The Engineering Unlocks Behind DeepSeek | YC Decoded

there's a new AI model in town Chinese AI company deep seek recently made ways when it announced R1 an open source reasoning model that it claimed achieve comparable performance to open AI 01 at a fraction of the cost the announcement Unleashed a wave of social media panic and stock market chaos Nvidia losing nearly $600 billion doll in market cap today alone but for those following AI developments closely deep seek on R1 didn't come out of nowhere the company has been publishing its research and releasing its model weights for months following a pth similar to meta's Lama model this is in contrast to other major AI Labs like open AI Google deep mine and anthropic that have close weights and publish more limited technical reports what's changed is just that now the broader public is actually paying attention so let's decode what the real deel elements here are where they come from and why they [Music] matter first of all it is important to distinguish between two relevant models here deep seek R1 and deep seek V3 deep seek V3 which was actually released this past December is a general purpose based model that achieves comparable performance to other base models like open ai's GPT 40 anthropics claw 3. 5 Sonet and Google's Gemini 1. 5 deep seek R1 which is released at the end of January is a reasoning model built on top of deep seek V3 in other words deeps took B3 and applied various algorithmic improvements to it in order to optimize its reasoning ability resulting in R1 a model that's achieved comparable performance to open a01 and Google Flash 2.

0 on certain complex reasoning benchmarks but many of the algorithmic Innovations responsible for r1's remarkable performance were actually discussed in this past December V3 paper or even before that in deep seeks V2 paper which was published in May 2024 or the Deep seek math paper which came out February 2024 V3 stitches together many of these Innovations which were designed primarily with compute and training efficiency in Mind One Way de seek optimized for efficiency and got more floating Point operations per second or flops from the gpus was by trading V3 natively in 8bit floating Point format rather than the usual 16bit or 32-bit format this is not a new idea many other labs are doing it too but it was key for getting such massive memory savings without sacrificing performance a crucial enhancement is their fp8 accumulation fix which periodically merges calculations back into a higher position fp32 accumulator to prevent small numerical errors from compounding the result far more efficient training across thousands of gpus cutting cost while maintaining model quality but why does this efficiency matter given its Hardware constraints and US exports controls on the sale of gpus to China deep seek needed to find a way to get more training and more bandwidth from their existing cluster of gpus you see at AI Labs these gpus which do number crunching and matrix multiplication to train these models are actually sitting idle most of the time at fp8 it is typical to only see around 35% model flops utilization or mfu meaning gpus are only being utilized at Peak potential about a third of the time the rest of the time these gpus are waiting for data to be moved either between caches or other gpus this is nvidia's key Advantage it is not just about gpus it is about an integrated solution they've been building for over a decade that includes the networking with infinity band software with Cuda and developer experience essentially Nvidia provides a deeply integrated system that lets AI researchers program GPU cluster L is a distributed system and and closer to what Jensen describes is one giant GPU another clever way deep seek makes the most out of their Hardware is their particular implementation of a mixture of experts architecture deep seek V3 has 671 billion modern parameters but only 37 billion are activated for a given token prediction by contrast the largest and most capable Lama 3 Model doesn't use a mixture of expert architecture so it activates its full 405 billion for each token prediction in other words V3 activates 11x fewer parameters for each forward pass saving tons of computation mixture of experts isn't A New Concept but it's been challenging to train models with this architecture efficiently deep seek introduced novel techniques that stabilize performance and increase GPU utilization additionally to overcome key performance bottlenecks V3 makes use of multi head latent attention or MLA which deep seek first revealed with its V2 paper which was published in May 2024 MLA is a solution designed to tackle KV cat storage limitation one of the biggest sources of bam overhead in large models instead of storing full key and value matrices MLA manages to compress press them down into a latent representation reconstructing them only when needed this helped the B2 model reduce his KV cach size by 93. 3% and boosted its maximum generation throughput to 5. 76 times finally unlike traditional models that predict only the next token V3 makes use of multi- token prediction or MTP MTP enables V3 to anticipate multiple future tokens at each each step this densifies training signals providing more feedback per step for better data efficiency and faster learning it also improves representation planning allowing the model to pre-plan sequences for smoother more coherent outputs during inference MTP modules can be repurposed for speculative decoding reducing sequential processing steps and significantly speeding up generation taken alog together this makes V3 one of the most impressive based models on the market and it's been out for some time now however the recent release of deep seeks R1 reasoning model is what really made waves most llms can be improved by being prompted to think step by step but what sets reasoning models apart is that they are specifically trained to break down hard problems and think about them for paragraphs at a time in September open AI showed the power of this new approach with 01 this achieved state-of-the-art results in math coding and science benchmarks with R1 deep seek took a similar approach and published The Secret Sauce open Ai and deep seek achieve their impressive results through reinforcement learning a technique to shape an llms Behavior based on feedback and reward signals modern llms use some variation of reinforcement learning with human feedback AKA RL HF or reinforcement learning from AI feedback AKA RL AI F to improve their models usefulness and Alignment but reasoning models apply RL specifically towards the task of thinking step by step through complex problems so how did deep seek apply RL to get a reasoning model at a high level they assemble a bunch of problems with verifiable outputs especially in math and coding problems and then design a training pipeline to get the model to think for a bit and output the correct answers but they don't give the model any external examples of how to think whether from humans or Ai and Grading process was extremely simple rather than using a complex AI to give the model fine grain feedback deep seek uses simple rules to evaluate the model's final output on accuracy and formatting they use these output scores to update their model through a novel technique they published in February 2024 called group Rel relative policy optimization or grpo remarkably with this process alone deep seek saw reasoning emerg over thousands of RL steps the model learned skills like extended Chain of Thought and even experience a aha moment where it recognized its own mistakes and backtracked to correct its reasoning this model was r10 one of the first large models to achieve top tier results purely through enforcement learning pure RL has long been a subject of Investigation in Western research Labs such as deep Minds alphao which simulated thousands of random games of self-play to beat Lisa do the world's top go player in 2016 in 2019 open aai achieved notable success using reinforcement learning to train a robotics hand to solve a Rubik's Cube and beat a top human team in competitive Dota 2 but unconstrained by human examples R1 Z's thinking steps suffer from poor readability switching between English and Chinese at random so deep seek introduced a cold start phase fine-tuning on structured reasoning examples before RL to get our one this eliminated the language mixing issues and made outputs far more comprehensible the results are impressive R1 achieves comparable performance to 01 on certain math and coding benchmarks but the pace of innovation is speeding up just 2 weeks after R1 was released open AI release 03 Mei which outperforms R1 and 01 on key benchmarks so if R1 didn't actually come out of nowhere what explains the hype cycle one explanation is the sheer accessibility of deep seeks model ran is freely accessible through their website and app and it is free to download run locally and customized also because of all the efficiency improvements it offers near state-of-the-art performance at a fraction of the price of other reasoning models another explanation is that a lot of the hype cycle didn't actually have to do with the specific algorithmic improvements that we describ but with misconceptions around v3's alleged $5.

5 million in training cost there's some important fine print here the 5. 5 million figure refers only to the cost of the final training run for V3 it doesn't include any of the training cost of R1 or the associated R&D or Hardware operating expenses which are presumably in the hundreds of millions given the extreme algorithmic optimizations here that 5.