Altman Praises DeepSeek

Introduction

Sam Altman, OpenAI’s CEO, publicly praised DeepSeek’s R1 AI model on X (formerly Twitter), calling it “an impressive model, particularly around what they’re able to deliver for the price”. While acknowledging the model’s capabilities, Altman also asserted that OpenAI will “obviously deliver much better models” and expressed excitement about having a new competitor

He highlighted the model’s efficiency, which stems from its use of less advanced Nvidia H800 chips and a training budget under $6 million—an achievement that contrasts sharply with the resource-intensive methods employed by OpenAI and other U.S. tech giants.

While Altman acknowledged the invigorating nature of competition, he reaffirmed OpenAI’s commitment to leveraging greater computational power to push the boundaries of AI innovation. He emphasized that OpenAI would soon release models that surpass R1 in both capability and impact, underscoring his belief that increased compute remains essential for achieving breakthroughs in AI

Altman outlook on Open AI for AGI

OpenAI has outlined a five-step roadmap to achieve Artificial General Intelligence (AGI) by the end of the decade. The company is currently transitioning from the first to the second stage, with the initial level focusing on conversational AI capabilities already achieved through models like ChatGPT. The subsequent stages involve developing:

“Reasoners” - AI models that can solve problems across various topics at a PhD level

“Agents” - AI systems capable of taking independent actions

“Innovators” - AI that can aid in inventing new ideas

“Organizations” - AI systems that can autonomously perform all functions of an organization

While OpenAI’s CEO Sam Altman has expressed confidence in their ability to build AGI, he has also tempered expectations, suggesting that the impact of AGI might be more gradual than previously anticipated.

Altman now describes the transition to AGI as a “long continuation,” emphasizing that the world may not change dramatically overnight upon its achievement.

Competitive Stance


Despite praising R1, Altman emphasized OpenAI’s commitment to its research roadmap and belief that “more compute is more important now than ever before to succeed at our mission”.

He hinted at upcoming releases and maintained confidence that OpenAI will continue to lead in AI innovation.

The praise signals a notable moment of inter-company acknowledgment in the rapidly evolving AI landscape, with Altman recognizing the potential of a Chinese AI startup while maintaining OpenAI’s competitive edge.

DeepSeek cost effectiveness will change competition and market

DeepSeek R1’s cost-effectiveness is reshaping the AI market by challenging traditional cost structures and intensifying competition. With training costs of just $6 million—far below the $100 million spent on models like GPT-4—R1 demonstrates that high performance can be achieved without massive budgets, undermining the notion that larger models and more data are always superior.

This affordability has disrupted tech markets, causing a 3% drop in Nasdaq and significant losses for companies like Nvidia, as investors reevaluate the financial requirements for AI development. By offering API costs up to 40 times lower than competitors, R1 is democratizing access to advanced AI, empowering smaller businesses and startups to adopt cutting-edge technologies.

The model’s success pressures established players like OpenAI to innovate more aggressively while fostering a shift toward efficiency-driven AI development. This trend could lead to broader adoption of cost-effective AI solutions, transforming both market dynamics and accessibility.

Mixture of experts architecture unique

DeepSeek R1’s Mixture of Experts (MoE) architecture is unique due to its selective activation of parameters, enabling high efficiency and scalability. The model consists of 671 billion parameters spread across specialized expert networks, but only 37 billion are activated during a single forward pass. This sparsity minimizes computational costs while maintaining performance comparable to or better than traditional dense models.

A gating network determines which experts to activate based on the input, ensuring that only the most relevant components are used. This design allows for specialization within the model, where different experts focus on specific tasks, such as grammar or creative writing. As a result, R1 achieves improved accuracy and clarity with reduced resource consumption.

Additionally, this architecture supports scalability by allowing more experts to be added without a proportional increase in computational overhead, making it ideal for multi-domain applications

Reinforcement learning play in DeepSeek R1's training

Reinforcement learning (RL) plays a central role in DeepSeek R1’s training, enabling the model to enhance its reasoning and decision-making capabilities without heavy reliance on traditional supervised fine-tuning. The key aspects of RL in DeepSeek R1 include:

Group Relative Policy Optimization (GRPO): This in-house RL method trains the model by sampling multiple outputs for a given input and assigning rewards based on predefined rules, such as accuracy and format adherence. This avoids issues like reward hacking and reduces training costs.

Reasoning Enhancement: RL is used extensively in the second phase of training to improve the model’s performance in tasks requiring logical reasoning, such as math, coding, and science. Rule-based rewards guide the model to produce accurate and well-structured answers.

Iterative Refinement: Later phases incorporate rejection sampling and diverse RL tasks. Outputs are filtered for quality before being used for further fine-tuning, ensuring the model aligns with human preferences while maintaining reasoning strength.

This approach allows DeepSeek R1 to autonomously refine its problem-solving strategies, making it more efficient and capable of handling complex tasks at scale.

DeepSeek R1's reinforcement learning approach compare to traditional supervised learning methods

DeepSeek R1’s reinforcement learning (RL) approach differs significantly from traditional supervised learning methods in several key areas:

Learning Process:

Reinforcement Learning: DeepSeek R1 uses Group Relative Policy Optimization (GRPO), a rule-based RL framework where the model learns through trial and error, guided by reward systems. This allows it to independently discover reasoning patterns and solve problems without relying on pre-labeled datasets.

Supervised Learning: Relies on large labeled datasets to map inputs to known outputs, which can limit the model’s ability to generalize beyond the training data.

Data Requirements:

RL in DeepSeek R1: Reduces dependency on massive labeled datasets, mitigating ethical concerns like data privacy and bias. Instead, it generates synthetic data through rejection sampling and self-verification processes, refining its reasoning capabilities iteratively.

Supervised Learning: Requires extensive labeled data, which can be expensive and time-consuming to collect.

Capabilities:

RL in DeepSeek R1: Enhances reasoning by exploring diverse solution spaces, enabling the development of coherent chains of thought and adaptability to complex tasks like coding and logic.

Supervised Learning: While effective for structured tasks like classification or regression, it may struggle with dynamic problem-solving or tasks requiring sequential decision-making.

By integrating RL with a multi-stage training process, DeepSeek R1 achieves greater autonomy and efficiency compared to traditional supervised learning methods while addressing challenges like readability and domain-specific performance

tasks benefit the most from DeepSeek R1's reinforcement learning approach

DeepSeek R1’s reinforcement learning (RL) approach excels in tasks requiring reasoning and precision, particularly in:

Mathematical Computations: RL enhances the model’s ability to solve and explain complex math problems, making it highly effective for educational and research applications.

Code Generation and Debugging: R1 uses RL to refine its logic and problem-solving skills, enabling it to generate accurate code snippets, debug errors, and explain coding concepts.

Scientific Explanations: RL helps the model articulate complex scientific concepts clearly, ensuring accuracy and readability.

These tasks benefit from RL because it allows the model to self-correct, validate reasoning, and adapt through trial-and-error learning, ensuring outputs are both accurate and logically sound.

DeepSeek R1's reinforcement learning approach compare to traditional supervised learning methods

DeepSeek R1 be integrated into customer service chatbots

DeepSeek R1 can significantly enhance customer service chatbots through its advanced features and seamless integration capabilities. Key benefits include:

Contextual Understanding: Its sophisticated natural language processing (NLP) allows chatbots to understand nuanced queries, maintain context over long conversations, and provide accurate, human-like responses.

Multilingual Support: R1 handles multilingual queries effectively, making it ideal for global businesses needing culturally nuanced communication.

Real-Time Learning: The model adapts dynamically to user interactions, improving accuracy and responsiveness over time.

Cost Efficiency: Its lightweight architecture reduces operational costs while maintaining high performance, making it accessible for businesses of all sizes.

Customizability: Developers can tailor R1 to specific industries or tasks, such as resolving customer complaints, managing support tickets, or providing product recommendations.

With these capabilities, R1 enables chatbots to deliver efficient, scalable, and personalized customer service experiences.

Conclusion

Altman’s comments came after DeepSeek stunned the tech world by revealing R1’s development cost was less than $6 million—a tiny fraction of what U.S. tech companies typically spend.

The model’s cost-effectiveness was so significant that it triggered a massive market reaction, with Nvidia experiencing a record one-day loss of $593 billion in market value.

DeepSeek R1 has garnered attention for its remarkable cost efficiency compared to industry leaders like OpenAI’s o1 model.

The Chinese AI startup’s model is estimated to be 20 to 40 times cheaper to run than comparable models from OpenAI.

Specifically, DeepSeek R1’s API pricing is set at $0.55 per million input tokens and $2.19 per million output tokens, significantly undercutting OpenAI’s rates of $15 and $60 per million tokens, respectively.

This dramatic price difference is attributed to DeepSeek’s innovative approach, which emphasizes software-driven resource optimization over hardware dependency, allowing the model to deliver high performance at a fraction of the cost