DeepSeek
If anyone’s been following tech this week, they know my prediction from last week came true:
But the real news is DeepSeek — a model developed by an AI company out of China. And the reasons for this are threefold:
DeepSeek is outperforming its American peers in most tasks
(ChatGPT, Llama, Claude)
It’s completely open source
Which means anyone can use it
And anyone can see its source code (even competitors like Google and Meta)
It achieved this new level of “performance” in a very interesting way
And this is what I want to focus on
In the research paper that accompanies the DeepSeek release, they explain that what enables the model’s unprecedented ability to “reason” is a greater emphasis on something called ‘reinforcement learning’. What this means:
All other models = supervised fine tuning
exorbitant amounts of data
lots of time to train
effectively "telling" the model what is right and wrong
(which gets regurgitated to the user when asking questions)
DeepSeek = reinforcement learning
less "telling the model what is right/wrong”
and more “letting it think for itself”
Consider these anecdotes from the paper:
“This improvement is not the result of external adjustments but rather an intrinsic development within the model”
”DeepSeek-R1-Zero naturally acquires the ability to solve increasingly complex reasoning tasks by leveraging extended test-time computation…allowing the model to explore and refine its thought processes in greater depth.”“[E]mergence of sophisticated behaviors as the test-time computation increases. Behaviors such as reflection—where the model revisits and reevaluates its previous steps—and the exploration of alternative approaches to problem-solving arise spontaneously.”
What this means is that DeepSeek is simply better at solving problems than its peers. And not only that, but they’ve also shown that this new level of “performance” can be achieved using significantly less data than other models. Which means less hardware. (And goes back to the point I was trying to make in my last Substack — that these models are getting more efficient, and inherently less dependent on Nvidia GPUs).
Lots to think about. Many implications. Here’s a few on my mind:
Reflecting on mistakes enables enhanced “reasoning”
Making DeepSeek open source is huge. They don’t care if competitors adopt their strategy
DeepSeek illustrates that human intellect isn’t purely “memory”, but rather contains some semblance of “problem solving”. Seems related to subconscious thought. Same argument made by Cormac McCarthy here [LINK]
Shared theme across Google’s Titan and DeepSeek — the notion of “working” memory. Taking time to enhance a skill and then iterating on it.
Complex problem —> Reason —> Reflect —> Find Fault —> Improve
References:
The DeepSeek research paper and source code (surprisingly digestible, unlike the “Titans'“ one out of Google) [LINK]