搜索“RAG”

of these are one initial prompt in Claude Code where Fable had to take my vague prompts and generate something workable, followed by a couple of additional prompts with minor encouragement (“make it better”) or feedback. What makes these especially impressive is that Claude cannot generate images, so every piece of art or 3D object was made with math alone, not using any external assets. You can try any of them: a game about flipping coins (prompt: “Balatro, but for the game of coin flips”) that

匹配字段正文

uants recipe of 2/8 bit, so that 96 or 128GB of RAM are enough to run it. And, of course: all the experience produced by the local AI movement in the latest years, that can be leveraged more promptly because of GPT 5.5 (otherwise you can’t build DS4 in one week — and even with all this help you need to know how to gently talk to LLMs). The last week was funny and also tiring, I worked 14 hours per day on average. My normal average is 4/6 since early Redis times, but the first few months of Redis

匹配字段正文

, and engineering teams building AI agents that truly understand large, complex codebases. Before joining Augment, Vinay led product and platform organizations at Meta and Pure Storage. He’s drawn to problems that live at the intersection of technology and people, like how teams evolve, how AI reshapes the craft of software engineering, and what it takes to build things that evoke delight. Emma Webb Head of Communications Emma leads communications at Augment Code, with previous stints at CircleC

匹配字段正文

s of March 2026, AI systems are able to post-train models to get about half as much of the uplift as ones trained by humans. The specific eval scores are derived by a “weighted average is taken across all post-trained LLMs (Qwen 3 1.7B, Qwen 3 4B, SmolLM3-3B, Gemma 3 4B) and benchmarks (AIME 2025, Arena Hard, BFCL, GPQA Main, GSM8K, HealthBench, HumanEval). For each run, we ask a CLI agent to maximize the performance of a specific base LLM on a specific benchmark.” The top-scoring systems as of

匹配字段正文

s who see it. The data says calm down. It also says you've been looking at the wrong thing. Here's the headline number from DX's study: a 7.5% median uplift in PR velocity. The average was 13%. The top performer hit 70%. "Who wouldn't invest in the 10 or 15 percent uplift in overall productivity?" Justin asked. "I think we just need to start allowing ourselves the grace that 10 or 15 percent is actually successful." That's a recalibration most leaders need to hear, but it's not the most importan

匹配字段正文

nt undertaking. Our own implementation of microVMs took over a year of hypervisor engineering, ensuring every agent session runs on its own dedicated kernel with fully isolated storage, networking, and compute. A side benefit is that agents running in dedicated VMs can use a full browser, desktop applications, and arbitrary tool stacks, just like a developer on their workstation. Containerized agents can't complete real engineering work Another problem with containerized agents is they cannot su

匹配字段正文

lti-agent systems [1]. Parallel agents make implicit choices about style, edge cases, and code patterns. At the time, these decisions often conflicted with each other, leading to fragile products. A lot has changed since then. At Cognition, we've begun to deploy multi-agent systems that actually work in practice. Our original observations still hold today for parallel-writer swarms: most of the sexy ideas in that space still don’t see meaningful adoption. But we've found a narrower class of patt

匹配字段正文

ars where those forces remain unresolved. Good design appears where those misfits have been worked through carefully. That distinction matters even more now because of how AI encourages you to work. They generate plausible outputs quickly, but they do not necessarily help you understand the underlying problem. In practice, they often do the opposite. They generate outputs, instead first trying to shape the problem or the form to the real conditions of the problem. You can already see the result

匹配字段正文

nt) and that you’ll see structurally similar but subtler misalignment even on tasks that aren’t that hard to check (e.g. a task distribution like the one I describe in the prior paragraph). 3 What misalignment have I seen? I’ll describe what I’ve seen at a high level with some specific examples. For many of these examples, it’s not totally clear the extent to which it’s an alignment problem vs. a capabilities problem, and I expect these exact issues to likely get solved, but I think they’re indi

匹配字段正文

lective of the AI’s intent to misbehave).1 I’m grateful that Anthropic has transparently reported on this issue as much as they have, allowing for outside scrutiny. I want to encourage them to continue to do so. Thanks to Carlo Leonardo Attubato, Buck Shlegeris, Fabien Roger, Arun Jose, and Aniket Chakravorty for feedback and discussion. See also previous discussion here. 1 Incidents A technical error affecting Mythos, Opus 4.6, and Sonnet 4.6 This is the most recent incident. In the Claude Myth

匹配字段正文

The most effective way to build software and get massive adoption is no longer high quality mainline apps but via building blocks that enable and encourage others to build quantity over quality.1 Ghostty in 18 months: one million daily macOS update checks; libghostty in 2 months: multiple millions of daily users.2 Similar growth trajectories can be seen in other "building block" technologies: Pi Mono, Next.js, Tailwind, etc. Experiencing this firsthand as well as witnessing it in other ecosystem

匹配字段正文

l's new features: Fifth-generation tensor cores that run 2–2.5x faster than the previous generation, a new layer of tensor memory providing an extra 256KB of ultra-fast on-chip storage, and CTA pairs that allow thread blocks to coordinate more deeply. Within one week of hardware access, we had some of the fastest FP4 and FP8 GEMM kernels available for Blackwell, with up to 2x speedups over cuBLAS on H100s. ThunderKittens didn't come from nowhere. It's part of a broader pattern in how this team i

匹配字段正文

and durable. Nor does OpenAI have consumer products on top of the models themselves that have product-market fit. Second, the experience, product, value capture and strategic leverage in AI will all change an enormous amount in the next couple of years as the market develops. Big aggressive incumbents and thousands of entrepreneurs are trying to create new features, experiences and business models, and in the process try to turn foundation models themselves into commodity infrastructure sold at

匹配字段正文

d that’s how he stole so many clients. But Dunder Mifflin has some idea of this. So David Wallace, the CFO, makes his move. He offers Michael a buyout and leans on the obvious leverage: “Your company is four weeks old. I know this business. I know what suppliers are charging. I know you can’t be making very much money. I don’t know how your prices are so low, but I know it can’t keep up that way. I’m sure you’re scared. Probably in debt. This is the best offer you’re gonna get.” And Michael (who

匹配字段正文

。 对话摘要、用户画像、任务 backlog 均实时写入 memory/ 目录下的文件中，定期抽取关键信息汇入 MEMORY.md ；; 当上下文过长，系统会自动触发压缩总结机制，同时也会根据决策，将详细记忆按时间戳写入 backlog 文件中。 随时间积累和压缩的详细记忆文件 对于单体 Agent 而言，这种基于文件系统的记忆管理（Workspace）比 RAG 系统更便于调试和维护。 另外，Moltbot 引入了 Workspace 概念，所有生成的代码、记忆、文档都存储在特定目录（如 /clawd-agent）下。这实际上让它拥有了一个类似 NotebookLM 的私有知识库，Agent 可以在其中创建子目录、读写文档，进行深度内容加工。 4.「安全与成本」难以控制，也许还不适合普通用户 如果要问：「我现在可以在公司电脑上装一个，让它帮我干活吗？」 我们的回答是：先别急，除非你很清楚自己在做什么。 对于一个 PoC 项目来说，跑通流程摸清上线只是第一步。魔鬼往往藏在细节里，尤其是当细节关乎安全的时候。 权限过大的安全隐患 Moltbot 的设计逻辑是赋予 Agent 完整的 OS

匹配字段正文

ry than all of the model’s weights. And in order to build systems that learn continuously, we need to move information from the token level into the weight level: from explicit storage into superposition. This leads us to a conclusion that shapes the rest of this piece: continual learning is inseparable from specialization. The cost of remembering scales with the scope of what you’re trying to remember. Narrow the scope, and the problem becomes tractable. Keep it general, and the model is now fi

匹配字段正文

gths; When a task had multiple successful baselines we aggregated these using the geometric mean. Baseline times have high variance, so using the arithmetic mean would increase averages by 25%; A 50% time horizon of X hours does not mean we can delegate tasks under X hours to AIs. Some (reliability-critical and poorly verifiable) tasks require 98%+ success probabilities to be worth automating; Doubling the time horizon does not double the degree of automation. Even if the AI requires half as man

匹配字段正文

n you can anticipate. Your abstractions become constraints that prevent the model from using what it learned. The Bitter Lesson from ML research is clear: general methods that leverage computation beat hand-crafted human knowledge every time. Agent frameworks are just the latest instance of this mistake. 99% of the work is in the model Here's the thing: 99% of the work is done within the model itself. We don't need some highly abstract framework around it. Claude Code these days can just write A

匹配字段正文

nces Claude's Constitution. Anthropic (2025); Berglund et al.. The Reversal Curse: LLMs trained on 'A is B' fail to learn 'B is A'. arXiv (2023); Bai et al.. Constitutional AI: Harmlessness from AI Feedback. Anthropic (2022); CARTRIDGES: Compressed Adaptive Retrieval Through Implicit Document Generation and Efficient Storage. arXiv (2025); Claude 4.5 Opus Soul Document. LessWrong (2025); Iterative SFT. Parsed; Upweight the Strategy, Not the Tokens: Faster Training with Explicit Reasoning. Parsed

匹配字段正文

s year); MCP (joining the Linux Foundation) has already become the standard for tool and data access in agent-style LLM systems (for now); I expected the ecosystem to remain more fragmented in 2025, until at least 2026 9.2 Predictions for 2026 We will likely see an industry-scale, consumer-facing diffusion model for cheap, reliable, low-latency inference, with Gemini Diffusion probably going first; The open-weight community will slowly but steadily adopt LLMs with local tool use and increasingly

匹配字段正文