PlainAI AI explained plainly

DeepSeek · 2025 · 01 · 20 · Model · ~2 min read

DeepSeek launched R1

A Chinese lab released an AI that thinks before it answers — about as well as OpenAI's best — and you can download it for free. It briefly tanked Nvidia's stock and got everyone asking whether AI is still mostly an American thing.

What's actually new

  • You can download it. Until R1, all the AIs that pause to think were locked behind paid APIs from US labs.
  • It's cheap. DeepSeek says they trained the base for around US$5.6 million — much less than what most American labs spend.
  • It does well on hard problems. Maths, coding, science questions — about as well as OpenAI's o1 on many tests.

If you want more

Worth knowing~30s
  • The cost number is real, but only partly. The US$5.6M is for the final training run. It doesn't count earlier base-model training, salaries, or buildings. The full picture is much bigger — still cheaper than American labs, just not by 100x.
  • It loses some tests. Comparison videos pick the wins. On longer documents and harder reasoning tasks, OpenAI's o1 still wins.
  • 'Open' here means weights only. You can run it. You can't see the data it was trained on or the recipe used to get there.
Who should care~20s

Researchers who want to run a thinking AI on their own hardware. Solo developers tired of paying API bills. Teachers showing students that AI isn't only built in California. Companies modelling what AI will cost them next year.

What to do about it~20s

Try it on OpenRouter or Hugging Face. Don't move important workflows to it yet — running it locally still needs serious GPUs, and the helper tools are younger than the OpenAI versions you might be used to.

Honest take~45s

The benchmarks were never the real story. The real story is that a Chinese lab shipped a thinking AI in the open, loud enough to scare markets. The Nvidia crash was an over-reaction — these models still need GPUs to run. But the bigger picture is real: the leading edge of AI isn't only an American thing anymore, and the open-source side just got a flagship. Whether that's good or bad depends entirely on what people build with it.

Other recent model updates

Sources

Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com