AI 投资人的五条 maxi 判断更新

Jared Sleeper @JaredSleeper原推发布于 2026.10.11

在 X 上打开原推

Updates + additions to my AI maxi/mini takes. Maxi takes: 1) Cybersecurity-related AI spend will go to the moon as every public-facing app will need regular pen-testing. We see this driven by great startups like @Xbow et al. and cybersecurity continues to be top of mind for the world's enterprises in our survey data. I personally mash the security scan button on @DevinAI regularly and should probably get around to having it auto-run every day. 2) AI for code willingness-to-spend has another big leg up as companies roll out faster models that are enormous compute hogs. See: Astra Ultrafast. SWEs don't want to wait for things to get done, they just want them done. This requires ungodly amounts of compute. Intelligence may saturate but speed will not. 3) Enterprise/individual adoption disparity remains enormous (and growing!). There is enormous room for diffusion alone to drive more usage. My trainer recently told me that he listens to a 20-minute podcast that @Muse makes for him every morning with updates on all of the topics he cares about. Pretty amazing. 4) Video/image generation has finally reached a point of utility, but has a long way to go still before we're engaging with real-time 3D worlds in 4K. @burkaygur and the @fal team dropped a model (H3 Max) that generates a frontier-quality five-second clip in less than five seconds just before Labor Day, and that might genuinely be the ChatGPT moment for video gen. I'm seeing this in real time at my portfolio company @airHQ, where AI usage is exploding, especially in image-to-video use cases. 5) We're getting to the point where agents can take on longer/more goal-oriented tasks and humans are no longer bottlenecks on spend. Companies like @sailresearchco are really pressing on the infrastructure for this, and a combination of long-running VMs and well-defined goal setting can easily balloon tokens/person. 6) My most maxi take of all is that we're going to continue to see incremental inference spend being driven by "dark inference" use cases that break the typical investor frameworks- companies pursuing advanced R&D spending millions on tokens, HFT firms, etc. If AI can crush unsolved math problems, it isn't hard to imagine that ~50% of all spend on frontier R&D should be tokens... and we're a long, long way from that. Mini takes: 1) Vibecoding is officially intelligence-saturated, and the word is slowly getting out. I almost exclusively use @DevinAI's fusion model, which is excellent and cost-effective, and (aside from speed) I can't imagine it being much better at what I'm asking it to do. 2) More and more anecdata is starting to emerge about enterprise spend rationalization/moves to open source for some of the AI native ISVs. The reality is, "enterprise" as SaaS investors think of it probably isn't a huge driver of long-term AI spend- it is going to be dirt cheap for AI to replace the sum total of all intelligence that went into managing, say, a big company's ServiceNow instance. My gut is that enterprise AI spend (at least the RPA-adjacent spend many investors think of when they imagine this) disappoints on a relative basis. 3) The compute shortage, as it stands today, is STILL largely driven by post-training spend from Ant/OAI/Meta/Google/etc. as well as a slew of neolabs and VC-backed companies. This spend is highly vulnerable to capital markets/animal spirits and could reverse in months if perceived returns on training spend fall and VCs stop backing incremental labs and/or OAI/Ant/hyperscalers rationalize spend. This is the most likely scenario for a compute glut. I think @GavinSBaker is right in the long run that open-weight models are good for infra, but I do think the current supply/demand dynamic is driven by a "race to the frontier" and the assumption that being at the frontier is incredibly valuable... I don't think you want to be around for the AI trade if that assumption breaks. 4) Still likely that distillation attacks of one kind or another are a meaningful share of OAI/Ant rev. This ties neatly into #3. 5) Models eat themselves by designing cheaper versions that achieve the same goals. So far, this is arguably where RSI has shown up the most. Jev is one example- I know companies where it cut spend on certain AI use cases by 90%+. In other words, AI tools themselves are accelerating the march toward cheap/ubiquitous intelligence and hastening potential saturation, especially because... 6) Frontier models are now being regularly delayed due to security risks. If the most valuable products are too dangerous to release AND #5 continues to play out, that is powerful deflationary pressure on the AI stack.