Air Street Capital · consulting-deck
2018 Air Street Capital The State of AI Report 2018
156 pages · 3 arc beats · 20 loops
2018 Air Street Capital The State of AI Report 2018
Air Street Capital arc beats above · slides in the middle · loops below · scroll → 20 LOOPS
1102030405060708090100110120130140150156
Deck intelligence map
3 coverage by narrative range · generated from this deck JSON
Narrative range 153 total
Metadata
Components
Metrics
Tools
Frameworks
Beats
Loops
The Facts (What) 66 slides 100% 66/66 slides 100% 66/66 slides · 381 hits — 0/66 slides
92.4% 61/66 slides · 107 hits 16.7% 11/66 slides · 12 hits 100% 66/66 slides · 130 hits 90.9% 60/66 slides The Implications (So What) 81 slides 100% 81/81 slides 100% 81/81 slides · 521 hits — 0/81 slides
95.1% 77/81 slides · 137 hits 4.9% 4/81 slides 100% 81/81 slides · 162 hits 96.3% 78/81 slides The Action (Now What) 6 slides 100% 6/6 slides 100% 6/6 slides · 25 hits — 0/6 slides
83.3% 5/6 slides · 9 hits — 0/6 slides
100% 6/6 slides · 12 hits 16.7% 1/6 slides Slide inventory
156 every slide · same image gating as the playbook
02
The slide serves as a front-matter/introductory page for a consulting-style report.summarize
04
Part of a larger report series, likely an introductory slide.establish_context
Open slide detailBeat · The Facts (What)
05
The slide serves as a foundational reference for technical terms.establish_context
Open slide detailBeat · The Facts (What)
07
The slide uses a process flow diagram to illustrate the concept of transfer learning.present_framework
08
The slide uses a technical diagram to illustrate the concept of transfer learning in AI.present_framework
09
The slide uses ROC curves to compare algorithm performance against human experts.illustrate_case
10
The slide uses two bar charts to demonstrate the performance advantage of GPUs in AI training.establish_context
11
The slide uses two charts to demonstrate the scaling efficiency of GPU clusters in AI model training.analyze_data
12
The slide uses three distinct charts to demonstrate scaling laws: a conceptual framework, a loss-reduction chart, and a parameter-growth chart.analyze_data
13
The slide illustrates the Information Bottleneck theory, showing the evolution of layers (L1-L5) through phases A-E.present_framework
14
The chart shows a log-scale progression of compute usage for various AI models from 2013 to 2018. The right side depicts a flywheel effect.analyze_data
15
The slide uses growth multipliers (+15X and +5X) to highlight the rapid adoption of GPU-based AI development.quantify_impact
16
The slide illustrates that while GPUs are used for ML, a significant portion of their core area is dedicated to other tasks, yet they consume substantial power.analyze_data
17
Data presented in two tables (Table II and Table III) comparing resource allocation for training vs inference.analyze_data
18
The chart is a classic visualization of the end of Dennard scaling.analyze_data
19
The slide highlights the architectural shift towards memory-centric design (75% RAM) and demonstrates significant performance gains in inference speed.present_solution
20
The chart uses multipliers (81x, 4x, 3.3x, 6.1x) to emphasize the price disparity between standard CPUs and specialized AI hardware.quantify_impact
21
The slide uses a line chart to show training convergence and a bar chart to show cost efficiency.quantify_impact
22
The slide features a photograph of a TPUv3 pod.establish_context
23
Slide from State of AI 2018 report.illustrate_case
24
The chart uses a combination of a bar chart for absolute revenue and a line chart for percentage share.quantify_impact
25
The chart uses a green line to represent stock price over time, with a specific annotation for the 'Deep learning works starts to work' inflection point.quantify_impact
26
The slide uses a combination of bulleted strategic highlights and a dual-axis bar/line chart to show financial performance.analyze_data
27
The slide uses a simple table format to segment the AI hardware market landscape.summarize
29
The slide uses a combination of qualitative news clippings and quantitative financial data to support the thesis.quantify_impact
30
The slide illustrates how AI models interpret visual data by identifying objects as 'nouns'.present_framework
31
The slide uses four examples of AI mislabeling images to demonstrate the lack of 'common sense world models' in AI.frame_problem
32
Slide from state.ai 2018 report.illustrate_case
33
The diagram shows a cyclical process involving taxonomy creation, crowd-acting, model training, and customer testing, with feedback loops.present_framework
34
The slide highlights the capability of TwentyBN-trained networks compared to standard image-trained networks.illustrate_case
35
DeepMind research slide from 2018 State of AI report.illustrate_case
36
The diagram shows the standard RL feedback loop: Agent receives observations and rewards, and performs actions.present_framework
37
The slide uses two charts: a line chart showing training progress over 40 days and a bar chart comparing peak Elo ratings of various Go engines.illustrate_case
38
The slide includes a screenshot of the game interface and a comparison table of training infrastructure and performance metrics.illustrate_case
39
The slide illustrates the 'World Models' architecture (Ha & Schmidhuber, 2018).illustrate_case
40
The chart uses a humorous, hand-drawn style to illustrate the sudden surge in research interest.establish_context
41
The slide uses screenshots of Google Translate to show how machine learning models can perpetuate societal stereotypes.illustrate_case
42
The slide uses external media headlines and a personal anecdote to illustrate the problem of algorithmic bias.illustrate_case
43
The slide uses a matrix to categorize different types of algorithmic bias observed in various tech products.diagnose
44
The comic is a reference to the 'AI-Box Experiment' thought experiment.illustrate_case
45
The slide uses a visual example of a dog playing a guitar to demonstrate how explainability techniques (like saliency maps or feature attribution) reveal what aillustrate_case
46
The slide showcases four examples of AI-generated explanations for image-based questions, using heatmaps to highlight visual evidence.illustrate_case
47
The chart is a lollipop chart showing feature importance.diagnose
48
Shows two examples: a panda misclassified as a gibbon, and a macaw misclassified as a bookcase.illustrate_case
49
The slide demonstrates adversarial machine learning, specifically targeted adversarial patches.illustrate_case
50
The slide uses a visual comparison to demonstrate a security vulnerability in AI.illustrate_case
51
The slide uses a chronological sequence of technical diagrams to show the progression of AI model complexity.illustrate_case
52
The slide shows a bar chart of model accuracy over time (2013-2016) and a bubble chart showing accuracy vs. operations (complexity).analyze_data
53
The slide shows two neural network cell architectures and a scatter plot comparing accuracy vs. computational cost.illustrate_case
54
The slide explains the technical mechanism of federated learning using OpenMined's framework.illustrate_case
55
The slide uses a process flow diagram to explain a technical concept.present_framework
57
The slide uses a choropleth map to represent talent density, supplemented by callouts for specific regional insights.size_opportunity
58
The slide uses a color-coded choropleth map to visualize talent density.size_opportunity
59
The map highlights the US, Canada, UK, and China in red, with flow lines connecting the US to the other three countries.illustrate_case
60
The chart uses house-shaped bars to represent headcount. Data points: Facebook (300), Tencent (400), Baidu (450), Google (1400), IBM (900), Microsoft (1000).compare_peers
61
The data is presented as a simple list of counts and institution names.analyze_data
62
The chart illustrates the increasing concentration of AI research talent at Google/DeepMind.quantify_impact
64
The slide highlights the concentration of AI research talent in a few top-tier academic institutions and major tech corporations.analyze_data
65
The chart uses a 100% stacked bar format to compare salary growth trends by country.analyze_data
66
The slide uses external media citations to validate the trend of rising AI talent costs.cite_precedent
67
The slide uses screenshots of news articles to provide evidence of industry trends regarding talent compensation and litigation.illustrate_case
69
The chart shows a slight upward trend in female attendance at the NIPS conference over three years.analyze_data
Open slide detailBeat · The Facts (What)
71
The chart on the left shows patent counts by company (Microsoft, Google, Amazon, Facebook, Apple). The chart on the right shows R&D spend by company, categorizeanalyze_data
72
The slide uses a comparative structure to show the parity in AI service offerings across the three major cloud providers.present_solution
73
The slide uses two distinct metrics (GitHub stars and research citations) to validate the thesis of TensorFlow's market leadership.analyze_data
74
Includes a process diagram illustrating the drug development lifecycle from preclinical to clinical studies.frame_problem
76
The slide uses a line chart for longitudinal spending data and a stacked area chart for age-related health morbidity.establish_context
77
The slide uses a case study approach to frame the problem statement for AI application in healthcare.illustrate_case
78
The slide uses a list-based structure to categorize AI applications in healthcare, supported by company logos and a biological diagram.illustrate_case
79
The slide shows a workflow for training and testing a deep learning model for tumor detection in pathology slides, alongside examples of medical imaging (OCT anillustrate_case
80
The chart uses a horizontal bar layout where each bar is composed of company logos.analyze_data
81
Includes screenshots of SenseTime's surveillance software interface.illustrate_case
82
The slide uses news snippets to illustrate the ethical tensions in AI development for defense.illustrate_case
83
The slide uses a visual collage of data-related icons and text labels to illustrate the breadth of personal data collection.establish_context
84
Includes a counter-style infographic for 2017 data breaches and a stacked bar chart showing consumer confidence in institutions.establish_context
85
Part of the State of AI 2018 report.summarize
86
Part of the state.ai 2018 report.establish_context
87
Part of the State of AI 2018 report.summarize
88
The slide showcases technical capabilities of satellite imagery providers (Planet, Orbital Insight, Descartes Labs) using RGB and NDVI analysis.illustrate_case
89
Part of the 'state.ai 2018' report series.diagnose
90
Part of a larger report series; slide focuses on practical applications of ML in a specific industry vertical.summarize
91
The slide uses a 'Why now?' framing to explain the rapid adoption of robotics in logistics.diagnose
93
The slide showcases GreyOrange, RightHand Robotics, and 6 River Systems.illustrate_case
94
The slide uses two charts to explain the economic rationale for increased automation in blue-collar sectors.analyze_data
95
Part of a larger report series; uses a list-based structure to categorize industry applications.illustrate_case
97
The slide uses a combination of a line chart showing the food production gap and a pie chart showing investment distribution.frame_problem
98
Slide from state.ai 2018 report.analyze_data
99
Part of the State of AI 2018 report.summarize
100
Includes company logos and photographic examples of agricultural robotics.illustrate_case
101
The slide uses a structured layout to present key statistics justifying the need for autonomous technology.frame_problem
102
Part of the 'state of AI 2018' report.illustrate_case
103
The slide uses a navigation bar at the top and a report identifier at the bottom right.illustrate_case
104
Part of the state.ai 2018 report.analyze_data
105
Part of a larger report series (state.ai 2018).present_solution
106
The slide uses two charts to illustrate the inefficiency of current work patterns.frame_problem
107
The slide uses a list-based structure to categorize automation technologies.present_framework
108
The slide uses a 'Why now?' and 'Where and how is ML being used effectively?' structure to frame the industry trend.illustrate_case
110
The slide provides context for subsequent data slides by defining the sample size and recruitment methods for Pew and Brookings surveys.establish_context
111
The chart is a diverging stacked bar chart showing likelihood of job replacement.analyze_data
112
Data source: Pew Research Center. The chart compares two metrics: 'Lost a job' and 'Had pay or hours reduced' across 13 demographic categories.analyze_data
113
Data source: Pew Research Center. The chart uses a diverging bar format to show binary sentiment (likely vs not likely).analyze_data
114
Data source: Pew Research Center. The chart shows a consistent trend where those impacted by automation hold stronger opinions or higher levels of concern/enthuanalyze_data
115
Source: Brookings Institute, State of AI 2018 report.analyze_data
116
Source: Brookings Institute, State of AI 2018 report.analyze_data
117
Source: Brookings Institute, State of AI 2018 report.analyze_data
118
Data source: Brookings Institute. Part of the State of AI 2018 report.analyze_data
119
Source: Brookings Institute, State of AI 2018 report.analyze_data
120
Data source: Brookings Institute. Slide from state.ai 2018 report.analyze_data
121
Data source: Brookings Institute. The chart shows a significant portion (35%) of respondents do not know or provided no answer.analyze_data
122
Data source: Brookings Institute. Part of the State of AI 2018 report.analyze_data
123
The slide uses a single data point (4.1%) to contrast with the narrative of automation-driven unemployment.analyze_data
124
The chart highlights the stagnation of routine cognitive and manual jobs compared to the growth of non-routine cognitive roles.analyze_data
125
Source: The New York Times, stateof.ai 2018 report.analyze_data
126
The chart shows recession periods as grey vertical bars and the average weeks of unemployment as a red line.analyze_data
127
Source: The Brookings Institute, State of AI 2018 report.analyze_data
128
The chart highlights recessionary periods with vertical grey bars.analyze_data
129
The chart uses a U-shaped distribution of income volatility against annual income levels.analyze_data
130
The slide uses a 'Don't worry' vs 'Worry' dichotomy to frame the debate on AI and labor.frame_problem
131
The chart uses a back-to-back bar configuration to compare job volume growth against relative wage levels.analyze_data
132
The chart distinguishes between historical data (solid blue) and forecast data (lighter blue/grey).analyze_data
133
Source: International Robotics Federation. Part of the State of AI 2018 report.compare_peers
134
The chart shows a massive divergence between the computer subsector and the rest of manufacturing.diagnose
135
The chart shows a significant decline in retail employees in 2016 compared to Amazon's growth.analyze_data
136
Part of the 'State of AI 2018' report, specifically the 'Politics' section.present_solution
138
The table compares the cost and value of providing specific services (UBS) versus a UBI equivalent.quantify_impact
139
The slide highlights the competitive nature of national AI strategies, specifically focusing on China's 2017 announcement.establish_context
140
The slide uses a photo of Emmanuel Macron to anchor the case study on French AI policy.illustrate_case
141
The slide uses a photo of Lee Sedol playing AlphaGo to contextualize the South Korean AI focus.establish_context
142
Part of a larger report series; focuses on EU policy.summarize
143
The slide uses a photo of a government official (Matt Hancock) to contextualize the policy announcement.illustrate_case
144
The table compares China and USA across multiple metrics, highlighting US dominance in most areas except for mobile users (data).compare_peers
146
The slide uses two photo-based case studies to illustrate political intervention in the semiconductor industry.illustrate_case
147
The chart uses a combination of bar charts (for absolute dollar values) and line charts (for percentage of total imports).analyze_data
148
The slide uses a red bar and a callout to emphasize the disparity between China and the U.S. in 2017.frame_problem
149
The slide uses a downward arrow annotation to highlight the impact of U.S. policy on M&A activity.analyze_data
150
The slide uses a sequence of images showing Theresa May to illustrate the progression of synthetic video manipulation.frame_problem
Open slide detailBeat · The Implications (So What)
152
The slide uses a numbered list format to present forward-looking statements.summarize