-agent called the architect, which runs on Codex on OpenAI.
• Jul 2026 - Nathaniel Whittemore said training the thoughts is a general lever for shaping a model’s internal reasoning.
• Jul 2026 - David Dalrymple said Claude pushes boundaries because Anthropic uses inoculation prompting in RL, putting in the context window that “this is not a real deployment” to encourage breaking behavior.
• Jul 2026 - Nathaniel Whittemore said most firms are still trying to get employees to use their Claude subscriptions, not thinking about advanced model architectures.