How is DeepSeek so good, what makes it better than others?

I get that language models need a lot of work to train, with big teams handling the data and making sure everything runs smoothly.

So how is DeepSeek able to compete with big US companies that have way more money and resources? Even if they found a more efficient way to do things, I don’t get how they pulled this off.

What makes you think they don’t have the resources?

They’re backed by a trading company that pulls in billions.

Grayden said:
What makes you think they don’t have the resources?

They’re backed by a trading company that pulls in billions.

Because they only spent $6 million on it while OpenAI spent $100 million.

@DanBurn
They spent six, but that doesn’t mean they only had six.

@DanBurn
Training costs and data costs aren’t the same. Data can be reused across different models.

Grayden said:
What makes you think they don’t have the resources?

They’re backed by a trading company that pulls in billions.

What are you comparing exactly?

Grayden said:
What makes you think they don’t have the resources?

They’re backed by a trading company that pulls in billions.

I thought they were a much smaller company.

Dru said:

Grayden said:
What makes you think they don’t have the resources?

They’re backed by a trading company that pulls in billions.

I thought they were a much smaller company.

They are not small. Check out their background, they have serious funding behind them.

China has been putting out strong models since early 2023. I remember when DeepSeek Coder dropped, and it was already ahead of what we had at the time.

They have really good data filtering and training methods, so I figured they’d take over the open-source space eventually. I just didn’t expect them to catch up to the closed-source giants this fast.

So yeah, I’d say their edge is in how they handle and refine data. Now they’re getting a ton of real user input just like OpenAI did, and that’s only going to make them even better.

It’s good, but their website is always overloaded. I keep getting ‘DeepSeek is experiencing high traffic, check back later.’

Alden said:
It’s good, but their website is always overloaded. I keep getting ‘DeepSeek is experiencing high traffic, check back later.’

Same here. It was taking forever to load, so I got bored and opened another tab. Came back and saw that message.

A lot of research is public, so new models can just build on what’s already out there.

If they’re this good, should we be worried about data privacy?

Teo said:
If they’re this good, should we be worried about data privacy?

Depends on how they handle user data, but yeah, that’s always a concern.

Been using it for a while now, and I think it’s great because they trained it well on quality data.

My only worry is that they might start locking the best features behind a paywall later.

Johnstone1 said:
Been using it for a while now, and I think it’s great because they trained it well on quality data.

My only worry is that they might start locking the best features behind a paywall later.

Doubt it. They open-sourced a lot of their stuff, so they’d get backlash if they suddenly paywalled key features.

Interesting, I might try it out.

They have some really skilled engineers.

Baylen said:
They have some really skilled engineers.

And they’re really good at reverse engineering too.

Raine said:

Baylen said:
They have some really skilled engineers.

And they’re really good at reverse engineering too.

Reverse engineering what exactly? They built their own model from scratch.