Instinct 微调的浏览器模型会被 Jev 取代吗

Vincent Nallatamby @VNallatamby原推发布于 2026.09.18

在 X 上打开原推

Instinct finetuned a kimi model for browser use, is it useless now with Jev? as a startup founder, knowing if it's worth finetuning your own model has been a recurring question for the past 3 years. interestingly, in just 1 month we had both the argument and its counter. instinct was massive partly because their finetuned Kimi model was super fast and efficient for browser-use. it was a really interesting wedge for them, and worked beautifully just 1 month later, Jev seems to be a potential paradigm change for computer-use, and i'd not be surprised if everyone starts adopting it in the coming months (including Instinct), potentially making the use of a finetuned kimi model useless for browser-use. so was finetuning a good strategy for instinct? i mean, clearly yes, the results are obvious. but we should see it as it is: as a wedge, not as a moat. the finetuning wedge strategy worked for instinct, but i'm sure there are dozen of startups out there who spent millions ultimately for nothing I believe a possible good strategy for startups is: 1. finetune a model to push the frontier on a new capability, 2. productize it asap and uses your model as a wedge, 3. as soon as other models are better on your core capability, switch asap, 4. keep building a proprietary dataset, 5. when you're big enough (like millions of users), kickstart again the finetuning effort Ultimately what matters is your ability to time the market. if you're too early, you'll spend too much time on the research and your product will suck anyway. if you're too late, you'll end up getting commoditized in 3 months If you're building at the applicative layer, you need to be 6 months ahead of the frontier, max