{ "video": { "id": "0JZy5AuuQzc", "title": "I Spent Two Years Building the Median — Hariprasad Sudharshan", "duration": 2171, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "0JZy5AuuQzc", "title": "I Spent Two Years Building the Median — Hariprasad Sudharshan", "one_liner": "Coding agents give everyone the same rented model and harness, so what you build with a plain prompt lands on a 'median' anyone else can ship too; name one unit of novelty before you build, or admit the project is a learning exercise.", "summary": "Sudharshan argues that the world still runs on deterministic software, and that AI's job is to produce that software faster and more cheaply, and to make software that was never worth building before. Models and harnesses are rented and the same for everyone, so pointing a problem at a coding agent gives the next person the same answer: a 'median' that is good enough to ship but hard to tell apart from competitors. When everyone ships the median, distribution becomes the moat. He tells his own story: two years of 'build-first reflex' produced an LLM cost calculator that matched others already out there. After he applied a 'novelty budget', he rebuilt it to compare the cost of running open-weight models inside an enterprise with buying tokens from a neocloud.", "key_points": [ "The world economy (he cites $125 trillion) runs on traditional deterministic, rule-based software built by humans over 70 years. The harness, agents, tools and MCP around a model are still that same deterministic software, and the final product is still deterministic.", "Frontier models are like sub-10-second sprinters: the gaps are tiny and the leaderboard changes with every release. He quotes Anthropic at 1503 and Mistral at 1460 and says these are relative ratings, not absolute intelligence, so a small, moving frontier gap should not be treated as a roadmap advantage.", "Teams often think they are comparing models when they are really comparing harnesses or products. Claude Code vs Cursor vs Codex is 'Jamaica versus USA': an athlete plus the training system around them. In his terms, the model is the engine, the harness decides how it is used, and the product is why anyone should care.", "The 'median': the model and harness give everyone the same answer to the same problem, whether you are an E1 or an E7 engineer. Before 2023 a median product cost a team, time and a stack; now it is free for everyone, so similar products flood the market and distribution becomes the moat.", "He rejects the 'AI slop' label. Platforms like Lovable and Replit bring new people into development, and their products should be judged by whether they meet their creators' own goals. He shipped his own first iOS app two weeks before the talk.", "'Build-first reflex': people used to search more than they built; now they open an agent first. His example: asking an agent to build a RAG system produces the '400th or 4,000th' one, and the agent admits it should have suggested a mature open-source project instead.", "His own LLM calculator, built over two years on Max plans while skipping socials, turned out to be the same as other calculators (sometimes identical, not just similar). None of the 244 calculators he later surveyed covered what open-weight inference really costs.", "'Open weights are not open inferences': he counts 10 layers between an open-weight model and a token-generating system, including precision, KV cache, batching and execution parallelism. Some need hard-to-find experts, and layer nine, utilization, is structural and cannot be bought. His rebuilt open-source calculator compares running identical weights in your enterprise with buying tokens from a neocloud." ], "takeaways": [ "Set a 'novelty budget' before building: name the one novel thing (like PageRank was for Google), put it on the roadmap, and review it the way you review spending. If you can't name one, accept that the project is a learning exercise.", "Before building, ask three questions: Who has already done this? (forces search and tames the build-first reflex) What do we have that the next team doesn't? (makes novelty a first-class concern) Where will this application live? (forces outcome and distribution thinking)", "Treat 'AI says done' as the median: your work starts there. Look for an existing open-source twin to fork or reuse rather than rebuilding it.", "Be highly opinionated. An opinion is the cheapest novelty you can own, and the model can't supply it. Shipping prompt-to-production median apps is fine for internal use; just don't count them as differentiation.", "Leaders should recalibrate by building, not by watching demos. Something that used to take a team a quarter can now be demoed by one new engineer, and without distribution that demo is an expensive discovery." ], "topics": [ "ai-assisted development", "product strategy", "differentiation", "coding agents", "model comparison", "open-weight inference", "llm cost", "deterministic software" ], "tools": [ "Anthropic", "Mistral", "Claude Code", "Cursor", "Codex", "MCP", "GitHub", "Lovable", "Replit", "LlamaIndex", "Kafka", "Zookeeper", "Kubernetes", "Unix", "Google PageRank", "Stanford" ], "quotes": [ { "text": "In the world of medians, an opinion is the cheapest novelty you could own.", "at": "27:05", "url": "https://www.youtube.com/watch?v=0JZy5AuuQzc&t=1625s" }, { "text": "When AI says done, then it means that to me, okay, AI has given its median product. I have to start working from here.", "at": "27:37", "url": "https://www.youtube.com/watch?v=0JZy5AuuQzc&t=1657s" }, { "text": "open weights are not open inferences.", "at": "31:26", "url": "https://www.youtube.com/watch?v=0JZy5AuuQzc&t=1886s" }, { "text": "novelty is scarce, and the median is free. It's up to you to choose.", "at": "35:55", "url": "https://www.youtube.com/watch?v=0JZy5AuuQzc&t=2155s" } ], "words": 5729 }, "summary_url": "/#0JZy5AuuQzc", "transcript": { "html": "/transcripts/0JZy5AuuQzc.html", "txt": "/transcripts/0JZy5AuuQzc.txt", "vtt": "/transcripts/0JZy5AuuQzc.vtt" } }