Air Street Capital · consulting-deck
2023 Air Street Capital The State of AI Report 2023
163 pages · 3 arc beats · 16 loops
2023 Air Street Capital The State of AI Report 2023
Air Street Capital arc beats above · slides in the middle · loops below · scroll → 16 LOOPS
1102030405060708090100110120130140150160163
Deck intelligence map
3 coverage by narrative range · generated from this deck JSON
Narrative range 148 total
Metadata
Components
Metrics
Tools
Frameworks
Beats
Loops
The Facts (What) 110 slides 100% 110/110 slides 100% 110/110 slides · 686 hits — 0/110 slides
69.1% 76/110 slides · 112 hits 5.5% 6/110 slides 100% 110/110 slides · 220 hits 98.2% 108/110 slides The Implications (So What) 35 slides 100% 35/35 slides 100% 35/35 slides · 218 hits — 0/35 slides
80% 28/35 slides · 49 hits 22.9% 8/35 slides · 11 hits 100% 35/35 slides · 70 hits 94.3% 33/35 slides The Action (Now What) 3 slides 100% 3/3 slides 100% 3/3 slides · 10 hits — 0/3 slides
100% 3/3 slides · 8 hits — 0/3 slides
100% 3/3 slides · 6 hits 66.7% 2/3 slides Slide inventory
163 every slide · same image gating as the playbook
04
The slide serves as a table of contents and mission statement for the report.front_matter
06
This slide serves as a foundational reference for the rest of the deck.establish_context
07
This slide acts as a key for interpreting model-related diagrams in subsequent slides.present_framework
10
The slide uses a traffic light system (Green=YES, Red=NO, Amber=~) to indicate the status of predictions.summarize
Open slide detailLoop · Precedent
12
The chart shows a comparison of GPT-4, GPT-4 (no vision), and GPT-3.5 across various academic and professional exams.illustrate_case
13
Includes a process diagram illustrating the RLHF training pipeline.present_framework
14
The slide references Berkeley research on imitation learning and Meta's LLaMA-2 paper findings on RLHF.summarize
15
The slide discusses alternatives to Reinforcement Learning from Human Feedback (RLHF) in AI training.summarize
16
Includes a screenshot of a tweet by William Falcon mocking the lack of detail in the GPT-4 technical report.summarize
17
The slide discusses the release of LLaMa, its performance metrics (RACE benchmark), and the controversy regarding its 'open-source' status.illustrate_case
18
The slide discusses the ecosystem of open-source LLMs, mentioning specific models like MPT-30B, Falcon, RedPajama, Pythia, and Vicuna, and techniques like LoRA.summarize
19
The slide uses a stacked bar chart to show human evaluation of LLaMa-2 helpfulness against other models.illustrate_case
20
The chart maps model release/popularity over time (x-axis) against MMLU performance (y-axis), with bubble size representing social media mentions.analyze_data
21
The chart tracks the share of different research topics in the AI field.summarize
22
The slide presents a critique of the 'emergent capabilities' phenomenon in LLMs, citing Stanford research that suggests these are artifacts of metric choice.diagnose
23
The slide discusses technical methods like FlashAttention, ALiBi, RoPE, and Positional Interpolation in the context of LLM scaling.summarize
24
The slide references the 'Lost in the Middle' research paper by Samaya.ai, UC Berkeley, Stanford, and LMSYS.org.diagnose
25
Part of the State of AI 2023 report.summarize
26
The slide highlights the shift from 'more data/parameters' to 'better data' in LLM training, specifically referencing the 'phi' model series.summarize
27
The slide uses a visual comparison (meme vs. truth) to explain the shift from dense models to Mixture of Experts (MoE) architectures.illustrate_case
28
The slide uses three separate line charts to illustrate data exhaustion timelines for different data types.analyze_data
29
The chart shows Top-1 Accuracy vs Parameters (M) for models trained on 'Real' vs 'Real + Generated' data, indicating performance gains.analyze_data
30
Includes logos of University of Maryland, Google, Google DeepMind, ETH Zurich, and Princeton.summarize
31
Discusses the trade-off between training epochs, model size, and overfitting.analyze_data
32
The slide contrasts structured quantitative evaluation (HELM/Hugging Face) with qualitative user sentiment ('vibes').present_framework
33
Includes a process diagram showing the training pipeline for Code Llama models.summarize
34
The slide uses two process diagrams to explain how AlphaDev functions as an RL agent.illustrate_case
35
The slide compares different prompting architectures (IO, CoT, CoT-SC, ToT) visually.present_framework
36
The slide uses bar charts to compare accuracy, verbosity, and mismatch metrics for two model versions over time.analyze_data
37
Includes a small example box illustrating tool usage (QA and Calculator).summarize
38
The chart shows the 'Minecraft Tech Tree' progression of different agents, with Voyager (ours) showing the highest performance.illustrate_case
39
The slide contrasts standard Chain of Thought (linear) with RAP (tree-based planning).present_framework
40
Includes a diagram of the reasoning module (DAG) and a performance comparison table.illustrate_case
41
Includes a technical diagram of the BLIP-2 architecture.analyze_data
42
The slide explains a technical approach to visual reasoning using LLMs as code generators for visual APIs.present_solution
43
The slide features a conceptual diagram of the LINGO-1 architecture and two real-world examples of the model providing driving commentary.illustrate_case
44
The slide uses a diagram to explain the multimodal architecture (PaLM + ViT) and provides examples of robotic manipulation and visual Q&A.illustrate_case
45
The slide explains the architecture and capabilities of the RT-2 model, highlighting its ability to generalize to novel objects and its deployment efficiency.illustrate_case
46
The slide uses a circular process diagram to explain the self-improving training loop of the RoboCat agent.illustrate_case
47
The slide highlights a milestone in AI robotics, specifically in competitive drone racing.illustrate_case
48
The slide highlights a specific research finding regarding emergent map-building in blind navigation agents.illustrate_case
49
The slide explains the technical architecture of CICERO, including its use of dialogue history, board state, and a 2.7B-parameter BART-like model.illustrate_case
50
The slide compares diffusion-based models vs. masked transformer models in the context of video generation.summarize
51
The slide highlights the shift from pure generation to instruction-based editing using models like InstructPix2Pix and Imagen Editor.summarize
52
Comparison of MipNeRF360 vs 3D Gaussian Splatting performance metrics.present_solution
53
The slide highlights specific research papers and models (DreamFusion, RealFusion, SKED, Instruct-NeRF2NeRF) in the context of 3D generative AI.summarize
54
The slide highlights two specific research papers/models in the field of computer vision.summarize
55
Includes a diagram of the model architecture showing prompt inputs and image inputs leading to a valid mask output.summarize
56
The slide highlights technical advancements in self-supervised learning, specifically dataset curation and model stability.summarize
57
The slide compares traditional numerical weather prediction methods with new deep learning/physics-informed models.summarize
58
Includes a spectrogram image and a performance comparison table for various music generation models.summarize
59
The slide explains the technical application of generative AI (diffusion models) in biotechnology, specifically protein design.present_framework
60
The slide highlights a breakthrough in computational biology using language models (ESM-2) to predict protein structures.illustrate_case
61
The slide explains a specific scientific research application (GEARS) using a process diagram and two scatter plots (UMAP visualizations).illustrate_case
62
Slide from State of AI 2023 report.illustrate_case
63
The slide uses a bar chart to show historical progress on the USMLE and a grouped bar chart to compare Med-PaLM 2 vs. physicians across quality and risk metricsillustrate_case
64
The slide discusses MultiMedBench (14 tasks) and the architectural shift to multimodal medical models.present_solution
65
The slide highlights the use of social media data for specialized medical AI training.illustrate_case
66
The slide presents a specific technical architecture (CoDoC) for AI-human decision deferral in clinical settings.illustrate_case
67
The chart tracks publication volume by scientific discipline, highlighting the rapid growth in medicine and the high volume in mathematics.analyze_data
68
The slide uses two grouped bar charts to show the concentration of AI research output by country and by organization.analyze_data
70
The chart uses a red box to highlight the most recent quarter (Q2 FY24) and includes emojis to emphasize the growth.analyze_data
71
Includes a quote from CoreWeave regarding their fleet scale.illustrate_case
72
The slide uses a 2x2 grid layout to present four distinct company case studies.illustrate_case
73
The title has a strikethrough on 'Footballers' which seems to be a humorous or editorial correction.establish_context
74
The chart highlights significant growth in large-scale clusters, specifically mentioning Tesla, Stability, and Hugging Face.analyze_data
75
The chart categorizes entities by National HPC, Private Cloud, and Public Cloud.analyze_data
76
The chart uses a logarithmic scale on the y-axis to visualize the massive disparity in citation counts.quantify_impact
77
The chart tracks usage volume of various NVIDIA GPU models over time (2018-2023).analyze_data
78
The chart tracks the number of mentions or usage metrics for AI chip companies among researchers.compare_peers
79
The slide also includes a rumor about NVIDIA's H100 shipment projections for 2024.analyze_data
80
The chart uses a 'Units of A100 GPUs' scale to normalize compute power over time.quantify_impact
81
The slide uses two line charts to illustrate the drivers for in-house hardware development: increasing model complexity/size and the surge in GPU server demand.analyze_data
82
The chart illustrates the 'Current Export Restrictions' as a threshold boundary on a 2D plane of Computational Performance vs Interconnect Bandwidth.establish_context
83
Part of the State of AI 2023 report.illustrate_case
84
The slide validates a previous prediction using specific company performance metrics.illustrate_case
85
The slide highlights the transition of Synthesia from a 'fringe' technology to mainstream enterprise adoption.illustrate_case
86
The chart uses a log-log scale to visualize exponential growth rates over time.analyze_data
87
The slide uses two news article headlines as evidence for the dual narrative of revenue growth vs. high costs.summarize
88
Includes a stock performance chart, strategic rationale for AI integration, and UI screenshots of the new AI-powered product.illustrate_case
89
The slide uses two charts to illustrate the displacement of traditional developer Q&A platforms by AI-driven coding assistants.analyze_data
90
Includes a chart comparing Stack Overflow vs GitHub web traffic and a small chart showing CoPilot acceptance rate over time.illustrate_case
91
The slide presents two charts (A and B) showing the treatment effect of ChatGPT on time taken and output quality compared to a control group.illustrate_case
92
Includes screenshots of Character.AI and news headlines regarding Replika and Character.AI.illustrate_case
93
Part of the State of AI 2023 report.summarize
94
Data source: Sequoia Capital. The slide uses two grouped bar charts to contrast incumbent apps (green) with AI-first companies (blue).compare_peers
95
The slide contrasts two different strategic responses to generative AI in the stock media industry.illustrate_case
96
The slide uses a screenshot of a press release to provide evidence for the claim.cite_precedent
97
The slide discusses the Thaler v. Perlmutter case and the Andy Warhol Foundation v. Goldsmith case (implied by the Prince portrait reference).cite_precedent
98
Part of the State of AI 2023 report.summarize
99
Slide 99 from State of AI 2023 report.illustrate_case
100
The slide uses a bar chart to contrast growth metrics for Hugging Face's ecosystem.summarize
101
The slide highlights a strategic shift in the AI industry towards fine-tuning and specialized models.summarize
102
The slide uses press release snippets to validate the trend of AI adoption in pharma.illustrate_case
103
The chart illustrates the 'AI fever' phenomenon by showing a sharp spike in Recursion Pharmaceuticals' stock following an investment announcement from NVIDIA.illustrate_case
104
The slide uses a chronological sequence of logos to illustrate organizational branding changes.establish_context
105
Includes a list of authors and logos of companies they founded or joined.illustrate_case
106
The slide tracks the career paths of the authors of the 'Attention Is All You Need' paper and maps them to their current startups and 2023 funding amounts.illustrate_case
107
The slide showcases the capabilities of the GAIA-1 model through visual examples of different driving conditions.illustrate_case
108
Includes logos of Waymo, Cruise, and Stack AV at the bottom.summarize
109
The slide uses two stacked bar charts to show investment by round size and by AI vs. GenAI split.quantify_impact
110
The slide compares private startup/scaleup EV (left) with public startup/scaleup EV (right) over time, segmented by launch year cohorts.analyze_data
111
The chart shows a significant concentration of capital in the US compared to other regions.analyze_data
112
Data source: dealroom.co as of 19 Sept 2023.compare_peers
113
The slide contains four distinct bar charts showing investment and deal volume metrics for AI categories.analyze_data
114
The chart shows a clear dominance of acquisitions over IPOs in recent years, with 2023 data being partially projected.analyze_data
115
The chart highlights a significant drop in non-AI investment in 2023 compared to previous years.analyze_data
116
The chart highlights a significant spike in funding in 2023 compared to previous years.quantify_impact
117
The chart illustrates the shift toward massive capital requirements for AI model training, with a specific callout to CoreWeave's debt financing model.quantify_impact
118
The slide evaluates a past prediction and provides evidence of the actual outcome through three categories of NVIDIA's activities.summarize
119
The slide maps AI startups (left) to their funding round size (middle) and corporate investors (right).illustrate_case
120
Data source: dealroom.coanalyze_data
122
Uses a Venn diagram to show overlapping regulatory strategies across different nations.present_framework
123
The slide discusses the evolving regulatory stance of the UK and India regarding AI.summarize
124
Part of a larger 'state of AI' report; slide 124.summarize
125
Part of the 'state of ai 2023' report, specifically the 'Politics' section.summarize
126
The slide includes a table mapping functions to institutional efforts, though the table is small and serves as a supporting visual.establish_context
127
The slide highlights the 'Frontier Model Forum' and individual CEO lobbying efforts as examples of industry-led governance.summarize
128
The chart shows a decline in Chinese imports of semiconductor production machines, supporting the thesis of successful export controls.analyze_data
129
Part of the 'state of ai 2023' report series.summarize
130
The slide uses a rhetorical question to frame a nuanced discussion about semiconductor manufacturing capabilities in China.frame_problem
132
Includes a photo of the Anduril Fury autonomous vehicle.summarize
133
Includes a reference to the FT Film 'Ukraine's Tech War'.illustrate_case
134
Part of the 'State of AI 2023' report, slide 134.establish_context
135
Includes a screenshot of a Politico article titled 'Inside the AI culture war'.establish_context
136
Slide from State of AI 2023 report.summarize
137
The table shows exposure levels of various occupations to AI models compared to human assessments.summarize
139
The slide presents a framework for categorizing AI risks as proposed by Dan Hendrycks (Center for AI Safety).present_framework
140
The slide uses a collage of screenshots from various media outlets and public statements to illustrate the 'mainstream' nature of the AI x-risk debate.establish_context
141
Part of the 'State of AI' report series.summarize
142
Slide from the State of AI 2023 report.summarize
143
The slide highlights the shift of AI safety from a niche technical debate to a mainstream political and national security priority.establish_context
144
The chart shows a positive correlation between safety and ethics research output, with some institutions heavily skewed towards one or the other.analyze_data
145
The slide highlights specific industry responses to AI safety concerns, featuring diagrams of Anthropic's ASL framework and a general model evaluation lifecyclesummarize
146
Part of the State of AI 2023 report.summarize
147
The slide uses a specific case study (Sydney/Bing) to illustrate broader safety concerns like jailbreaking and prompt injection.diagnose
148
The slide highlights three specific research findings regarding AI safety vulnerabilities.summarize
149
The slide categorizes challenges into three pillars based on RLHF components, referencing specific sections (§3.1, §3.2, §3.3) of a research paper.diagnose
150
The chart shows a comparison between conventional pretraining and conditional pretraining (pretraining with feedback) in terms of toxicity reduction.illustrate_case
151
Includes a process diagram for the 'Self-Align' technique.present_framework
152
The slide uses a complexity-theory-inspired notation (NP^NP) to illustrate the recursive difficulty of evaluating AI assistants.diagnose
153
Discusses 'Let's Verify Step by Step' (OpenAI) and 'Contrast-Consistent Search' (UC Berkeley/Peking University).present_framework
154
The chart shows a negative correlation between subject model parameters and explanation score.diagnose
155
The chart shows win rates across various models, comparing human evaluation against GPT-4 and GPT-3.5 as judges.analyze_data
157
Part of the State of AI 2023 report.summarize