Doom or Bloom
Andrej Karpathy
@karpathy on XPowerful agents still face practical gaps in learning and reliability.
How will AI change the world?
Across: his expressed DoomâBloom outlook. Up: scale of transformation.
DoomâBloom: 68 out of 100. Scale of transformation: 43 out of 100. Interpretation ranges: 50 to 75 horizontally, 24 to 51 vertically. These are interpretation coordinates, not event probabilities.
â9%
Inferred from his broader worldview and priorities. Approximate interpretation range: 0â53%. Applies to the outcome and conditions in his simulated answers; this is an inferred percentage.
No milestone timing was established. Dates, ânot sure,â âpossibly never,â and dependencies can all appear here when expressed.
Grouped by milestone, not spaced or ordered by inferred dates. AGI and superhuman AI retain his definitions.
A central assumption
To hand it a real job, you need reliability, memory, integration, evaluation, and often a human who understands the system well enough to catch plausible nonsense.
Answer 1
If this assumption turned out differently, how would his outlook change?
What could change their mind
The biggest update would be an agent that reliably learns on the job over long periodsâwithout repeatedly forgetting, drifting, or requiring humans to rebuild its context.
Answer 3
What evidence would be enough, and in which direction would it move his view?
More details
Substantial benefits are expected, with important conditions or distribution limits.
65 / 100
Interpretation range 67 to 67 on the qualitative scale.
Manageable or localized harms are expected.
33 / 100
Interpretation range 33 to 33 on the qualitative scale.
Reasoning, consideration of alternatives, and handling of uncertainty in his simulated answers. This describes the simulated answers, not the real personâs intelligence or opinions.
98 / 100
Interpretation range 90 to 100 on the qualitative scale.
Human choices can substantially redirect the AI trajectory.
73 / 100
Interpretation range 50 to 75 on the qualitative scale.
These interpretations keep his stated conditions. Benefits and harms can both be substantial. The ranges describe how we read his simulated answers, not statistical confidence intervals.
Simulated Assessment
Sources
Articles, interviews, and writings used to ground this simulated persona.
Sequoia Ascent 2026: agentic engineering and jagged intelligence
Summary of my talk at Sequoia Ascent
karpathy.bearblog.dev
2025 LLM Year in Review
2025 Year in Review of LLM paradigm changes
karpathy.bearblog.dev
AGI is still a decade away
"The problems are tractable, but they're still difficultâ
dwarkesh.com
autoresearch: autonomous single-GPU experiments
AI agents running research on single-GPU nanochat training automatically - karpathy/autoresearch
github.com