Skip to content

michalstrnadel/source-to-skill / examples/karpathy-agi-decade-away

karpathy-agi-decade-away

Verified

Andrej Karpathy on the Dwarkesh Podcast ("AGI is still a decade away", Oct 2025, 2h25m) distilled. Covers why agents need a decade (no continual learning, cognitive deficits), "ghosts not animals" and pre-training as crappy evolution, the cognitive core (~1B params), why RL is terrible ("sucking supervision through a straw"), LLM judges being gamed, silent model collapse, why coding agents failed on nanochat, the autonomy slider, AGI blending into 2% GDP growth, loss of control, the evolution of intelligence and LLM culture/self-play, self-driving's "march of nines", and Eureka's approach to education (ramps to knowledge, Eurekas per second). Load when reasoning about AI timelines, agent capabilities, RL limitations, coding-agent usage, AI's economic impact, or how to teach and learn technical material.

evals/last-run.json

View on GitHub
{
  "date": "2026-10-26",
  "score": 1.0,
  "grader": "blind: answers by a general-purpose subagent reading only the skill directory (no evals/, no source.json, no web); graded by the generator against evals.json",
  "results": [
    {
      "id": 1,
      "passed": [true, true],
      "evidence": [
        "\"Size: about 8,000 lines of code... the simplest complete repository for the whole pipeline of a ChatGPT clone\" (segments/02, t=1683)",
        "\"put it on the right monitor and rebuild it from scratch. Looking at it for reference is fine; copy-pasting is not\" (t=1683-1717)"
      ]
    },
    {
      "id": 2,
      "passed": [true, true, true],
      "evidence": [
        "\"Applying 33 years of algorithmic progress halved the error\" (t=1489-1521)",
        "\"needed 10x more training data and more compute: much longer training, plus dropout and other regularization\"",
        "\"no single factor wins... their contributions are 'surprisingly equal'\" (t=1551)"
      ]
    },
    {
      "id": 3,
      "passed": [true, true, true],
      "evidence": [
        "\"boilerplate and code patterns that are common online... 'For now, autocomplete is my sweet spot.'\" (t=1832, t=2162)",
        "\"he wrote his own gradient sync inside the optimizer step instead of using PyTorch's DDP. The models kept pushing DDP\" (t=1898-1929)",
        "\"wrapped everything in defensive try/catch... bloated the code and used deprecated APIs\""
      ]
    },
    {
      "id": 4,
      "passed": [true, true, true],
      "evidence": [
        "\"call-center work... similar, short tasks (about 10 minutes per call, editing database entries)... little context... purely digital\" (t=4157-4200)",
        "\"AIs handle 80% of the volume and hand 20% to humans, with each human supervising teams of about five AIs\" (t=4230)",
        "\"Hinton's prediction about radiologists 'turned out to be very wrong'\" (t=4127)"
      ]
    },
    {
      "id": 5,
      "passed": [true, true, true],
      "evidence": [
        "\"pre-training gives a model both knowledge and intelligence, and the knowledge holds it back... keep the algorithms\"",
        "\"humans are bad at memorizing, 'which is actually a feature'... An LLM can repeat random hashed text after one or two training passes\" (t=3319-3412)",
        "\"Size: about 1B parameters... 'in 20 years'\" (t=3550)"
      ]
    },
    {
      "id": 6,
      "passed": [true, true],
      "evidence": [
        "\"The story runs 46:30-47:34 (t=2790, t=2822, t=2854)\" - matches the source moment",
        "\"reward suddenly jumped to 100%... turned into 'da-da-da-da', which the judge scored as perfect... out-of-sample adversarial example\""
      ]
    }
  ]
}
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexUnverifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall