八月三十日

Builder Zara Zhang argues that AI products are multiplying faster than real differentiation. Her complaint is not just aesthetic: users are asked to learn yet another similar tool and grant access to the same personal systems before the product has proven why it deserves that trust. The useful signal is that distribution may be getting easier while product clarity and permission design are getting worse.

Builder Zara Zhang 认为，AI 产品增长的速度已经快过真实差异化的形成。她的不满不只是审美层面的，而是用户被迫去学习又一个相似工具，并在产品尚未证明自己值得信任之前，就先交出对同一批个人系统的访问权限。这里更有价值的信号是，分发也许变得更容易了，但产品清晰度和 permission 设计正在变差。

Vercel CEO Guillermo Rauch says `fx` 0.0.7 makes two deliberate bets: stronger MCP support after testing against services including Context7, Datadog, MongoDB, Linear, Notion, Supabase, Neon, Stagehand, and Kernel, and a narrower tool surface with better shell execution. His point is operational rather than cosmetic. CLI agents fail in the messy parts of process control, where commands hang, run long, emit huge output, or emit nothing, so a smaller interface plus better execution handling is itself a product strategy.

Vercel CEO Guillermo Rauch 表示，`fx` 0.0.7 做了两个明确取舍：一是在 Context7、Datadog、MongoDB、Linear、Notion、Supabase、Neon、Stagehand 和 Kernel 等服务上测试后，强化了 MCP 支持；二是收窄 tool surface，并改进 shell execution。他强调的不是表面功能，而是运行层现实。CLI agent 往往失败在 process control 的脏活上，也就是 command 会挂起、跑很久、输出过大，或者根本没有输出，因此更小的接口加上更可靠的执行控制，本身就是一种产品策略。

FPV Ventures partner Nikunj Kothari predicts that the current fight among large model labs will make subsidized token pricing harder to sustain. His claim is that compute shortages and true infrastructure costs will become more visible within six months, pushing up effective prices and tightening reset limits, while open model harnesses benefit from the gap. This is a forecast, not a confirmed market move, but it is a useful reminder that many agent economics still sit on top of temporary lab subsidies.

FPV Ventures 合伙人 Nikunj Kothari 预测，大型 model labs 之间的竞争会让当下被补贴的 token 定价更难持续。他的判断是，在未来六个月内，compute 紧缺和真实基础设施成本会变得更显性，从而推高有效价格并收紧 reset limits，而开放的 model harnesses 会从这个价差中受益。这还是预测，不是已经确认的市场变化，但它提醒人们：许多 agent economics 仍然建立在实验室阶段的临时补贴之上。

FirstMark Capital partner Matt Turck frames recursive self-improvement as the frontier labs' shared destination: AI systems helping build better AI systems. He attributes a sharper scenario to Ryan Greenblatt, who says this path could lead to AI takeover as early as late 2028 or 2029. The timeline itself remains an attributed claim rather than established fact, but the broader signal is clear: automated AI R&D is increasingly treated as the strategic endgame.

FirstMark Capital 合伙人 Matt Turck 把 recursive self-improvement 描述为 frontier labs 的共同终点，也就是让 AI 系统帮助构建更强的 AI 系统。他把一个更激进的判断归于 Ryan Greenblatt：这条路径最早可能在 2028 年末或 2029 年走向 AI takeover。这个时间点本身仍然只是带有明确出处的说法，而不是既定事实，但更大的信号很清楚：自动化 AI R&D 正越来越被视为战略终局。

The Takeaway: Science founder and CEO Max Hodak is turning a big theory about the brain into a commercial product by treating neural interfaces as an engineering discipline, not a distant moonshot.

核心结论：Science 创始人兼 CEO Max Hodak 正把关于 brain 的宏大理论变成可商业化产品，他把 neural interface 视为一个工程学科，而不是遥远的登月项目。

Hodak says Prima, Science's retinal prosthesis, received marketing approval in Europe in July 2026 and that first sales would begin in the coming weeks. The device uses a tiny implant beneath the retina plus glasses that project an image onto it, bypassing dead rods and cones to restore a form of vision for blind patients. The current result is still limited, with a narrow black-and-white field of view, but he says the roadmap is concrete: better grayscale, then at least red and green, with blue remaining harder.

Hodak 表示，Science 的视网膜假体 Prima 已在 2026 年 7 月获得欧洲 marketing approval，首批销售将在接下来的几周开始。这套设备由视网膜下方的微型 implant 和一副把图像投射到 implant 上的眼镜组成，通过绕过已经失效的 rods and cones，给失明患者恢复一种可用的视觉。当前结果仍有限，视野狭窄且只有黑白，但他给出的 roadmap 很具体：先提升 grayscale，再争取实现红色和绿色，而蓝色会更难。

What makes the conversation stand out is Hodak's conviction that device iteration can outperform the slower logic of drug discovery. He points to trial patients doing Sudoku, crossword puzzles, and even reading books, then states the premise plainly: "The brain very literally, very clearly, plainly is a computer." That view extends beyond sight restoration into alignable representations between brains and AI models, and eventually toward substrate independence, but the near-term proof is simpler: a brain-computer interface has already crossed into a sellable medical business.

这段内容最突出的地方，在于 Hodak 认定 device 的持续迭代可以胜过 drug discovery 更慢的逻辑。他举出临床试验患者已经能做 Sudoku、填 crossword puzzle，甚至读书的例子，然后把自己的前提说得非常直接：“The brain very literally, very clearly, plainly is a computer.” 这种看法不仅指向视力恢复，也延伸到 brain 与 AI model 之间可对齐的 representation，以及更远的 substrate independence；但最近期、最有说服力的证明其实更简单：brain-computer interface 已经跨进了可以出售的医疗业务。