- AI For All
- Posts
- The Jig is Up
In partnership with
Hello readers,
Welcome to the AI For All newsletter! A Chinese startup called DeepSeek humiliated Silicon Valley and sent the stock market into freefall. Shell-shocked, OpenAI accused DeepSeek of stealing their IP. Can I play the world’s smallest violin? Let’s find out!
The Jig is Up
My reaction to the DeepSeek news
The release of DeepSeek-R1 triggered an existential crisis across the AI industry, wiping $600 billion in market cap from NVIDIA and sending US tech shares off a cliff. DeepSeek even displaced ChatGPT as the #1 app in the App Store and is attracting business users in droves. How could the tables turn so decisively in China’s favor?
DeepSeek-R1 is competitive with OpenAI’s o1 while having a 90 to 95% cheaper API and supposedly being developed for only $6 million using a modest arsenal of 50,000 GPUs, most of which may have been NVIDIA H800s, which fall within US export restrictions. DeepSeek might not have even needed to smuggle chips into China.
One often meets his destiny on the road he takes to avoid it. In trying to forestall China’s AI development, the US simply forced companies like DeepSeek to innovate out of necessity, leading to R1’s numerous optimizations that have Mark Zuckerberg convening war rooms to reverse-engineer R1 (while pretending not to care).
I want to emphasize: R1 is not a leap in model capability, just efficiency. R1, like all other “reasoning” models, does not reason reliably. What R1 shows is that you don’t need billions of dollars in GPUs and data centers to build these models. This is not just bad for NVIDIA, but dries up what was already a shallow moat for companies like OpenAI, lowering the barrier to entry for competitors and further commoditizing LLMs.
Centering the “AI race” on hardware was always misguided. Everyone should want more efficient models. Look at what a brain can do with 20 watts. What we need is better software, which we won’t get until we move on from LLMs. Maybe R1 will prompt Silicon Valley to actually innovate again (it’s been a while after all).
Dashing such hopes is Masayoshi Son, who is in talks to invest $25 billion into OpenAI and lead a $40 billion investment round. Son, who lost billions investing in WeWork, is determined to become something of an omen for OpenAI. Additionally, Sam Altman urged continued spending on AI infrastructure in response to DeepSeek.
Lastly, OpenAI complained that DeepSeek stole its data to develop R1 in an amusing case of irony. OpenAI also released ChatGPT Gov, and I don’t care. I hope the two remaining federal employees post-DOGE can find a use for it.
🔥 Rapid Fire
AI power needs threaten billions in damages for households per Bloomberg
Environmental impact of AI becoming clearer while Big Tech emissions soar
Google struggles to attract developers and its own employees to Gemini
Developers say that AI software engineer Devin is bad at its job after tests
US Copyright Office says wholly AI-generated content is not protected
French government’s AI chatbot suspended after getting answers wrong
Hugging Face launches Open-R1, a fully open reproduction of DeepSeek-R1
Mistral releases Mistral Small 3 latency-optimized 24B-parameter model
Alibaba Cloud releases Qwen2.5-1M, Qwen2.5 VL, and Qwen2.5-Max models
Convergence launches Proxy AI assistant to challenge Operator and Claude
Receive Honest News Today
Join over 4 million Americans who start their day with 1440 – your daily digest for unbiased, fact-centric news. From politics to sports, we cover it all by analyzing over 100 sources. Our concise, 5-minute read lands in your inbox each morning at no cost. Experience news without the noise; let 1440 help you make up your own mind. Sign up now and invite your friends and family to be part of the informed.
🔍️ Industry Insights
“There are a host of straightforward ways to reduce hallucinations. A model with more parameters that has been trained for longer tends to hallucinate less, but this is computationally expensive and involves trade-offs with other chatbot skills, such as an ability to generalize. Training on larger, cleaner data sets helps, but there are limits to what data are available.”