2025 Andrej Karpathy YC School presentation

Andrej Karpathy
arc beats above · slides in the middle · loops below · scroll → 3 LOOPS
SETUP TENSION ANALYSIS EVIDENCE RESOLUTION APPENDIX
HOVER FOR DETAILS · CLICK A SLIDE FOR FULLSCREEN · STEP 10
No image bound to this slide. Wire up imgproxy to display the rendered slide JPEG.

Deck intelligence map

7
coverage by narrative range · generated from this deck JSON

Slide inventory

131
every slide · same image gating as the playbook
01
Slide 1
The slide includes a photo of the speaker on stage and a link to external notes.front_matter
Open slide detailBeat · Attention
02
front_matter
Open slide detailBeat · Attention
03
transition
Open slide detailBeat · Attention
04
The slide content is captured from a photograph of a presentation screen.transition
Open slide detailBeat · Attention
05
The slide uses a metaphorical 'map' to represent software development clusters on GitHub, likely based on repository co-occurrence or dependency data.other
Open slide detailBeat · Situation & Context
06
The slide uses a visual metaphor to contrast imperative programming with neural network weight matrices.establish_context
Open slide detailBeat · Situation & Context
07
The slide uses a metaphor of 'maps' to contrast the discrete, clustered nature of code repositories with the dense, interconnected nature of AI models.compare_options
Open slide detailBeat · Situation & Context
08
The slide uses a visual metaphor to contrast the structure of traditional software development with modern AI model development.compare_options
Open slide detailBeat · Situation & Context
09
The slide uses a 'Software 2.0' framework popularized by Andrej Karpathy.present_framework
Open slide detailBeat · Situation & Context
10
The slide captures a presentation by Andrej Karpathy at AI Startup School, illustrating the shift in programming paradigms.present_framework
Open slide detailBeat · Call to AdventureLoop · Golden Circle
11
The slide illustrates the evolution from explicit programming to data-driven training to prompt engineering.illustrate_case
Open slide detailBeat · Call to AdventureLoop · Golden Circle
12
The slide illustrates the evolution of programming paradigms from explicit rules to data-driven training to prompt engineering.compare_options
Open slide detailBeat · Call to AdventureLoop · Golden Circle
13
The slide uses a metaphorical 'map' visualization to represent the shift from Software 1.0 (GitHub code) to 2.0 (HuggingFace models) and 3.0 (LLM prompts).present_framework
Open slide detailBeat · Call to AdventureLoop · Golden Circle
14
This is a direct screenshot of a social media post used as a quote.cite_precedent
Open slide detailBeat · Call to AdventureLoop · Golden Circle
15
The slide is a screenshot of a tweet used as a visual aid during a presentation.illustrate_case
Open slide detailBeat · NeedLoop · Cost Of Inaction
16
The slide illustrates the transition from explicit rule-based programming to learned representations (BEV Net) for self-driving cars.present_framework
Open slide detailBeat · NeedLoop · Cost Of Inaction
17
The diagram uses a Voronoi-like or territory-expansion visual metaphor to show 3.0 displacing 2.0 and 1.0.frame_problem
Open slide detailBeat · NeedLoop · Cost Of Inaction
18
The slide captures a presentation by Andrej Karpathy. The diagram on the screen illustrates the shifting market share/importance of software paradigms 1.0, 2.0,summarize
Open slide detailBeat · NeedLoop · Cost Of Inaction
19
transition
Open slide detailBeat · NeedLoop · Cost Of Inaction
20
establish_context
Open slide detailBeat · Problem (Identify pain)
21
The slide content is presented as a list of analogies. The right side of the image shows a social media post containing a photo of the slide being presented by present_solution
Open slide detailBeat · Problem (Identify pain)
22
Uses a metaphor-analogy framework to explain the business model of LLMs.establish_context
Open slide detailBeat · Problem (Identify pain)
23
The slide uses a metaphor-analogy framework to explain the evolving nature of the AI industry.present_framework
Open slide detailBeat · Problem (Identify pain)
24
Uses a metaphor-analogy framework to explain the economics of AI infrastructure.present_framework
Open slide detailBeat · Problem (Identify pain)
25
The slide uses a metaphor/analogy to explain the economics of AI infrastructure.present_framework
Open slide detailBeat · Solution (Provide relief)
26
Uses a visual analogy comparing OS logos to LLM logos to support the thesis.establish_context
Open slide detailBeat · Solution (Provide relief)
27
The slide content is presented both as text on the left and as a projected slide on the right, which includes logos of various operating systems and a diagram osummarize
Open slide detailBeat · Solution (Provide relief)
28
This is a famous conceptual framework by Andrej Karpathy describing the LLM as the central processing unit of a new operating system.present_framework
Open slide detailBeat · Solution (Provide relief)
29
The slide features a diagram titled 'LLM OS' which maps traditional computer architecture components (CPU, RAM, Disk, I/O) to LLM-based equivalents.present_framework
Open slide detailBeat · Solution (Provide relief)
30
The slide uses a visual analogy to explain that LLM applications are becoming 'operating systems' for models, similar to how IDEs are OS-agnostic.compare_options
Open slide detailBeat · Solution (Provide relief)
31
The slide uses historical photos to illustrate the centralized nature of early computing.establish_context
Open slide detailBeat · Discoveries
32
The slide uses historical photos to illustrate the era of centralized computing.establish_context
Open slide detailBeat · Discoveries
33
The slide uses screenshots of social media posts (X/Twitter and LinkedIn) as evidence of a trend towards local AI compute.illustrate_case
Open slide detailBeat · Discoveries
34
The slide content is a summary of social media posts displayed on the screen behind the speaker.establish_context
Open slide detailBeat · Discoveries
35
Uses the 'Corporate needs you to find the differences' meme template from The Office.illustrate_case
Open slide detailBeat · Discoveries
36
Uses the 'three-headed dragon' meme format to contrast institutional vs consumer technology adoption.frame_problem
Open slide detailBeat · Discoveries
37
References Andrej Karpathy's 'Power to the people' blog post; uses a meme-style diagram to illustrate the reversal of technology diffusion.present_framework
Open slide detailBeat · Discoveries
38
summarize
Open slide detailBeat · Discoveries
39
summarize
Open slide detailBeat · Discoveries
40
transition
Open slide detailBeat · Discoveries
41
The slide uses the 'Attention Is All You Need' Transformer architecture diagram to ground the philosophical claim.establish_context
42
The slide includes a photo of a speaker presenting a slide that mirrors the text on the left. The content references a specific X (Twitter) post.summarize
43
The slide uses a cultural reference (Rain Man) to anchor the concept of encyclopedic memory.other
44
The image uses a visual metaphor for AI hallucination (the spectral cat) in a library setting.filler
45
The slide uses a photo of a student looking frustrated at a desk with a whiteboard showing '2+2=5' in the background, serving as a metaphor for AI inaccuracy.frame_problem
46
Uses a metaphor/analogy to explain AI technical limitations.establish_context
47
The slide uses a metaphor (anterograde amnesia) to explain technical limitations of LLMs.frame_problem
48
Used to illustrate the concept of memory loss in a relatable, pop-culture context.filler
49
The speaker is using these films as a metaphor or hook to introduce a topic related to memory or cognitive patterns.establish_context
50
The slide uses an AI-generated image of a person in a library with a book labeled 'TRUST ME' to represent the vulnerability of AI models to prompt injection.frame_problem
51
The slide uses a collage of AI-generated images to metaphorically represent the 'psychology' of LLMs.summarize
52
The slide uses a metaphorical approach to explain LLM behavior, contrasting high-level knowledge (savant) with specific failure modes (cognitive issues).summarize
53
transition
54
transition
Open slide detailLoop · Jobs To Be Done
55
illustrate_case
Open slide detailLoop · Jobs To Be Done
56
The slide uses a visual split to contrast traditional IDEs with LLM-integrated ones, followed by a list of technical implementation steps.illustrate_case
Open slide detailLoop · Jobs To Be Done
57
The slide is a transcription of a presentation slide shown on a screen in the background.illustrate_case
Open slide detailLoop · Jobs To Be Done
58
The slide uses a screenshot of a Perplexity search result to illustrate the underlying mechanics of the platform.illustrate_case
Open slide detailLoop · Jobs To Be Done
59
The slide content is also visible on the screen in the background photo.illustrate_case
Open slide detailLoop · Jobs To Be Done
60
The slide uses a comparison of two complex software UIs to frame the challenge of AI integration.frame_problem
Open slide detailLoop · Jobs To Be Done
61
The diagram uses a circular flow to represent the human-AI loop, with specific annotations for optimizing each phase.present_framework
62
The slide contains a photo of a presentation screen showing a circular process diagram.present_framework
63
The slide content is a transcript of a speaker's question regarding the future of software autonomy and LLM accessibility.frame_problem
64
The illustration serves as a visual metaphor for human-in-the-loop AI systems.front_matter
65
present_solution
66
The slide content is duplicated between the left-hand text and the projected screen image.present_solution
67
The slide uses a 'before-after' framing to show how to improve prompt engineering for AI coding tasks.illustrate_case
68
The slide uses a visual metaphor of a slider to represent increasing levels of vehicle autonomy.illustrate_case
69
This appears to be a placeholder or a transcript-based slide.other
70
illustrate_case
71
The slide uses a slider metaphor to represent increasing levels of vehicle autonomy.illustrate_case
72
The slide uses a personal narrative to frame the historical context of autonomous vehicle development.establish_context
73
Uses code-like syntax (any/all) to illustrate the difference in reliability standards between demos and products.summarize
74
Uses programming-style syntax (works.any vs works.all) to illustrate the reliability gap.frame_problem
75
The slide uses a provocative, minimalist style to introduce a core thesis.establish_context
76
Attributed to 'Andrej', likely Andrej Karpathy.establish_context
77
summarize
78
Uses a futuristic UI aesthetic to frame the comparison.present_framework
79
The slide uses a pop-culture metaphor to explain the spectrum between AI as a tool (augmentation) and AI as an independent actor (agent).illustrate_case
80
The slide uses a metaphor to explain AI product design principles.present_solution
81
Uses a checkmark/cross comparison format to contrast hype vs reality.compare_options
82
Uses a simple binary comparison list format to define scope.compare_options
83
The slide is minimalist, containing only a headline, an emoji, and a parenthetical question.summarize
84
The slide is a screenshot of a social media post.illustrate_case
85
The slide shows a speaker on stage with a projection of a Wikipedia page about 'Vibe coding'.summarize
86
The slide is a projection at a Y Combinator event. The content is a direct quote from a social media post.illustrate_case
87
This is a screenshot of a Wikipedia page.establish_context
88
The slide consists of a screenshot of a tweet and a photograph of children using computers, likely at the event mentioned in the tweet.illustrate_case
89
illustrate_case
90
The slide shows a 'before' (text menu) and 'after' (visual menu) transformation using the menugen.app tool.other
91
The slide features a photo of a presentation at Y Combinator's AI Startup School, showing a real-world example of an app (MenuGen) built using AI coding tools.illustrate_case
92
This is a UI screenshot of a SaaS product.present_solution
93
The slide uses emojis as bullet points to categorize tasks by difficulty/frustration level.summarize
94
The slide uses emojis to indicate the difficulty or frustration level of various tasks.summarize
95
The slide is structured as a technical documentation page with three columns of instructions.plan_implementation
96
The slide uses a real-world example from a presentation to critique poor documentation practices and highlight better alternatives.illustrate_case
97
The slide uses a screenshot of a tweet as a primary source of evidence for the 'problem' of integration complexity.frame_problem
98
transition
99
establish_context
100
The slide is presented by a speaker at a Y Combinator event.present_framework
101
The slide appears to be a markdown-rendered document or a technical summary page.summarize
102
The slide shows two distinct types of documentation: marketing/onboarding (left) and technical implementation (right).present_solution
103
The slide uses two screenshots to demonstrate the 'llms.txt' and '.md' pattern for AI-friendly documentation.illustrate_case
104
The slide features screenshots of a tweet and a documentation page to demonstrate the concept of LLM-friendly documentation.illustrate_case
105
This appears to be a screenshot from a video presentation about the Manim library.front_matter
106
The slide contains a screenshot of a tweet and a link to the original post.illustrate_case
107
The slide contrasts Vercel's documentation approach (cURL) with Stripe's MCP implementation.compare_options
108
The slide features two distinct examples: a Twitter post about cURL commands and a technical overview of the Stripe MCP server.present_solution
109
The slide uses a before-and-after visual comparison to demonstrate the utility of the Gitingest tool.illustrate_case
110
The slide features a photo of Andrej Karpathy presenting at a conference, with a screenshot of GitIngest as an example of a context builder.present_solution
111
Shows how raw code repositories are transformed into structured system architecture diagrams.establish_context
112
The slide features a photo of a presentation by Andrej Karpathy at AI Startup School, showing examples of context builders like Devin DeepWiki.summarize
113
The slide shows a screenshot of the AI agent interface performing a task on TripAdvisor.present_solution
114
The slide uses a custom diagram to represent the evolution of software paradigms (1.0, 2.0, 3.0) and a circular flow diagram for human-AI interaction.present_framework
115
The slide uses a custom diagram to represent the evolution of LLM capabilities and a circular flow diagram for human-AI interaction.present_framework
116
The slide uses a spectrum-like visual to contrast human-in-the-loop augmentation with autonomous agent capabilities.present_framework
117
summarize
118
The slide uses a social media thread as a testimonial/case study to emphasize the importance of speed and 'doing things' rather than waiting.illustrate_case
120
The slide presents a conceptual evolution of software development paradigms, citing social media posts as evidence.summarize
121
transition
122
transition
123
transition
124
The slide uses a personal anecdote to illustrate the 'hype cycle' or the gap between initial prototype success and full-scale deployment.summarize
125
The slide references external social media posts as sources for its conceptual framework.present_framework
126
transition
127
transition
128
transition
130
appendix
131
The slide contrasts unrealistic expectations (marked with X) with realistic, achievable milestones (marked with checkmarks).frame_problem