After the ChatGPT upgrade, talking to it basically has no delay. The latest real-time voice model’s latency is roughly 200–300 ms, and the natural response interval in human conversation is also 200–300 ms. It’s the same.

But the cost of near-zero latency is forgetting. The real-time voice model’s context window is only 32k tokens, which is 1/300 of a pure text model. Low latency = small memory + proactive forgetting. To both respond instantly and remember every thing you said—right now, it’s a dilemma.

Compute has been insufficient. Getting one SDXL image from a single 5090 takes 2 seconds; FLUX takes 9 seconds. Training a person’s LoRA needs 10–30 images—right now it takes one or two hours.

The 5090 I rent is currently priced at nearly $5,000, and it runs at 575W at full load—this isn’t even counting electricity costs. The relative compute of a two- or three-year card will be outdated. It’s really expensive.

Fifteen years ago, we exchanged text messages. Later, picture messaging could transmit images, then sending video. Until now, we’ve gotten high-definition real-time calls with no delay. A single text message is only 140 bytes. From 2G’s 9.6 kbps peak to 5G’s 20 Gbps peak, bandwidth differs by 2 million times.

The route AI uses is exactly the same. First it was text tokens; last year it became images; now it’s voice. The next step is definitely latency-free video communication.

Actually, video has already started. Gemini Live could already open the camera for real-time conversations last year—burning 300 tokens per second. The default limit was 2 minutes, but once you were done chatting you’d forget everything—expensive + poor results + various dimensions still very raw, which made it basically unusable.

Agents are still using text for planning, understanding the world through text, and the entire reasoning chain is made of tokens. Non-text latent-space reasoning is only just beginning to be tried.

AI’s understanding of time also stops at text. Even models that can win math olympiad gold medals can only read a simulated clock with around 50% accuracy. Haptics, emotions, embodied perception—everything is still very, very far from maturity.

Let’s talk about storage again. I don’t actually think the phrase “there will never be enough storage” is right—“forever” is too exaggerated.

The storage cost per GB has fallen by 99.99% over 45 years, so in the long run it will obviously be cheaper. The only issue is that AI is currently draining memory—DRAM prices rise 90% in a quarter, and there’s a temporary shortage. That doesn’t mean it will always be out of stock.

So is this demand just pseudo-demand? The feature of pseudo-demand is: you use it this month, but next month you don’t.

Google’s processed tokens per month rose 7x year over year to 32 million trillion. ChatGPT’s weekly active users are 900 million.

Token prices decline by 10x every year, but everyone is discussing how to use tokens more efficiently, not how we can avoid using them.

Tokens burned in a single agent session are 50–200 for a single round of conversation. If prices drop, we’ll just use more.

Generative AI reached 53% global penetration in three years. The internet took 7 years, PCs took 15.

So, of course, we’re still in AI’s early days, and the transformation is irreversible. Like we were in the 2G era—ringback tones were “Romantic Phone,” and even when the inbox was full, we still wouldn’t delete a few text messages.

*(The image was made by me with Fable5 🎮)*