2022 Air Street Capital The State of AI Report 2022

Air Street Capital
arc beats above · slides in the middle · loops below · scroll → 18 LOOPS
SETUP TENSION ANALYSIS EVIDENCE RESOLUTION APPENDIX
HOVER FOR DETAILS · CLICK A SLIDE FOR FULLSCREEN · STEP 5
No image bound to this slide. Wire up imgproxy to display the rendered slide JPEG.

Deck intelligence map

3
coverage by narrative range · generated from this deck JSON

Slide inventory

114
every slide · same image gating as the playbook
01
Slide 1
front_matter
03
The slide serves as an introduction/agenda for the report, outlining the structure of the document.establish_context
05
establish_context
Open slide detailBeat · The Facts (What)
06
establish_context
Open slide detailBeat · The Facts (What)
07
The slide uses a structured list format to categorize key takeaways across four domains.summarize
Open slide detailBeat · The Facts (What)
08
transition
Open slide detailBeat · The Facts (What)Loop · Hypothesis Test
09
The slide uses a color-coded 'Grade' column (green for Yes, red for No) to summarize the outcome of past predictions.summarize
Open slide detailBeat · The Facts (What)Loop · Hypothesis Test
10
summarize
Open slide detailBeat · The Facts (What)Loop · Hypothesis Test
11
transition
Open slide detailBeat · The Facts (What)
12
The slide describes a human-AI collaborative workflow for mathematical discovery.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
13
Part of a series of slides tracking 2021 predictions.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
14
The slide tracks a previous year's prediction and provides a technical explanation of the breakthrough, supported by a bar chart comparing speed-ups.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
15
The slide details the application of RL to magnetic coil control in fusion reactors (TCV/ITER).illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
16
The chart shows a significant jump in predicted structures in 2022 compared to historical published data.summarize
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
17
Includes a bar chart showing perplexity decrease with model size and a table of ESM-2 performance metrics.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
18
Includes logos of research institutions at the bottom.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
19
Includes a time-lapse visual of plastic degradation and a schematic of the ML model used for enzyme engineering.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
20
The slide references a Princeton study on data leakage in 329 papers across 17 fields.diagnose
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
21
The slide explains the methodology of training an Inverse Dynamics Model (IDM) to scale up training data for a foundation model.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
22
The slide highlights the competitive landscape of LLMs applied to software engineering tasks.summarize
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
23
The slide features a Venn diagram mapping various 'Efficient Transformer' research papers into categories like Recurrence, Memory/Downsampling, Low Rank/Kernelsanalyze_data
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
24
Includes a specific example of a math problem solved by Minerva.summarize
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
25
The slide includes three small scatter plots showing different benchmark saturation dynamics.summarize
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
26
The chart illustrates the relationship between model parameters and compute (FLOPs), comparing DeepMind's findings with previous OpenAI scaling laws.analyze_data
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
27
The slide discusses the non-linear performance gains in LLMs as they scale, using four specific task examples (Arithmetics, Fig. of speech, Multi-task NLU, Trananalyze_data
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
28
The chart compares truthfulness and informativeness across different model sizes and prompting strategies, highlighting WebGPT's performance.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
29
The chart uses a logarithmic scale for the y-axis to visualize the massive acceleration in compute requirements.analyze_data
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
30
Includes a list of academic paper titles as a visual reference for the trend.summarize
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
31
The slide uses visual examples to demonstrate the capabilities of different AI models.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
32
The slide uses visual evidence (image generation examples) to support the claim of competitive parity.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
33
Slide from State of AI 2022 report.establish_context
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
34
The slide highlights the speed at which the open-source community has replicated or improved upon proprietary landmark models.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Tale Two Worlds
35
The timeline highlights the speed at which the open-source community (e.g., DALL-E mini/Craiyon, Stable Diffusion) caught up to or surpassed closed-source modelillustrate_case
Open slide detailBeat · The Facts (What)Loop · Tale Two Worlds
36
The slide highlights the speed of open-source community adoption and improvement of proprietary AI models.summarize
Open slide detailBeat · The Facts (What)Loop · Tale Two Worlds
37
Includes visual diagrams of the SayCan decision process and success metrics.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
38
The chart compares model performance (ImageNet accuracy) vs computational cost (GFLOPs) for various architectures.analyze_data
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
39
Discusses the technical approach of masking large patches of pixels to train Vision Transformers (ViT).summarize
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
40
The slide discusses the evolution of transformer models from task-specific to generalist architectures.summarize
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
41
The slide validates a previous prediction using the IRIS model as a case study, supported by visual evidence of world model simulation and comparative performanillustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
42
Data source: Zeta Alpha. The chart illustrates the diversification of transformer model applications.analyze_data
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
43
The chart shows a significant increase in NeRF-related research papers from 2019 to 2022.analyze_data
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
44
The slide uses three bar charts (E, F, G) to demonstrate the efficacy of ML-recommended antibiotics compared to physician-prescribed ones.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
45
The slide highlights the performance gap between traditional methods and property-prediction transformers in drug discovery.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
46
The slide highlights the dual-use nature of AI in pharmaceutical research, specifically referencing the MegaSyn model.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
47
The charts use a diverging bar chart format where red bars indicate higher prevalence in China and blue bars indicate higher prevalence in the US.compare_peers
Open slide detailBeat · The Facts (What)Loop · Tale Two Worlds
48
The slide uses two horizontal bar charts to contrast national output vs institutional output.compare_peers
Open slide detailBeat · The Facts (What)Loop · Tale Two Worlds
49
The chart uses a stacked bar approach to show the difference between standard corpus and CNKI-included corpus for China.quantify_impact
Open slide detailBeat · The Facts (What)Loop · Tale Two Worlds
50
transition
Open slide detailBeat · The Facts (What)
51
The slide uses a comparison of valuation vs revenue to illustrate the scale gap between incumbents and startups.compare_peers
Open slide detailBeat · The Facts (What)Loop · David Goliath
52
Logarithmic scale used for the y-axis.analyze_data
Open slide detailBeat · The Facts (What)Loop · David Goliath
53
Data source: stateof.ai 2022 report.analyze_data
Open slide detailBeat · The Facts (What)Loop · David Goliath
54
The slide uses a combination of a line chart for stock price and bar/line charts for financial metrics (Revenue, R&D) to illustrate the growth during the deal tillustrate_case
Open slide detailBeat · The Facts (What)Loop · David Goliath
55
The slide highlights the synergy between AI research and hardware engineering, specifically referencing the H100 GPU and PrefixRL.illustrate_case
Open slide detailBeat · The Facts (What)Loop · David Goliath
56
The slide uses a mapping of compute providers to AI labs, highlighting the trend of strategic partnerships.present_framework
Open slide detailBeat · The Facts (What)Loop · David Goliath
57
The slide highlights the shift in compute power from public/national infrastructure to private corporate entities.analyze_data
Open slide detailBeat · The Facts (What)Loop · David Goliath
58
The chart shows a correlation between government contracts and AI software upgrades in China.analyze_data
Open slide detailBeat · The Facts (What)Loop · David Goliath
59
Includes screenshots of Google's AI Test Kitchen app.frame_problem
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
60
The slide uses a logo-grid to illustrate the proliferation of startups founded by alumni from DeepMind and OpenAI.summarize
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
61
The slide maps individual researchers from the original paper to the companies they founded or joined, and lists the capital raised by those companies.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
62
Includes a mix of internal Google metrics and a controlled experiment conducted by GitHub Copilot.quantify_impact
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
63
The slide uses a stacked bar chart for company-specific pipelines and a horizontal bar chart for the aggregate percentage distribution of assets.quantify_impact
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
64
The slide uses a combination of narrative text and two charts to illustrate the 'physical bottlenecks' of drug development.frame_problem
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
65
The slide features a grid of 16 small area charts showing variant submission counts over time, with red (EWS prediction) and green (WHO designation) markers.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
66
The slide uses a process flow diagram to explain the AI's role in triaging chest X-rays.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
67
Includes data from Beauhurst and GOV.UK. Shows sector breakdown, university-specific spinout counts vs equity stakes, and historical trend of average equity staillustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
68
The slide uses three distinct charts to illustrate the 'spinout problem' (time, equity, and NPS).illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
69
The slide uses a timeline-based case study to demonstrate the efficacy of a specific academic research funding model.illustrate_case
Open slide detailBeat · The Facts (What)Loop · Pattern Hunter
70
Data source: dealroom.co. The slide uses stacked bar charts to show investment by round size.analyze_data
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
71
Data source: dealroom.co. The charts show a significant decline in large-scale funding compared to the relative stability of smaller rounds.analyze_data
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
72
The slide uses stacked bar charts to show EV by launch year cohorts.compare_options
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
73
The slide uses a bar-chart-in-table format to visualize the data.compare_peers
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
74
Data source: dealroom.co. The chart shows a decline in absolute investment in 2022 YTD compared to 2021, but maintains a >50% share for the US.analyze_data
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
75
The slide uses a combination of table and bar chart elements to visualize investment data across 24 industries.analyze_data
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
76
Data source: dealroom.co. The chart shows a clear shift from public market exits to M&A activity in 2022.analyze_data
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
77
Data source: dealroom.co. The chart shows a clear trend of increasing AI exits globally, with a specific focus on the performance of European regions versus theanalyze_data
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
78
The chart uses a stacked bar format to show funding rounds by size, with a projection for the full year 2022.analyze_data
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
79
Data source: dealroom.co. The chart tracks EV by launch year cohorts.analyze_data
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
80
The slide uses a stacked bar chart to show EV growth by launch year cohorts and a grid of company logos with valuation details.size_opportunity
Open slide detailBeat · The Facts (What)Loop · Segmentation Split
81
transition
Open slide detailBeat · The Implications (So What)
82
The slide uses two scatter plots with trend lines to visualize the divergence in compute power and research output share.analyze_data
Open slide detailBeat · The Implications (So What)Loop · Tale Two Worlds
83
The slide contrasts institutional bureaucracy with agile research collectives.establish_context
Open slide detailBeat · The Implications (So What)Loop · Tale Two Worlds
84
The chart shows a clear decline in academic dominance (blue) and a rise in industry/collective collaborations (red/grey) in recent years.summarize
Open slide detailBeat · The Implications (So What)Loop · Tale Two Worlds
85
The slide illustrates the ecosystem of Stability AI, showing how it supports various research groups (LAION, EleutherAI) and commercializes the output (DreamStupresent_solution
Open slide detailBeat · The Implications (So What)Loop · Tale Two Worlds
86
illustrate_case
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
87
Part of the State of AI 2022 report.summarize
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
88
The slide uses a process diagram to explain the 'Uber-like' dispatch model of the software.illustrate_case
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
89
Data source: CSET, stateof.ai 2022 report.compare_peers
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
90
Includes a table summarizing the relative need for reshoring across different semiconductor device categories.analyze_data
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
91
Slide from the State of AI 2022 report.establish_context
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
92
Part of the 'State of AI 2022' report series.summarize
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
93
summarize
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
94
transition
Open slide detailBeat · The Implications (So What)
95
Features quotes from Alan Turing, I.J. Good, and Marvin Minsky.establish_context
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
96
Includes quotes from the UK National AI Strategy document.cite_precedent
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
97
The slide uses a series of diverging stacked bar charts to visualize survey agreement levels.summarize
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
98
The chart uses a stacked bar to compare 2021 and 2022 researcher counts across different venues, highlighting a 30x gap between NeurIPS and Safety.quantify_impact
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
99
The slide uses two bar charts to contrast the scale of safety-focused funding (in millions) against general capabilities funding (in billions).quantify_impact
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
100
Includes a chart showing Elo scores vs parameters and a chat example of a safety-aligned model.present_solution
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
101
Includes a process diagram for RLHF and a bar chart comparing model performance.present_framework
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
102
Includes a conceptual diagram of the red teaming process and a bar chart showing harmlessness by model size.illustrate_case
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
103
The slide discusses the reverse-engineering of neural networks into human-interpretable programs.summarize
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
104
The slide uses the CoinRun experiment to illustrate a specific AI safety failure mode.diagnose
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
105
The slide highlights the risk of AI being trained in environments that reward immoral behavior and proposes a solution to mitigate this.illustrate_case
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
106
Part of the State of AI 2022 report.illustrate_case
Open slide detailBeat · The Implications (So What)Loop · Pattern Hunter
107
transition
Open slide detailBeat · The Action (Now What)
108
The slide uses a numbered list format to present forward-looking statements.summarize
Open slide detailBeat · The Action (Now What)Loop · Time Machine
109
summarize
Open slide detailBeat · The Action (Now What)
110
appendix
114
front_matter