Helloand welcome to 9to5Neural.AI is moving fast. We keep you up to date.We mentioned last week that American AI firms were facing intense competition from DeepSeek, a Chinese AI firm. DeepSeek has affected Wall Street today as NVIDIA’s stock dropped 17%. Let’s look at DeepSeek and NVIDIA’s reaction, as well as the larger picture of AI development.
What is DeepSeek?
DeepSeek, a Chinese AI company, was born from a hedge fund named High-Flyer. Liang Wengeng, a Chinese citizen, founded the company in Hangzhou, Zhejiang in 2023. Wengeng cofounded High-Flyer ( ) seven years earlier, focusing AI investments.
DeepSeek started training its models even before the U.S. Government restricted China’s ability to access American AI chips. The company will therefore have a good supply of NVIDIA graphics cards from before the restrictions were imposed.
DeepSeek was still forced to operate within the constraints of limited NVIDIA hardware access. This constraint may well have forced DeepSeek into focusing on the innovations it touts. Its V3 model
DeepSeek’s ability to compete with OpenAI’s Brand new o3model. ChatGPT O3 is the successor of o1, perhaps because O2 is a well-established UK phone carrier.
DeepSeek’s model is almost as competitive, but requires fewer resources. It also costs a fraction of OpenAI’s cost to operate.
DeepSeek arrived at this point by focusing on distilling models, rather than spinning them up using the same strategy used by American companies. DeepSeek benefits greatly from the work done by AI firms that we are familiar with. DeepSeek is also a product of the AI firms we already know. Due to the need to optimize existing models it is necessary to focus on distillation. Export of American AI chips to China is restricted by the United States.
DeepSeek training method
This is just the beginning. What happens next remains to be seen, but I believe that OpenAI and other American AI companies will prioritize model distillation in order to reduce operation costs and remain competitive. DeepSeek’s achievements are not unique to American AI firms. Now that the competition is here, it’s only a matter if you prioritize model efficiency.
But prioritizing the distillation of models isn’t what helped DeepSeek reach the AI race. DeepSeek also relies on AI training AI. American AI firms still use Human-labeled data is important in human-in-the loop training . The benefit of the AI-training – AI method has the advantage of being scalable, as it requires much less human input. The challenge is that mistakes can be amplified. This makes AI alignment checks harder. Alignment is a way to say that our AI models reflect and operate in accordance with our values.
The AI models are unbiased because of the fine-tuning that is done under supervision and the reinforcement learning based on human feedback. We make sure that the data is accurate.
I don’t anticipate a radical shift in the way American AI firms ensure quality data, but I do expect to see a significant movement towards AI training AI. OpenAI and other similar firms have always had this goal in mind; DeepSeek’s pressure may have pushed them to move faster.
If you follow DeepSeek you’ll probably come across a The research paper for their newest model gives a $6 million figure . The claim is that V3 has been developed for less than $6 million using NVIDIA H800 hardware. This claim is true, but it also ignores the investment costs associated in training earlier models. Not to mention NVIDIA’s supply acquired before U.S. AI chips export restrictions.
There is also another figure to consider: $600 billion. NVIDIA lost $600 billion in market capital today alone. This is the result of investors being scared by DeepSeek models that are cheaper to train and run, resulting in less opportunity for NVIDIA’s growth than expected.
This is a shortsighted and overreaction. My thinking is that DeepSeek has shown a great efficiency when it comes to how current AI models can developed. Great! This could reduce the time required to develop the next major AI model.
In short, adding more NVIDIA GPUs to the problem will likely still be the solution for pushing forward AI technology. We might even get further and faster now. Remember: The AI race is not about where we are right now, but moving forward.
AI isn’t solved
which leads to OpenAI’s massive Stargate ProjectStargate is a Texas building that’s stuffed to the gills full of compute. Say future AI models will be able to achieve more with less computing. This simply means that Stargate’s AI models can do more with the compute they have.
The gap between where these companies want to go in AI and where we currently are is real. DeepSeek’s impact may be that it forced other AI companies to prioritize different goals. We’ll have to wait and see what DeepSeek does next before we can judge whether they are a more innovative company.
Some other notes.
NVIDIA (19459057) found the silver lining of DeepSeek’s works with this statement released today:
DeepSeek represents an excellent Al advance and a perfect demonstration of Test Time Scaling. DeepSeek’s research shows how new models can easily be created by leveraging readily-available models, and a compute system that is fully compliant with export controls. Inference requires a large number of NVIDIA GPUs, and high-performance network. We have three scaling laws now: pre-training, post-training (which continue), and a new test-time scale.
We’re building an improved airplane mid-flight but we still require jet fuel to fly.
NVIDIA’s stock is still up 93% over the past year and 1,782% in the last five years. OpenAI is likely to be more generous when it comes to ChatGPT o3 mini due in part to DeepSeek’s competition.
After publishing on Monday, OpenAI boss Sam Altman responded on X to the attention DeepSeek is garnering:
deepseek’s r1 is an impressive model, particularly around what they’re able to deliver for the price. we will obviously deliver much better models and also it’s legit invigorating to have a new competitor! we will pull up some releases.
but mostly we are excited to continue to execute on our research roadmap and believe more compute is more important now than ever before to succeed at our mission. the world is going to want to use a LOT of ai, and really be quite amazed by the next gen models coming.
look forward to bringing you all AGI and beyond.
Fair summation of DeepSeek’s achievement, and obviously is doing a lot of work in that sentence.
President Trump addressed the DeepSeek effect on Monday, per Reuters–
DeepSeek AI, a Chinese company’s release, should be a wakeup for our industries to focus on winning.
Having read about China and the companies there, I was impressed by one company in particular that had developed a much cheaper and faster AI method. I see that as an asset.
That’s a good thing, because you will be doing it too, so you will not be spending as much and you should get the same results, hopefully.
Our ideas are always the best. We’re always the first. So I’d say that this is a positive and could be a very positive development. You’ll save billions of dollars and come up with the same solution.
The AI industry is the next NASA.
DeepSeek (19459057) has slowed new account creation due to a large cyber attack that is impacting the service. This message is currently displayed across the top of Chat.deepseek.com :
Due large-scale malicious attacks against DeepSeek’s service, registration may be busy. Please wait and try again. Users who are registered can log in normally. We appreciate your support and understanding. We were able, however, to create a brand new account on Monday after several hours of trying.
You may have also seen a viral social media post that claimed installing DeepSeek for iOS gave the Chinese AI firm access to your iPhone’s personal data, including emails and messages. Fortunately, this is not how iOS works. Sign in with Apple can be used to create an account, and it can generate a throwaway e-mail address for added security. DeepSeek can see what you type into the chatbot.
DeepSeek still suggests that you talk about math, coding and logic problems when asked what happened at Tiananmen in 1989. However, Perplexity appears to have solved that problem.
Find out more about the latest AI developments in the next issue of 9to5Neural, only on 9to5Mac. The previous issue can be read here.
Top iPhone accessories
Anker USB-A/C Chargers
Add 9to5Mac into your Google News feedBeats Solo Buds (Black)
Ailun iPhone 16 screen protector
Anker 621 magnetic portable charger
Anker USB-A/C Chargers
Apple MagSafe Charger (1m]
Beats
FTC: we use auto affiliate links that earn us income. More.
