LLM is the Next JPEG
Let me tell you a story.
In the 1990s, if you wanted to send a photo to someone, you needed to consider many things: file format, compression algorithm, color depth. Each software had its own format. Then JPEG appeared.
Suddenly, no one cared about image formats anymore. JPEG became infrastructure. You wouldn't say "I made a picture with JPEG," you'd just say "I sent a picture."
LLMs are going down the same path.
When Technology Becomes Air
"LLMs are commodities. Apple is happy to buy tokens from LLM companies, but Apple is a company that sells differentiated products." — @deuteronormative
This statement is straightforward. If you're Apple, you don't generate your own electricity, you buy it from the grid. You don't make your own tires, you buy them from Michelin. Now, you won't train your own LLM either, you'll buy tokens from the cloud.
This isn't to say LLMs aren't important. Electricity is important. Tires are important. But they are infrastructure, not differentiating factors.
The Winner of the Cost War
Alibaba just released Qwen 3.5:
- 397 billion parameters, 17 billion activated
- 60% cheaper than Qwen 3
- 8x faster
- Token price is 1/18th of Gemini 3 Pro
This isn't a technological breakthrough, it's a price war. LCD TVs also dropped in price like this back in the day. The first company to break below $1000 wasn't the one with the best technology, but it was the winner.
Pragmatic Advice
What does this mean if you're a developer?
-
Don't train your own models. Unless you're OpenAI, Anthropic, or Alibaba, training models is burning money. Use APIs.
-
Focus on price rather than parameters. 397 billion parameters sounds cool, but your users don't care. They care about response speed and cost.
-
Be prepared to migrate. LLMs are commodities, which means they are replaceable. Use GPT today, Claude tomorrow, and Qwen the day after. Your architecture should support this kind of switching.
Interesting Paradox
The people who understand LLMs the most don't talk about LLMs much.
"Andrej Karpathy wrote a mini GPT in 240 lines of pure Python. No TensorFlow. No PyTorch. Just math. It shows that LLMs are not magic—they are just next token prediction."
When you understand that "next token prediction" is all there is to this technology, a lot of the hype disappears. This isn't a put-down. A microwave is just heating water molecules, but it changed the kitchen.
Next Steps
LLMs will become like JPEG: ubiquitous, no one discusses them, but indispensable.
Until then, smart people will choose the cheapest vendor in the price war. Because when technology is commoditized, the only thing that matters is cost.
