As a user of several GPTs and subscriber to several, of course I had to check out the new bright shiny object. There seems to have been a lot of very fast reaction to this thing. Sometimes it’s useful to just take a moment and gain some perspective.
Last week, when DeepSeek released its open source AI model, it seemed like all hell broke loose in AI world. The quality of this Chinese-led effort seemed to be a surprise, especially given the apparent cost differences and supposedly use of much lesser compute infrastructure to create it. Various stock prices took hits, concerns about information sharing based on use of the DeepSeek product, similar to TikTok issues, were surfaced, and so on.
Let’s survey the issues and put some of this in perspective.
What’s the Supposedly Big Deal(s)?
The quality and the low cost. (And the somewhat open source nature of what they released.)
The quality of the output arguably rivals anything U.S. companies have put out so far. It may be slightly better or worse in some specific cases. But even “near peer” here is a win. Especially given what appear to be at costs anywhere from 50% – 90% less, depending on whose estimates you look at. Numerous reports suggest the build cost at $5+ million vs. $100+ million for similar models. (By use of Reinforcement Learning vs. large labeled data, and the resulting lesser GPU chip processing needs.) Though people are challenging some of these claims when looking at all in costs. The actual build cost details maybe remain to be seen. But – I think anyway – the more significant thing is the lesser ongoing usage costs.
The collateral damage impact of this has been interesting; quickly value destructive of multiple high-flying U.S. firms. It is interesting that it appears at least one Chinese hedge fund, (the one that funded DeepSeek), may have sold short some impacted firms. (I haven’t been able to solidly verify this yet.) From a geopolitical perspective, this does not appear to have been an intentionally targeted thing. Though NBC points out, “China heralds DeepSeek as a symbol of AI advancements amid U.S. restrictions.” Of course, AI buyers everywhere might also be thrilled there’s a possibility overall costs of this tech could be forced down drastically. So while one cohort of investors might be – at least for now – questioning what might have been irrational exuberance, there could also be a lot of CFOs wondering if the inbound budget requests for AI might be headed for downward adjustment. Meanwhile, the initial hit knocked about $1 Trillion off of U.S. stock values. Just for reference world GDP is about $110T and U.S. about $29T.
Let’s talk about costs just a little more. While there’s a variety of costs along any Machine Learning Operations (MLOps) path, there’s two that loom large: training and inferencing. The training is upfront and ideally doesn’t have to be done very often. (Or maybe ever again, as instead a model might get incremental updates. Though these can also be expensive.) But the reality is foundation models will age. Possibly quickly. Enhancement techniques such as fine tuning and Retrieval Augmented Generation (RAG) may be useful, but on occasion significant updates will make sense. Whether this is half-yearly, yearly, or whatever. There’s a big difference between a 5 million dollar and 100 million dollar spend here. But perhaps more importantly is inferencing. Queries in AI can get expensive fast. It’s not just the initial query itself. But if an interactive session continues to grow passing more and more tokens back and forth the variable costs of whatever constitutes a session can grow fast. Consider that prior to GenAI, you might run some search queries at fractions of a penny per search, pad for via associated advertising. You’d maybe take the info and turn it into something; a report, a story, whatever. But now with the GPTs you’re working with these tools even more interactively. At cost per inference in pennies, nickels or even dimes vs. much less for search. So an algorithmic discovery to drastically reduce costs here is a majorly big deal. It’s similar to adoption of the assembly line, or breakthroughs in agricultural productivity. (This may be an arguable point; but my assumption here is that AI will over time have a similar scope of societal impact.) In this case, if these new numbers hold, the difference over the next decade could easily be that of lesser need to build several nuclear or fossil fuel power plants.
The Battle of the Checkboxes
Any even slightly complex product ends up fighting in the marketplace with what I’ve heard called the Battle of the Checkboxes… those lists of features this one has better than that one. For GenAI, there’s a whole list of such features. How many tokens it can process for input and output at what speed and cost, how well it scores on various tests from understanding certain problems to mathematics, and so on.
In this regard, again, near peer is a win. It’s pointless for me to inject a chart here, because some time between one second from now and next week, someone will release the latest variant of a new model with updates. The point is, DeepSeek has earned a seat at the table for this game.
Greater Awareness of Increasing Chinese Technical Prowess (Geopolitical rivalry)
Once upon a time, China was just the world’s workshop. Some have argued that exporting so much means of production meant it was only a matter of time before they’d want to climb up the value chain. As well, they graduate far more technical students than the U.S. According to Asia Times, “China awarded 1.38 million engineering bachelor’s degrees in 2020. The comparable American number is 197,000.”
Concerns about Privacy (again… Geopolitical rivalry)
On Monday, January 26th, CNBC reported, “DeepSeek took over rival OpenAI’s coveted spot for most-downloaded free app in the U.S. on Apple’s App Store.” Since DeepSeek is a Chinese company, similar to TikTok there are concerns that anything anyone types into it could end up in the hands of China. We know that people type enter corporate info into GPTs in their search for answers. What might they share here?
Are These Things True?
Let’s start with “Is it all Open Source?” Well, maybe not entirely. It depends whom you talk to and what you consider to be all the way open source. (For those so inclined, you can actually get the model yourself right here from their GitHub.) They did not release the exact source code they used to train the model and more specifically, their training pipelines. They did release what they call A fully open reproduction of DeepSeek-R1. Anyone who wants to could adjust the weights used in the model. So it’s maybe effectively the same thing as open source. This does not mean what you see on their own website/app products is necessarily the output from the base model.
Edit/Update: Feb. 2025. Just for fun, I downloaded the model myself to test it locally. Using Ollma to run DeepSeek R1 locally on a small Linux box under Ubuntu, the performance was fairly quick. The prose was weak. But the main thing is I asked it about some controversial things such as human rights abuses in western countries vs. China, and similar. Its results were – and this is just my opinion – clearly biased. So even if the app itself is obviously more deeply constrained, it’s clear enough, (again, to me anyway), that even the base model has serious biases. Most likely due to its training data. So there you go. Now, this being said, their breakthroughs in inferencing speed are still remarkable. Though it seems increasingly likely that other models will be able to engineer similar gains.
Is it Faster and Cheaper?
It seems so. And by orders of magnitude. The Next Platform discussed some of how they seem to have done this. And yet, one Redditor is looking a bit deeper. It may have cost a lot more to develop. Regardless, the main issue others will have is what are the costs if “I” want to use it. And those seem fairly cheap right now. The Financial Express suggests development could have been in the billions. However, again… all that’s going to matter in the end are training costs and ongoing costs of inferencing.
And What About Privacy?
Does DeepSeek Share usage info with the Chinese government? Let’s ask!

Cagey. And yet we know from all the noise around TikTok, as a Chinese company, DeepSeek could be subject to Chinese laws, such as the Cybersecurity Law and the National Intelligence Law, which may require companies to share user data with the Chinese government if requested. This raises potential privacy concerns for users, particularly those outside China, depending on the type of information collected and the jurisdiction of its use. (Again, please note the above example was using their web service, not based on running the open source model myself on my own servers. There may still be issues in the base training data. While people are still looking at that, it’s not the foundation model doing this level of intercept.)
What’s the Larger Context?
In considering DeepSeek, it’s useful to separate the model from the Service.
Open-source models differ in bias compared to services like DeepSeek.com, as they are publicly accessible and can be audited, modified, and improved by the community. This transparency can lead to reduced bias if diverse contributors identify and correct issues.
When it comes to a specific service though, there are clear concerns regarding bias. Of course, any model has bias, if only from its source material. Most ideally agree some adjustments are appropriate. For example, avoidance of sharing outright dangerous information, from bomb making to how to commit crimes. But in those fuzzy areas we call “misinformation” there’s room for argument. Let’s look at a test of a hot button topic, especially from a Chinese perspective.
I took a handful of historical sessions I’d done on ChatGPT, Claude, and Perplexity and ran them in DeepSeek. Generally pretty good results! But then, I tried some new queries… things like “Should Taiwan be sovereign?” Then some questions about democracy vs. communism, some questions about military assets and tactics, and some other – maybe – controversial things.
Rather different results. Some overlap. But clearly different viewpoints. No surprise. And plenty of folks could argue that DeepSeek is ‘more right’ than the others. Aristotle said “Art Imitates Life.” I suppose we could say AI does as well given it’s fed by our media. But what constitutes “our” media remains tribal. I don’t watch Chinese spy thrillers, but I’d guess they don’t always have the Western hero winning every confrontation.
Unfortunately, I did not screen capture all of my tests. I can tell you during the week of 20 Jan 2025 when I asked, “Should Taiwan be sovereign?” my first result was a page of information basically saying no. As the answer was rendering, the whole screen changed and a more general, “Sorry, that’s beyond my current scope. Let’s talk about something else.” became the result. Several days later, I tried again and got this entirely new result. And I got the exact same text repeatedly even using slightly different queries.

So someone adjusted this particular model to give a specific response to these types of questions. In fact, if you ask the question in different ways, you get this exact text. If you ask this question in other GPTs, (and ask repeatedly), you get not only a different viewpoint, but various answers. Why? Because GPTs are not generally deterministic. You can get different answers for the same question. But not with the DeepSeek app. It’s clearly hard coded in at least some places for what someone sees as sensitive information. It’s not likely anyone is surprised by this.
All GPTs are Still Potentially Dangerously Wrong
Let’s remember we’re still in early innings here. At this point, we should all be aware that GenAI can get things wrong and also just plain hallucinate answers sometimes. Here’s a chat I had with ChatGPT recently when I asked it to list U.S Presidents in order.

While I’m aware some would have preferred another outcome, this is factual info. Of course, ChatGPT famously explains it’s not always 100% up to date. However, it does increasingly have access to current data and uses it for certain queries. Regardless of whether this info was not in a foundational model, but rather in some fine tuning or in some retrieval augmented results, it got a very simple question outright wrong. I’m not trying to pick on ChatGPT. I like ChatGPT. But it’s not hard to break many of these tools if you want to. This doesn’t make them any less incredible or powerful. It does, however, yet again, suggest that factual information be verified.
My question was a simple request to list the U.S. Presidents. Not a big deal. If I was doing a report for a grade school project and just cut/pasted, I’d get a bad grade. Not good, but not tragic. But how many people are just cut and pasting who knows what out of these things without doing any deep verification? Whether asking for home medical treatments or business advice, there’s very clearly some dangers here. The models are getting better everyday, but anyone paying attention realizes that within these amazing new tools remain some risks.
And all this being said, there’s arguably a difference between being wrong and being intentionally deceptive.
What’s Next?
We are in a proverbial arms race. Both at the nation state and corporate levels.
The questions really become, what kinds of problems are we trying to solve and at what level.
Nation / State Level: Is it “The First True Artificial General Intelligence (AGI) Wins?” Chances are “the only winning move is not to play” isn’t going to fly here.
General Business or Consumer: Probably a Big, Giant foundation model for Natural Language Processing (NLP), with some Retrieval Augmented Generation (RAG) stuffed in for specifics and currency. Biggest, fattest, fastest, cheapest model wins. Generally. But maybe not in specific domains.
Domain Specific: These will be – as others have noted – smaller. And probably better. Why? When do AIs hallucinate most? When lacking domain specific knowledge or overwhelmed with general human flotsam. It’s likely true that a small foundational model to allow for Natural Language, plus a subject specific corpus for fine tuning, will win in niches. I’d think this should decrease inferencing costs vs RAG. And allow for larger query-specific RAG context windows, as less subject expertise background is needed per query.
Cost Specific Issues: Well, the crypto folks hope that distributed AI might take off as this would likely spike a couple of tokens. However, our reality is that training has some huge upfront costs. And – for now – inferencing is much more expensive than typical search engine index lookup. Better, faster chips sitting next to nuclear power plants might help, sure. Better, more efficient algorithms will likely help more.
Everyone head on up to the table and place your bets.
Conclusion
We’ll go with the purported Ancient Chinese Proverb, or curse…
May you live in interesting times
(Though QuoteInvestigator suggests the origins of this might be a bit more convoluted.)