{ "video": { "id": "LZuWZRze3MU", "title": "The New Primitives: Building AI Native Software — Kwindla Kramer, Daily", "duration": 1274, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "LZuWZRze3MU", "title": "The New Primitives: Building AI Native Software — Kwindla Kramer, Daily", "one_liner": "Kwindla Kramer (Daily, Pipecat) places today's agent-building inside the 80-year arc from Vannevar Bush's 1945 'As We May Think' to argue that agents are the web pages of 1995 — a real but early primitive — and that the real prize is the fully AI-native software that comes after, which he demos as Gradient Bang, a multiplayer game running hundreds of inference calls at every moment.", "summary": "Kramer opens by noting Pipecat is the most widely used voice-agent framework — open source, vendor neutral, used by AWS, Nvidia, Anthropic and thousands of startups — then spends most of the talk on history rather than voice. He walks the decades (1950s programming languages and compilers, 1960s interactivity and Sketchpad, 1970s relational databases and Smalltalk, 1980s personal computer and VisiCalc, 1990s multimodal web, 2000s mobile, 2010s cloud) to show what had to be built between the stored-program computer and the personal computer, and reads Apple's 1987 Knowledge Navigator as the midpoint of that timeline. His argument: we're at the 'web pages in 1995' stage, moving through what he calls 'agents++' (Nadella's multimodal harnesses, progressive tool disclosure, org-level harnesses) toward AI-native software, and we can now actually build the Memex, Jarvis and Knowledge Navigator for real. He closes with Gradient Bang, a massively multiplayer game built with LLMs at the core of every interaction, as a canvas for discovering what that software looks like.", "key_points": [ "Pipecat, from Daily, is presented as the most widely used framework for building voice agents today — open source and vendor neutral, used by AWS, Nvidia, Anthropic and thousands of startups, scale-ups and enterprises.", "The central analogy: agents today are what web pages were in 1995 — Kramer was a baby programmer then writing HTML by hand, web servers and search in C, and authoring tooling in Perl, and was as excited about HTML then as about agents now; web pages persist but we mostly talk about web and native mobile applications instead.", "'Agents++' is the intermediate step already visible: a Satya Nadella clip (No Priors / Latent Space crossover) describes a loop across models, data and tools, every product (GitHub Copilot, Security Copilot) as a multimodal harness with tool access and progressive disclosure of tools for token efficiency, fed rich context; Kramer adds copilots embedded in every piece of software and organization-level harnesses.", "Vannevar Bush's 1945 'As We May Think' predicted screen document display, scanning/OCR, speech-to-text and text-to-speech, programming languages, hypertext, search engines, data networks, something like the GoPro, something like the Kindle store, voice interfaces and brain-computer interfaces — and Kramer revises Bush's abacus→calculating-machine timeline into abacus → stored-program computer → personal computer (40 years) → AI agents era (40 years after that).", "VisiCalc is his counter-argument to AI mass-unemployment fear: it didn't put accountants out of business, it made far more accounting-like work possible and created categories of work nobody could conceive of when a screen's worth of calculations took a roomful of people — and most of the room still uses its direct descendant in Sheets or Excel.", "With hindsight, he says the single most important thing about the web was that it was multimodal from the very beginning — text, audio, video and data belonged together, more so than the GUIs of the 1980s — and the web was a conscious attempt to make Knowledge Navigator real.", "Apple's 1987 Knowledge Navigator concept video already showed a foldable tablet, touchscreen, a conversational voice assistant with strong personality, global and personal information access, real-time video generation, real-time computer vision, video-call integration, delegation of complex tasks for autonomous execution, and what we'd now call continual learning; it splits the 80-year timeline neatly in half.", "Minority Report's gestural interface was screen special effects but based on John Underkoffler's MIT Media Lab PhD work; Kramer co-founded a startup with Underkoffler in 2006 to commercialize it and showed their 2012 demo reel, six years in — the same multimodal, multi-device, multi-screen, multi-player connected computer he's still working on 15 years later.", "Tavus recently released a reimagined Knowledge Navigator video built entirely on real, available technology — the full four minutes is one take, completely real.", "Gradient Bang, a massively multiplayer game built with friends, has LLMs at the core of every interaction with hundreds of inference calls happening at every moment — impossible even a year ago — and demonstrates asynchronous non-blocking context compression, long-running sub-agents that share context, progressive skills loading, dynamic UI generation, and conversational voice." ], "takeaways": [ "Don't treat the agent as the destination. Build agents, but ask what the equivalent of 'web applications' is for this era — the AI-native software that stands to agents as today's internet does to 1995 web pages.", "Structure products as multimodal harnesses looping over models, data and tools, with progressive disclosure of tools to stay token-efficient and rich context feeding the loop — the Nadella framing Kramer endorses.", "Use the 1987 Knowledge Navigator video and Tavus's 2026 all-real remake as a concrete, buildable spec: watch both in full, and note that every capability in the original is now implementable.", "Treat multimodality as foundational rather than a feature: text, audio, video and data belong in one system, which is the property Kramer credits for the web's importance.", "Build a deliberately over-ambitious canvas project (his is a game with hundreds of inference calls per moment) to discover AI-native patterns — async non-blocking context compression, context-sharing long-running sub-agents, progressive skills loading, dynamic UI generation — that you won't find by incrementally improving an existing agent." ], "topics": [ "agents", "voice-agents", "ai-native-software", "computing-history", "multimodal", "hci", "context-engineering", "sub-agents" ], "tools": [ "Pipecat", "Daily", "AWS", "Nvidia", "Anthropic", "GitHub Copilot", "Microsoft Security Copilot", "Microsoft", "Apple", "Macintosh System 2", "Windows 1.0", "VisiCalc", "Google Sheets", "Microsoft Excel", "Smalltalk", "Sketchpad", "Tavus", "Gradient Bang", "No Priors", "Latent Space" ], "quotes": [ { "text": "I was as excited about HTML in 1995 as I am about agents today.", "at": "04:27", "url": "https://www.youtube.com/watch?v=LZuWZRze3MU&t=267s" }, { "text": "Because VisiCalc didn't put accountants out of business. Instead, it made much much much much more accounting-like work possible.", "at": "10:18", "url": "https://www.youtube.com/watch?v=LZuWZRze3MU&t=618s" }, { "text": "the single most important thing about the web was that it was multimodal from the very beginning.", "at": "12:35", "url": "https://www.youtube.com/watch?v=LZuWZRze3MU&t=755s" }, { "text": "At every moment in the game, there are hundreds of inference calls happening and we couldn't have built anything like this even a year ago.", "at": "19:25", "url": "https://www.youtube.com/watch?v=LZuWZRze3MU&t=1165s" } ], "words": 3791 }, "summary_url": "/#LZuWZRze3MU", "transcript": { "html": "/transcripts/LZuWZRze3MU.html", "txt": "/transcripts/LZuWZRze3MU.txt", "vtt": "/transcripts/LZuWZRze3MU.vtt" } }