AI turns text from flat to 3D. Our systems are still flat.

AI is good at something most writing tools never were.

Give it messy notes, mixed Chinese and English, half-finished thoughts. It can pull the point across sentences, rebuild the frame, and tell you what is strong and what is weak in the same paragraph. Text stops being a flat line you scroll. It becomes something you can turn and inspect.

That is the part people feel first. It is also not the hard part.

Original post: x.com/Howard_Pengg/status/2116509497137193264

Also posted: Chinese · Threads

What agents are bad at

The personal agents I use are worst at time order.

Ask one to keep a long conversation straight and it starts pruning context. The longer the chat, the more often it drops the wrong thing. Cross-sentence extraction is easy. Chronology is not.

That gap matters once agents stop being demos and start sitting next to real systems: wallets, exchanges, operating systems, anything that has a sequence of actions that cannot be reshuffled.

If an agent cannot hold time, it cannot be trusted with anything that depends on order.

Why the Binance hack made me stop

Binance was hit for about $6.3 billion. The company's own account points to a compromised DevOps laptop, then access into a service container and unauthorized withdrawals. There is no evidence so far of a private-key leak. Roughly $3 billion was frozen. Binance said it would cover losses from SAFU and company funds.

I am not saying the EVM caused that theft. I am saying the episode made me look at the stack we still run.

EVM. Operating systems. The defaults around who can move what, when, and with whose approval. Most of that stack was designed for humans at keyboards, with human reaction times and human review loops.

We are walking into a world where a lot of the actors around those systems will be agents. Volume goes up. Speed goes up. The cost of a wrong hop goes up with both.

Human-first was a design choice

None of this is abstract if you run agents every day.

You see them lose the thread in a long chat. You see them retry. You see them act on a stale summary because the useful detail got pruned ten turns ago.

Now put that same class of actor next to money movement, deploy pipelines, or any shared machine. The old assumption was that a person would notice. That assumption is the design. It is also the vulnerability.

So the question I keep coming back to is simple:

Does everything need a redesign for an agent-surrounded era, or do we keep bolting agent wrappers onto human-first systems and hope review catches up?

I think the second path runs out. The first one is expensive. Expensive is still cheaper than pretending chronology and authority models built for people will hold when the callers are not people.

What I am watching

• Whether agent products treat time order as a first-class problem, not a context-window footnote.
• Whether money and infra stacks add agent-aware controls (scoped authority, sequenced approvals, kill switches that match machine speed) instead of another chatbot on top.
• Whether "human in the loop" stays a slogan or becomes a real gate with a latency budget that matches how fast agents move.

AI already made text three-dimensional. The systems underneath are still drawn for a flat, human-paced world. That mismatch is the story, not the model demo.