JEV 创始人:可靠性优先,拒绝 benchmark 疲劳,机器原生才是未来JEV Creator: Reliability Over Benchmarks, Machine-Native AI Is the Future
The Takeaway: AI 的真正潜力不在于更会讨好人的聊天机器人或刷榜模型,而在于打造可靠、机器原生、可组合的决策系统,让软件真正掌控智能。
JEV(TypeSave)CEO Jago 曾在 OpenAI 工作,如今以“系统一”决策模型冲击行业。他坚持认为,当前 LLM 的大部分调用最终应服务计算机而非人类——风格和散文只是极少数场景,机器需要的是高基数枚举和位级操作。JEV 的目标正是从这一认知核心反向工程。他坦承模型仍远未成熟,存在大量 jaggedness(能力锯齿),但故意公开不完美特性,让开发者自行判断可靠性需求。
Jago 对 benchmark 和“形状崇拜”极度警惕:新形状容易,真正有用且可靠的智能才难。他批评行业过度承诺、快速失望的循环已经让开发者对新型号产生疲劳,并希望未来更多协作——软件公司指出 jaggedness,模型方修复,共同解锁科幻用例。他对 coding agent 的“KV cache 暴政”尤其不满:当前设计几乎都围绕可追加状态,限制了路由、子代理共享上下文、语义压缩等更优雅架构。他畅想依赖类型的语义版本、组织级语义安全检查、脱离 KV cache 的多代理共享状态等。
对 Anthropic 等实验室,他既认可 coding agent 的价值,也怀疑部分估值建立在对 AI 能力边界的误解上——“让他们治癌”可能最终像“让他们吃蛋糕”。他估计 coding agent 对工程生产力的平均提升约 20-40%,且高度依赖场景与用户技能。安全问题上,他区分 ASI 与生物/网络风险,认为真正自主智能仍需重大突破,当前更多是工具增强而非失控代理。最终他呼吁开发者重新获得 agency,一起把 AI 变成可玩、可组合、真正赋能创造的力量。
“the let them cure cancer is going to be like Marie Antoinette’s, like, let them eat cake.”
JEV(TypeSave)CEO Jago 曾在 OpenAI 工作,如今以“系统一”决策模型冲击行业。他坚持认为,当前 LLM 的大部分调用最终应服务计算机而非人类——风格和散文只是极少数场景,机器需要的是高基数枚举和位级操作。JEV 的目标正是从这一认知核心反向工程。他坦承模型仍远未成熟,存在大量 jaggedness(能力锯齿),但故意公开不完美特性,让开发者自行判断可靠性需求。
Jago 对 benchmark 和“形状崇拜”极度警惕:新形状容易,真正有用且可靠的智能才难。他批评行业过度承诺、快速失望的循环已经让开发者对新型号产生疲劳,并希望未来更多协作——软件公司指出 jaggedness,模型方修复,共同解锁科幻用例。他对 coding agent 的“KV cache 暴政”尤其不满:当前设计几乎都围绕可追加状态,限制了路由、子代理共享上下文、语义压缩等更优雅架构。他畅想依赖类型的语义版本、组织级语义安全检查、脱离 KV cache 的多代理共享状态等。
对 Anthropic 等实验室,他既认可 coding agent 的价值,也怀疑部分估值建立在对 AI 能力边界的误解上——“让他们治癌”可能最终像“让他们吃蛋糕”。他估计 coding agent 对工程生产力的平均提升约 20-40%,且高度依赖场景与用户技能。安全问题上,他区分 ASI 与生物/网络风险,认为真正自主智能仍需重大突破,当前更多是工具增强而非失控代理。最终他呼吁开发者重新获得 agency,一起把 AI 变成可玩、可组合、真正赋能创造的力量。
“the let them cure cancer is going to be like Marie Antoinette’s, like, let them eat cake.”
The Takeaway: The real potential of AI lies not in more agreeable chatbots or benchmark-optimized models, but in reliable, machine-native, composable decision systems that let software truly own intelligence.
Jago, CEO of TypeSave and creator of JEV, previously at OpenAI, is pushing “system one” decision models into the industry. He argues that the vast majority of future LLM calls should serve computers rather than humans—style and prose are rare edge cases; machines need high-cardinality enums and bit-level operations. JEV works backward from that cognitive core. He is candid that the models are still far from ready, full of jaggedness, and deliberately ships imperfect properties so developers—not the lab—can decide the reliability profile they need.
He is deeply skeptical of benchmarks and “shape worship”: inventing a new shape is easy; making intelligence actually useful and reliable is hard. The industry’s cycle of over-promise and under-deliver has already created fatigue among developers toward new model releases. He wants far more collaboration—software companies calling out specific jaggedness, model teams fixing it, jointly unlocking sci-fi use cases. He is especially frustrated by the “tyranny of the KV cache” in coding agents: nearly every design decision is constrained by append-only state, blocking elegant routing, shared multi-agent context, semantic compaction, and more. He dreams of semantic dependent typing, organization-level semantic security checks, and multi-agent shared state freed from the cache.
On Anthropic and the labs, he acknowledges coding agents’ value while suspecting some valuations rest on misunderstandings of AI’s actual boundaries—“let them cure cancer” may end up like “let them eat cake.” He estimates coding agents deliver a 20-40% average productivity boost, highly dependent on use case and user skill. On safety he separates ASI discussions from bio/cyber risks, arguing true agency is still multiple major breakthroughs away; current systems are powerful tools, not runaway agents. He ends by urging developers to reclaim agency and co-create AI that is playable, composable, and genuinely generative.
“the let them cure cancer is going to be like Marie Antoinette’s, like, let them eat cake.”
查看原文 →
Jago, CEO of TypeSave and creator of JEV, previously at OpenAI, is pushing “system one” decision models into the industry. He argues that the vast majority of future LLM calls should serve computers rather than humans—style and prose are rare edge cases; machines need high-cardinality enums and bit-level operations. JEV works backward from that cognitive core. He is candid that the models are still far from ready, full of jaggedness, and deliberately ships imperfect properties so developers—not the lab—can decide the reliability profile they need.
He is deeply skeptical of benchmarks and “shape worship”: inventing a new shape is easy; making intelligence actually useful and reliable is hard. The industry’s cycle of over-promise and under-deliver has already created fatigue among developers toward new model releases. He wants far more collaboration—software companies calling out specific jaggedness, model teams fixing it, jointly unlocking sci-fi use cases. He is especially frustrated by the “tyranny of the KV cache” in coding agents: nearly every design decision is constrained by append-only state, blocking elegant routing, shared multi-agent context, semantic compaction, and more. He dreams of semantic dependent typing, organization-level semantic security checks, and multi-agent shared state freed from the cache.
On Anthropic and the labs, he acknowledges coding agents’ value while suspecting some valuations rest on misunderstandings of AI’s actual boundaries—“let them cure cancer” may end up like “let them eat cake.” He estimates coding agents deliver a 20-40% average productivity boost, highly dependent on use case and user skill. On safety he separates ASI discussions from bio/cyber risks, arguing true agency is still multiple major breakthroughs away; current systems are powerful tools, not runaway agents. He ends by urging developers to reclaim agency and co-create AI that is playable, composable, and genuinely generative.
“the let them cure cancer is going to be like Marie Antoinette’s, like, let them eat cake.”