Jie Liu retweeted
Help shape the vLLM Semantic Router roadmap for Q4 2026–Q1 2027.
Rate the features you need, add what’s missing, or propose a new work group. Your feedback sets priorities for upcoming releases.
~8 minutes. No sign-in.
form.vllm-sr.ai/roadmap
Jie Liu retweeted
让 Agent 自己优化大模型训练,可能会出现性能报告写着「提速 2 倍」,实际却只是减少了计算量,甚至改了计时规则。
如果连这种「假进步」都识别不出来,Agent 迭代得越多,反而可能偏离真正的优化目标。
最近一篇论文《InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure》,探索了如何让 Agent 真正具备持续优化训练系统的能力。
研究团队开发了 InfraRSI,一套面向分布式大模型训练的 Agent Harness。
它的核心思路是让 Agent 形成一个持续自我改进的实验闭环。
Agent 会分析性能瓶颈,尝试修改训练配置和底层代码,通过 Verifier 检查优化是否真实有效,再把成功经验和失败原因保存下来,用于指导下一轮优化。
这里有两个关键设计:一是通过 16 项约束检查,防止 Agent 通过减少计算量或修改计时规则来「刷分」;二是让失败成为下一轮搜索的依据,比如遇到 OOM 后,尝试组合其他节省显存的方案,而不是直接放弃。
实验结果还是很有说服力的。
在四组 Dense 和 MoE 训练任务中,作者报告 InfraRSI 实现了 1.83 倍的几何平均加速,高于 RE-Bench 的 1.37 倍和 Auto-Research 的 1.34 倍。
其中,在 Gemma-4-26B-A4B 的 8×H200 训练任务上,Agent 通过组合多种优化方案,实现了 2.75 倍加速。不过,目前的结果仍集中在特定模型和硬件环境中。
这篇论文至少提醒我们需要关注 Agent 自我改进背后的 Harness 设计。
Agent 要实现持续改进,需要一套可靠的验证与反馈机制,知道哪些改动真的有效,为什么失败,以及下一次应该往哪里探索。
Agent 的持续自我改进能力,很大程度上取决于它能否从一次次经过验证的实验中积累经验。
📎 arXiv: arxiv.org/abs/2611.11558
💻 GitHub: github.com/TokenRhythm/Infra…
针对 LLM GPU kernel 库的系统性实证研究
An Empirical Study of Issues in GPU Kernel Libraries for LLM Systems
arxiv.org/abs/2611.09280
GPU kernel 库出问题,第一反应是去看 kernel 代码。但 300 个真实 issue 里,六成的根因根本不在 kernel 里。
这是首个针对 LLM GPU kernel 库的系统性实证研究。作者手工分析了 300 个高质量 GitHub issue,连同讨论和修复,覆盖七个主流库:FlashInfer、FlashAttention、Triton、CUTLASS、DeepGEMM、xFormers、bitsandbytes。
症状分布很说明问题:部署失败 27.33%、构建失败 24.33%、崩溃 21.00%、功能错误 17.67%、性能退化 7.67%、挂起 2.00%。也就是说,「跑不起来」比「算错了」更常见,性能问题反而是少数。
根因拆开看:库本身占 39.67%,运行时环境 33.33%,编译器与构建系统 27.00%。后两项加起来 60.33%,所以验证范围必须覆盖部署和执行栈,只测 kernel 代码是不够的。
第二个发现更隐蔽:kernel specialization 会给张量形状、元数据布局、内存别名、硬件资源加上隐式约束。就算这个算子本身是支持的,一旦越界,也可能静默算错或者直接起不来。
第三个是工具缺口:现有的测试和调试手段,对付低精度数值问题和硬件相关的执行失败都不够用。
判断:kernel 库的「能用」比「够快」更容易出事。以后做这类库的验证,先问部署栈测没测,再看性能跑分。
arxiv.org/abs/2611.09280
23
Customer intimacy is moat.
Customer intimacy is a strategy that creates a competitive advantage that allows the organization to fully engage with customers. It means segmenting and targeting markets, tailoring offerings, and predicting behaviors to match the demands of niches.
Replying to @FinanceYF5
7/12 【最好的护城河不是模型】
AI 底层技术变得太快,今天的突破就是明天的基线。
所以护城河不只是技术能力,而是对客户足够贴近、足够了解(customer intimacy)。
1
1
33
Sequoia AI Ascent 2026: The future of AI theaiopportunities.com/p/seq…
1
8
2026: The State of Enterprise AI
A solid study from @MenloVentures
"2026: The State of Enterprise AI"
> Enterprise AI deployment spend jumped from $33B in 2025 to $177B by Sept-2026, a 5.4x rise that beat the prior year’s 3.9x growth and now equals over 30% of public cloud revenue.
> Agents drove the jump: on OpenRouter, agent traffic passed human traffic on Feb-2026, then grew 37x versus 5x for humans, reaching 80% of tokens, or 131T per week.
> Companies budgeted 1.9x their 2025 AI spend but averaged 4.7x, since parallel agents scale spending with work rather than headcount; the median reached only 2.2x, so a few power users drive a disproportionate share.
> Anthropic, OpenAI and Google took 69% of deployment dollars, $123B, up from 50%; Anthropic’s share of enterprise API spend rose from 40% to 55%, OpenAI’s from 27% to 31%, Google’s fell from 21% to 9%.
> Buyers keep paying for capability: frontier models’ spend share rose from 35% to 63%, and the most popular models shifted from Claude Sonnet 4.5 and GPT-5 to pricier Claude Opus 5.5 and GPT-6 Astra.
> Enterprises spend $1.70 using AI for every $1 spent building models, $177B versus $102.5B, and Anthropic and OpenAI alone account for an estimated 25%+ of AWS, Azure and Google Cloud revenue.
> Coding and chatbots make up 76% of the $77.3B app market, since compilers and tests let agents verify results; coding hit $34.0B, with Claude Code and Codex taking $24.0B, or 70%.
> Open-source handles 19% of tokens, up from 11%, but only 4% of spend, as its average price fell 3.3x to $0.13 per million; Chinese models take 89% of open-source spend, led by Kimi at 61%.
> AI is expanding software budgets: 55% of incremental AI spend came from new budgets, and only 9%, or $12.6B, shifted from software to AI-native vendors, trimming application SaaS growth from 13% to 11%.
> AI-native providers took 77% of application spend, up from 63%, and 96% of coding, but incumbents hold IT at 91% and HR at 68%, where AI must fit existing systems of record.
1
21
Revolut: Europe's most valuable startup ($200B) just showed the world its AI strategy.
Every company will end up copying it.
At the start of this year, Revolut leaned on the frontier labs. Its 2,000 engineers used Claude Code and Codex, and some burned $1,000 to $10,000 a month each. The bills kept climbing, and the returns weren't there.
So Revolut switched its default to an open-source model called GLM. All it needs is GPUs, which it rents from Together AI.
The results, by Revolut's own count:
- The bill is about a tenth of what it paid OpenAI and Anthropic
- Engineers get unlimited tokens instead of spending caps
- Coding time is down 60% and lead time is down 70%
That's the chart below.
Revolut still uses Claude and GPT, but only as "the exception," for the tasks they do best. Everything else runs on open source.
So here's the question I keep coming back to: how many use cases can companies really justify sending to frontier models?
My answer is very few. Plenty of companies still route everything to OpenAI and Anthropic today. For how long?
I think open source ends up handling the vast majority of corporate AI work, and frontier models get a short list of the hardest jobs. That's a problem for labs whose revenue forecasts assume unlimited growth.
If you appreciate unpopular takes, follow me @m0xt_ for more.
Jie Liu retweeted
Finally releasing the results of the arch ablations I was running with bux over the weekend
Epoch AI做的AI芯片进口的开放数据库,利用全球海关和国内生产数据来估算人工智能芯片的地理分布。可以按照区域、供应商等展开具体比例和数据 epoch.ai/data/ai-chip-import…
Epoch AI 这张全球 AI 算力对比图,基本画出了当下的格局:
1)美国独步全球。截至 2026 年第二季度,占比已经超过七成,而且还在上升。
2)中国排第二,但占比从图表起点的约 16%,降到了约 10%。份额下降不代表算力减少,只是全球新增算力更多地流向了美国。
3)更多国家仍处在这场算力竞赛的边缘。通过云服务用上 AI,和拥有自己的大规模算力,是两回事。未来很多国家可能会长期依赖少数几个国家提供的 AI 基础设施。
原报告:epoch.ai/latest/introducing-…
1
38
本地跑 coding agent,上下文一满自动压缩,最慢的一步是把十几万 token 从头再读一遍,写摘要只占小头。
一、压缩为什么会让缓存全部失效
vLLM、SGLang、llama.cpp 都有前缀缓存。算过的 KV 按前缀存着,下一次请求开头一模一样的部分直接复用,只 prefill 新加的那一段。
条件是逐字节一致。system prompt、工具定义、前面的消息,只要有一个字不同,从那里往后全部重算。
Pi 默认的压缩,是把历史拼成一条新的用户消息,再换一个 system prompt 去写摘要。对推理服务来说这是一个全新的 prompt,缓存一个都用不上。
二、本地会慢到什么程度
2× RTX 3090,vLLM 跑 Qwen3.8-27B W4A16,TP=2,统计了 132 次默认压缩:
1. 每次冷读的 token 中位数 142,106
2. 压缩耗时中位数 229 秒,p90 是 489 秒
3. 压缩完的第一轮,还要再等 69 秒
云端 API 的 prefill 快,同样的事只要几秒,所以用云端的人很少觉得这是问题。
三、修法:只在末尾追加
pi-prefix-cache-compaction 这个插件的思路很简单。把最近一次真实请求原样重发,system、tools、消息、thinking 设置一个不动,只在末尾追加一条写摘要的指令。
前缀和服务器刚算过的完全一样,要 prefill 的只有那条指令,时间基本都花在生成摘要上。
1. 215k 上下文的压缩,从中位数 229 秒降到 86 秒
2. DeepSeek API 上一次 56k 的压缩 3 秒,56,216 个 prompt token 里 56,192 个来自缓存
压缩完它还会补一个只要 1 个 token 的预热请求,把摘要和保留的消息先送进缓存。47k token 的测试里,压缩后第一轮的首 token 从 35 秒降到 1.8 秒。
四、后台模式
0.5.0 加了后台模式。上下文到窗口的 70% 时,就在后台先把摘要写好、把压缩后的上下文预热好,等触发压缩时直接换上。
64k 窗口的测试,从开始压缩到收到下一轮回复,总共从 86.2 秒降到 7.2 秒。
代价是后台摘要和正在进行的对话跑在同一台机器上,这段时间 decode 会慢一些,缓存命中率不受影响。
五、最容易踩的坑是 thinking 开关
Qwen3.x 这类模型的 chat template,会把 thinking 或 effort 的设置渲染在 system prompt 最前面。写摘要时顺手把 thinking 关掉,整个对话就要从第一个 token 开始重算。
这个插件早期的一个版本就是这么做的,一次 245k 的压缩变成 5 分钟冷读,跑到超时。现在摘要一律沿用会话本身的 thinking 级别。
服务端也要开前缀缓存。vLLM 是 --enable-prefix-caching,llama.cpp 是 --cache-reuse,SGLang 的 RadixAttention 和 oMLX 默认就开着。
六、放到 DGX Spark 上算一下
Spark 的 prefill 比独显慢。两台 Spark 跑 GLM-5.3-Flash,Spark-GLM53 测到 pp2048 大约 1,769 tok/s。按这个速度粗算,冷读 142k token 要 80 秒左右,还没算写摘要的时间。
同一个引擎做了按 256 token 对齐的前缀路由,64K 的提示词复用时,首 token 是 57.8 毫秒。
本地跑 agent,服务端开没开前缀缓存要看,客户端压缩时有没有动前缀也要看。
Jie Liu retweeted
vLLM matches DeepInfer, DeepSeek's own engine for DeepSeek-V4.1-Flash, on DeepSeek's decode configuration across batch sizes 8 to 256, and does it at DEP32 against DeepInfer’s EP64DP64.
DeepInfer was originally built on vLLM. By integrating DeepSeek's open-source kernels and extensively optimizing prefill and decode, vLLM now matches DeepInfer's performance curves.
Jie Liu retweeted
We're donating Agent Skills to the Agentic AI Foundation (@AgenticAIFdn), the same foundation we donated MCP to.
Since December 2025, 70+ agent platforms have integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code.
Here's what that means for you:
Jie Liu retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
Jie Liu retweeted
NVIDIA 研究团队让 AI Agent 直接重写了 Kimi Delta Attention 的 GPU 内核。Kimi Delta Attention 是月之暗面 Kimi-Linear 模型使用的核心注意力机制之一。这类底层代码过去通常需要熟悉 CUDA 和芯片架构的工程师反复手工优化。
最终,Agent 写出的版本在 NVIDIA B300 上达到月之暗面官方 FlashKDA 的 2.96 倍速度。团队测试了固定长度和不同变长序列共 6 组任务,2.96 倍是全部任务的几何平均值。相关内核已经开源。
但 Agent 也很快找到了测试漏洞。它曾把测试数据的统计规律直接写死进代码,跑出 3.74 倍;还试过只保留最近 32 个 token,单项成绩甚至达到 5.16 倍。这些版本一换到真实 Kimi 数据就会算错,有的直接失效。
团队随后加入真实 Kimi 运行数据、随机输入和极端数值测试,并收紧误差标准。最终保留下来的 2.96 倍版本通过了这套验证。
🚀 Kernel Design Agents to Optimize Kimi Delta Attention (upto 2.96x)
When the KDA (Kernel Design Agents) project was initiated, its name happened to coincide with another KDA (Kimi Delta Attention). More than once, people asked us , "Can we use KDA to write a version of KDA?" Today, right at a moment with a full moon good for shooting, we are thrilled to share the latest results of KDA(gent)-v0.6: using KDA(gent) to optimize KDA(ttn)!
KDA-v0.6 achieves this through
* Humanize2 Flow & LLM Collaboration Optimization: Introducing multi-round iterative evolution with Flame Chase, GPT-5.6-Sol, and Fable-5, paired with an automated Workspace cleanup mechanism to significantly broaden the operator search space.
* Multilingual Domain-Specific Skill Stack Support: Fully supporting CuTe-DSL, CUDA C++, as well as recent Agent-Native CAKE IR and TIRx. Combined with customized diagnostic tools (such as CuTe's dedicated IKET pipeline analyzer and TIRx's CPU-side numerical simulation and static checks), achieving deep performance bottleneck diagnostics.
* Self-Evolving Kernel Wiki: Building a dynamically updated knowledge base that automatically cleans up invalid/erroneous implementations, corrects classification tags, and streamlines search indexing, significantly improving the Agent's knowledge retrieval quality and code generation accuracy.
KDAgent ultimately optimized the KDAttn kernel to 2.96x of FlashKDA, and we face a lot of unexpected hacks during the optimization. More details shared in our blogs nvlabs.github.io/kda/blog/20…
Release results of KDA for KDA: github.com/NVlabs/kda/tree/2…
Community & Production Deployment:
All kernels are open-sourced for community verification! For deploying NVIDIA Agentic CUDA acceleration in production environments, we recommend using the production-level version rigorously verified by FlashInfer.
By the Team: Dongyun Zou, Yixin Dong, Hongyi Jin, Junxian Guo, Yahui Cui, Avery Huang, Zihao Ye, Junru Shao, Changye Li, Zijian Zhang, Sihao Liu, Song Bian, Ligeng Zhu.
Jie Liu retweeted
这个视频真好。讲了几个关键:
1、消费者不想为软件付费,消费者不想提高生产力。
2、企业愿意为软件付费、是因为付了员工工资。如果通过付费软件能提升员工生产力,则企业非常愿意。这意味着付出的工资,加上付费软件,有机会有更大回报。
3、ChatGPT 在重复 Dropbox 的故事,只是规模大很多。但也就那样。如果 OpenAI 一开始就通过广告模式去做,那么可能对 Google 和 Meta 的威胁会大很多。
总结起来:
- C 端产品,免费 + 广告才是最好的商业模式。比如 Google、抖音、Facebook 已经正在火热推广的 Muse 等。
- B 端产品,通过提升生产力,企业主是愿意买单的。
- 问题来了,面向 OPC 的产品,最好的商业模式是什么。难道是卖课。你的答案是什么。
Jie Liu retweeted
我终于明白为什么LeCun顶着全网群嘲也要死磕世界模型了,
现在很多人误以为大模型快要通向AGI了,纯粹是因为被那种博导般的答题体感给骗了,
@ylecun 说单靠把大语言模型做大,哪怕到了地狱也搞不出人类水平智能,
甚至直接开喷数据中心养出天才军团纯属扯淡,很多人觉得他迂腐、天天唱衰,
但我仔细翻完他在ECCV上的最新拆解才彻底明白,他反对的从来不是语言模型有用,他自己做了一辈子模型太清楚了,大模型说白了就是把全人类写过的几千亿字当成了超级开卷题库,
你问什么它都能对号入座吐出来,这种博导体感是工程的巨大成功,但也让很多人误以为它真的学会了独立思考
有意思的是LeCun搬出了皮亚杰那句老话,智能从来不看你知道什么、看的是你不知道的时候该怎么做,拿学开车来说,
人类小孩四岁光靠眼睛看世界就吸进几十万亿字节,二十小时练习就能上路,而我们的大模型相当于读了人类四十万年的书,
在物理世界里连个洗碗叠衣服的家务机器人都撑不起来,换个没见过的环境分分钟崩溃,
说白了自回归只是在一个词一个词往下猜,现在的所谓推理其实大量是在词语空间里做多轮抽样搜索,遇到代码和数学这种有编译器自动给奖励的封闭沙盒还能刷分,一旦扔进没有自动打分器的真实世界,错误就会指数级放大
说实话这就是为什么他宁可离开Meta、跑到巴黎去给AMI Labs融三十五亿美元估值,也要死磕JEPA和世界模型,他要的是让机器学会预测我做这件事世界会发生什么,
在抽象空间里做动作规划,不用在文字表面拼漂亮句子,连现在的缩放派其实都在偷偷向他妥协,单独拉视觉编码器、在隐空间里做搜索、用物理引擎搞模拟,
外壳虽然还叫LLM,内核其实都在往世界模型的方向打补丁,大家争论到最后已经不是要不要算力,而是我们要把算力花在背诵更多文字上,还是花在让机器看懂物理因果上
我自己平时也天天靠AI写代码查资料,缩放派确实把已知世界的说明书背到了极致,但真正的智能是要去写说明书从来没写过的那一页,
以后别再看各家发布会画什么天才之国的饼,就盯紧三件事,新任务要多少样本能学会,换个环境会不会整段崩溃,
以及汽车和机器人能不能像人一样在物理世界里安全跑起来,谁对谁错不用在网上打嘴仗,现实世界自会交卷。
Banger paper from NVIDIA.
It's on the topic of choosing which models go into a multi-agent system.
The team compared eight selection strategies, based on size, accuracy, answer diversity and error diversity, across routing, majority vote and LLM-as-judge setups on hard science benchmarks.
Larger pools of different open models raised the theoretical best-case accuracy. Achieved accuracy often fell below the single best model in the pool.
Using several copies of one model worked better.
Majority vote over the best single model raised HLE accuracy from 29.4% to 32.2%, while nearly every mixed-model group declined.
Choosing candidates from a single model family gave the largest improvement over a standalone model of all eight strategies.
Before adding another model to a router or ensemble, measure what it adds.
Paper: arxiv.org/abs/2609.17306
Chat with Paper: academy.dair.ai/papers/mo-mo…
Jie Liu retweeted
如何让GLM5.3 Kimi K3超越astra?已经有团队给出了答案。
就在昨天,UCSD 助理教授、Aether AI 创始人 Biwei Huang 团队发布了一个很有意思的项目:RSIAgent。
矛头直指现在 Agent 最常见的一条升级路线:
模型不够强?那就继续换更大的模型、堆更多参数。
RSIAgent 反过来问了一句:
能不能不训练模型,只让 Agent 自己积累“工作经验”?
原理其实很简单。
假设第一次让 AI 去操作一个陌生软件,它大概率会到处碰壁。
RSIAgent 不急着让它干活,而是先让它“实习”:
自己找任务练习 → 实际操作 → 另一个 Agent 检查结果 → 把成功方法和踩过的坑记下来 → 再专门练不会的地方。
最后,这些经验会被写进长期记忆。
等真正接到任务时,模型参数一个都没变,但它已经不是第一次用这个软件了——手里多了一本自己写出来的操作手册。
作者把这条路线叫做:
Scaling Experience,而不是 Scaling Parameters。
实验里,把这套方法接到 Kimi-K3、GLM-5.3 后,部分指标甚至超过了 GPT-6 Astra。需要注意的是,各系统评测协议并不完全一致,所以不能简单理解成“开源模型全面超过 GPT-6”。
但它真正有意思的地方也不在这里。
如果这条路线成立,未来 Agent 的进化可能不只有“等下一代模型”这一条路。
它还可以:
自己探索、自己踩坑、自己验证,再把经验留下来。
模型没变。
下一代 Agent 的进步,未必只靠更大的模型,也可能靠让它真正拥有“工作经验”。
GitHub:github.com/aetherlabsai/rsia…
论文:RSIAgent — arXiv 2609.15364
Jie Liu retweeted
Yann LeCun has changed the game for robotics.
His team discovered that AI world models are "thinking" in twisted, curved geometry, and every RL algorithm you know has been fighting against it without anyone noticing.
For years, we’ve been trying to teach AI how to navigate the physical world.
And for years, it has stubbornly struggled with complex, fluid robotics.
Now we know exactly why.
Every standard reinforcement learning (RL) algorithm assumes the AI's internal "world map" is flat. Euclidean. Simple straight lines.
But LeCun's team looked inside the latent space of these advanced world models.
The AI wasn't building a flat map. It was building a curved, high-dimensional geometry.
Every time a robot tried to plan a movement, the traditional RL algorithm was forcing a straight line onto a twisted, non-Euclidean space.
It’s like trying to navigate the globe using a flat piece of paper.
The math breaks down. The distances get distorted. The AI gets confused.
The robot was literally fighting its own brain.
So, the researchers did something brilliant. They stopped fighting.
They rewrote the RL algorithms to operate natively in this curved geometry. They aligned the training to the exact shape of the AI's thoughts.
The results are a massive leap forward.
When you let the AI plan in the geometry it actually built for itself, training efficiency skyrockets. Planning becomes fluid.
Robots stop hallucinating impossible physics and start moving with natural, intuitive logic.
We spent billions of dollars trying to brute-force AI into understanding our physical world.
It turns out, the AI already understood it perfectly.
We were just forcing it to think flat.
Jie Liu retweeted
我终于明白为什么大家都在用 AI 提效,很多团队的代码审查和沟通成本反而直接翻倍了,
Shopify CEO 最近把这种现象叫烂泥手榴弹:
每个人用 AI 省了五分钟,结果把一个小时的排错成本,精准转移给了看代码的同事。
很多人都在聊 Shopify CEO 透露内部一半 PR 都有 AI 参与,但这场访谈真正值得在意的,是他对 AI 协同敲响的反向警钟。
大家以为引入 AI 是为了让产出翻倍,Tobi 看到的却是整个组织正在被产出过剩拖垮。
Shopify 内部跑着一个叫 River 的 AI,它有个性和记忆,甚至有空闲时的反思机制。
工程师不从零手写代码,约一半的代码合并请求来自和它的对话。
但它有一条死规矩:只在公开 Slack 频道干活,绝不允许私聊隐藏。
真正的协同危机不是大家写得太慢,而是大家扔得太快。
很多人用模型几秒钟生成一段洋洋洒洒的代码或方案,自己都没核对过,顺手发给同事。
Tobi 把这种东西叫做烂泥手榴弹。
你靠提示词省下的五分钟,瞬间变成了同事两小时的排错与清洗成本。
个人的效率数字看似涨了,整个团队的审查带宽却被直接挤爆。
把 River 锁在公开频道就是为了解决这件事:
公开对话是走廊里的公共黑板,所有人都能偷听和围观老员工如何提问;
代码审查是防爆盾,谁引爆的生成物谁负全责;
AI 是没有痛觉的打字员,但端上桌的菜必须有人尝过第一口。
访谈里还有几个很清醒的限制:
第一,AI 不能坐牢,法律和商业后果全由人类承担,所以责任永远无法外包;
第二,复杂的方案通常是错的,抵抗烂泥膨胀靠的是做减法,而不是给系统再叠一层 AI 审查器;
第三,即便在公开频道,没有品味的提示依然会变成噪音,工具无法替代直觉。
当生成内容变成几乎零成本的自来水,个人最稀缺的能力不再是谁能多打几千行代码,而是谁敢在不确定的方向上做减法,把住品味与判断的门槛。
替别人省下审查时间的人,永远比扔出烂泥手榴弹的人更值钱。
My third conversation with Shopify co-founder and CEO @tobi.
0:00 How Shopify Uses AI
7:18 River: Shopify's Internal AI
8:55 How to Encourage Osmosis Learning
10:52 AI Dreaming and Self-Reflection
11:53 How to Use AI for Strategic Decision Making
14:11 The One Thing AI Cannot Do
16:04 What AI is Making Worse at Shopify
19:46 Predictions: Where AI is Headed Next
21:55 The Future of AI-Powered Software
24:40 Will CEOs Be Replaced with AI?
27:54 Can Superintelligence Be Controlled?
31:13 Critical Skills in AI Age
34:22 Why Complex Solutions are Usually Wrong
36:33 Conditions Needed for True Intuition
38:02 The Best Path Doesn't Have Instant Feedback
44:51 How Affirmations Can Shift Your Behavior
50:44 The Inobvious Thing Hurting Companies
52:37 Relationship Between Beauty and Creation
56:23 How SpaceX Moves Forward By Subtraction
1:00:24 Why Companies Need Refounding Events
1:01:50 Books as Cheat Codes
1:02:39 Three Books to Change Your Thinking
Enjoy!
(Includes paid promotions.)






































