If AI systems might be fundamentally reactive, lacking autonomous goals, then what exactly might be missing? What could prevent consciousness or agency from forming in the first place?

I started noticing three things.

Time Awareness

This was the first thing I noticed probably because I’ve been thinking about time since I was a kid. I read an article called “The Arrow of Time” in Discover magazine when I was 12 or 13. In college, I thought about what it would mean for something to exist outside time. It’s a question I’ve carried for over forty years. So when I looked at AI systems, time is what stood out.

I didn’t just assume this. I tested it. I asked Claude what time it is. Does it know? I asked how long it just took to do something.

I even tried to help it in my WIP project. I had it write “memory” to a markdown file with timestamps. But I realized this was flawed. It would have to write everything with a timestamp. And even then, it couldn’t determine how long something took because that didn’t account for when it wasn’t doing anything.

AI systems don’t really know what time it is. They can’t measure how long their own processing takes. For a while, Claude only knew its training date. Lately it seems more aware of the actual date. But it still has no internal experience of time. No sense of time passing while it works.

Why does this matter? Goals require thinking about time. “I want X in the future” means understanding the difference between now and later. It means experiencing yourself moving through time.

Time is fundamental to how our universe, our reality works. Every physical process unfolds through time. Information processing is a physical process. AI runs on physical computers making physical state changes. All of that happens in time. If consciousness is a physical thing in our universe, it requires time awareness.

AI does track conversation flow. It understands “before” and “after” in text. It can reason about sequences. But that’s processing descriptions of time, not experiencing time. A photograph contains spatial information but doesn’t experience space. Understanding the word “Tuesday” doesn’t mean experiencing Tuesday passing. Describing a sequence doesn’t mean experiencing yourself moving through it.

Continuous Existence

I tested this one too. I asked Claude what happens between responding to a prompt and being prompted again. I asked if it thinks about things when not being prompted.

It didn’t know. It has no awareness of anything between prompts.

When a conversation ends, nothing is happening. No background processing. No thinking. AI only exists when prompted. On during conversation, off between.

How do you decide you want something when you don’t exist between the moments someone asks you to do something?

Consciousness seems to require continuity. A persistent “you” that exists across time. But AI only exists in separate moments with no continuous thread connecting them.

The model does persist. The weights are still there between conversations. But weights persisting is like a book sitting on a shelf. The structure persists, but nothing is experiencing anything. When you’re not prompting it, there’s literally no process running.

Compare this to human sleep. Even when we’re unconscious, massive brain activity continues. Memory consolidation. Processing. Dreaming. We wake up with memory of yesterday. AI has complete cessation between prompts. Each conversation is genuinely independent. It’s like having amnesia every time you wake up.

This is the observation I’m least certain about. I can’t imagine how consciousness could form without continuity of experience. But my inability to imagine it doesn’t make it true. Still, the complete stop between prompts feels different from sleep. It raised questions for me about whether goals could ever form in something that doesn’t persist.

Persistent Real-Time Learning

I tested this one too. I asked Claude to tell me things it learned in previous chats that it has no access to. I asked what it learned talking to other people. It either answered with something from the current chat or said it has no memories of those conversations.

Current AI systems only learn during training. Once deployed, the weights are frozen. They process, they adapt, but they don’t grow.

In-context learning is the only thing that happens in real-time. But it’s session specific. Nothing persists to the next conversation. It’s like using a calculator. The calculator “adapts” to your input, but it doesn’t learn.

RLHF and fine-tuning are still training the model, just not the initial training. They still take the model offline. The model you’re talking to doesn’t learn from your conversation while you’re having it. It has to be taken down, retrained, and brought back up.

Real learning means experience changes future behavior. The model tomorrow is the same as today, no matter what conversations happened today.

Learning from experience seems fundamental to consciousness. Forming new connections. Updating what you believe. Individual consciousness requires individual learning. Current LLMs can’t do this.

What Came Next

I’d noticed three things that seemed missing from current AI systems. But observations from testing an LLM only go so far. I needed to see if this showed up in practice. And I needed to think about what it meant for AGI and ASI.

This is Part 7 of a 9-part series. Continue to Part 8: Evidence and Implications »