Why is Deepseek being so good a bad thing?

I don't trust their claims. They've gotten their hands on more advanced GPUs than has been publicly acknowledged.

You’re raising some great questions that touch on both skepticism and the potential upside of Deepseek’s claims. Let’s break it down systematically:

1) Do we trust the claims being made?

• The Skepticism: The claim that Deepseek was built for $5 million versus the billions spent by others (e.g., OpenAI, Google) raises eyebrows. It’s fair to ask how they achieved this when others have invested heavily in compute resources, data collection, and engineering talent. Did they leverage existing open-source tech? Did they cut corners on data quality or fine-tuning? The lack of transparency on specifics might make people cautious.

• Verification: Until there’s third-party validation or benchmarks (e.g., OpenAI’s evals or Hugging Face integrations), it’s hard to fully trust their claims. A demo or peer-reviewed performance metrics could help clear this up.

2) Isn’t this a good thing long-term?

• The Case for Optimism:
• Cost Efficiency: If Deepseek genuinely developed something comparable to models like GPT-4 for a fraction of the cost, it democratizes AI development. It means smaller players (startups, researchers, and even governments) could create competitive AI without needing to burn billions.

• Open Source: Open-source AI can foster innovation, as developers globally can adapt, improve, and customize models to suit their needs. For instance, if Deepseek’s codebase is open and permissively licensed (like Apache 2.0 or MIT), others could fork it, tweak it, and deploy it independently. That’s a huge deal for transparency and accessibility.

• Energy Efficiency: A model that uses significantly less power could help reduce the environmental footprint of AI, a growing concern as AI adoption scales.

3) Data Privacy and Trust in a Chinese Company

• Concerns About Chinese Companies: Many will be wary of using models tied to Chinese organizations, especially for sensitive data, given geopolitical tensions and concerns about surveillance or data misuse. However:
• If the model is truly open source and self-hostable, users don’t have to interact with Chinese servers or infrastructure. They can run it locally or on private cloud instances, mitigating concerns.
• Companies can even adapt or harden the code against potential vulnerabilities, as the open-source nature allows scrutiny of its workings.

4) Open Source Models: Can People Adapt the Code?
• Explanation of Open Source:
• If Deepseek is open source, it means the source code (and possibly model weights) is available for public use and modification. Developers can:
• Fine-tune the model on custom datasets.
• Adapt it for niche use cases (e.g., legal AI, medical AI).
• Audit the code for security or performance enhancements.
• The true “openness” depends on the licensing. For instance, permissive licenses like Apache 2.0 allow for both commercial and personal use, while restrictive licenses may impose limits.
• Adaptability: If all pieces (architecture, weights, and training recipes) are included, anyone could theoretically replicate or improve on their work.
5) If This Was Announced Without Context, Would Everyone Be Excited?
• Hypothetical Reaction: Two weeks ago, if someone said, “Hey, a new AI model is coming out, costs 1/50th of GPT-4, and uses way less power,” the reaction likely would have been overwhelmingly positive. It aligns with the goal of making AI more accessible and sustainable.
• Current Concerns: The mixed reception arises because:
• It’s a Chinese company, which triggers geopolitical and privacy concerns.
• Bold claims with limited evidence naturally provoke skepticism.

1. Trusting the Claims: The lack of transparency makes the $5M claim hard to fully trust without third-party validation or detailed technical disclosures.

2. Good Long-Term Impact: If the cost and energy savings are real, and the model is truly open source, it’s a win for AI accessibility and innovation. Even if people don’t trust the company, the open-source aspect ensures others can adapt the tech independently.

This could be a game-changer—but the burden of proof lies on Deepseek to demonstrate what they’ve achieved.

I would also like to know more about the DeepSeek chip.

Businesses using AI at a very fundamental level that effects the user experience/interface and how internal components of their business function and are integrated. This is mission critical stuff for a business and I would think that they would be interested in getting a product with very good support, inegration across platforms, and reliability.

With much mission critical operability at stake, will they trust a free Open Source product with little or no track record?

Add to that the fact that AI is still growing and evolving and being rolled out, so the demand for chips that make AI possible may very well offset the emergence of new players in the field like DeepSeek.

But clearly the field is evolving and for simple functions like search, maybe it may prove useful.

China doesn’t build anything. This will essentially the best parts of the current AI “borrowed” and slapped together. That is how they roll. It is good for consumers, it is bad for companies investing billions as they get nothing.

1) We don't yet, and there is plenty of speculation that they are lying.

2) It is in the long term, but in the short term it completely throws the projected economics of the entire space out the window, which is bound to cause major turmoil. Nvidia's future earnings projections are suddenly in jeopardy, OpenAI and Anthropic (which have billions in investment from Microsoft, Google, Amazon etc) etc etc. This is bound to get political since just 6 days ago OpenAI announced the Stargate project: (https://openai.com/index/announcing-the-stargate-project/) , which has plans to spend $500B on datacenters for AI-related technology. This was heavily touted by Trump as a big "invest in America" and, if DeepSeek's claimed advantages are real, might be DOA since its hard to justify that scale of investments if we suddenly need 20x less compute

Ask Deepseek to tell you more about the famous picture of a man holding grocery bags standing in front of a tank.

Assuming that Deepseek really did what they say they did, it is a good thing for consumers of AI that someone has been able to do it on a shoestring budget. AI proponents claim that it is ok if AI takes all our jobs because the massive boost in worker productivity will mean that prices will crash and everything will be almost free (of course this more for the service side of the economy and not eggs and oil, etc.). If AI is cheap, this prediction may actually come to fruition. The biggest risk of AI is that it will be held by a small number of companies who will extract a huge premium for their product in an uncompetitive market (just like cell phones, etc.).

The main downside of Deepseek is that it is also likely that Deepseek is lying and that they have stolen or misappropriated AI tech from US companies and are just making a cheap copy in order to drive US AI firms out of business. This is what China does. They did it/tried to do it with steel, cell phones, EVs and solar panels among other things. That brings up an even longer discussion about intellectual property laws. Folks like Dean Baker of CEPR and others have long argued that IP laws hinder innovation and promote monopolistic behavior among large companies.

The other downside is that long before this, there has been a fear of a growing tech bubble centered in AI but fanning out into stuff like crypto, Tesla, etc. If you were looking for something that would get everyone running for the exits, this may be it. And a tech bubble bursting is going to be bad for the broader economy.

Or their model isn't very good.

Or it's just a matter of time until the American companies rework their models and get competitive.

Not really a reason to panic.

It's a good thing. They developed a new costsaving training method and shared it with the world (both the methods and the model itself).

To my knowledge they’re open about having built upon llama?

Apparently the open source model is totally uncensored and will tell you everything about 64 event if you run it yourself rather than through their web app.

Just came to say Wejo crowdsourced and the community responded, and I for one am better informed than I was a few hours ago.

Isn't it Nvidia's own H800 (cheaper and less advanced) because of the ban on chip exports to China?

ft.com

US imposes export controls on chips for AI to counter China

American semiconductor industry warns move will aid competitors while EU has also protested against the new rules

Open link

Anyway, my thoughts:

1. We don't, really.

2. It's good for the public (well, AI end users) but not for people who may have a vested interest in AI being expensive. Presumably it could result in AI being cheaper and more widely accessible, lower cost of access.

I don't agree with it being slapped together. I think it works in a different way. It's quite creative. It's fundamentally different. The tweaks are in the way it works rather than throwing more processing power at it. Agree with last sentence.

It’s good for the field and possibly for humanity if you believe in an AGI future.

If you like monopolar US hegemony (I do!) it’s a clear sign that China is at parity with the US in another field that’s critical for the next 30 years. People dismissing this as “copying” are wrong. DeepSeek has lots of clear algorithmic improvements. Who cares if the CCP invested $50 billion or not secretly — This is a state of the art model.

The bad news is for the heavy investors in far more expensive AI linked to expensive chips. This threatened the possibility of their market collapsing, though it may well turn out to be overblown.

What is the energy cost?

Deepseek will be monitized as everything else is. Its back door to CPUs will be found eventually