Air Street Capital · consulting-deck
2021 Air Street Capital The State of AI Report 2021
188 pages · 3 arc beats · 22 loops
2021 Air Street Capital The State of AI Report 2021
Air Street Capital arc beats above · slides in the middle · loops below · scroll → 22 LOOPS
1102030405060708090100110120130140150160170180188
Deck intelligence map
3 coverage by narrative range · generated from this deck JSON
Narrative range 184 total
Metadata
Components
Metrics
Tools
Frameworks
Beats
Loops
The Facts (What) 148 slides 100% 148/148 slides 100% 148/148 slides · 917 hits — 0/148 slides
50.7% 75/148 slides · 152 hits 8.8% 13/148 slides · 14 hits 100% 148/148 slides · 296 hits 77% 114/148 slides The Implications (So What) 29 slides 100% 29/29 slides 100% 29/29 slides · 171 hits — 0/29 slides
69% 20/29 slides · 37 hits 3.4% 1/29 slides 100% 29/29 slides · 58 hits 96.6% 28/29 slides The Action (Now What) 7 slides 100% 7/7 slides 100% 7/7 slides · 21 hits — 0/7 slides
57.1% 4/7 slides · 7 hits — 0/7 slides
100% 7/7 slides · 14 hits 57.1% 4/7 slides Slide inventory
188 every slide · same image gating as the playbook
03
The slide serves as an introduction and table of contents for the report structure.establish_context
04
Includes a logo grid of contributing organizations and a list of individual reviewers.other
05
This slide provides foundational definitions for the rest of the report.establish_context
Open slide detailBeat · The Facts (What)
07
The slide uses a structured list format to categorize key findings across four pillars.summarize
Open slide detailBeat · The Facts (What)
09
The slide acts as a retrospective review of predictions made in the previous year's report.summarize
11
Includes technical diagrams of the ViT architecture and a performance chart showing ImageNet accuracy over fine-tuning steps.illustrate_case
12
Includes a line chart comparing model performance, a table of COCO detection results, and conceptual diagrams for RegNet and SwAV.summarize
13
The slide showcases visual evidence of unsupervised object segmentation via attention maps.illustrate_case
14
The slide uses two line charts to show the reduction in Word Error Rate (WER) over time for speech recognition and visual examples of 3D point cloud tasks.summarize
15
Includes a technical diagram of the Perceiver architecture and a performance comparison table against BERT on the GLUE benchmark.present_solution
16
The chart compares 'Frozen Pretrained Transformer' vs 'Full Transformer' vs 'Full LSTM' across 7 benchmarks.analyze_data
17
The slide highlights a shift in research focus back to non-transformer architectures like MLP-Mixer and convolutional models.summarize
18
The slide explains the technical mechanism of NeRF and mentions GIRAFFE as a generative variant.summarize
19
The chart shows CASP14 performance metrics comparing AlphaFold (G427) against other participants.illustrate_case
20
Includes a technical diagram of the RoseTTAFold architecture and two bar charts comparing performance benchmarks.illustrate_case
21
The slide uses a combination of a process diagram and two charts (bar and line) to validate the efficacy of AI-generated proteins.illustrate_case
22
The slide uses a metaphor of 'grammaticality' (viability) vs 'semantic change' (antigenic variation) to explain viral evolution.illustrate_case
23
The slide highlights a specific AI application in structural biology.illustrate_case
24
The chart shows the exponential growth of cryo-EM structures in the Protein Data Bank (PDBe) over time, segmented by resolution.summarize
25
The slide describes a specific machine learning architecture (likely 'CPA' or Compositional Perturbation Autoencoder) for drug discovery.illustrate_case
26
The slide contrasts a traditional 'Brute Force' approach with a 'MolPAL' (Molecular Pool-based Active Learning) approach.illustrate_case
27
The slide compares various machine learning approaches for predicting Buchwald-Hartwig reaction yields using bar charts and line graphs.illustrate_case
28
The slide uses a timeline-like visual to show the progression of DeepMind's algorithms (AlphaGo, AlphaGo Zero, AlphaZero, MuZero) and their capabilities.summarize
29
The slide compares DreamerV2 against IQN, Rainbow, C51, and DQN across four performance metrics (Gamer Median, Gamer Mean, Record Mean, Clipped Record Mean) oveillustrate_case
30
The slide highlights the XLand environment and Population Based Training as key enablers for zero-shot generalization in RL.illustrate_case
31
The slide uses three line charts with trend lines to show the impact of AI on human performance.illustrate_case
32
The slide includes two charts: a scatter plot showing the decline in the number of runs over time and a density plot showing the distribution of reported mediandiagnose
33
Includes a comparison diagram of existing methods vs. ClipBERT and a performance table.illustrate_case
Open slide detailBeat · The Facts (What)
34
The slide highlights the trade-off between model capacity and language resource density in ASR.analyze_data
Open slide detailBeat · The Facts (What)
35
Includes a process diagram of the GSLM architecture.present_framework
Open slide detailBeat · The Facts (What)
36
The slide explains the technical shift from GANs to diffusion models in AI research.summarize
37
Includes a technical diagram of the contrastive learning architecture and the loss function formula.illustrate_case
38
The slide explains the CLIP architecture using a three-step process diagram.illustrate_case
39
Includes visual examples of DALL-E outputs for specific text prompts.illustrate_case
40
The slide details the VLiD architecture, showing training and inference workflows.illustrate_case
41
The slide uses a visual example of a prompt-to-code workflow.illustrate_case
42
Includes a visual example of a coding problem (H-Index) and a bar chart comparing model performance across difficulty levels.diagnose
43
The slide contrasts formal theorem proving datasets with the new MATH dataset, highlighting the difficulty of the latter even for humans.illustrate_case
44
Includes a qualitative example of model responses to tricky questions as model size increases.diagnose
45
The slide references a CMU survey of 60+ papers. It includes a table comparing four paradigms: Fully Supervised (Non-Neural), Fully Supervised (Neural), Pre-trapresent_framework
46
Includes a tweet as a case study example.illustrate_case
47
The slide uses data from LAMA and SuperGlue benchmarks to illustrate performance variability in prompting.diagnose
48
The slide includes two technical diagrams illustrating the architecture of the ERNIE 3.0 model.summarize
49
The slide showcases the M6 model's capabilities in multimodal tasks (image-to-text and text-to-image) using examples from the M6-Corpus.illustrate_case
51
The slide highlights the unreliability of human benchmarks in NLP, specifically referencing GPT-2 and GPT-3 evaluation.diagnose
52
The slide highlights the lack of geographic diversity in training data for medical AI, leading to potential clinical bias.analyze_data
53
The slide uses two charts to illustrate demographic imbalances: a population pyramid by ethnicity and age, and a stacked bar chart showing sex distribution by ediagnose
54
The slide uses a scatter plot to compare sensitivity vs false positive rate across various studies.diagnose
55
The slide highlights a specific research finding regarding algorithmic bias in medical imaging.illustrate_case
56
The slide uses data from Papers With Code to track trends in AI research reproducibility and framework adoption.analyze_data
57
Includes a diagram of the data cascade process and a summary table of triggers and impacts.frame_problem
58
The slide uses a diagram to propose a documentation framework for CommonCrawl-based datasets.diagnose
59
Includes specific examples of legal NLP (legalBERT/CaseHOLD) and malware detection (SoReL-20M).summarize
60
The slide highlights the trade-off between dataset size and training efficiency, showcasing specific techniques like SEALS and SVP.analyze_data
61
The chart shows the rapid growth in submissions vs. accepted papers, illustrating the pressure on the review system.frame_problem
62
The slide compares code provision rates (initial vs camera-ready) and acceptance rates across different author affiliations.analyze_data
63
Data from a randomized prospective social media trial in the Annals of Thoracic Surgery.analyze_data
64
The slide uses two charts: a bar chart for absolute keyword frequency and a diverging bar chart for growth/decline in usage.analyze_data
65
The slide explains the application of GNNs to physical dynamics, specifically mesh-based simulation, highlighting efficiency gains and generalization capabilitiillustrate_case
66
The slide uses a world map to visualize the percentage reduction in negative ETA outcomes across various global cities.illustrate_case
67
The chart shows a bubble plot comparing ROC-AUC score vs GPU Memory (GB) for various GNN models.diagnose
68
Technical slide explaining a specific AI research methodology (L3P).present_framework
69
The slide highlights the success of Chinese entities in AI research benchmarks.illustrate_case
70
The slide highlights a specific application of AI in meteorology, comparing DGM performance against PySTEPS and Axial Attention models.illustrate_case
Open slide detailBeat · The Facts (What)
71
The slide uses a process flow diagram to show the transition from raw video data to actionable digital biomarkers.illustrate_case
Open slide detailBeat · The Facts (What)
72
The slide uses a causal diagram (DAG) to illustrate confounding variables and two horizontal bar charts to show vaccine effectiveness over time for AstraZeneca illustrate_case
Open slide detailBeat · The Facts (What)
73
The chart compares Accelerator Years, Energy Consumption (MWh), and Net CO2e (metric tons) for models Meena, T5, GPT-3, Gshard-600B, and Switch Transformer.analyze_data
Open slide detailBeat · The Facts (What)
74
Includes a code snippet demonstrating JAX syntax and a logo/library comparison graphic.summarize
Open slide detailBeat · The Facts (What)
76
Data sources: Adzuna, OECD.AIanalyze_data
77
The chart uses a logarithmic scale on the y-axis to show growth across orders of magnitude.analyze_data
78
The chart uses a dual-line approach to distinguish between actual historical data and future projections.analyze_data
79
The slide uses a table to demonstrate that the increase in PhD graduates in China is not due to a proliferation of low-quality programs, but rather growth in elanalyze_data
80
The slide combines a bar chart showing international study destinations and a table ranking top employers.analyze_data
81
The chart shows a clear divergence where elite institutions and tech firms are increasing their share of research output compared to lower-tier universities.analyze_data
82
Contains three distinct charts: a net flow area chart, a line chart of transition shares by university rank, and an interaction plot of citation rank over time.analyze_data
83
The slide uses a horizontal bar chart to show top universities losing faculty and a stacked bar chart to show the trend of departures over time.analyze_data
84
The chart tracks annual growth rates (%) from 2009 to 2018 for four categories: total students, applied sciences students, applied sciences professors, and totaanalyze_data
85
The chart uses a dual-axis approach: a line chart for funding and a bar chart for enrollment.analyze_data
86
The slide uses screenshots of Twitter threads to provide anecdotal evidence of the problem.illustrate_case
87
The slide highlights the concentration of corporate funding in academic AI research.quantify_impact
88
The slide highlights the trend of private-public partnerships in scientific research infrastructure.illustrate_case
89
Includes three stacked bar charts showing racial/ethnic breakdown, gender breakdown, and Hispanic/non-Hispanic breakdown.analyze_data
Open slide detailBeat · The Facts (What)
90
The slide highlights the rapid growth of AI-related roles and degrees compared to the general population and average degree growth.quantify_impact
Open slide detailBeat · The Facts (What)
93
The slide highlights Exscientia's performance metrics (compounds synthesized and time-to-hit) against industry benchmarks.illustrate_case
94
The slide uses a combination of process flow diagrams and Kaplan-Meier survival curves to validate the efficacy of the AI platform.illustrate_case
95
Includes a process flow diagram and comparative bar charts showing efficiency gains.illustrate_case
96
The slide illustrates a technical process for drug discovery using GNNs and active learning.illustrate_case
97
The slide compares a 'Simulated tradition approach' vs 'ML approach' using box plots for two metrics: Potency (nM) and Protease stability (% intact).illustrate_case
98
The slide uses three visual examples of computer vision in action (heatmap, PPE detection, driving safety).illustrate_case
99
Includes a chart of billion-dollar disaster events, a satellite image of a typhoon, and a UI screenshot of the Tractable app.illustrate_case
100
The slide highlights a specific AI application in the energy sector.illustrate_case
101
The slide features a line chart comparing antibiotic usage and two heatmaps comparing prediction error rates between industry standards and AI models.illustrate_case
102
The slide presents two distinct findings: a heatmap of bacterial correlations with health markers and ROC curves for coffee consumption prediction.analyze_data
103
The slide highlights a specific medical AI application in ophthalmology, detailing the problem (AMD), the solution (computer vision system), and the validation illustrate_case
104
The slide highlights the use of multi-armed bandits for balancing exploration and exploitation in a public health context.illustrate_case
105
Includes specific clinical metrics on sensitivity, specificity, transfer time reduction, and length of stay improvements.illustrate_case
106
The slide uses two circular process diagrams to contrast 'Fixed data, evolving model' (model-centric) with 'Fixed model, evolving data' (data-centric).present_framework
107
The slide explains the concept of dynamic benchmarking using the Dynabench platform as a case study.present_solution
108
The slide highlights the importance of robust benchmarks for real-world ML applications.establish_context
109
The slide references a specific research paper/finding regarding underspecification in ML models.diagnose
110
The slide uses a PRISMA-style flow diagram to show the systematic review process of ML papers.illustrate_case
111
Includes a trend chart showing the increase in automated labels over time and four visual examples of computer vision use cases.analyze_data
112
The slide tracks a specific past prediction and provides context on why the deal is stalling.illustrate_case
113
The slide uses a value chain decomposition to show regional concentration and a waterfall-style chart to estimate the cost of localization.frame_problem
114
Includes a stock price chart and a photo of an EUV machine.illustrate_case
115
The slide uses a stacked bar chart for regional capex and a line chart for lead times.analyze_data
116
The slide uses a waterfall-style bar chart to show regional production declines and an area chart for historical context.analyze_data
117
Slide from State of AI 2021 report.illustrate_case
118
The chart shows the decline of US semiconductor manufacturing share from 1990 to 2020, with projections to 2030.establish_context
119
The slide highlights the growth and platform adoption of AI-first cybersecurity firms, supported by stock performance data.analyze_data
120
The slide uses a grouped bar chart to show ARR growth across six quarters for three major enterprise software companies.analyze_data
121
The slide includes a revenue growth chart, a data pipeline diagram, and a list of output channels.illustrate_case
122
Data source: Koyfin. Stock price data as of 30 Aug 2021.analyze_data
123
Slide from State of AI 2021 report.summarize
124
The slide contrasts a multi-stage modular pipeline (sensors -> perception -> planning -> control) with a simplified end-to-end neural network approach.compare_options
Open slide detailBeat · The Facts (What)
125
The slide illustrates a reinforcement learning (RL) pipeline for AVs and provides a line chart showing the reduction in interventions per 1000 miles as trainingillustrate_case
Open slide detailBeat · The Facts (What)
126
The slide uses a screenshot from the AI City Challenge to illustrate the computer vision tasks mentioned.summarize
Open slide detailBeat · The Facts (What)
129
The slide uses screenshots of Google Sheets, Google Maps AR, and MediaPipe to demonstrate AI integration.illustrate_case
Open slide detailBeat · The Facts (What)
132
The slide highlights advancements in AI-driven image enhancement from two major industry players.illustrate_case
Open slide detailBeat · The Facts (What)
133
The slide features a process flow of the app's functionality and a bar chart showing subscriber growth.illustrate_case
Open slide detailBeat · The Facts (What)
134
The chart shows the learning curve of the RL agent compared to a baseline 'best engineered' solution.illustrate_case
Open slide detailBeat · The Facts (What)
135
The slide uses a series of line charts to demonstrate the accuracy of the forecasting model against actual sales data.illustrate_case
Open slide detailBeat · The Facts (What)
136
The slide highlights the impact of AI-driven financial products on merchant performance and usage patterns.illustrate_case
Open slide detailBeat · The Facts (What)
139
Includes a diagram illustrating the cohort assignment process.diagnose
Open slide detailBeat · The Facts (What)
140
The slide uses the 'Thinker' statue pose as a visual metaphor for AI invention.establish_context
Open slide detailBeat · The Facts (What)
141
The slide uses a mix of currency units ($, €) and includes a logo grid for company examples.analyze_data
142
The chart uses a 2.1x growth annotation for the EU+UK segment between 2020 and 2021.quantify_impact
143
Data source: dealroom.co. The chart shows a clear trend of increasing capital concentration in larger funding rounds over time.analyze_data
144
Data source: dealroom.co. The chart shows a clear growth trend for both sectors, with SaaS consistently maintaining a higher enterprise value.quantify_impact
145
The chart tracks enterprise value growth across four valuation buckets. The right side provides a logo grid of notable companies with their respective valuationsize_opportunity
146
Data source: dealroom.co. Slide 146 of State of AI 2021 report.analyze_data
148
The title text claims 2/3 for USA and 1/3 for EU+UK, but the chart shows China as a distinct, albeit small, segment, suggesting the title is a simplified summaranalyze_data
149
The chart shows a significant spike in 2021 YTD compared to previous years, dominated by USA-based exits.quantify_impact
150
The slide uses horizontal bar charts to visualize industry-specific exit data.quantify_impact
151
Data source: dealroom.co. Part of the State of AI 2021 report.analyze_data
152
Data source: dealroom.co. The chart shows a clear upward trend in corporate M&A activity within the AI sector.quantify_impact
154
The slide uses a narrative approach to document a specific event in AI ethics history.illustrate_case
155
The chart shows cumulative probability of training a transformative model over time, with a median estimate of 2052.quantify_impact
156
The chart shows trust ratings for various corporate entities, with OpenAI and Google/Microsoft scoring higher than others.summarize
157
The chart distinguishes between industry labs (blue) and academic labs (red).quantify_impact
158
The chart distinguishes between industry labs (blue) and academic labs (red).compare_peers
159
Includes a screenshot of a tweet by Owain Evans as supporting evidence.summarize
160
The slide discusses corporate governance and ethical oversight shifts within the Google/DeepMind relationship.summarize
161
Slide from State of AI 2021 report.summarize
162
The slide highlights Anthropic's positioning as a safety-focused entity in the AGI landscape.summarize
163
The slide highlights the tension between closed-source commercial AI and open-source community-driven research.illustrate_case
164
The slide uses a treemap to visualize the composition of 'The Pile' dataset.illustrate_case
165
The slide describes a specific organizational model for AI research, comparing it to decentralized approaches like EleutherAI.establish_context
166
Part of the 'State of AI 2021' report, specifically the Politics section.summarize
167
Includes a photo of Margrethe Vestager, European Commissioner for Competition.establish_context
168
Includes a satirical comic strip illustrating public/political sentiment toward regulation.frame_problem
169
Part of the 'State of AI 2021' report, section on Politics.establish_context
170
The slide uses a map of China as a background for a chronological timeline of AI ethics initiatives.establish_context
171
The slide highlights a contradiction between stated ethical guidelines and state-sponsored surveillance practices.frame_problem
172
The map is a satirical illustration of fragmented state-level laws, not a factual representation of actual legislation.frame_problem
173
The slide uses a map to visualize the status of privacy legislation across the US, with a specific focus on the regulatory strength of California vs. Virginia.analyze_data
174
The slide uses a visual categorization of agencies based on ownership/usage of facial recognition systems.frame_problem
175
The slide highlights the emergence of AI as a 'force-multiplier' in modern warfare.summarize
176
The slide highlights the transition of AI from research/games to military production.illustrate_case
177
The slide highlights a specific milestone in military AI development, focusing on the Skyborg Vanguard program.summarize
178
Part of the 'state of ai' report series.summarize
179
Includes logos of Anduril, Shield AI, and Rebellion Defense at the bottom.summarize
180
Includes a quote from Microsoft CEO regarding the ethical stance on military contracts.establish_context
181
The slide highlights the strategic use of must-pass military legislation to advance broader national AI policy.establish_context
183
Part of the 'State of AI 2021' report.summarize