DeepSeek V3 is a general-purpose model released in December 2024, while R1 is a reasoning model built on top of V3. The host says R1 reached performance comparable to OpenAI o1 and Google Flash 2.0 on some complex reasoning benchmarks. Many of the techniques behind the models appeared in DeepSeek papers published earlier in 2024.
The Engineering Unlocks Behind DeepSeek
#Open Models#Language Models#Model Compression#Reasoning Models#Reinforcement Learning#Model Architectures
Y Combinator host explains the engineering behind DeepSeek’s V3 and R1 models, and what their efficiency gains mean for AI builders.
V3 was trained in FP8, with periodic FP32 accumulation to limit numerical errors while using less memory. Its mixture-of-experts design has 671 billion parameters but activates 37 billion per token, and multi-head latent attention reduces KV cache storage. Multi-token prediction adds training signals and can also support faster generation.
DeepSeek used reinforcement learning on problems with verifiable answers, especially in math and coding, scoring final answers with simple accuracy and formatting rules. Its GRPO method updates the model from those scores without supplying examples of how to reason. Reasoning emerged over thousands of steps, but the initial R1-Zero outputs mixed languages and were hard to read, so DeepSeek added a structured reasoning fine-tuning phase.
The host notes that R1 performed comparably to o1 on some math and coding benchmarks, while OpenAI’s o3-mini soon outperformed both on key benchmarks. R1’s free access and lower cost helped draw attention, but the reported $5.5 million for V3 covered only its final training run, not R1 training, research, or hardware operating costs. The host says a Berkeley lab reproduced key R1-Zero techniques in a smaller model for $30.
The host argues that DeepSeek shows there is room for new companies to improve GPU workloads, inference software, tooling, and AI-generated kernels. Lower intelligence costs could benefit consumer and business applications. The episode closes with a YC Spring Batch application deadline of February 11.
Takeaways
- Measure model efficiency across the full workload, including memory, networking, and GPU utilization.
- Mixture-of-experts models can reduce per-token computation by activating only part of their parameters.
- Use verifiable tasks and simple outcome-based rewards when training models to solve problems.
- Treat reported training-run costs as partial figures unless they include research, hardware, and later training.
- Falling inference costs create opportunities for products that were previously too expensive to run.
Mentioned in this talk
DeepSeek
OpenAI
Nvidia
Anthropic
Meta
Google DeepMind
Y Combinator
o1
DeepSeek-R1Claude 3.5DeepSeek-V3Gemini 1.5GPT-4oGoogle Flash 2.0
React Native Testing LibraryDeepSeekMath
R10
AlphaGoFLOPsFP32FP8InfiniBandCUDAmixture-of-expertsMulti-head latent attentiono3-minispeculative decodingMulti-token predictionreinforcement learning from AI feedbackReinforcement learning from human feedbackReinforcement learningGRPOLee SedolSearch inside this talk
What was claimed 23
- metricDeepSeek-R1: Claimed performance comparable to OpenAI o1 at a fraction of the cost
“it claimed achieved performance comparable to OpenAI o1 at a fraction of the cost”
YouTube Unverified1 source· Talk
- metricNvidia: Nearly $600 billion in market capitalization lost in one day
“Nvidia losing nearly $600 billion in market cap in a single day”
YouTube Unverified1 source· Talk
- launchedDeepSeek launched DeepSeek-R1
“DeepSeek recently made waves when it announced R1, an open-source reasoning model”
YouTube Verified4 sources· Talk
- featureDeepSeek: Publishes research and releases model weights
“publishing its research and releasing its model weights for months”
YouTube Unverified1 source· Talk
- metricDeepSeek-R1: Performance comparable to OpenAI o1 and Google Flash 2.0 on certain complex reasoning benchmarks
“performance comparable to OpenAI o1 and Google Flash 2.0 on certain complex reasoning benchmarks”
YouTube Unverified1 source· Talk
- featureDeepSeek-R1: Reasoning model optimized for complex reasoning
“applied various algorithmic improvements to optimize its reasoning ability”
YouTube Unverified1 source· Talk
- usesDeepSeek-R1 uses DeepSeek-V3
“DeepSeek R1...is a reasoning model built on top of DeepSeek V3”
YouTube Unverified1 source· Talk
- launchedDeepSeek launched DeepSeek-V3general-purpose base model
“DeepSeek V3, which was actually released this past December”
YouTube Unverified1 source· Talk
- usesDeepSeek-V3 uses FP8
“training V3 natively in 8-bit floating-point format”
YouTube Unverified1 source· Talk
- uses
“DeepSeek V3 has 671 billion model parameters, but only 37 billion are activated”
YouTube Unverified1 source· Talk
- metricDeepSeek-V3: 671 billion total parameters; 37 billion activated per token prediction
“671 billion model parameters, but only 37 billion are activated for a given token prediction”
YouTube Unverified1 source· Talk
- uses
“V3 makes use of multi-head latent attention, or MLA”
YouTube Verified2 sources· Talk
- metricReact Native Testing Library: Maximum generation throughput increased to 5.76 times
“boosted its maximum generation throughput to 5.76 times”
YouTube Unverified1 source· Talk
- metricReact Native Testing Library: Reduced KV cache size by 93.3%
“the V2 model reduce its KV cache size by 93.3%”
YouTube Unverified1 source· Talk
- uses
“V3 makes use of multi-token prediction, or MTP”
YouTube Unverified1 source· Talk
- usesIntroduced the method in DeepSeekMath
“the algorithm DeepSeek used”
7:38inReinforcement Learning for AgentsYouTube Verified3 sources· Talk
- featureGRPO: Uses model output scores to update the model
“They use these output scores to update their model through a novel technique”
YouTube Unverified1 source· Talk
- uses
“one of the first large models to achieve top-tier results purely through reinforcement learning”
YouTube Unverified1 source· Talk
- featureDeepSeek-R1: Cold-start fine-tuning on structured reasoning examples improves output comprehensibility and addresses language mixing
“fine-tuning on structured reasoning examples before RL to get R1”
YouTube Unverified1 source· Talk
- metrico3-mini: Outperforms R1 and o1 on key benchmarks
“o3-mini, which outperforms R1 and o1 on key benchmarks”
YouTube Unverified1 source· Talk
- featureDeepSeek-R1: Freely accessible through DeepSeek's website and app, and free to download, run locally, and customize
“R1 is freely accessible through their website and app, and it's free to download, run locally, and customize”
YouTube Unverified1 source· Talk
- metricDeepSeek-V3: $5.5 million for the final training run only; excludes R1 training, R&D, and hardware operating expenses
“the $5.5 million figure refers only to the cost of the final training run for V3”
YouTube Unverified1 source· Talk
- metricY Combinator: $500,000 investment for accepted applicants to the first Spring Batch
“If you're accepted, you'll receive $500,000 in investment”
YouTube Unverified1 source· Talk
