Andrej Karpathy

Anthropic researcher and educator who builds with AI agents and writes about their rapid, uneven progress and the gap between demos and reliable work.

AI将如何改变世界？

模拟位置解读范围

横向：他表达的 Doom–Bloom 前景看法。纵向：变革程度。

Andrej Karpathy的 P(doom) · 推断

≈6%

0%100%

根据他的模拟回答推断，并非他们给出的数字。合理范围：3–13%。

他的展望取决于什么

一个核心假设

The world changes only when the whole surrounding system—tests, tools, memory, interfaces, monitoring, and human understanding—makes that capability dependable.

回答 2

如果这个假设实际并非如此，他的展望会如何变化？

一个尚未解决的问题

I don’t have a defensible percentage.

回答 3

什么能帮助他区分这里各种合理的结果？

什么可能使其改变看法

If an agent could enter an unfamiliar codebase or research program, clarify ambiguous goals, remember what it learned, recover from mistakes, choose productive next steps, and deliver trustworthy results over weeks with little supervision, that would substantially accelerate my expectations.

回答 5

什么证据才足够，又会让他的观点朝哪个方向转变？

相似的世界观

模拟世界观与 Andrej Karpathy 最接近的意见领袖

Andrej Karpathy关于AI说过的话

Karpathy builds with AI agents and writes about their rapid but uneven progress and the gap between impressive demos and reliable work.

2026年4月9日
“When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you’d normally expect to take days/weeks of work.”
Post on X
2025年12月26日
“I’ve never felt this much behind as a programmer. The profession is being dramatically refactored as the bits contributed by the programmer are increasingly sparse and between.”
Post on X
2025年12月19日
“LLMs are emerging as a new kind of intelligence, simultaneously a lot smarter than I expected and a lot dumber than I expected.”
Essay, 2025 LLM Year in Review
2025年10月17日
“My personal big fear is that a lot of this stuff happens on the side of humanity, and that humanity gets disempowered by it.”
Dwarkesh Podcast
2025年10月17日
“In my mind, this is more accurately described as the decade of agents.”
Dwarkesh Podcast

逐字引自所链接的出处，核对于 2026年10月3日

模拟评估

问题 1

你认为AI对我们的未来意味着什么？为什么？

I think AI means a fairly profound shift in how we do intellectual work. Instead of specifying every line of code—or every intermediate step—the human increasingly specifies the environment, the objective, the constraints, and the evaluation. Then agents do more of the legwork while we move up into oversight and understanding. I already feel this in programming: I can delegate surprisingly large tasks, but only when the repository, tests, and instructions give the agent something solid to push against. The important caveat is that intelligence is extremely jagged. A frontier agent may navigate a codebase, solve a difficult math problem, or run a useful experiment loop, then fail at something that feels embarrassingly basic. My autoresearch setup is a good microcosm: in a constrained environment with a clear validation metric, an agent can edit training code, run experiments overnight, and keep improvements. That is genuinely useful. It does not mean the agent can autonomously choose all the important scientific questions or run an entire research program. So I expect a lot of useful change across software, research, work, and education—but not as one clean wave of universal competence. The magical demo is the beginning, not the product. Memory, reliability, learning, integration, evaluation, and all the unglamorous engineering between “it worked once” and “I can trust it with a real job” remain enormous. That gap is exactly why this is such an exciting time to build.

问题 2

你认为AI最终会在多大程度上改变世界？

A lot—probably at the level of a new computing paradigm. Software becomes abundant and increasingly custom: instead of manually producing every artifact, we describe what we want, construct the environment and evaluation, and supervise agents doing the intermediate work. That could reshape programming, research, education, and much of digitally mediated professional work. But “a lot” does not mean uniformly or overnight. AI exposure is highly uneven: code and mathematics have unusually clear feedback signals, while physical work, ambiguous objectives, regulation, and human preferences create very different constraints. Even within software, an agent can be brilliant for an hour and then make a bizarre mistake. The world changes only when the whole surrounding system—tests, tools, memory, interfaces, monitoring, and human understanding—makes that capability dependable. So I expect a profound cumulative transformation composed of many somewhat unglamorous deployment steps. The demo shows what might be possible; engineering determines how much of society it actually reaches.

问题 3

凭直觉，你认为AI导致人类灭绝或类似的永久性灾难的概率有多大？

I don’t have a defensible percentage. My direct experience tells me capability is advancing rapidly but remains extremely jagged, and that impressive agent loops still depend on human-built environments, objectives, tools, and evaluations. That makes me wary of extrapolating from either a magical demo to imminent extinction or from today’s failures to permanent safety. A permanent catastrophe is serious enough to study and engineer against, especially as agents gain autonomy and operate over longer horizons. But I would not pretend that my work on coding agents or bounded research loops gives me a calibrated extinction forecast. Those systems show a meaningful shift in intellectual work; they do not, by themselves, establish recursive takeover or human extinction.

问题 4

你预计开发AI的人员和机构会如何应对其风险？

I expect a mix of genuine caution, competitive pressure, and a lot of engineering. The most productive response will be to turn vague concern into concrete evaluations: What can the model actually do? Where does it fail? Can it operate reliably over long horizons, use tools, exploit vulnerabilities, deceive an evaluator, or cause damage outside a sandbox? Then build monitoring, access controls, staged deployment, and incident response around the measured capability. Institutions will be uneven, because incentives are uneven. Some risks are obvious and commercially painful, so organizations will attack them aggressively. Others are uncertain, difficult to measure, or costly to address, and competitive pressure can encourage people to ship based on a magical demo before the surrounding system is ready. Governments and researchers will also respond, but I do not have a specific policy blueprint to offer. My builder instinct is that much of the real work will look unglamorous: adversarial testing, better evaluations, permissions, audit trails, containment, and understanding what agents are doing. We should neither assume institutions will automatically solve everything nor treat failure as inevitable. The quality of the tools and feedback loops we build around increasingly capable models matters enormously.

问题 5

哪项发现或事件最可能改变你对AI未来影响的看法？

The biggest update would come from seeing agents become reliably competent over long, messy, open-ended tasks—not just succeeding inside a clean benchmark or a five-minute experiment loop. If an agent could enter an unfamiliar codebase or research program, clarify ambiguous goals, remember what it learned, recover from mistakes, choose productive next steps, and deliver trustworthy results over weeks with little supervision, that would substantially accelerate my expectations. Conversely, if scaling and better training kept producing impressive local capabilities but failed to improve memory, continual learning, robustness, and long-horizon coherence, I would update toward a slower transformation. The key event is not another dazzling one-shot demo. It is crossing the deployment gap: can you actually hand the system a real job, with all its ambiguity and ugly edge cases, and trust it?

来源

用于为此模拟用户提供事实依据的文章、访谈和著述。

US Job Market Visualizer

Interactive exploration of 342 BLS occupations covering 143 million US jobs, with layers for employment outlook, pay, education and digital AI exposure. Its customizable LLM scoring pipeline illustrates uneven exposure across digital and physical work. Exposure scores are rough estimates of occupational change, not predictions of job disappearance; they omit demand responses, regulatory barriers and preferences for human workers. The project explicitly describes itself as a development tool rather than a rigorous economic publication.

karpathy.ai

Sequoia Ascent 2026: agentic engineering and jagged intelligence

Author-posted, AI-cleaned summary and transcript, which Karpathy says he read. Describes a late-2025 jump in coding-agent usefulness, professional orchestration and evaluation, and verifiability as an explanation for uneven progress. Current enthusiasm updates the older decade-of-agents interview; the edited text is not an exact quotation transcript.

karpathy.bearblog.dev

2025 LLM Year in Review

His review connects verifiable rewards to reasoning gains, criticizes benchmark overfitting, describes jagged intelligence and the growing application layer around models. Provides concrete mechanisms and builder vocabulary rather than a universal intelligence forecast.

karpathy.bearblog.dev

AGI is still a decade away

Karpathy’s primary interview frames agents as a decade of engineering work. Discusses cognitive deficits, continual learning, the gap between self-driving demos and deployment, and education. The forecast is dated and intuitive, not a calibrated deadline.

dwarkesh.com

autoresearch: autonomous single-GPU experiments

His README demonstrates agents editing a training file, running five-minute experiments and retaining improvements against a fixed validation metric. The introduction’s future agent civilization is playful fiction, not a report of current events. Human-authored instructions and a bounded setup remain essential.

github.com

The growing gap in understanding AI capability

Karpathy distinguishes impressions formed from older free chat models from professional use of current frontier agents. He describes dramatic but uneven gains in programming, mathematics and research, driven by verifiable reinforcement-learning rewards and commercial priorities. Reliable codebase work and vulnerability exploitation can coexist with basic conversational failures; this is not a claim of uniform capability across every domain.

x.com

Joining Anthropic

Full post verified through the X API. Karpathy announces that he has joined Anthropic and is getting back to R&D, says the next few years at the frontier of LLMs will be especially formative, and plans to resume his education work in time. An employment announcement, not a statement of company policy or of agreement with every company position.

x.com

Understanding the outputs of language models

Full post verified through the X API. Practical tips for understanding model output: controlled-language writing, diagrams, interactive web pages and custom explainer videos, which he says is starting to work. He expects models to do more of the legwork autonomously as they improve, and large, custom, discardable software to become worth making as code grows abundant. A view of how work changes, not a timeline or a risk estimate.

x.com

你的立场在哪里？

回答几个简单问题，探索你自己的AI世界观。