Air Street Capital · consulting-deck
2020 Air Street Capital The State of AI Report 2020
177 pages · 3 arc beats · 22 loops
2020 Air Street Capital The State of AI Report 2020
Air Street Capital arc beats above · slides in the middle · loops below · scroll → 22 LOOPS
1102030405060708090100110120130140150160170177
Deck intelligence map
3 coverage by narrative range · generated from this deck JSON
Narrative range 165 total
Metadata
Components
Metrics
Tools
Frameworks
Beats
Loops
The Facts (What) 120 slides 100% 120/120 slides 100% 120/120 slides · 742 hits — 0/120 slides
87.5% 105/120 slides · 207 hits 12.5% 15/120 slides · 20 hits 100% 120/120 slides · 240 hits 96.7% 116/120 slides The Implications (So What) 41 slides 100% 41/41 slides 100% 41/41 slides · 235 hits — 0/41 slides
75.6% 31/41 slides · 58 hits 4.9% 2/41 slides 100% 41/41 slides · 82 hits 97.6% 40/41 slides The Action (Now What) 4 slides 100% 4/4 slides 100% 4/4 slides · 14 hits — 0/4 slides
100% 4/4 slides · 6 hits — 0/4 slides
100% 4/4 slides · 8 hits — 0/4 slides
Slide inventory
177 every slide · same image gating as the playbook
03
The slide serves as a table of contents and mission statement for the report.front_matter
07
The slide uses a structured list format to categorize insights into four distinct pillars.summarize
09
The slide uses a traffic-light color coding system (Green/Yellow/Red) to grade the accuracy of past predictions.summarize
11
The chart shows a peak in late 2019 followed by a decline to 15% by June 2020.analyze_data
12
The slide acts as a showcase/screenshot of the Papers With Code website interface.establish_context
13
The slide uses a line chart to show the percentage of PyTorch mentions across various AI conferences over time.analyze_data
14
Data source: Papers With Code. The chart shows a clear shift from other frameworks to PyTorch between 2017 and 2020.analyze_data
15
The chart highlights the exponential jump in model size, particularly with GPT-3.analyze_data
16
The charts demonstrate power-law relationships between model performance (test loss) and three key scaling factors: compute, dataset size, and parameter count.analyze_data
17
The slide uses a mix of bulleted data points and narrative paragraphs to illustrate the cost barrier to entry for AI model training.quantify_impact
18
The chart displays three metrics across three model sizes (37.5B, 150B, 600B weights).analyze_data
19
The table highlights the diminishing returns of scaling ML models, specifically noting that reducing ImageNet error to 1% would cost over $100 quintillion.analyze_data
20
Charts illustrate scaling laws in machine learning, specifically the relationship between model size (parameters), data (tokens), and compute budget.analyze_data
21
The chart shows a positive correlation between model size and translation performance across varying levels of training data availability.illustrate_case
22
The slide uses two charts: a log-scale scatter plot for compute eras and a linear scatter plot for training efficiency.analyze_data
23
The slide uses a combination of a data table, a line chart showing F1 score vs data points, and a bar chart showing intent accuracy across low/high data regimesillustrate_case
24
Includes screenshots of Hugging Face model interface and a Twitter post by Sharif Shameem demonstrating GPT-3 code generation.illustrate_case
25
The slide highlights the capability of TransCoder to translate code without expert knowledge, achieving high success rates in unit tests.illustrate_case
26
Includes a visual representation of the DrRepair process flow and a code snippet example.illustrate_case
27
The slide uses two different chart types (bar-vertical and bar-horizontal) to illustrate the same trend of AI models surpassing human performance.analyze_data
28
The slide uses a grouped bar chart to compare performance across three categories (Commonsense, Linguistics, Knowledge) for four model sizes (Small, Medium, Laranalyze_data
29
The slide uses a visual example of image completion to demonstrate the transformer's ability to process non-textual data.illustrate_case
30
The 2020 data point includes an annualized projection indicated by a dashed box.quantify_impact
31
The slide uses a visual comparison to show how techniques from computer vision are being applied to biological research.illustrate_case
32
Includes a chord diagram of biological interactions and two line/contour plots showing drug perturbation effects.illustrate_case
33
The slide uses a combination of medical imaging segmentation and a predictive line chart to demonstrate early intervention capabilities.illustrate_case
34
Includes ROC curves comparing AI performance against human readers (radiologists).illustrate_case
35
Features a photo of Judea Pearl, a pioneer in causal inference.present_framework
36
The charts compare algorithm accuracy vs doctor accuracy for associative (top) and counterfactual (bottom) models.diagnose
37
The slide contrasts standard Shapley values (which assume feature independence) with ASV, demonstrating the difference on an income classifier model.present_solution
38
The slide uses a combination of box plots for distribution visualization and a summary table for performance metrics.illustrate_case
39
The slide uses a chemical reaction diagram to illustrate the synthesis process and a bar chart to show the progression of model accuracy.illustrate_case
40
The slide contrasts traditional molecular vectorization with GNN-based graph representation.present_framework
41
The slide uses a funnel diagram to contrast the scale of AI-driven screening vs conventional screening.illustrate_case
42
The slide features a technical diagram of the PNA architecture and a box plot comparing its performance against MPNN, GAT, GCN, and GIN models.illustrate_case
43
The slide compares GNN-based hit rates (72%, 33%, 16%) against a ~1% industry standard, supported by a process diagram and stacked bar charts.illustrate_case
44
The chart shows performance (mean squared error) of different models (UniRep fusion, baseline, Doc2Vec, RGN) across various protein tasks.illustrate_case
45
Includes logos for ZOE, King's College London, and MGH.illustrate_case
46
The slide highlights the 'COVID Moonshot' initiative, showcasing the speed of AI-driven molecule design compared to traditional methods.illustrate_case
47
The slide explains a technical process involving pose estimation and image-to-image translation to create realistic sports video.illustrate_case
48
The slide highlights the shift from traditional CNN-based architectures like Faster R-CNN to transformer-based approaches in computer vision.present_solution
49
The slide contrasts traditional visible-only depth/ground prediction with the new 'Footprints' method that infers hidden geometry.present_solution
50
The slide illustrates a technical process flow for AI training data generation.present_framework
51
The slide describes a technical process involving CNNs and geometric cues for image processing.illustrate_case
52
The slide showcases performance metrics across 8 different tasks from the DeepMind Control Suite, comparing Dreamer against other agents.illustrate_case
53
The slide features a visual demonstration of future state prediction samples (wait, nudge, turn right, go straight) using semantic segmentation and optical flowillustrate_case
54
The slide explains the LoRRA architecture and provides two examples of its application.illustrate_case
55
The slide showcases the capabilities of SinGAN across five tasks: Paint to image, Editing, Harmonization, Super-resolution, and Animation.illustrate_case
56
Charts show performance trade-offs between accuracy (COCO AP) and efficiency (parameters, latency).analyze_data
57
The chart shows the progression of algorithm complexity and accuracy, starting from an empty program to a complex evolved algorithm.analyze_data
58
The chart uses a dual-axis representation (bar for total count, line for percentage share).quantify_impact
59
The slide displays screenshots of training loss/accuracy graphs for different platforms.illustrate_case
60
The slide uses a diagram to illustrate the technical workflow of federated learning and encrypted inference.present_framework
61
The slide compares CNNs and TICK-GP models on digit recognition and compares SVGP vs VISH on flight delay prediction.summarize
62
The slide evaluates a past prediction against actual outcomes.illustrate_case
64
The slide highlights the trend of academic talent moving to industry, specifically in the AI sector.analyze_data
65
The slide contrasts a positive industry-academic partnership (DeepMind/Cambridge) with a critical perspective on dual affiliation (Berkeley faculty op-ed).frame_problem
66
The slide uses a timeline diagram and a regression table to illustrate the causal link between professor departures and student entrepreneurship.analyze_data
67
The slide highlights a specific institutional investment in AI talent.illustrate_case
68
Slide from State of AI 2020 report.illustrate_case
69
The slide validates a previous prediction by showcasing the successful launch of MBUZAI.illustrate_case
70
The slide tracks the origin of AI research talent, specifically those with undergraduate degrees from China.quantify_impact
71
The chart visualizes talent migration patterns for AI researchers.analyze_data
73
The slide highlights the role of H1B visa sponsorship in career outcomes for AI PhD graduates.quantify_impact
74
Data source: CSET. Part of the state of AI 2020 report.analyze_data
75
The chart uses a waterfall-style visualization to show cumulative distribution of talent origins.analyze_data
76
Includes references to specific presidential proclamations and data on foreign student retention.analyze_data
77
The chart uses fractional counts for paper authorship.compare_peers
78
The chart displays the Publication Index position gains for top 20 organizations at ICML 2020 vs 2019.analyze_data
79
The chart illustrates the surge in interest in AI/NLP education at Stanford over two decades.quantify_impact
80
Data source: Indeed.com US data. Timeframe: 2015-2018.quantify_impact
81
The chart uses Exponential Moving Averages (EMA) for the data series.analyze_data
83
The slide highlights the efficiency gains of AI in drug discovery, noting a 20% reduction in candidate testing compared to traditional methods.illustrate_case
84
The slide uses a simple timeline structure to demonstrate the validation of AI in pharma.illustrate_case
85
The slide illustrates the 'platform strategy' model where parent AI companies spin off specific drug assets into independent entities.illustrate_case
86
The slide features a logo grid of the project partners.illustrate_case
87
The chart shows a correlation between predicted and true binding affinity scores with a Spearman r of 0.88.illustrate_case
88
The slide uses a process flow diagram to illustrate the transition from manual input to automated analysis.illustrate_case
89
The slide uses a hub-and-spoke diagram to show data inputs, line charts for time-series metabolic data, and a scatter plot for model validation.illustrate_case
90
The slide illustrates the FDA's proposed TPLC regulatory framework for AI/ML-based software as a medical device (SaMD).present_framework
91
The slide highlights a specific industry response to a identified problem (low quality in 20,000 studies).summarize
92
The slide discusses the impact of CMS reimbursement on AI adoption in healthcare, specifically for Viz.ai.summarize
93
Source: NCSL, Morning Brew. Part of the 'state of ai 2020' report.establish_context
94
The slide uses a timeline-like structure to present three specific regulatory milestones for Waymo, Nuro, and AutoX.summarize
95
The chart uses a bar for human mileage and text-based values for AV mileage to emphasize the massive disparity.quantify_impact
96
The slide uses a line chart to illustrate the lack of reliability in 'miles per disengagement' as a metric for AV progress, specifically calling out Baidu's datanalyze_data
97
Includes a quote/excerpt from the official Amazon acquisition announcement.illustrate_case
98
The slide uses a bubble-based visualization to represent funding scale, with lines connecting company logos to their respective bubbles.quantify_impact
99
Includes images of a test vehicle and a control center dashboard.illustrate_case
100
Part of the State of AI 2020 report.illustrate_case
101
Includes a product portfolio diagram for Velodyne and a sample LiDAR point cloud visualization for Luminar.compare_options
102
The chart illustrates the gap between initial expectations of exponential growth and the reality of diminishing returns in AI capability over time.diagnose
103
The slide uses a dual-column timeline format to display chronological releases of datasets.summarize
104
The slide uses a process flow diagram combined with gauge charts to show the distribution of ML usage across the autonomous driving stack.present_framework
105
Includes logos of Lyft, Waymo, and Uber at the bottom.summarize
106
Part of the 'state of ai 2020' report series.summarize
107
The slide highlights the rapid evolution of specialized AI hardware, specifically focusing on Graphcore's technological advancement.quantify_impact
108
The slide uses a visual comparison of hardware units to illustrate cost efficiency.compare_options
109
The chart shows performance gains for Mask R-CNN, SSD, ResNet-50, Transformer, and GNMT models.quantify_impact
110
The slide highlights the performance leap of the Ampere architecture.quantify_impact
111
The slide uses a combination of a bar-line chart for GitHub project growth and a line chart for search trends.summarize
112
The slide lists 12 specific regulatory requirements for AI applications.summarize
113
The chart uses a diverging stacked bar format to compare two distinct metrics (cost decrease vs revenue increase) across the same set of business functions.analyze_data
114
The slide uses a bubble-heatmap style table to visualize AI adoption across different industry sectors.analyze_data
115
The chart shows ARR growth from $1M in 2015 to $360M in 2019.quantify_impact
116
The slide highlights the practical application of deep learning in voice assistants, contrasting performance with legacy systems.illustrate_case
117
The slide highlights the value proposition of Tractable's AI in the insurance industry, specifically for auto claims.illustrate_case
118
The slide highlights a specific product capability (Tinyclues) within the context of industry trends in AI/ML automation.present_solution
119
The slide uses a radar chart to compare four companies across five ESG pillars, supported by a corresponding data table with conditional formatting.quantify_impact
120
The chart shows a sharp increase in COVID-19 related phishing emails compared to general phishing emails during early 2020.illustrate_case
121
The slide uses a visual callout technique to map specific document errors to their locations on an ID card.illustrate_case
122
The slide highlights the efficiency gains of AI over manual compliance processes, specifically in adverse media coverage.illustrate_case
123
The slide uses a bar chart to show time reduction and a precision-recall curve to show performance maintenance.illustrate_case
124
The slide highlights the rapid adoption of open-source AI research into large-scale production environments.illustrate_case
125
Part of the 'Industry' section of the State of AI 2020 report.illustrate_case
126
The slide uses a process flow diagram to illustrate the transition from a 3D file to a finished part via automated CNC programming.illustrate_case
127
The chart shows a clear upward trend in the usage of various transformer models over time.summarize
128
The slide highlights adoption across six major industries with specific metrics on downloads and contributors.illustrate_case
129
The chart uses a combination of bar and line graphs to represent capital raised and deal count respectively, with 2020 data being annualized.analyze_data
Open slide detailBeat · The Facts (What)
131
The slide highlights academic and investigative work that brought AI ethics into the mainstream.establish_context
132
The slide uses a color-coded map to visualize the prevalence of facial recognition technology globally.establish_context
133
The slide uses the 'tip of the iceberg' metaphor to imply broader, unrecorded instances of algorithmic bias.illustrate_case
134
Part of the 'State of AI 2020' report, specifically the 'Politics' section.illustrate_case
135
The slide uses a bar chart to contrast the scale of Clearview's database against traditional law enforcement databases.illustrate_case
136
The slide highlights corporate policy shifts by Microsoft, Amazon, IBM, and Apple regarding facial recognition.summarize
137
The slide uses a grid of photos to illustrate the 'Original' vs 'Balanced' datasets.summarize
138
The slide uses a specific case (Ed Bridges vs South Wales Police) to illustrate the shift toward proactive regulation of AI technologies.illustrate_case
139
The slide highlights the intersection of policy and corporate influence in AI regulation.illustrate_case
140
The slide uses a specific case (Professor Guo Bing) to illustrate a broader trend of privacy regulation in China.illustrate_case
142
Slide from State of AI 2020 report.establish_context
143
The slide uses screenshots of the OpenAI Playground interface to provide empirical evidence of model bias.illustrate_case
144
The chart is a t-SNE or similar dimensionality reduction visualization showing clusters of game commands.illustrate_case
145
Includes an annotated image of a military vehicle with various electronic warfare modules.establish_context
146
Part of the 'State of AI 2020' report.illustrate_case
147
Slide from State of AI 2020 report.establish_context
148
Highlights the dual-use nature of AI techniques migrating from gaming to military applications.illustrate_case
149
Slide from State of AI 2020 report.summarize
150
Includes a logo of the CIA, likely as an example of an actor with specific operational AI principles.establish_context
151
The slide highlights a gap between AI research practices and more stringent standards in life sciences.summarize
152
The slide highlights the 'People + AI Guidebook' framework, which breaks down responsible AI into six key areas.present_framework
153
Part of a larger report series (state.of.ai 2020).establish_context
154
Includes a line chart showing quarterly smartphone shipments for Samsung vs Huawei from 2015 to 2020.summarize
155
Part of the 'State of AI 2020' report series.establish_context
156
Includes a line chart of financial metrics and a timeline of semiconductor process nodes.compare_peers
157
Includes a line chart comparing cumulative total returns of SMIC and TSM.establish_context
158
Slide from the State of AI 2020 report.summarize
159
The chart shows the number of 300mm semiconductor fabs by region from 2011 to 2019.establish_context
160
The slide uses a list format to present a chronological sequence of events, categorized by outcome (blocked, allowed, pending).summarize
161
The slide highlights the trend of 'AI Nationalism' through specific regulatory changes in three major economies.establish_context
162
The slide uses a standard consulting layout with a top navigation bar highlighting the 'Politics' section.establish_context
163
The slide uses a bar chart to illustrate the growth in federal non-defense AI R&D spending.quantify_impact
164
Includes a photo of Chuck Schumer at a podium.establish_context
165
The slide highlights a specific government initiative in China to decentralize AI development.establish_context
166
Includes a cultural reference (Wolf Warrior movie poster) to illustrate the geopolitical context.establish_context
167
The map uses color-coding for strategy status (complete vs forthcoming) and dot colors for the focus of public transformation.establish_context
168
The chart shows a widening gap between labor tax rates and equipment/software tax rates over time.diagnose
169
The slide uses a series of stacked bar charts to visualize survey data from 1,872 enterprises regarding AI-induced workforce contraction.analyze_data
170
The slide highlights a specific academic/industry collaboration (Tackling Climate Change with Machine Learning).illustrate_case
172
The slide uses a numbered list format to present forward-looking statements.summarize
Open slide detailBeat · The Action (Now What)