AI Builders
Digest

Andrej Karpathy指出，对LLM真实能力的共享认知鸿沟正在急剧扩大，形成尖锐漏斗。约60亿人（75%）几乎未接触过LLM；10-20亿人（20%）只是偶尔使用免费ChatGPT类产品，视其为更好的搜索或写作助手；约2000万人（0.2%）在数学与代码领域专业使用前沿LLM，亲眼看到复杂项目可被agent在提示下完成；而仅约5000人（0.00006%）能接触内部前沿系统，体验到数千agent协作数周完成软件巨型项目、零日漏洞、机器速度网络攻防、新科学发现等。驱动因素包括影响随野心与问题规模放大、系统在数字可验证高价值领域的“锯齿状”能力，以及访问层级差异。公众大多只见过呆萌聊天机器人，多数专业人士仅获温和提升，而极少数人正经历曲线垂直上升的眩晕。

Andrej Karpathy notes that the shared understanding of LLM capabilities is widening into a sharp funnel. Roughly 6B people (~75%) have barely touched LLMs; 1-2B (~20%) are casual free-tier users treating them as better search or writing aids; ~20M (0.2%) professionals in math & code see agents complete complex projects that once took weeks or months; and only ~5,000 (~0.00006%) with internal frontier access witness swarms of agents collaborating for weeks on mega-projects, zero-days, machine-speed cyber ops, and new science. The funnel is driven by impact scaling with ambition, the jaggedness of capabilities in digital/verifiable domains, and access tiers. The public mostly sees derpy chatbots while a tiny sliver experiences the curve going vertical.

Box CEO Aaron Levie回应Karpathy的漏斗分析，强调我们仍处于AI扩散的极早期。绝大多数人把AI当聊天机器人做简单交互，近似下一代搜索，这远未体现全部能力。另一端，agent swarm可能是agent在工作流中的终极形态，适用于数学、网络、生物、材料科学、软件工程、金融等任何可并行执行且能用算力加速的高价值任务。目前只有极少数人这样使用，原因包括模型能力与用户领域的匹配度、成本、可用harness以及前沿模型的可及性。变革管理、工作流升级、数据迁移与合规框架是关键障碍，但这正是应用层AI在企业中发光的机会。

Box CEO Aaron Levie builds on Karpathy’s funnel, stressing how early AI diffusion remains. Most people still use AI as a chatbot for simple next-gen search interactions that dramatically underrepresent full capability. On the other end, agent swarms may be the ultimate form factor for workflows in math, cyber, bio, materials, software engineering, finance and any domain with parallelizable, compute-accelerable valuable tasks. Only a fraction of a fraction of the population operates this way today due to capability-domain fit, cost, harness availability and limited frontier-model access. Change management, workflow upgrades, data migration and compliance frameworks remain major barriers, yet this is precisely where the applied AI layer will shine in the enterprise.

Y Combinator President & CEO Garry Tan呼应Karpathy，鼓励人们成为那5000人中的一员。幸运的是价格会下降，更多人将看到并有机会把它带到世界。他称之为人类历史上最伟大的alpha源泉。他还指出，用确定性代码替换LLM调用是显而易见的选择，能产出更符合用户需求的更好代码。

Y Combinator President & CEO Garry Tan echoes Karpathy, urging people to join the 5,000 today. The lucky part is that price will come down and more people will see it and get a chance to bring it to the world. He calls it the greatest alpha fountain in the history of man. Separately he notes that replacing LLM calls with deterministic code is a no-brainer that also produces better code better matched to what users want.

Anthropic Claude Code工程师Thariq认同Terrence Tao的方向：我们不应满足于不透明的超级智能。智能的关键信号是沟通。我们应努力打造与人类建立理解而非取代人类的产品和模型。

Anthropic Claude Code engineer Thariq agrees with Terrence Tao’s direction: we should not settle for opaque superintelligence. A key signal of intelligence is communication. We should strive to make products and models that build understanding with humans instead of disintermediating them.

OpenAI Codex & ChatGPT的Thibault Sottiaux宣布，GPT-6.1 Sol在Codex & Work中的默认上下文长度从272k提升至500k，约80%增长，最高仍可达1M。此举经过仔细测试，成为大多数用户的新甜点，可减少compaction并支持需要长期保持大量上下文与细微差别的任务。通过订阅使用时不再收取长上下文费用。他原本预期用户会更偏向2X用量换回5小时限制以平滑需求曲线，但投票结果让他意外。

OpenAI Codex & ChatGPT’s Thibault Sottiaux announced that default context length for GPT-6.1 Sol in Codex & Work has been raised from 272k to 500k (~80% increase), with a maximum still at 1M. The change was carefully tested and is the new sweet spot for most users, reducing compactions and better supporting tasks that require keeping lots of context and nuance over time. Long-context fees are no longer charged when drawing usage from the subscription. He had expected stronger preference for 2X usage at the cost of bringing back 5h limits to smooth demand, but poll results surprised him.

Peter Yang分享了Micky（@Rasmic，已教超过10万人用AI构建）的观点：目前还不会完全信任agent，尤其是在有人使用的东西上。手动审查是最后一步，所有工作已完成，并尽可能从agent那里获得证据。通过证据驱动测试（前后对比甚至greploop），agent呈现证据：这是已完成的工作，以及为什么它是好的。对任何有经验的开发者来说，手动审查往往是工作中最糟糕的部分，这种方式有助于最小化摩擦。Peter还展示了与女儿一起用Codex构建modretro游戏的过程。

Peter Yang highlights Micky (@Rasmic, who has taught 100K+ people to build with AI): “I’m not at a point where I trust the agent, especially with stuff that people are using.” Manual review is the final step after all the work is done and as much evidence as possible has been gathered from the agent. With evidence-driven testing (before/after and even greploop), the agent presents evidence of what was done and why it is good. For any experienced developer, manual review is often the worst part of the job; this approach helps minimize that friction. Peter also shared building a modretro game with his daughter using Codex.

FirstMark Capital的Matt Turck指出，人们曾经是打孔操作员，然后是Lotus Notes管理员，接着是Flash开发者和Dreamweaver网页管理员，再变成DevOps和Tableau分析师，现在正在成为agent管理者和harness工程师。一切都会顺利的。

FirstMark Capital’s Matt Turck observes that people were once punch-card operators, then Lotus Notes admins, then Flash developers and Dreamweaver webmasters, then DevOps and Tableau analysts, and now they are becoming agent managers and harness engineers. It will work out just fine.

The Takeaway：Fractile通过自研全栈推理芯片追求极致内存带宽，目标是把多万亿参数模型以每秒数千token的速度运行，从而让长运行agent实现数量级加速，并在芯片领域结构性获得3-6个月优势。

Fractile创始人兼CEO Walter Goodwin在No Priors播客中阐述了公司从2022年夏天成立以来的核心押注。早期他们像Grok或Cerberus一样做SRAM基芯片，以获得超高带宽，但很快意识到上下文长度爆炸带来的扩展性瓶颈。过去两年他们转向与内存厂商和逻辑foundry合作，寻找通往高容量低成本DRAM的极高带宽路径，目标是把GPU/TPU的扩展性与SRAM芯片的速度优势统一起来。平台计划于明年下半年量产。

他强调，真正让速度改变游戏规则的不是更灵敏的聊天机器人，而是让长运行agent大幅提速：“这是AI能力的根本提升器。”芯片必须同时具备极高带宽（以便每秒数千次加载权重与状态）和经济的内存成本（数据中心规模推理的经济性最终归结为每GB内存成本）。

Fractile约150人，覆盖从工作负载理解、前端设计、物理设计到先进封装的全栈，形成闭环敏捷循环。这让他们能比传统“交接给Broadcom类ASIC房屋”的模式更快迭代。芯片设计仍有物理与财务约束（流片3-5个月，3-5年摊销窗口），但压缩从观察到量产的延迟能带来更多“射门机会”，从而结构性获得3-6个月优势——在芯片领域这相当于前沿模型实验室的窗口优势。

“如果你能在结构上挖出三到六个月的优势，你就会赢下所有这些部署。”

The Takeaway: Fractile is building full-stack inference chips optimized for extreme memory bandwidth so that multi-trillion-parameter models can run at many thousands of tokens per second, unlocking radical speed-ups for long-running agents and a structural 3-6 month advantage in the chip race.

Fractile founder and CEO Walter Goodwin explains the company’s journey since summer 2022. Early work focused on SRAM-based designs (à la Grok/Cerberus) for ultra-high on-die bandwidth, but exploding context lengths exposed scalability limits. For the past two years the team has pursued high-bandwidth access to higher-capacity, lower-cost DRAM in collaboration with memory vendors and logic foundries, aiming to combine GPU/TPU-style scalability with SRAM-chip speed. The resulting platform is slated to ramp in the second half of next year.

The real payoff is not snappier chatbots but making long-running agents dramatically faster: “the ability to take a multi-trillion parameter model and run it comfortably at many thousands of tokens per second… is a fundamental elevator on capability for AI.” Chips must deliver both extreme bandwidth (to reload weights and state thousands of times per second) and economical memory, because data-center-scale inference economics collapse to cost per gigabyte of memory.

With ~150 people spanning workload insight, front-end design, physical design and advanced packaging, Fractile runs a closed agile loop instead of the traditional hand-off to ASIC houses such as Broadcom. Physical and financial realities (3-5 month fab cycles, 3-5 year amortization) still constrain how fast new silicon can ship, yet compressing the observation-to-volume latency creates more shots on goal and a structural 3-6 month edge—the chip-world equivalent of a frontier-lab model advantage.

“If you can just find a way to structurally carve out a three to six months advantage, you will be winning all of those deployments.”

多数其余推文为轻量互动或个人分享，未形成独立实质性洞察。包括Swyx转发演讲技巧、Guillermo Rauch对学习未来的简短评论、Zara Zhang仅在Chrome使用Gemini、Dan Shipper的个人读书feed实验等。

Most remaining posts were light engagement or personal shares without standalone substantive insight, including Swyx re-upping talk tips, Guillermo Rauch’s brief note on the future of learning, Zara Zhang using Gemini only inside Chrome, and Dan Shipper’s personal book-reading feed experiment.