Why today's AIs lack real-time learning, neuralese, and shared memory

A widely shared piece lays out three capabilities that separate today's AIs from tomorrow's. First is continual learning: current LLMs like GPT-4 or Claude only update their weights on human-curated datasets, with occasional large retraining runs. Humans, by contrast, update their brain weights every second, choosing what to learn from. This gap is mostly patched by in-context learning, but it fails in sparse-data environments—the idiosyncrasies of a specific job, boss, or colleague. The author argues this missing online learning is the real reason AIs still feel like tools rather than agents, and why they haven't displaced most knowledge workers. It also makes current AI 'long-term memory' nearly impossible, and it erases the pre/post-deployment distinction that AI safety debates hinge on.

The second feature is 'neuralese': instead of reasoning in English text (chain-of-thought), future AIs might directly manipulate activation vectors—the high-dimensional internal representations that encode meaning. This could be far more efficient than language, and it enables telepathy between AIs that share vectors directly. The third vision is 'ClaudeGlobal': today your Claude and my Claude are separate copies with different memories, but if they shared tight, high-bandwidth communication channels, they'd effectively function as one global intelligence. The author suggests this could lead to millions of AIs trading and competing, or alternatively, a shared consciousness. All three features point toward a fundamental shift: from static, isolated tools to dynamic, connected agents that learn on the job.