A few thoughts on DeepSeek’s latest release 1. What impressed me most is its a… · @damnang2 · TwKit

@damnang2

A few thoughts on DeepSeek’s latest release 1. What impressed me most is its ability to create tasks and environments for reinforcement learning at scale. DeepSeek says it improved performance by increasing the diversity and quality of training tasks and strengthening verification, while retaining established methods of supervised learning, reinforcement learning and distillation. To me, this reflects an important shift in AI competition. Even with the same learning algorithm, performance depends on which problems a model tackles and how accurately its results are rewarded. I expect companies with stronger training systems to achieve greater performance gains from additional compute. 2. Failures in actual use are becoming valuable competitive assets. DeepSeek recreates failures encountered by employees and partners in their daily work and turns them into new reinforcement learning tasks. Weaknesses discovered during use become specific targets for subsequent training. This is why I believe access to publicly available models alone will struggle to close the gap between companies. You can download and run a model, but you do not automatically gain the experience of understanding where customers struggle or the environments needed to reproduce those failures. As this process repeats, companies embedded more deeply in actual workflows will improve faster. 3. Agents are also helping build the environments they learn in. DeepSeek assigns different agents to construct training environments, solve problems, review results and fix errors. The process also checks for leaked answers and opportunities to exploit the reward system. This makes verification even more important. Automatically generating more tasks can also generate more flawed tasks and misleading rewards. The ability to produce training environments must improve alongside the ability to keep them reliable. DeepSeek’s strength is that it has also built the infrastructure to support this process across millions of sandboxes. 4. Memory optimization makes broader use of agents economically viable. According to the report, the global KV cache footprint falls to roughly one quarter of that of the previous Flash model, while the persistent KV cache footprint falls to roughly one eighth. These figures refer to KV caches. They do not mean that total memory requirements, including model weights, fall by the same proportions. I do not interpret this as a structural negative for overall HBM demand. Lower memory requirements per task create room for longer tasks and more agents running simultaneously. My expectation is that, over time, usage growth driven by lower costs will more than offset those savings. AI will become affordable for tasks that are currently too expensive, and running multiple agents in parallel to attempt solutions and verify results will become more common. I see this optimization as a technology that makes that expansion economically possible.

曝光 3101 · 评论 0 · 点赞 8 · 书签 1