This week in AI: DeepSeek’s Lunar New Year firecracker

Regular readers know we don’t tend to use our regular Engine blog for actual news; that would make it a newsletter, and you get too many of those already. It's all too easy to confuse fact and opinion in a world where TikTok has become a news source for many, and this medium is firmly in the op-ed space. But things went a little crazy this week so we thought we’d re-cap the main events, and take a look at (a) why you should care and (b) why you don’t need to care.

DeepSeek, R1 and the market impact

DeepSeek caused a bit of a stir when it demonstrated that you could build AI on par with OpenAI's models for a fraction of the cost — around 5% of the cost, in fact. $6 million instead of several billion, although that claim has since been disputed. There were shockwaves:

• NVIDIA's stock plummeted by 17%, resulting in a loss of around $600 billion in market capital — the largest single-day drop in US history.

• NVIDIA's CEO, Jensen Huang, saw his personal net worth decline by around $21 billion. That’s the same as the entire GDP of Guyana. It’s two fewer 590ft superyachts, according to Perplexity. But tbf those come with a lot of ongoing insurance, staffing and docking liabilities, so Jensen might have been eyeing up something more practical, or maybe something with more social value.

• The tech-heavy Nasdaq fell by 3.1%, with many AI-heavy stocks dropping by more than 8.7%.

Ouch. Happy Lunar New Year!

Adding to the drama, DeepSeek's app overtook ChatGPT as the number one free app in Apple's App Store. As if that wasn't enough, DeepSeek also launched Janus Pro 7B, a new vision model capable of understanding and creating images. Meta were reported to be scrambling war rooms of engineers to work out how DeepSeek did it all.

One article I read called this a David and Goliath story, but that seems to me a rather US-centric view, perhaps not fully grasping China’s relative economic capacity (or land mass, technical capability and population size).  

The aftermath and ChatGPT: Timing is everything

Success comes at a price. DeepSeek had to limit new users - temporarily - due to large-scale malicious attacks. Analysts were divided in their reactions, with some declaring the end of an era, while others dismissed the panic as overblown.

Amidst the fallout, OpenAI made some big plays. They announced the upcoming release of the o3-mini, a new model expected to be released in a few weeks. Then, in another big strategic move, OpenAI announced ChatGPT Gov, a version of ChatGPT designed specifically for US government agencies. This announcement came at a crucial time, as concerns about DeepSeek's privacy policies were growing. The White House is looking into the national security implications of DeepSeek, and the US Navy has already banned its use.

ChatGPT Gov aims to position OpenAI as the secure, American-made alternative. This move is seen as OpenAI drawing a line in the sand about AI security and sovereignty.

But there was also controversy when it was revealed that OpenAI had quietly funded EpochAI, the organisation behind a benchmark that o3 had recently excelled in. This raised questions about the transparency and ethics of OpenAI's practices. That makes it trickier to argue that your edge over foreign competition is trust.

Pot. Kettle. Black.

Admist this week of AI beef, the irony was lost on the CEOs taking a hit, but not on anyone else. The possibility of ChatGPT losing its job to an AI deserves at least a wry smile. Open AI seem miffed that DeepSeek has some striking design similarities to ChatGPT. They’re even more miffed that they may have used ‘distillation’ to extract knowledge from Open AI to train their own AI. Memes abound.

Note: I’m not saying Open AI are stealing data necessarily. That’s for the courts to decide in the landmark New York Times (and friends) lawsuit and elsewhere. But it’s hard to expect sympathy when you complain that someone else used the stuff you obtained without permission and didn’t even ask your permission.

What do I need to worry about. Really.

Data security. DeepSeek collects extensive personal data including device information, keystroke patterns and your content. Data is stored on servers in China, subject to Chinese data regulation and government access. It also has a different take on some news and information to other sources. Try asking it about Jensen’s birthplace: Taiwan.

It’s not alone with the reliability problem: ChatGPT has it’s own accuracy woes. Use Perplexity for your free search AI: in my view is it’s still the easiest user-friendly option for verifying sources.

But it’s a timely reminder to check you know where your data is coming from and where it’s going: avoid sharing sensitive information, verify facts before you use them, and understand the implications of where data is stored and who can access it (the government access issue applies in the US, not just China and elsewhere).

You also said I shouldn’t worry – but it feels like I really should

Good point. Here’s the thing:

1. DeepSeek’s arrival on the scene is a solid reminder AI doesn’t have to be a billion-dollar solution

We’ve known this a while – we build solutions priced to work for charities and small businesses as well as big-hitters. Don’t bet the farm on One Big Thing. Explore your options – a lot of that money spent on enterprise solutions is probably going on armies of programmers and sales teams, and swanky offices, and they’re under pressure from big investors to make their money back with interest. You can literally shop local for world class AI.

2. You can keep your data secure: use the right tools for the right job

Private RAG – retrieval augmented generation – GPTs can use all the latest functionality but remain secure to your system. Work out what you want to do and buy an AI assistant to do it – don’t chuck your valuable data into a free online AI because your company hasn’t given you the right tools yet.

3. The supplier, platform or app you pick stands between you and all this chaos – let them deal with it

We currently – mainly – use ChatGPT as our underlying tech, but we can build in Azure or AWS Bedrock in whatever country a customer needs their data to be processed. We might use a different LLM or GPT, depending what meets the security criteria and needs at the time. This should be how it works with any tool you buy – in the same way you don’t worry about changes in the cloud infrastructure when you’re happily typing away in Word.

But we’ll keep you posted 😊

‍