Join institutional investors, analysts and strategists from the world's largest banks: Subscribe today
Already have an account? log in
Will DeepSeek Deep-Six The Mag7?
"More investment does not necessarily lead to more innovation."
The quote's from quant fund manager Liang Wenfeng, and it better not be true in the context of the AI arms race. If it is (true in that context), the biggest tech companies on planet Earth and, by extension, a top-heavy US equity market walking around on very tall stilts, are in a whole lot of trouble.
Liang's reasoning holds up. If innovation's the exclusive purview of the companies with the most money, then "large companies woul
https://docsbot.ai/models/compare/gpt-4o/deepseek-v3 This is an elegant depiction of deepseek-v3 (released last month) vs. gpt-4o. Deepseek is not able to handle image processing but for text processing it is close to as capable (if not more) for a fraction of the cost.
Perhaps we don’t need nuclear powerplant data centers running Blackwell chips, we need American software development companies to write more efficient code.
This may be the first genuine crisis of the Trump administration, I’m sure tariffs will solve this problem in 24 hours or some such nonsense.
There is a lot of money at stake so we can expect a tsunami of stories seeking to debunk DeepSeek. Especially from those who have already poured trillions into building out ever-larger data centers.
Trust. The Keyword.
That is partly addressed by being open source.
Agreed, open source makes this a bigger threat. See Linux beating Unix for reference.
If I know your real-life identity (derke, i know who you work for and who you are), will you be comfortable to seek indepth questions from me, if you know who i am?
The point is that open source coding can be replicated and offered by others. Such as Uncle Elon!
Question of the day: if AI is so wonderful at coding, why didn’t it create something similar on its own?
I mean, obviously the answer is “it’s not”.
I’ve played around with several GPT’s and tried to get them to build an app purely from prompts. It takes lots of re-prompting, lots of error handling, and ultimately what you end up with is something that barely does what you originally asked for. It’s more an exercise in patience than a game changer putting software developers out of business.
But, as with all things (see EV batteries) innovation will eventually make them better.
Fabulous.
This market reaction might really be an (over)reaction to one of my long time greatest fears: not if, but when China simultaneously embraces both capitalism and democracy, “look out below”.
My favorite Marc Andreessen expression is, “100%, 100%”
Brilliant article and exactly what I was looking for today. This is why I subscribe. Thanks, H.
Curious your thoughts on Michael Every’s theory of Trump using economic statecraft.
I don’t open his emails anymore. He’s too silly for my tastes, and his (implicit) pretensions to being unflappable went completely out the window in October of 2023, when he lost his mind over Israel-Gaza. Also, he seems to be something of a 2020 US election-denier and someone who can’t get over Hunter Biden’s laptop. So, you know, suffice to say he and I wouldn’t get along. But he shouldn’t feel bad. I don’t get along with hardly anyone, and I’m fully aware that I’m the problem, not other people.
Love it…
If AI is the future and requires investment to develop and a disruptor company comes along and makes it cheaper, isn’t that a net positive for the market as a whole?
I view the US AI “industry” just like pharma. They are a bunch of huge companies spending a fortune for block buster apps they can charge a fortune for. I don’t suppose anyone will believe me, but about ten years ago my daughter and her small tech team working inside an ad agency, were trying to build an app that would serve as a universal platform connector that would allow her firm’s sales agents to connect their laptops and such to connect to any customer’s platform. Instead of achieving just that goal, she and her four-person tech team created an app that could create other apps made to order. She did a demo seminar for several customers where she asked them to create a set of parameters/specifications they would want from a new app. Each participant did that, entered their lists at the program prompts and she then asked the participants to hit enter. Each got their desired working app in less than ten minutes. They then fed those apps into the universal program connector and watched their apps log on to the platform of their choice and run their newly created app. The team went to their top management team and were told that the company wasn’t interested in developing this product package because they thought no one would pay to buy the product. I have friends in the venture capital community. They saw the demo and committed to join me to offer an amount equal to the firm’s annual sales just to get all the rights to the two products and hire the team. The company wouldn’t sell and they never allowed release of the package.
Last week we were already getting the pushback counterfactuals – that DeepSeek was (supposedly) actually trained on 50,000 owned GPUs plus untold more rented in the cloud and/or DeepSeek (supposedly) isn’t really as good as o1.
I say, follow the money. Does a small Chinese quant fund have many USD billions to spend on training its model? If the current iteration of DeepSeek reasoner is merely 90% as good as o1, how are the promised trillion dollar economics of AI training any less threatened?
I think it is likely that AI will not be a monopoly or oligopoly like search or social media, because there is no monopoly on intelligence, human or artificial, American or Chinese. There will be many models competing, new ones emerging all the time, no one model with a sustainable lead. Suppose AI models end up like airlines – a service that is indispensable but fungible and thus not very profitable. That is not great for the Model Builders (OpenAI, Anthropic, etc).
For the Infrastructure Providers (NVDA, AVGO, ORCL, etc), the focus will shift even faster from the compute demand for training (building AI models) to the compute demand for inference (using AI models). The jury remains out here, I think, because we (or at least I) don’t know if DeepSeek r is any less compute-intensive to run than o1 or other models. DeepSeek pricing at 5 cents/1MM token input vs OpenAI at $15/1MM tokens should be worrisome for OpenAI, but maybe that is just “promotional” pricing, not reflective of actual compute and capital costs – who knows. However, we also don’t know that the GPU architecture, huge amount of HBM memory, leading-edge processor nodes, etc (thought to be) necessary for training are optimally efficient for inference. After all, AMZN has two separate processors Traininum and Inferentia. Perhaps HuaWei is perfectly capable of making inference processors; it is already making chips comparable to NVDA’s China-sold chips (H20 etc).
As @derek noted, this should be good for AI Consumers, from social media, cloud providers, software vendors (CRM, AAPL, META, AMZN, MSFT, GOOG, etc) to end users.
The disparate moves today seem generally consistent with this.
DeepSeek just released its image generation AI model.
One of the hallmark programs of the curriculum at the university at which I plied my calling for most of my career was its “Teaching English as a second language” program. Many international students would come to us, graduate, and sometimes stay stay awhile. One of the largest groups to come were the Chinese. We had a nice MBA program and many Chinese students would hang around to take that as well. One or two of them really liked our MBA program and suggested we start a new version for potential students off campus in Hong Kong and we checked it out and agreed. That program soon grew to exceed the size of our US on-campus program. A Chinese private education provider who paid us to bring in our program with our faculty each quarter, then asked us to expand to add a site in Shanghai. We did that and now we have two Chinese programs in China and I can tell you the Chinese students we have there think better, write English better and are generally smarter than our native students. I spent 15 years on the board of a US die casting firm with a newly established location in China. From there we made parts for Bendix, Deere, Chrysler, and other major US manufacturers exporting products their goods from China. It took awhile to adjust to the politics but the workers and managers were very good and we made a good deal of money. At present, Deere sends some of their Chinese managers to get their MBAs at our Chinese sites.
We in America like to have people somewhere that we can designate as stupid but dangerous enemies. In the Cold War era that role was was filled by Russia. (Frank Burns in MASH always said the Russians wanted to steal our toilets.) Then 9/11 came along and the enemies role was played by Iran and other selected Arab “terrorists. (Even though the Kingdom paid for the training and delivery of 9/11 they are exempt because they still buy our weapons and sell us oil.) Now it appears China will be “it,” with a side order of Arabs. We just can’t seem to get over ourselves. That’s nearly always been our loss.
love to see Liangs shorts going into last weekend. Bootleg ChatGPT