@xiongchun007 I have a real understanding of this direction. When I was working on my own product and it was working with different models, I almost collapsed just adapting to the tool call schemas of each vendor. Different models have vastly different understandings of parameter formats. I feel that if Harness really wants to become independent, the first major obstacle will be how to strike a balance between universality and the details of each vendor's model.
@xiongchun007 This is the dumbest suggestion I've ever seen. DeepSeek will ultimately only make money selling MaaS. Their cost-cutting capabilities are all in model infrastructure; selling SaaS offers no advantage. Liang Sheng's team is extremely valuable; making small profits from SaaS won't recoup their investment. Harness is an extremely serious matter; it's not just software, but the final rule layer for a specific model to maximize its strengths and minimize its weaknesses. The more targeted it is, the stronger its effectiveness.
@xiongchun007 It's too complex. Open-source models should be compatible with other models. Acquiring a good team to develop it would be more flexible than spinning it off independently.
@xiongchun007 Liang Sheng: Xiao Wang, who sits across from me, made this suggestion a while ago, but I emphasized that AGI is the main focus of our company, and everything else is trivial and not worth picking up.
@xiongchun007 This move is a bit like giving your own son to your godfather, but the godfather just happens to need a capable steward. Going independent would allow them to take on more work, not be limited by the ceiling of the model brand, and having a stake in the design also provides a backup plan. The only question is whether the team is willing to leave the big tree and fly solo.
@xiongchun007 Have you ever developed a product? Why would anyone use your Harness? Any player with a decent scale could make their own. What a pipe dream you have of dominating the market!
@xiongchun007 The product manager and the group's strategy chief both believe that tools should primarily serve the group's strategic goals, not the various LLM (Local Management Model) tools on the market. Allocating resources to these LLM tools would be counterproductive. If Liang Sheng's true goal was AGI (Aggregate Girdling), he wouldn't have done it this way. A product manager certainly wants a product to grow and become powerful, but compared to AGI, that product is just a tiny tool.
@xiongchun007 I don't need your advice. Those big model companies are smarter than you. Much of the time, decisions are made based on their big model thinking. They don't make decisions lightly. If any big model company goes under, that means the AI bubble is starting to burst.
@xiongchun007 If I believed you, Deepseek wouldn't be far from going bankrupt. The harness product is an important traffic entry point, mainly forming a data flywheel.
@xiongchun007 Post-training of the model requires training specifically for your own harness product. If you're developing a standalone harness product, what's the difference between it and open-source software like Pi and OpenCode?
@xiongchun007 Hey bro, why do I have to relay your suggestion just because I made one on x? And why do you only mention the advantages and not the disadvantages when making suggestions?
original · zh
@xiongchun007 不是大哥，为啥你在 x 上提个建议就得给你转达啊。。。而且提建议不说缺点只说优点吗？
@xiongchun007 Actually, nobody in DS wants to do such tedious work as harvesting, and harvesting is easy to surpass. Newly established companies without the support of top geniuses from DS will probably only amount to mediocre work.
@xiongchun007 It will go bankrupt on its first day of independence. Harness needs to access the code repository of DeepSeek's basic model to achieve native deep adaptation, and much of it is jointly developed. Going independent will make it lose its competitiveness.
@xiongchun007 You still don't understand why DeepSeek wanted to create its own Harness instead of using Pi, Codex, etc. It's to create a flywheel through collaborative training. The Harness shell, as a standalone product, isn't nearly as valuable as you imagine.
@xiongchun007 The core goal is still how to better accomplish the task. Personally, I believe that general-purpose tools for all mainstream LLM models cannot achieve this goal better.
@xiongchun007 The very first sentence doesn't fit their approach. They don't even make products. Saying that Deepseek Harness is an agent product is less accurate than saying it's part of a post-training pipeline.
@xiongchun007 That's an incredibly stupid suggestion. Deepseek got to where it is today precisely because of its self-restraint. Why bother with AGI when you're just jumping around with random ideas?
@xiongchun007 His goal is AGI (Automatic Gaining Intelligence), and harnessing is the core, which cannot be separated. His purpose is not to make money, but to enable AI to learn on its own, rather than making it a toy with limited context.
@xiongchun007 That's right. Each tool uses different standards, requiring numerous modifications to repetitive logic for each adaptation, a pure waste of development effort. Creating a universal, open Harness platform perfectly addresses this pain point of multi-model adaptation, making it incredibly attractive to developers.
@xiongchun007's suggestion is reasonable, but the practical constraint is that if DeepSeek Harness relies on the internal interfaces of the DeepSeek model, the maintenance cost will increase after it becomes independent. It would be better to first achieve loose coupling through an open-source license and plug-in design before considering corporatization.
@xiongchun007 General-purpose harnesses are certainly a separate category, but I personally believe that harnesses and their models should ideally be distinct parts of a whole, like the model being the brain and the harness being the hands and feet.
@xiongchun007 I guess your Uncle Liang made Hermess so that it would be comfortable for the company's employees to use. He doesn't care about the compatibility with mainstream LLMs. He only needs to consider his own model.
@xiongchun007 I've run a harvester evaluation program myself, and the pitfall I encountered was that normalizing the output format when comparing multiple models was far more difficult than I imagined. Just aligning the JSON structures returned by each provider took a week. In the end, I found the scoring template to be even more problematic: in several customer service scenarios, automated scoring and human judgment kept misaligning. How do you plan to verify the credibility of the scoring results for your harvester?
