OpenAI | Product Presentation Deck | 20 slides

OpenAI · 2017-08

arc beats above · slides in the middle · loops below · scroll → 2 LOOPS

Deck intelligence map

coverage by narrative range · generated from this deck JSON

Narrative range 9 total

Metadata

Components

Metrics

Tools

Frameworks

Beats

Loops

Problem 3 slides 100% 3/3 slides 100% 3/3 slides · 9 hits

— 0/3 slides

66.7% 2/3 slides

— 0/3 slides

100% 3/3 slides 100% 3/3 slides Agitate 2 slides 100% 2/2 slides 100% 2/2 slides · 5 hits

— 0/2 slides

50% 1/2 slides

— 0/2 slides

100% 2/2 slides 100% 2/2 slides Solution 4 slides 100% 4/4 slides 100% 4/4 slides · 11 hits

— 0/4 slides

25% 1/4 slides 75% 3/4 slides 100% 4/4 slides 100% 4/4 slides

Slide inventory

every slide · same image gating as the playbook

front_matter

Open slide detail

frame_problem

Open slide detailBeat · ProblemLoop · Cost Of Inaction

The slide displays the title, author list, and abstract of a well-known academic paper.front_matter

Open slide detailBeat · ProblemLoop · Cost Of Inaction

The slide uses a simple list to categorize failure modes, with a specific exclusion for non-ML problems.diagnose_problem

Open slide detailBeat · ProblemLoop · Cost Of Inaction

The slide shows a 'before' (left, dark/night mode) and 'after' (right, bright/day mode) of a racing game environment, likely used to test AI agent performance uillustrate_case

Open slide detailBeat · AgitateLoop · Cost Of Inaction

The slide uses a screenshot of a CNET article to provide concrete evidence of a failure in AI model performance.diagnose_problem

Open slide detailBeat · AgitateLoop · Cost Of Inaction

The slide explicitly references Goodhart's Law as a theoretical underpinning for the problem.frame_problem

Open slide detail

The image is a screenshot of a boat racing game.propose_solution

Open slide detailBeat · SolutionLoop · Jobs To Be Done

The slide features a visual representation of a simulated agent (likely a MuJoCo environment) used in the referenced research.cite_precedent

Open slide detailBeat · SolutionLoop · Jobs To Be Done

The slide shows a binary choice interface used for training AI models, specifically for reinforcement learning.present_framework

Open slide detailBeat · SolutionLoop · Jobs To Be Done

The diagram shows a standard RL loop (RL Algorithm <-> Environment) augmented with a Reward Predictor that receives human feedback.present_framework

Open slide detailBeat · SolutionLoop · Jobs To Be Done

The slide shows two screenshots of the same game (likely Enduro) with different agent strategies.illustrate_case

Open slide detailLoop · Jobs To Be Done

The images show Breakout, Pong, and Seaquest.other

Open slide detailLoop · Jobs To Be Done

The charts show reward over time (timesteps) for different training methodologies.analyze_data

Open slide detailLoop · Jobs To Be Done

The charts show learning curves for different training methods. The legend is consistent across all sub-plots.analyze_data

Open slide detail

The charts show learning curves for reinforcement learning agents. The legend is consistent across all plots.analyze_data

Open slide detail

The slide uses a reinforcement learning-style feedback loop diagram (Agent-Environment interaction) to explain how human intent is translated into technical actpresent_framework

Open slide detail

The slide uses a screenshot of a 2017 Economist article as a cultural/media reference to establish the context of the 'AI policy problem'.cite_precedent