This video details the significant leap forward made by Deepseek's new R1 0528 AI model, which has surprisingly achieved performance comparable to leading closed-source models like OpenAI's GPT-3 high and Google's Gemini 2.5 Pro on various benchmarks. The speaker introduces "AI forensics" using EQBench's "slop profiles" to infer that Deepseek likely achieved this by switching its training data from OpenAI's synthetic outputs to Gemini's. This development highlights the widespread industry practice of "knowledge distillation" and frames it within the intensifying US-China AI race. The video argues that readily available, high-performing open-source models like Deepseek R1 0528, offered at vastly lower costs, could fundamentally disrupt the profit margins of proprietary AI companies and influence global technological competition.
Summarizing the Deepseek R1 0528 Breakthrough and its Implications:
0:03 Deepseek R1 0528's Performance Leap: The new Deepseek R1 0528 model, released on May 28, 2025, marks a massive improvement over its predecessor. It now stands "on par" with OpenAI's GPT-3 high and is "slightly behind" it but "ahead of Gemini 2.5 Pro" on AIME 2024 and 2025 benchmarks, often beating Gemini 2.5 Pro on others.
1:17 Not Deepseek R2: This significant advancement is from the R1 series, not the anticipated R2 model, suggesting that even more powerful open-source models might be released in the future.
1:37 EQBench's AI Forensics: Sam P. from EQBench utilizes "slop profiles" (tendencies of AI models to use certain expressions/words) and bioinformatics tools to infer the "lineage trees" of LLMs, providing a novel method for AI forensics.
3:34 Shift in Training Data: EQBench's analysis suggests that the new Deepseek R1 model's "slop profile" aligns closely with Gemini technology, indicating that Deepseek likely switched from training on OpenAI's synthetic outputs to Gemini's.
4:23 "Knowledge Distillation": The video describes the industry's "open secret" of "knowledge distillation," where companies train their models on the synthetic outputs of other existing models, which can lead to models exhibiting behaviors or biases from their "source" models (e.g., non-OpenAI models claiming GPT architecture).
5:05 Intensifying US-China AI Race: The development is contextualized within the broader US-China competition in AI, with the US Department of Energy comparing AI development to the "next Manhattan Project."
5:50 China's Open-Source Strategy: A potential strategy for Chinese open-source AI is to "take the profit out of AI software" by providing free or very cheaply priced, high-performing open-source models, thereby disrupting the business models of US tech companies that rely on selling expensive closed-source AI.
7:23 Drastically Lower API Costs: Deepseek R1 0528 offers significantly lower API costs (e.g., $0.13-$0.55 input, $0.50-$2.20 output per million tokens) compared to OpenAI's GPT-3 ($2.50-$10 input, $40 output) and Gemini 2.5 Pro ($1.25-$2.50 input, $10-$15 output), making it "very hard to beat."
8:27 Deepseek Founder's Vision: Leang, Deepseek's founder, states that "modes created by closed source are temporary" and reaffirms their commitment to an open-source approach, anchoring their value in innovation and organizational culture.
8:48 Complex Geopolitical Dynamics: The video highlights the intricate relationship between governments and AI labs on both sides, mentioning potential US legislative changes to incentivize domestic software R&D and China's surging energy production for AI. It also notes that researchers often cooperate globally despite governmental competition.
11:11 Rising Stakes and Competition: The narrator concludes that the "wheels are turning ever faster" in the AI space, with increasing competition and interest, and that for many, AI is seen as an "existential win-lose condition," indicating rapidly escalating stakes.