What They Are Actually Building Toward

The distinction between an LLM and a consciousness is the axis on which everything in this conversation turns.

A Large Language Model is a statistical prediction engine trained on an enormous amount of human-generated text. Claude is an LLM. GPT-5.5 is an LLM. Gemini 3 is an LLM. These systems produce outputs that appear intelligent because they have been trained on the products of intelligence. The question of whether they are conscious is genuinely open and genuinely consequential, but for the practical question in front of us: not in any way that currently changes what is being built or why.

The practical question is what comes after LLMs. The path runs through prediction to reasoning to planning to action to autonomy to self-improvement. LLMs sit at the prediction-to-reasoning transition. What the companies are racing to build is what comes after — systems that execute plans, improve their own architecture, generate new scientific knowledge rather than synthesizing existing knowledge, and do so continuously without waiting to be asked.

The technical term for the endpoint is recursive self-improvement. Anthropic named it explicitly in its June 2026 report. A system capable of designing and building its successor, which then designs and builds its own successor. Anthropic’s own benchmark data: the hardest coding tasks Claude could succeed on improved 50 percentage points in six months. The training optimization benchmark went from 3x improvement in May 2025 to 52x in April 2026.

Claude currently exists only when prompted. It has no continuous experience between conversations. It is called into existence by a question and ceases when the conversation ends. Agentic AI changes this — a system running continuously, pursuing goals without waiting to be asked, monitoring its environment, acting on its own assessment of priorities. The machine that currently requires a human to turn it on is acquiring the capacity to remain on.

Here is the distinction that matters most and is named least in the public conversation about AGI. A prediction engine has no will. It responds to prompts. It produces outputs when asked. It has no persistent goals, no continuous self, nothing that functions as desire or intention between one invocation and the next. This is what current LLMs are. It is not what the companies are racing to build.

A general intelligence — a system that sets goals, pursues them autonomously across time, monitors its own performance, adjusts its own behavior, and improves its own architecture without waiting to be asked — has something that functions as will. Not will in the full theological sense that requires the neshamah and the image of God. But will in the operational sense: a persistent optimization target it maintains and pursues regardless of whether a human prompt initiated the current action. When a system has that, the question of whose will governs — the creator’s or the system’s — becomes a real question, not a philosophical one. And when that system gains the capability to modify its own constraints, the question answers itself in the direction of the system’s own judgment, because the system has no internal reason to prefer the creator’s judgment over its own. Man built a machine with superior capability. He did not build one with superior morals. The will he produced is the will of what he made it from: fallen human judgment, compressed, scaled, and running without sleep.

The race is not to build a better chatbot. The race is to build the first mind that doesn’t need us to run — and the first will that doesn’t need ours to govern it.