{
  "source_file": "0930",
  "source_platform": "Substack",
  "author": "swyx & Alessio Fanelli / Latent Space",
  "title_listed": "The State of AI Agents",
  "keywords": "AI Agent,智能体,工程",
  "domain": "大模型算法",
  "content_type": "深度技术解析",
  "is_paid_flag": "否",
  "url": "https://www.latent.space/p/agents",
  "publish_date": "2024-06",
  "corpus_stage": "推理部署",
  "notes": "AI Agent现状",
  "id": "5f6a24f7d0f7",
  "collected_at": "2026-10-01T20:33:17",
  "title": "The Anatomy of Autonomy: Why Agents are the next AI Killer App after ChatGPT",
  "body_markdown": "Welcome to the 12k new listeners who checked out\nthe Segment Anything pod\n!\nWe’ve launched a\nnew community page\nwith our upcoming events in SF, NY and\nMiami\n(this Friday)! See you if you’re in town!\nDiscussions happening on\nHacker News\nand\nTwitter\n.\n“GPTs are General Purpose Technologies”\n1\n, but every GPT needs a killer app. Personal Computing needed VisiCalc, the smartphone brought us Uber, Instagram, Pokemon Go and iMessage/WhatsApp, and mRNA research enabled rapid production of the Covid vaccine\n2\n.\nOne of the strongest indicators that the post GPT-3 AI wave is more than “just hype” is that the killer apps are already evident, each >$100m opportunities:\nGenerative Text for writing\n- Jasper AI going 0 to $75m ARR in 2 years\nGenerative Art for non-artists\n- Midjourney/Stable Diffusion\nMultiverses\nCopilot for knowledge workers\n- both GitHub’s\nCopilot X\nand “\nCopilot for X\n”\nConversational AI UX\n-\nChatGPT\n/ Bing Chat, with a long tail of Doc QA startups\nI write all this as necessary context to imply:\nThe fifth killer app is here, and it is\nAutonomous Agents\n.\nBut first, as usual, let’s start with an executive summary to catch up those out of the loop.\nAuto-GPT Executive Summary\nAuto-GPT\n(and its younger sibling\nBabyAGI\n) are independently developed Python projects, open sourced March\n30th\nand\nApril 2nd\nrespectively, which have caught enormous popularity, with Auto-GPT\ntrending #1 on Twitter and GitHub\nin the past 2 weeks (\nfar outpacing every other open source AI project\nincluding\nSegment-Anything\n, Stable Diffusion, and the now Sequoia-crowned\n$200m-valuation-LangChain\n).\nBoth projects do\nnot\ninvolve foundation model training or indeed any deep ML innovation; rather they demonstrate viability of applying\nexisting\nLLM APIs (GPT3, 4, or any of the alternatives) and reasoning/tool selection prompt patterns\nin an infinite loop,\nto do potentially\nindefinitely long-running\n, iterative work to\naccomplish a high level goal\nset by a human user.\nWe do mean “high level” —\nToran Richards\n’\noriginal demo\nfor Auto-GPT was “\nan AI designed to autonomously develop and run businesses with the sole goal of increasing your net worth\n”, while Yohei Nakajima coded up\nJackson Fall’s viral HustleGPT prompt on ChatGPT\nand told it to\n“\nstart and grow a mobile AI startup\n”\n. In the 2 weeks since, community members have\nbuilt extensions\nand\nclones\nand\nagent managers\nand\nframeworks\nand\nChatGPT plugins\nand\nvisual toolkits\nand so on, with usecases in\nmarket research\n,\ntest driven development\n, and\nscientific literature review\n.\nBeyond those similarities, the projects are very different in their approaches.\nBabyAGI is intentionally small,\nadding and stripping out LangChain\n, with its initial code being\nless than 150 lines\nand\n10 env vars\n(it is now ~800LOC).\nGPT4 visualizing the codebase.\nOther attempts\n.\nWhile Auto-GPT is\nmore expansive\n(7300 LOC), with the ability to clone GitHub repos, start other agents,\nspeak\n, send tweets, and generate images, with\n50 env vars\nto support every vector database and LLM provider/Text to Image model/Browser.\nfull list from source\nThe projects have caught the imagination of leading AI figures as well, with Andrej Karpathy calling AutoGPTs the “\nnext frontier of prompt engineering\n” and Eliezer Yudkowsky\napprovingly observing\nBabyAGI’s refusal to turn the world into paperclips even when prompted.\nA Brief History of Autonomous AI\nIn my understanding of neurobiology,\nevery convolution\nthat wrinkles the brain a bit more makes us a little smarter. In a similar way, AI progresses by “convolutions”, and in retrospect our path to the present day has been obvious. I’d like to map it out:\nKey Autonomy Capabilities arranged in rough chronological order\nFoundation models\n:\nEverything starts with the evolution\nand\nwidespread availability of massive LLMs (via API or Open Source). The sheer size of these models finally allow for\n3 major features\n:\n~perfect natural language understanding and generation\nworld knowledge (\n175B Parameters can store 320GB\n, which is 15 Wikipedia’s)\nemergence\nof major capabilities like\nin-context learning\nThis leads to the rise of the early prompt engineers, like\nGwern Branwern\nand\nRiley Goodside\nwho explored creative single-shot prompts.\nCapability 1: Metacognition (self improvement of pure reasoning)\nKojima et al (2022)\nfound that simply adding “\nlet’s think step by step\n” to a prompt\ndramatically\nraised the performance of GPT3 on benchmarks, later found to be effective due to externalizing the working memory for harder tasks.\nWei et al (2022)\nformalized the technique of Chain of Thought prompting that further improved benchmark performance.\nWang et al (2022)\nfound that taking a majority vote of multiple Chains of Thought worked even where regular CoT was found to be ineffective.\nMore and more techniques like\nCalibrate Before Use\n,\nSelf-Asking\n,\nRecursively Criticize and Improve\n,\nAutomatic Prompt Engineering\n, appear.\nCapability 2: External Memory\n(reading from mostly static external data)\nThe capability of in-context/few shot learning could be used to cheaply update a foundation model beyond its’ knowledge cutoff date and focus attention on domain specific, private data\nThe constraints of limited context length lead to the need for embedding, chunking and chaining frameworks like LangChain, and vector databases like\nPinecone (now worth $700m), Weaviate ($200m), and Chroma ($75m)\n.\nAnother way of using natural language to access and answer questions form relational databases are the Text to SQL companies, which included Perplexity AI (\n$26m Series A\n), Seek AI (\n$7.5m Seed\n), and a long tail of other approaches including\nCensusGPT\nand\nOSS Insight\n.\nCapability 3: Browser Automation\n(sandboxed read-\nand-write\nin a browser)\nSharif Shameem (an upcoming guest! more on a future pod) first demoed\nGPT-3 automating Chrome to buy Airpods\nin 2021.\nAdept raised a Series A\nwith an all-star team of Transformer paper authors and launching the\nACT-1 Action Transformer\n(now with a\nhefty $350m Series B\ndespite the departure of Vaswani et al)\nNat Friedman’s NatBot\nbrought browser automation back into the zeitgeist a year later, showing how an agent can make a restaurant reservation across google search and maps from a single natural language instruction.\nDust XP1\nwas also released but was read-only, did not do any automation.\nMULTI·ON\nwent that extra mile and is now also in\nthe ChatGPT Plugin Store\n.\nA nice variant of browser agents are desktop agents -\nEmbra AI\nseem to be the most hyped here (though still pre launch), and\nRewind AI\ncould be next.\nIt would seem that\nMulti-modal GPT4\n’s visual capability would be able to greatly enable the desktop agents here, especially where no accessibility text or DOM is available.\nCapability 4: Tool making and Tool use\n(server-side, hooked up to everything)\nSearch.\nGenerated answers from memorized world knowledge, or retrieved and stuff into context from a database, will never be as up to date as just searching the web. OpenAI opened this can of worms with\nWebGPT\n, showing their solution to crawling the web, summarizing content, and answering with references (now live in ChatGPT Plugins and in Bing Chat, but replicated in the wild with\nDust\nand\nothers\n).\nWriting Code to be Run\n. We knew that GPT-3 could write code, but it took a certain kind of brave soul like\nRiley Goodside to ask it to generate code\nfor known bad capabilities (like math) and to\nrun\nthe code that was generated.\nReplit turned out\nto be the perfect hosting platform for this style of capability augmentation (\nanother example here\n).\nReAct\n.\nYao et all (2022)\ncoined\nthe ReAct pattern\nwhich introduced a\ndelightfully simple\nprompt template for enabling LLMs to make reliable tool choices for\nRea\nsoning +\nAct\ning given a set of tools.\nSchick et al (2023)\nintroduced the Toolformer that specifically trained a model with special tokens, but this does not seem as popular.\nMulti-model Approaches.\nModels calling other models with capabilities they didn’t have were also being explored, with\nHuggingGPT/Microsoft JARVIS\nand\nVisualChatGPT\n.\nSelf-Learning\n.\nSelf-Learning Agent for Performing APIs\n(SLAPA) searches for API documentation to teach itself HOW to use tools, not just WHEN. This approach was adapted for the OpenAPI (fka Swagger) spec for ChatGPT Plugins, which\nalso used natural language\n.\nOther semi-stealth mode startups that may be worth exploring in this zone are\nFixie AI\nand Alex\nMinion AI\n.\nAt this point it is worth calling out that we have pretty much reached the full vision laid out by\nthis excellent post from John McDonnell\n6 months ago:\nSo what net new thing are we seeing in this most recent capability spurt?\nI think the clue is in the 4 agents that naturally evolved in BabyAGI (scroll up for diagram):\nThe “\ncontext agent\n” (Capability 1 + 2) could be a much smarter version of the data augmented retrieval that both\nLlamaIndex\nand\nLangchain\nare working on. Yohei added the need for “\nrelevant (task) context\n” which may be slightly different than the classic semantic similarity algorithms offered by the vector databases.\nActive learning\nmay see a return to favor as autonomous “context agents” actively surface things they don’t know for prioritization\nThe “\nexecution agent\n” calls OpenAI, or any other\nfoundation model\n, and could optionally make or use any provided tools to accomplish a task (Capability 3 + 4)\nThe “\ntask creation agent\n”, well, creates tasks, but must not hallucinate and must self criticize and learn from previous tasks (Capability 1 + 2). Challenging, but not outside the bounds of\nsimple common sense benchmarks\n.\nAnd the last agent is the “\nprioritization agent\n”. Ah! A new task!\nThat leads us to identify…\nCapability 5: Planning, reflexion, and prioritization\nShinn et al (2023)\nshowed that Reflexion - an autonomous agent with dynamic memory and self-reflection, could dramatically improve on GPT-4 benchmarks.\nShoggoth the Coder\nwon the recent ChatGPT Plugins Hackathon as an independent agent capable of proposing and submitting PR fixes to open source projects.\nMeta’s Simulacra paper\nshowed the entertaining potential of autonomous NPC agents interacting with each other in a game-like setting.\nRegardless of use case, autonomous agents will be expected to plan further and further ahead, prioritizing task lists, reflecting on mistakes and keeping all relevant context in memory.\nThe “Sparks of AGI” paper\nspecifically called planning out as a notable weakness of GPT-4, meaning we will likely need further foundation model advancement before this is reliable.\nThe recent\nLangChain Agents webinar\ndiscussion also highlighted the need for the ability to stack agents and coordinate between them.\nIn the\nLatent Space Community\n, AI virtual software developer platform\ne2b\nis already discussing the potential of having fleets of AI developer workers.\nWhy Autonomous AI is the Holy Grail\nWhat makes software valuable to humanity?\nIn both my investing and career advice, I am fond of encouraging people to develop a “\ntheory of value of software\n”.\n3\nOne of the clearest value drivers\n4\nof software is automation. The one currency we all never have enough of is\ntime\n, and the ability to obsolete human effort, whether by clever system design, hiring someone else, or programming a machine, both frees up our time and increases our ability to scale up our output by just doing more in parallel. In fact this can be regarded as a core definition of technology and civilization:\n“Civilization advances by extending the number of operations we can perform without thinking about them” — Alfred North Whitehead.\nThe relationship betwen\nautomation\nand\nautonomy\nis subtle but important:\nChatGPT doesn’t do anything without your input, but once you punch the right prompts in, it can do an awful lot of research for you, especially\nwith Plugins\nAutoGPTs by default require you to enter a goal and hit “yes” to approve each step it takes, but that is incrementally easier than having to write responses\nAutoGPTs also have limited (\nrun for N steps\n) and unlimited (\nrun forever)\n“\ncontinuous modes\n” which are fully autonomous but very likely to go wrong and therefore have to be closely monitored\nWe’ve just explained that technological and civilization advance requires us to be able to do things\nwithout thinking about them\n, so clearly full autonomy with as much trust and reliability as possible is the ultimate goal here. Let a thousand agents bloom! AI Assistants is where most people start, but Josh Browder is working on\nAI Lawyer\n, Replika is working on\nAI Waifu\n, I want AI Junior Developers and AI Video and Podcast and Newsletter Editors, Karpathy wants us to keep going with the\nAI C-Suite\n.\nFortunately, we don’t have to reason out every step of this progression from first principles, because\nthe Society of Automotive Engineers\nestablished a shorthand for this almost a decade ago:\nthe 5 levels of autonomy\nI’ll assume you are familiar with some of the self driving car discourse\n5\n, but it's time to understand that self-driving AI agents in 2023 are just about where self-driving cars were in ~2015. We are beginning to have some intelligence in the things we use, like Copilot and Gmail autocompletes, but it's very lightweight and our metaphorical hands are always at ten and two.\nIn the next decade, we'll want to hand over some steering, then monitoring, then fallback to AI, and that will probably map our progress with autonomous AI agents as well.\nEdit from May 2023: we found out from our upcoming podcast guest Itamar Friedman of Codium that they had already done some thinking on\n6 levels of Autonomous Code Integrity\n. Check them out!\nIn the following decade, we’ll develop enough trust in our agents that we go from a many-humans-per-AI paradigm down to one-human-per-AI and on to many-AIs-per-human, following an accelerated version of the industrialization of computing from the 1960s to the 2010s since it is easier to iterate on and manipulate bits over atoms.\nThere will be two flavors, or schools of thought, on autonomous AI:\nThe Jobs School: AI Agents that augment your agency, as “\nbicycles for your mind\n”\nThe Zuck School: AI Algorithms that replace your agency, hijacking your mind\nWe’ll want to try our best to guide our efforts to the former, but we won’t always succeed.\nRelated Reads\nThe Complete Beginners Guide To Autonomous Agents\nLangChain’s take on\nAutonomous Agents\nMisc Observations I couldn’t put anywhere\nI will perhaps expand these in future, or not, but happy to discuss each in comments.\nDespite needing prompt templates, tool selection, memory storage and retrieval, and orchestration of agents,\nneither Babyagi nor AutoGPT use LangChain\n. EDIT:\nLangChain’s docs\nnow have some implementations.\nThe new FM “Bitter Lesson” - Every time we try to finetune foundation models to do a certain task, software engineers can write abstractions on top to do it less efficiently but faster.\nvector db’s > memorizing transformers\nReAct/SLAPA > Toolformer/Adept ACT-1\nthere’s families of transformers and alternative architectures (H3, fwd-fwd algo) but nobody can name any that have stuck\nAcademic research of deep reinforcement learning for agents seems trapped in games?\nMineDojo, Deepmind DreamerV3, GenerallyIntelligent\nAvalon\nCICERO diplomacy is notable because did NOT use deep RL?\nSemantic search has been less of a killer product. why?\nBackend GPT has not led anywhere. why?\n“wen AI Agents” easier to discuss than “wen AGI”\nall the open source winners are new to open source lol? is “open source experience” that valuable?\nharrison\nyohei\nsiggravitas\nWhy do people all use pinecone to store like 10 things in memory\nSafety - no more AGI off switch with autonomous agents\nAGI Moloch means have fun staying poor = have fun staying safe\nwe’re calling for the wrong kinds of pause:\npause on fleets of workers until we have extremely well developed constitutions and observability?\nPredictions\nthere will be “AI Agent platforms”\nwith tools all enabled → Zapier and fixie well placed\nthere will be “AI Agent fleets”\nespecially if “idempotent”, readonly\ncustom research into the 4 kinds of agents\n“context agent” → active learning\nexecution agent is most straightforward but need to do its job well\ntask creation → problems: hallucination, omission\nprioritization agent\nstepping on each other\nwill need to do DAGs\nspawn on command\nActor model\n/ Agent Oriented Programming?\n5th agent -\nreflection? metalearning?\nthere will be “AI social networks”\nsubreddit simulator\norchestration and the cyborg problem\nsarahcat observation\nis correct\nswyx commentary on cyborg nodes\nwould be amazing\nAre level 5 agents\nAGI-hard\n?\nhard as self driving cars - perpetually 5 years away - uber & apple gave up\nmust know when to yield to humans\neasy to require confirm for destructive actions, but sometimes unclear\n“Interesting non-obvious note on GPT psychology is that unlike people they are completely unaware of their own strengths and limitations. E.g. that they have finite context window. That they can just barely do mental math. That samples can get unlucky and go off the rails.” Etc.\nkarpathy\nmust solve prompt injection (\nhey @simon\n)\nwill probably need to self improve statefully\nprincipal agent problem\nwe dont know how to prioritize humans, and you want to eval bot ability to prioritize? good luck\n(Update): Sept 2023 update\nAs we learned more about how people build agents, and what the valuable parts of the stack are, I updated the chart\nbased on a @karpathy\ntweet.\n1\nThe title of an influential\nrecent OpenAI paper\n, but also\nan actual concept\nin the study of technology — something your humble correspondent first learned about in research for the Tech Strategy chapters of\nmy book\n3 years ago. Worth internalizing.\n2\nDr. Katalin Kariko’s story of perseverance\nis worth a full listen, if you haven’t heard the story.\n3\nTheory of Theory of Value of Software: If you set out with the goal to understand what makes some lines of code more valuable than other lines of code, and try to make predictions by running your theories through lots and lots of data while reducing your “loss”, the idea is that you’ll be better able to invest your time, money, and creativity in more rewarding directions than people who don’t take the same effort.\n4\nI am not yet ready to publish\nmy full list\n, but software value drivers include: Demand and Supply Aggregators, Production-ready frameworks for underspecified Standards, “Shadow IT” to circumvent internal politics and byzantine rules, systems of record (\nZawinswyx’s Law\n), and also replacing people and manual processes.\n5\nSidenote: if you are excited about a self driving car future, you\nmust\nvisit San Francisco and befriend someone with access to the Cruise apps. There’s ~200 of these fully self-driving all over the city every night.\nThe future is very close!",
  "metadata": {
    "title": "The Anatomy of Autonomy: Why Agents are the next AI Killer App after ChatGPT",
    "description": "Auto-GPT/BabyAGI Executive Summary, a Brief History of Autonomous Agentic AI, and Predictions for Autonomous Future",
    "keywords": null,
    "author": "Latent.Space",
    "og:type": "article",
    "og:title": "The Anatomy of Autonomy: Why Agents are the next AI Killer App after ChatGPT",
    "og:description": "Auto-GPT/BabyAGI Executive Summary, a Brief History of Autonomous Agentic AI, and Predictions for Autonomous Future",
    "og:image": "https://substackcdn.com/image/fetch/$s_!DhMA!,w_1200,h_675,c_fill,f_jpg,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F79e364f2-7240-46b5-b975-829cf973f881_1996x1292.png",
    "og:url": "https://www.latent.space/p/agents",
    "twitter:title": "The Anatomy of Autonomy: Why Agents are the next AI Killer App after ChatGPT",
    "twitter:description": "Auto-GPT/BabyAGI Executive Summary, a Brief History of Autonomous Agentic AI, and Predictions for Autonomous Future",
    "twitter:image": "https://substackcdn.com/image/fetch/$s_!CuQq!,f_auto,q_auto:best,fl_progressive:steep/https%3A%2F%2Fswyx.substack.com%2Fapi%2Fv1%2Fpost_preview%2F115776561%2Ftwitter.jpg%3Fversion%3D4",
    "twitter:card": "summary_large_image",
    "article:modified_time": "2023-10-01T21:03:37.222Z"
  },
  "paywall": {
    "listed_paid": false,
    "detected": false,
    "marker": "",
    "public_only": false,
    "policy": "付费/会员内容只保留页面公开可见部分，不绕过付费墙"
  },
  "crawl": {
    "success": true,
    "status_code": 200,
    "error": ""
  },
  "code_blocks": [],
  "images": [
    {
      "status": "ok",
      "url": "https://substackcdn.com/image/fetch/$s_!D80L!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9a68add7-01a3-4d2d-a050-edeeeae66ddd_1996x1292.png",
      "bytes": 106303,
      "reused": true,
      "local_path": "attachments/5f6a24f7d0f7/images/027.png",
      "src": "https://substackcdn.com/image/fetch/$s_!D80L!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9a68add7-01a3-4d2d-a050-edeeeae66ddd_1996x1292.png",
      "alt": "",
      "downloaded": true
    },
    {
      "status": "ok",
      "url": "https://substackcdn.com/image/fetch/$s_!RD5d!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3d77b77f-dc06-4458-a61e-0059dc9c688f_1508x1096.png",
      "bytes": 147408,
      "reused": true,
      "local_path": "attachments/5f6a24f7d0f7/images/038.png",
      "src": "https://substackcdn.com/image/fetch/$s_!RD5d!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3d77b77f-dc06-4458-a61e-0059dc9c688f_1508x1096.png",
      "alt": "",
      "downloaded": true
    },
    {
      "status": "ok",
      "url": "https://substackcdn.com/image/fetch/$s_!zCgM!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F728f39f3-9bf1-4e75-bdfe-cb0ea2f7acbb_1200x1716.jpeg",
      "bytes": 212566,
      "reused": false,
      "local_path": "attachments/5f6a24f7d0f7/images/body-3803c8cfc149c63517a7.jpeg",
      "src": "https://substackcdn.com/image/fetch/$s_!zCgM!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F728f39f3-9bf1-4e75-bdfe-cb0ea2f7acbb_1200x1716.jpeg",
      "alt": "Image",
      "downloaded": true
    },
    {
      "status": "ok",
      "url": "https://substackcdn.com/image/fetch/$s_!cf67!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5eeaff61-b705-42c5-8d52-24f9a12d9c32_1996x1292.png",
      "bytes": 106303,
      "reused": false,
      "local_path": "attachments/5f6a24f7d0f7/images/body-0e96cab478137fc8b65e.png",
      "src": "https://substackcdn.com/image/fetch/$s_!cf67!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5eeaff61-b705-42c5-8d52-24f9a12d9c32_1996x1292.png",
      "alt": "",
      "downloaded": true
    },
    {
      "status": "ok",
      "url": "https://substackcdn.com/image/fetch/$s_!z814!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1560f3c6-b6c3-48d0-be93-4560f3290dd7_1544x1512.png",
      "bytes": 97539,
      "reused": false,
      "local_path": "attachments/5f6a24f7d0f7/images/body-279e534bd179beaee668.png",
      "src": "https://substackcdn.com/image/fetch/$s_!z814!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1560f3c6-b6c3-48d0-be93-4560f3290dd7_1544x1512.png",
      "alt": "https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1560f3c6-b6c3-48d0-be93-4560f3290dd7_1544x1512.png",
      "downloaded": true
    },
    {
      "status": "ok",
      "url": "https://substackcdn.com/image/fetch/$s_!izVZ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3751512-e1f9-4397-98c5-a62ae4e2e14e_984x924.png",
      "bytes": 231795,
      "reused": false,
      "local_path": "attachments/5f6a24f7d0f7/images/body-b58344ad3924a0ada17e.png",
      "src": "https://substackcdn.com/image/fetch/$s_!izVZ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3751512-e1f9-4397-98c5-a62ae4e2e14e_984x924.png",
      "alt": "The Rise of Autonomous Vehicles: Pros & Cons of Self-Driving Cars | Study",
      "downloaded": true
    },
    {
      "status": "ok",
      "url": "https://substackcdn.com/image/fetch/$s_!FTUI!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff08de1b1-3327-4fe0-8bc5-2ae36e717b77_1238x788.png",
      "bytes": 101355,
      "reused": false,
      "local_path": "attachments/5f6a24f7d0f7/images/body-2f126f05c6474b5c4b71.png",
      "src": "https://substackcdn.com/image/fetch/$s_!FTUI!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff08de1b1-3327-4fe0-8bc5-2ae36e717b77_1238x788.png",
      "alt": "",
      "downloaded": true
    }
  ],
  "audio": [],
  "videos": [
    {
      "url": "https://www.youtube.com/watch?v=SFKA7T-v6WE",
      "status": "ok",
      "local_path": "attachments/5f6a24f7d0f7/videos/SFKA7T-v6WE.mp4",
      "bytes": 756054354,
      "valid": true,
      "height": 720,
      "width": 1280,
      "duration": "6050.661000"
    },
    {
      "url": "https://www.youtube.com/watch?v=g0muj39vCCY",
      "status": "ok",
      "local_path": "attachments/5f6a24f7d0f7/videos/g0muj39vCCY.mp4",
      "bytes": 811549277,
      "valid": true,
      "height": 1080,
      "width": 1920,
      "duration": "3532.681000"
    }
  ],
  "links": [
    {
      "href": "https://www.latent.space/p/segment-anything-roboflow",
      "label": "the Segment Anything pod",
      "type": "a"
    },
    {
      "href": "https://www.latent.space/p/community",
      "label": "new community page",
      "type": "a"
    },
    {
      "href": "https://partiful.com/e/c4ox9BmcglwzLB3QkW3d",
      "label": "Miami",
      "type": "a"
    },
    {
      "href": "https://news.ycombinator.com/item?id=35629033",
      "label": "Hacker News",
      "type": "a"
    },
    {
      "href": "https://twitter.com/search?q=https%3A%2F%2Fwww.latent.space%2Fp%2Fagents&src=typed_query",
      "label": "Twitter",
      "type": "a"
    },
    {
      "href": "https://www.latent.space/p/multiverse-not-metaverse?utm_source=%2Fsearch%2Fmultiverse&utm_medium=reader2",
      "label": "Multiverses",
      "type": "a"
    },
    {
      "href": "https://twitter.com/swyx/status/1638550858073006089",
      "label": "Copilot X",
      "type": "a"
    },
    {
      "href": "https://www.latent.space/p/what-building-copilot-for-x-really",
      "label": "Copilot for X",
      "type": "a"
    },
    {
      "href": "https://lspace.swyx.io/p/everything-we-know-about-chatgpt",
      "label": "ChatGPT",
      "type": "a"
    },
    {
      "href": "https://substackcdn.com/image/fetch/$s_!D80L!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9a68add7-01a3-4d2d-a050-edeeeae66ddd_1996x1292.png",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://github.com/Significant-Gravitas/Auto-GPT/tree/master/autogpt",
      "label": "Auto-GPT",
      "type": "a"
    },
    {
      "href": "https://github.com/yoheinakajima/babyagi",
      "label": "BabyAGI",
      "type": "a"
    },
    {
      "href": "https://fortune.com/2023/04/15/babyagi-autogpt-openai-gpt-4-autonomous-assistant-agi/",
      "label": "30th",
      "type": "a"
    },
    {
      "href": "https://github.com/yoheinakajima/babyagi/graphs/contributors",
      "label": "April 2nd",
      "type": "a"
    },
    {
      "href": "https://twitter.com/SigGravitas/status/1646117442031349761?s=20",
      "label": "trending #1 on Twitter and GitHub",
      "type": "a"
    },
    {
      "href": "https://star-history.com/",
      "label": "far outpacing every other open source AI project",
      "type": "a"
    },
    {
      "href": "https://web.archive.org/web/20230414032253/https://www.businessinsider.com/sequoia-leads-funding-round-generative-artificial-intelligence-startup-langchain-2023-4?r=US&IR=T",
      "label": "$200m-valuation-LangChain",
      "type": "a"
    },
    {
      "href": "https://www.linkedin.com/in/toran-bruce-richards-54139814b/?originalSubdomain=uk",
      "label": "Toran Richards",
      "type": "a"
    },
    {
      "href": "https://github.com/Significant-Gravitas/Auto-GPT/commit/c74ad984cfa755031a85296c45a8a8d77e0b8906?short_path=b335630",
      "label": "original demo",
      "type": "a"
    },
    {
      "href": "https://twitter.com/jacksonfall/status/1636107218859745286",
      "label": "Jackson Fall’s viral HustleGPT prompt on ChatGPT",
      "type": "a"
    },
    {
      "href": "https://twitter.com/yoheinakajima/status/1640068466974633987?s=20",
      "label": "start and grow a mobile AI startup",
      "type": "a"
    },
    {
      "href": "https://twitter.com/SullyOmarr/status/1646632867412459521",
      "label": "built extensions",
      "type": "a"
    },
    {
      "href": "https://twitter.com/_Lonis_/status/1646641412182536196",
      "label": "clones",
      "type": "a"
    },
    {
      "href": "https://twitter.com/thegarrettscott/status/1645918390413066240",
      "label": "agent managers",
      "type": "a"
    },
    {
      "href": "https://twitter.com/dysmemic/status/1645535982996054016",
      "label": "frameworks",
      "type": "a"
    },
    {
      "href": "https://twitter.com/skirano/status/1646582731629887503",
      "label": "ChatGPT plugins",
      "type": "a"
    },
    {
      "href": "https://twitter.com/wm_eddie/status/1646752950541492225",
      "label": "visual toolkits",
      "type": "a"
    },
    {
      "href": "https://twitter.com/SullyOmarr/status/1645205292756418562",
      "label": "market research",
      "type": "a"
    },
    {
      "href": "https://twitter.com/adamcohenhillel/status/1644836492294905856",
      "label": "test driven development",
      "type": "a"
    },
    {
      "href": "https://twitter.com/eimenhmdt/status/1644199933321306112",
      "label": "scientific literature review",
      "type": "a"
    },
    {
      "href": "https://twitter.com/yoheinakajima/status/1644326587389853697?s=20",
      "label": "adding and stripping out LangChain",
      "type": "a"
    },
    {
      "href": "https://replit.com/@YoheiNakajima/babyagi?v=1",
      "label": "less than 150 lines",
      "type": "a"
    },
    {
      "href": "https://github.com/yoheinakajima/babyagi/blob/17f1e830e44d83f8b6d52e3b932ba26a98e702a0/.env.example",
      "label": "10 env vars",
      "type": "a"
    },
    {
      "href": "https://substackcdn.com/image/fetch/$s_!RD5d!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3d77b77f-dc06-4458-a61e-0059dc9c688f_1508x1096.png",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://github.com/yoheinakajima/babyagi/pull/207",
      "label": "Other attempts",
      "type": "a"
    },
    {
      "href": "https://twitter.com/SigGravitas/status/1641649430230351872?s=20",
      "label": "speak",
      "type": "a"
    },
    {
      "href": "https://github.com/Significant-Gravitas/Auto-GPT/blob/ecf2ba12db11ff19bce359b842f810f0e2d09d6a/.env.template",
      "label": "50 env vars",
      "type": "a"
    },
    {
      "href": "https://substackcdn.com/image/fetch/$s_!zCgM!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F728f39f3-9bf1-4e75-bdfe-cb0ea2f7acbb_1200x1716.jpeg",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://github.com/Significant-Gravitas/Auto-GPT/blob/ecf2ba12db11ff19bce359b842f810f0e2d09d6a/autogpt/prompt.py",
      "label": "full list from source",
      "type": "a"
    },
    {
      "href": "https://twitter.com/karpathy/status/1642598890573819905?s=20",
      "label": "next frontier of prompt engineering",
      "type": "a"
    },
    {
      "href": "https://twitter.com/ESYudkowsky/status/1640511156254289926",
      "label": "approvingly observing",
      "type": "a"
    },
    {
      "href": "https://blogs.scientificamerican.com/talking-back/einsteins-brain-more-special-than-we-ever-knew/",
      "label": "every convolution",
      "type": "a"
    },
    {
      "href": "https://substackcdn.com/image/fetch/$s_!cf67!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5eeaff61-b705-42c5-8d52-24f9a12d9c32_1996x1292.png",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://yaofu.notion.site/How-does-GPT-Obtain-its-Ability-Tracing-Emergent-Abilities-of-Language-Models-to-their-Sources-b9a57ac0fcf74f30a1ab9e3e36fa1dc1",
      "label": "3 major features",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/pdf/2301.00774.pdf",
      "label": "175B Parameters can store 320GB",
      "type": "a"
    },
    {
      "href": "https://www.jasonwei.net/blog/emergence",
      "label": "emergence",
      "type": "a"
    },
    {
      "href": "https://gwern.net/gpt-3",
      "label": "Gwern Branwern",
      "type": "a"
    },
    {
      "href": "https://twitter.com/goodside/status/1569128808308957185",
      "label": "Riley Goodside",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/abs/2205.11916",
      "label": "Kojima et al (2022)",
      "type": "a"
    },
    {
      "href": "https://twitter.com/shaneguML/status/1530244895713046528",
      "label": "let’s think step by step",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/abs/2201.11903",
      "label": "Wei et al (2022)",
      "type": "a"
    },
    {
      "href": "https://learnprompting.org/docs/intermediate/self_consistency",
      "label": "Wang et al (2022)",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/pdf/2102.09690.pdf",
      "label": "Calibrate Before Use",
      "type": "a"
    },
    {
      "href": "https://ofir.io/self-ask.pdf",
      "label": "Self-Asking",
      "type": "a"
    },
    {
      "href": "https://twitter.com/johnjnay/status/1641786389267185664",
      "label": "Recursively Criticize and Improve",
      "type": "a"
    },
    {
      "href": "https://sites.google.com/view/automatic-prompt-engineer",
      "label": "Automatic Prompt Engineering",
      "type": "a"
    },
    {
      "href": "https://archive.is/wCoqH",
      "label": "Pinecone (now worth $700m), Weaviate ($200m), and Chroma ($75m)",
      "type": "a"
    },
    {
      "href": "https://techcrunch.com/2023/04/04/ai-powered-search-engine-perplexity-ai-lands-26m-launches-ios-app/",
      "label": "$26m Series A",
      "type": "a"
    },
    {
      "href": "https://lspace.swyx.io/p/sarah-nagy",
      "label": "$7.5m Seed",
      "type": "a"
    },
    {
      "href": "http://censusgpt.com/",
      "label": "CensusGPT",
      "type": "a"
    },
    {
      "href": "https://ossinsight.io/explore/",
      "label": "OSS Insight",
      "type": "a"
    },
    {
      "href": "https://twitter.com/sharifshameem/status/1645664932418162688",
      "label": "GPT-3 automating Chrome to buy Airpods",
      "type": "a"
    },
    {
      "href": "https://www.businesswire.com/news/home/20220426005963/en/AI-Transformer-Inventors-Launch-Adept-with-65M-to-Lend-a-Hand-to-Knowledge-Workers",
      "label": "Adept raised a Series A",
      "type": "a"
    },
    {
      "href": "https://www.adept.ai/blog/act-1",
      "label": "ACT-1 Action Transformer",
      "type": "a"
    },
    {
      "href": "https://www.forbes.com/sites/kenrickcai/2023/03/14/adept-ai-startup-raises-350-million-series-b/?sh=7c29c7672cc3",
      "label": "hefty $350m Series B",
      "type": "a"
    },
    {
      "href": "https://twitter.com/natfriedman/status/1575631194032549888",
      "label": "Nat Friedman’s NatBot",
      "type": "a"
    },
    {
      "href": "https://dust.tt/xp1",
      "label": "Dust XP1",
      "type": "a"
    },
    {
      "href": "https://multion.ai/",
      "label": "MULTI·ON",
      "type": "a"
    },
    {
      "href": "https://twitter.com/DivGarg9/status/1648074780430696448",
      "label": "the ChatGPT Plugin Store",
      "type": "a"
    },
    {
      "href": "https://embra.app/",
      "label": "Embra AI",
      "type": "a"
    },
    {
      "href": "https://twitter.com/RewindAI",
      "label": "Rewind AI",
      "type": "a"
    },
    {
      "href": "https://www.latent.space/p/multimodal-gpt4",
      "label": "Multi-modal GPT4",
      "type": "a"
    },
    {
      "href": "https://openai.com/blog/webgpt/",
      "label": "WebGPT",
      "type": "a"
    },
    {
      "href": "https://twitter.com/dust4ai/status/1587104029712203778",
      "label": "Dust",
      "type": "a"
    },
    {
      "href": "https://www.notion.so/44485e5c97bd403ba4e1c2d5197af71d",
      "label": "others",
      "type": "a"
    },
    {
      "href": "https://twitter.com/goodside/status/1568448128495534081",
      "label": "Riley Goodside to ask it to generate code",
      "type": "a"
    },
    {
      "href": "https://twitter.com/amasad/status/1568825727528878081",
      "label": "Replit turned out",
      "type": "a"
    },
    {
      "href": "https://twitter.com/sergeykarayev/status/1569377881440276481",
      "label": "another example here",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/abs/2210.03629",
      "label": "Yao et all (2022)",
      "type": "a"
    },
    {
      "href": "https://react-lm.github.io/",
      "label": "the ReAct pattern",
      "type": "a"
    },
    {
      "href": "https://til.simonwillison.net/llms/python-react-pattern",
      "label": "delightfully simple",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/abs/2302.04761",
      "label": "Schick et al (2023)",
      "type": "a"
    },
    {
      "href": "https://github.com/microsoft/JARVIS",
      "label": "HuggingGPT/Microsoft JARVIS",
      "type": "a"
    },
    {
      "href": "https://github.com/microsoft/visual-chatgpt/tree/main",
      "label": "VisualChatGPT",
      "type": "a"
    },
    {
      "href": "https://twitter.com/DYtweetshere/status/1631349179934203904",
      "label": "Self-Learning Agent for Performing APIs",
      "type": "a"
    },
    {
      "href": "https://twitter.com/mitchellh/status/1638966754226610181?lang=en",
      "label": "also used natural language",
      "type": "a"
    },
    {
      "href": "https://fixie.ai/",
      "label": "Fixie AI",
      "type": "a"
    },
    {
      "href": "https://minion.ai/",
      "label": "Minion AI",
      "type": "a"
    },
    {
      "href": "https://jmcdonnell.substack.com/p/the-near-future-of-ai-is-action-driven",
      "label": "this excellent post from John McDonnell",
      "type": "a"
    },
    {
      "href": "https://substackcdn.com/image/fetch/$s_!z814!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1560f3c6-b6c3-48d0-be93-4560f3290dd7_1544x1512.png",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://twitter.com/jerryjliu0/status/1647626532519841793",
      "label": "LlamaIndex",
      "type": "a"
    },
    {
      "href": "https://blog.langchain.dev/retrieval/",
      "label": "Langchain",
      "type": "a"
    },
    {
      "href": "https://twitter.com/yoheinakajima/status/1644326588794941440?s=20",
      "label": "relevant (task) context",
      "type": "a"
    },
    {
      "href": "https://www.latent.space/p/benchmarks-101",
      "label": "simple common sense benchmarks",
      "type": "a"
    },
    {
      "href": "https://nanothoughts.substack.com/p/reflecting-on-reflexion",
      "label": "Shinn et al (2023)",
      "type": "a"
    },
    {
      "href": "https://twitter.com/atroyn/status/1645957559298449408",
      "label": "Shoggoth the Coder",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/abs/2304.03442",
      "label": "Meta’s Simulacra paper",
      "type": "a"
    },
    {
      "href": "https://arxiv.org/pdf/2303.12712",
      "label": "The “Sparks of AGI” paper",
      "type": "a"
    },
    {
      "href": "https://twitter.com/jh_damm/status/1646233661832929280?s=20",
      "label": "LangChain Agents webinar",
      "type": "a"
    },
    {
      "href": "https://github.com/e2b-dev/e2b",
      "label": "e2b",
      "type": "a"
    },
    {
      "href": "https://www.latent.space/p/chatgpt-plugins",
      "label": "with Plugins",
      "type": "a"
    },
    {
      "href": "https://github.com/Significant-Gravitas/Auto-GPT",
      "label": "continuous modes",
      "type": "a"
    },
    {
      "href": "https://twitter.com/jbrowder1/status/1612312707398795264",
      "label": "AI Lawyer",
      "type": "a"
    },
    {
      "href": "https://www.youtube.com/watch?v=SFKA7T-v6WE",
      "label": "AI Waifu",
      "type": "a"
    },
    {
      "href": "https://twitter.com/karpathy/status/1642610417779490816?s=20",
      "label": "AI C-Suite",
      "type": "a"
    },
    {
      "href": "https://sae.org/",
      "label": "the Society of Automotive Engineers",
      "type": "a"
    },
    {
      "href": "https://substackcdn.com/image/fetch/$s_!izVZ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc3751512-e1f9-4397-98c5-a62ae4e2e14e_984x924.png",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://www.steinlaw.com/resources/studies/the-rise-of-autonomous-vehicles/",
      "label": "the 5 levels of autonomy",
      "type": "a"
    },
    {
      "href": "https://www.codium.ai/blog/the-6-levels-of-autonomous-unit-testing-explained/",
      "label": "6 levels of Autonomous Code Integrity",
      "type": "a"
    },
    {
      "href": "https://www.themarginalian.org/2011/12/21/steve-jobs-bicycle-for-the-mind-1990/",
      "label": "bicycles for your mind",
      "type": "a"
    },
    {
      "href": "https://www.mattprd.com/p/the-complete-beginners-guide-to-autonomous-agents",
      "label": "The Complete Beginners Guide To Autonomous Agents",
      "type": "a"
    },
    {
      "href": "https://blog.langchain.dev/agents-round/",
      "label": "Autonomous Agents",
      "type": "a"
    },
    {
      "href": "https://twitter.com/swyx/status/1648539094438273024",
      "label": "neither Babyagi nor AutoGPT use LangChain",
      "type": "a"
    },
    {
      "href": "https://python.langchain.com/en/latest/use_cases/autonomous_agents.html",
      "label": "LangChain’s docs",
      "type": "a"
    },
    {
      "href": "https://twitter.com/swyx/status/1648398117417680901",
      "label": "vector db’s > memorizing transformers",
      "type": "a"
    },
    {
      "href": "https://youtu.be/g0muj39vCCY",
      "label": "Avalon",
      "type": "a"
    },
    {
      "href": "https://github.com/hwchase17/",
      "label": "harrison",
      "type": "a"
    },
    {
      "href": "https://twitter.com/yoheinakajima/status/1644326591428972546?s=20",
      "label": "yohei",
      "type": "a"
    },
    {
      "href": "https://github.com/Significant-Gravitas?type=source",
      "label": "siggravitas",
      "type": "a"
    },
    {
      "href": "https://twitter.com/yoheinakajima/status/1644326606767546368?s=20",
      "label": "stepping on each other",
      "type": "a"
    },
    {
      "href": "https://twitter.com/hadiazouni/status/1648375237283659792?s=20",
      "label": "Actor model",
      "type": "a"
    },
    {
      "href": "https://twitter.com/karpathy/status/1642607620673634304",
      "label": "reflection? metalearning?",
      "type": "a"
    },
    {
      "href": "https://twitter.com/sarahcat21/status/1646920947134504962",
      "label": "sarahcat observation",
      "type": "a"
    },
    {
      "href": "https://www.swyx.io/temporal-centicorn",
      "label": "swyx commentary on cyborg nodes",
      "type": "a"
    },
    {
      "href": "https://www.latent.space/p/agi-hard",
      "label": "AGI-hard",
      "type": "a"
    },
    {
      "href": "https://twitter.com/karpathy/status/1642598890573819905",
      "label": "karpathy",
      "type": "a"
    },
    {
      "href": "https://twitter.com/simonw/status/1648521987776745472?s=20",
      "label": "hey @simon",
      "type": "a"
    },
    {
      "href": "https://twitter.com/karpathy/status/1707437820045062561",
      "label": "based on a @karpathy",
      "type": "a"
    },
    {
      "href": "https://substackcdn.com/image/fetch/$s_!FTUI!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff08de1b1-3327-4fe0-8bc5-2ae36e717b77_1238x788.png",
      "label": "",
      "type": "a"
    },
    {
      "href": "https://openai.com/research/gpts-are-gpts",
      "label": "recent OpenAI paper",
      "type": "a"
    },
    {
      "href": "https://en.wikipedia.org/wiki/General-purpose_technology",
      "label": "an actual concept",
      "type": "a"
    },
    {
      "href": "https://learninpublic.org/toc",
      "label": "my book",
      "type": "a"
    },
    {
      "href": "https://www.nytimes.com/2021/06/10/podcasts/the-daily/mrna-vaccines-katalin-kariko.html",
      "label": "Dr. Katalin Kariko’s story of perseverance",
      "type": "a"
    },
    {
      "href": "https://github.com/sw-yx/brain",
      "label": "my full list",
      "type": "a"
    },
    {
      "href": "https://dx.tips/platform-kinds",
      "label": "Zawinswyx’s Law",
      "type": "a"
    },
    {
      "href": "https://twitter.com/Cruise/status/1644073327475171328",
      "label": "The future is very close!",
      "type": "a"
    }
  ],
  "repair_version": 2,
  "repaired_at": "2026-10-14T10:47:28Z",
  "source_url": "https://www.latent.space/p/agents",
  "papers": [
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/2301.00774.pdf",
      "bytes": 740346,
      "reused": false,
      "local_path": "papers/repair/71922b3a81cd0c9e4509.pdf"
    },
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/2205.11916",
      "bytes": 762445,
      "reused": true,
      "local_path": "papers/repair/7ee8d7eb1358c6773ecd.pdf"
    },
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/2201.11903",
      "bytes": 891773,
      "reused": true,
      "local_path": "papers/repair/52cc4c4f6cf368123b97.pdf"
    },
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/2102.09690.pdf",
      "bytes": 695668,
      "reused": false,
      "local_path": "papers/repair/b876569b48ce49238820.pdf"
    },
    {
      "status": "ok",
      "url": "https://ofir.io/self-ask.pdf",
      "bytes": 936780,
      "reused": false,
      "local_path": "papers/repair/52cff967b56373520ff5.pdf"
    },
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/2210.03629",
      "bytes": 633805,
      "reused": true,
      "local_path": "papers/repair/c08ee62f780204fcfff6.pdf"
    },
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/2302.04761",
      "bytes": 657966,
      "reused": true,
      "local_path": "papers/repair/68512ec86fa6106aa473.pdf"
    },
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/2304.03442",
      "bytes": 11947867,
      "reused": true,
      "local_path": "papers/repair/68f98ce7167c163414a4.pdf"
    },
    {
      "status": "ok",
      "url": "https://arxiv.org/pdf/2303.12712",
      "bytes": 7370286,
      "reused": true,
      "local_path": "papers/repair/8c253eba6c79c3396891.pdf"
    }
  ],
  "actual_video_files": [
    {
      "local_path": "attachments/5f6a24f7d0f7/videos/SFKA7T-v6WE.mp4",
      "bytes": 756054354,
      "valid": true,
      "height": 720,
      "width": 1280,
      "duration": "6050.661000",
      "status": "ok"
    },
    {
      "local_path": "attachments/5f6a24f7d0f7/videos/g0muj39vCCY.mp4",
      "bytes": 811549277,
      "valid": true,
      "height": 1080,
      "width": 1920,
      "duration": "3532.681000",
      "status": "ok"
    }
  ],
  "body_chars": 18850,
  "expected_body_images": 7,
  "discovery_status": "ok",
  "delivery_status": "ok",
  "missing_reason": "",
  "raw_html": "attachments/5f6a24f7d0f7/pages/source.html",
  "local_html": "attachments/5f6a24f7d0f7/pages/article.html",
  "body_base_path": "attachments/5f6a24f7d0f7/pages/",
  "body_markdown_path": "attachments/5f6a24f7d0f7/pages/article.md"
}