Air Street Capital · consulting-deck
2024 Air Street Capital The State of AI Report 2024
213 pages · 3 arc beats · 24 loops
2024 Air Street Capital The State of AI Report 2024
Air Street Capital arc beats above · slides in the middle · loops below · scroll → 24 LOOPS
1102030405060708090100110120130140150160170180190200210213
Deck intelligence map
3 coverage by narrative range · generated from this deck JSON
Narrative range 199 total
Metadata
Components
Metrics
Tools
Frameworks
Beats
Loops
The Facts (What) 145 slides 100% 145/145 slides 100% 145/145 slides · 870 hits — 0/145 slides
82.8% 120/145 slides · 211 hits 13.8% 20/145 slides · 21 hits 100% 145/145 slides · 290 hits 97.9% 142/145 slides The Implications (So What) 51 slides 100% 51/51 slides 100% 51/51 slides · 298 hits — 0/51 slides
78.4% 40/51 slides · 71 hits 7.8% 4/51 slides 100% 51/51 slides · 102 hits 96.1% 49/51 slides The Action (Now What) 3 slides 100% 3/3 slides 100% 3/3 slides · 7 hits — 0/3 slides
66.7% 2/3 slides · 4 hits — 0/3 slides
100% 3/3 slides · 6 hits — 0/3 slides
Slide inventory
213 every slide · same image gating as the playbook
07
This slide serves as a foundational reference for the rest of the deck.establish_context
08
This slide acts as a key for interpreting subsequent slides in the deck.establish_context
09
The slide uses a structured list format to categorize key takeaways across four main pillars of the AI landscape.summarize
Open slide detailBeat · The Facts (What)
11
The slide uses a traffic-light color coding system (Green=YES, Amber=~, Red=NO) to evaluate predictions.summarize
13
The slide highlights the convergence of performance between GPT-4, Claude 3.5 Sonnet, Gemini 1.5, and Grok 2.analyze_data
14
The slide discusses the shift from pre-training compute to inference compute in LLMs.summarize
15
Includes a Chatbot Arena chart and a screenshot of a viral YouTube video demonstrating o1's coding capabilities.illustrate_case
16
Includes a stacked bar chart showing human evaluation win/tie/loss rates against GPT-4o and Claude 3.5 Sonnet.summarize
17
The table uses a color-coded matrix (green for yes, orange for no/partial) to visualize the degree of openness for different LLM projects.compare_options
18
The slide uses two charts to demonstrate the impact of contamination: a bar chart showing performance gaps between standard benchmarks and clean test sets, and diagnose
19
Includes a small visual grid of examples showing 'Bad Question Clarity', 'Bad Options Clarity', 'No Correct Answer', and 'Wrong Groundtruth'.diagnose
20
The slide critiques the 'vibes-based' evaluation method of the LMSYS leaderboard, noting that GPT-4o and GPT-4o mini share top spots, potentially due to user pranalyze_data
21
The slide highlights the hybrid approach of neuro-symbolic AI in solving Olympiad-level math problems.illustrate_case
22
Discusses model pruning techniques (layer removal, knowledge distillation) and their impact on performance benchmarks.summarize
23
The slide uses an AI-generated image of two llamas wearing sunglasses labeled '8B' and '4B' to metaphorically represent model distillation.summarize
24
Includes a screenshot of a chat interface demonstrating Phi-3-mini-4k-instruct-q4 capabilities.summarize
25
Includes a visual comparison of image compression levels (32 tokens vs 256 tokens vs 65536 pixels).summarize
26
Mentions Stanford research on ReFT and LoReFT.present_solution
27
The slide highlights specific models like Mamba, Jamba, and Griffin as examples of this trend.summarize
28
The slide discusses model distillation techniques in AI research, specifically comparing token budget vs. accuracy.analyze_data
29
The chart shows a clear trend of Transformer dominance growing over time, with other paradigms remaining a small minority.analyze_data
30
The slide includes a process diagram for Nemotron-4-340B and a feature comparison table for various AI infrastructure providers.summarize
31
Includes a diagram illustrating the feedback loop of model collapse.diagnose
32
Includes a process flow diagram for the FineWeb pipeline.illustrate_case
33
The chart compares embedding performance vs generative performance, showing a clear separation between specialized embedding models and generative models.analyze_data
34
The slide discusses a specific technical improvement in Retrieval-Augmented Generation (RAG) systems.present_solution
35
The slide includes a schematic diagram of the Ragnarök framework and a list of example research questions.diagnose
36
The slide explains how DiLoCo allows training on loosely connected 'islands' of devices, reducing communication needs by 500x.summarize
37
The slide includes a table comparing various methods (CLIP, EVA-CLIP, OpenCLIP, etc.) across metrics like FLOPs and performance on ImageNet-1K and COCO datasetsanalyze_data
38
The slide highlights specific strategies (Multi-head Latent Attention, MoE, dataset curation) used by Chinese labs like DeepSeek and 01.AI.summarize
39
Includes a visual example of Qwen-2's vision capabilities analyzing an impressionist painting.summarize
40
The slide uses a 'before-after' framing to highlight the rapid advancement in AI vision capabilities.summarize
41
The slide explains technical improvements in AI image generation models (Stability AI).summarize
42
The slide uses visual examples of video generation to demonstrate the model's capabilities.illustrate_case
43
Part of the State of AI 2024 report.summarize
44
Part of the State of AI 2024 report.illustrate_case
45
Features official Nobel Prize announcement graphics for John J. Hopfield, Geoffrey E. Hinton, David Baker, Demis Hassabis, and John M. Jumper.illustrate_case
46
Includes a technical architecture diagram and a bar chart comparing success rates of AF3 against baselines.illustrate_case
47
The slide uses a bar chart to compare performance metrics of different AI models (Chai-1, AF3, RosettaFold) across different datasets.illustrate_case
48
The slide highlights the importance of in silico filtering in protein design and compares AlphaProteo's performance against previous methods across various targillustrate_case
49
Discusses the transition from specialized equivariant models to domain-agnostic diffusion models in molecular modeling.summarize
50
The slide explains the transition from Meta's research to the independent EvolutionaryScale company and the technical innovation of ESM3.present_solution
51
Includes a process diagram of the model training/inference and bar charts showing editing performance.illustrate_case
52
Includes logos of PoseCheck, PoseBusters, Polaris, and Inductive Bio at the bottom.frame_problem
53
The slide uses a hub-and-spoke diagram to illustrate how the MACE-MP-0 model integrates various material types.illustrate_case
54
The slide presents a technical case study on protein structure generation using AI models.illustrate_case
55
The slide illustrates the pipeline from data collection (UK Biobank/HCP) to model training (BrainLM) and downstream clinical prediction tasks.illustrate_case
56
The slide details the architecture and performance benefits of the Aurora model compared to traditional numerical weather prediction.illustrate_case
57
The slide explains a technical process for brain-to-image reconstruction using a specific neural architecture.illustrate_case
58
The slide highlights a medical application of AI/neural decoding, featuring a photo of the patient's interface and a line chart showing error rate reduction.illustrate_case
59
The slide highlights the limitations of current LLMs in generalization and the specific focus of the ARC benchmark on visual problem-solving.summarize
60
Includes a data table comparing LLM, Human, and No-rule performance across different game types and transition metrics.analyze_data
61
Includes a diagram of the Quiet-STaR algorithm.present_solution
62
The diagram illustrates the STRATEGIST framework, which uses a bi-level tree search combining high-level strategy learning with low-level self-play simulation.present_framework
63
The slide references 'grokking' (training beyond overfitting) and mechanistic interpretability findings from Ohio State University researchers.analyze_data
64
The slide explains the FunSearch methodology and provides a performance comparison table against standard heuristics.illustrate_case
65
The slide details a two-stage RL process (offline then online) for VLM training.illustrate_case
66
The slide contrasts Intelligent Go-Explore (using LLMs) with Classic Go-Explore (using hand-crafted heuristics) through a process diagram.illustrate_case
67
The slide features a process diagram showing a booking flow with failure and success states.summarize
68
The slide highlights the transition from research to application (indicated by the pencil-to-wrench icon).illustrate_case
69
The slide illustrates a technical process flow for automated RL task generation.illustrate_case
70
The slide uses a process flow diagram to explain the AI Scientist framework.illustrate_case
71
The slide describes a specific technical implementation (TestGen-LLM) and provides quantitative results (10% improvement, 73% acceptance rate).illustrate_case
72
The slide features a technical diagram of the Wayve Vision Language Action Model and a screenshot of the PRISM-1 simulation.illustrate_case
73
Part of the State of AI 2024 report.summarize
74
Uses a before-and-after framing to show the shift in industry sentiment.summarize
75
The slide uses a narrative structure to highlight four specific technical developments in robotics by Google DeepMind.summarize
76
The slide uses social media screenshots as evidence of community adoption.illustrate_case
77
The slide uses a visual metaphor of icons (pencil, image, robot) to represent the synthesis of generative models and robotics.summarize
78
The slide highlights two specific research directions: learning affordance from human video (CMU) and chain-of-thought reasoning (Berkeley/Stanford).summarize
79
The slide highlights a specific technical solution (HumanPlus) to a known industry bottleneck (data for imitation learning).illustrate_case
80
The slide highlights two specific research efforts: a Stanford/Columbia team focusing on gripper movement control and a UC San Diego team focusing on a two-partillustrate_case
81
Includes visual examples of Open-TeleVision and Bunny-Vision Pro systems.illustrate_case
82
The slide includes two process diagrams illustrating 'Self-training with search' and 'Uncertainty-guided search at inference'.present_solution
83
The slide highlights a specific research application of Stable Diffusion/U-Net for medical imaging.illustrate_case
84
The slide contrasts manual/brittle RPA processes with automated, foundation-model-based workflows.present_solution
85
The slide contrasts country-level publication proportions with organization-level year-on-year changes.analyze_data
87
Includes a US map overlay and photo of Jensen Huang.summarize
88
Slide from State of AI 2024 report.summarize
89
The slide uses a combination of bar charts to illustrate the financial performance gap between NVIDIA and its competitors.compare_peers
90
The chart uses a stacked bar representation for the challengers and a single bar for NVIDIA.compare_options
91
The slide uses screenshots of external articles to provide credibility to the skeptical viewpoint.establish_context
92
The chart uses a stacked bar approach to show the composition of clusters by type for each entity.analyze_data
93
The chart uses color coding to distinguish between Public Cloud, Private Cloud, and National HPC.analyze_data
94
Logarithmic scale chart showing hardware adoption trends from 2018 to 2024.analyze_data
95
Data source: Zeta Alpha, State of AI 2024 report.analyze_data
96
The chart tracks the number of AI research papers utilizing specific hardware systems.analyze_data
97
Dual-axis chart showing release intervals (bars) and performance growth (line).analyze_data
98
The slide compares historical NVLink specifications and mentions Tencent's Xingmai 2.0 as a competitive development.analyze_data
99
The slide uses a pie chart to visualize the root causes of infrastructure failures in large-scale AI training.illustrate_case
100
Slide from State of AI 2024 report.summarize
101
Includes a visual comparison of the Cerebras WSE-3 chip vs a standard GPU.summarize
102
Part of the 'State of AI 2024' report series.summarize
103
Includes a meme illustrating the 'whack-a-mole' metaphor.summarize
104
The slide uses screenshots of news articles as evidence for the described regulatory gap.summarize
105
The slide uses specific examples of illicit trade to illustrate the growing scale and complexity of semiconductor smuggling.illustrate_case
106
The chart displays revenue multiples for Anthropic, OpenAI, Perplexity, Runway, Cohere, Stability, and Character AI.quantify_impact
107
The slide uses headlines from external news sources to illustrate the industry trend of high burn rates in AI.summarize
108
The slide uses a line chart with annotated milestones to demonstrate the impact of strategic shifts on market valuation.illustrate_case
109
The slide uses color-coded bars to represent different model providers (OpenAI, Anthropic, Google, Meta, Mistral).compare_options
110
The slide uses bar charts to show cost per million tokens (implied) for various model generations.quantify_impact
111
The slide highlights price cuts of 76% and 86% for specific model iterations.analyze_data
112
The slide highlights the shift from static code snippets to interactive, real-time coding environments.illustrate_case
113
The slide uses two screenshots of media coverage to illustrate the shift toward product-centric AI development.summarize
114
The slide contrasts Mistral's success with Aleph Alpha's struggles, supported by a table showing model rankings.summarize
115
Includes a technical diagram comparing model architectures (Dense vs MoE vs Hybrid).analyze_data
116
The slide highlights antitrust concerns, specifically mentioning OpenAI/Microsoft, Anthropic/Google/Amazon, and NVIDIA's influence.establish_context
117
The slide uses news headlines as evidence to support the thesis of regulatory evasion through talent acquisition and licensing.summarize
118
The slide highlights the dominance of Github Copilot while showcasing the competitive landscape of AI coding startups.summarize
119
The chart shows preference for database vendors for RAG and LLM customization, highlighting the dominance of incumbents over specialized vector DBs.summarize
120
Includes a screenshot of Devin's workspace.summarize
121
Includes a screenshot of a Google AI Overview hallucination (glue on pizza) as a case study of reliability issues.summarize
122
Includes a screenshot of a tool used to check if YouTube videos were used for AI training.summarize
123
Part of the 'State of AI 2024' report, specifically the 'Industry' section.summarize
124
Part of the State of AI 2024 report.summarize
125
Includes references to external news coverage of the Cruise incident.illustrate_case
126
The slide uses a comparison frame to contrast current humanoid hype with the historical trajectory of autonomous vehicles.frame_problem
127
The slide uses social media screenshots as evidence of AI-generated visual artifacts in professional media.illustrate_case
128
Part of the 'State of AI 2024' report.illustrate_case
129
Slide highlights four specific AI startups: Sakana AI, H Company, Safe Superintelligence Inc., and Black Forest Labs.summarize
130
Uses news clippings as evidence for the difficulty of AI startup management.illustrate_case
131
The slide contrasts the rapid commercial success of a specialized player (ElevenLabs) with the cautious, risk-averse strategy of generalist frontier labs regardsummarize
132
Data for 2024 is partial (up to Sept 1).quantify_impact
133
Data sourced from US corporate fintech Ramp.analyze_data
134
Data source: Stripe. Comparison of time to $1M and $30M revenue milestones.quantify_impact
135
Includes a screenshot of the PolyAI website as a visual example.summarize
136
Includes a social media testimonial from Andrej Karpathy and a screenshot of the Kyutai Moshi interface.summarize
137
Includes specific mentions of Harvey, Klarna, and various law firms.summarize
138
Part of the 'State of AI 2024' report deck.summarize
139
The slide uses a technical architecture diagram to illustrate Apple's on-device vs server-side AI stack.analyze_data
140
The slide discusses the implications of Apple's internal foundation model work on their OpenAI partnership, using Unsloth's performance as a proxy for technicalanalyze_data
141
Includes a pipeline table and a photo of a lab facility.illustrate_case
142
The slide highlights the competitive landscape of AI video generation by showcasing visual outputs and financial backing of key players.illustrate_case
143
The slide uses a bar chart to illustrate the price disparity between proprietary models (Runway, Luma, etc.) and open-source/cheaper alternatives.compare_options
144
The slide highlights the convergence of fine-tuned image models and video generation tools.illustrate_case
145
The slide summarizes clinical progress in personalized cancer vaccines, highlighting specific patient outcomes and trial data.summarize
146
The slide contrasts early market failure with later success in the smart glasses category.illustrate_case
147
Includes a screenshot of a Marques Brownlee YouTube review and a pull-quote style review summary.illustrate_case
148
The slide uses two stacked bar charts to illustrate the growth of AI investment, highlighting the dominance of the US and the rise of Generative AI.quantify_impact
149
The chart illustrates a significant shift in market valuation, with public companies now accounting for the vast majority of the total AI market value.quantify_impact
150
The slide presents two horizontal bar charts side-by-side to contrast historical prevalence with recent funding activity.analyze_data
151
The chart illustrates a clear shift in funding composition, with the largest round category ($250M+) becoming the dominant share in 2023 and 2024.analyze_data
152
The chart displays four categories: Acquisition, SPAC IPO, IPO, and Buyout across 11 years.analyze_data
153
The slide highlights the trend of big tech companies acquiring AI startups to bring back key talent (acqui-hiring).summarize
155
Mentions 10^26 FLOPs threshold for mandatory safety reporting.establish_context
156
Part of the State of AI 2024 report.summarize
157
Includes a process diagram for high-risk AI system compliance.summarize
158
The slide uses news headlines as evidence to support the narrative of regulatory friction.establish_context
159
The slide highlights the tension between AI model data needs and privacy regulations like GDPR.summarize
160
Part of the State of AI 2024 report.summarize
161
Includes a screenshot of a DeepSeek-V2 chat interaction as a case study example of refusal behavior.summarize
162
Includes a screenshot of a Federal Register document as supporting evidence.establish_context
163
The slide highlights the gap between theoretical performance (e.g., Huawei Ascend 910B) and practical manufacturing reality (e.g., SMIC yield issues).summarize
164
Uses a before-and-after framing to contrast media narratives.illustrate_case
165
Includes an AI-generated illustration of robotic fish.summarize
166
The slide uses a narrative approach to explain geopolitical and financial shifts in AI funding.establish_context
167
Part of the #stateofai 2024 report.summarize
168
The slide uses a Goldman Sachs projection chart to illustrate the growth in power demand.analyze_data
169
Part of the State of AI 2024 report.establish_context
170
The slide discusses the emergence of defense tech startups like Anduril and Helsing, contrasting their successes with the overall small scale of the sector relasummarize
171
Part of the 'State of AI' report series.summarize
172
The slide summarizes viewpoints from Daron Acemoglu, Goldman Sachs, and Noah Smith, and references a UBI trial funded by Sam Altman.summarize
173
The slide uses a screenshot of a social media post as a visual example of bot-related misinformation.summarize
174
Includes a screenshot of a tweet by Leopold Aschenbrenner containing a chart on compute scaling.establish_context
176
Uses a before-after framing to illustrate a cultural shift in the AI industry.summarize
177
Includes a screenshot of a tweet by Matthew Yglesias discussing Sam Altman's views on existential risk.summarize
178
The slide tracks the progression of AI safety summits from Bletchley Park to Seoul and notes the non-binding nature of these agreements.summarize
179
Slide from State of AI 2024 report.summarize
180
Includes a screenshot of the ARIA website regarding their 'Safeguarded AI' programme.summarize
181
Part of the 'state of ai 2024' report.summarize
182
Part of the State of AI 2024 report.summarize
183
The slide uses a specific example of an LLM agent performing SQL injection to illustrate the threat.illustrate_case
184
The slide describes two specific automated red teaming methodologies: Rainbow Teaming and AdvPrompter.present_framework
185
The slide describes a specific technical method (CrossMax) to improve model robustness.illustrate_case
186
Includes a diagram illustrating a 'sleeper agent' attack mechanism.diagnose
187
Includes a technical diagram (Figure 1) illustrating the transformation of model outputs into accuracy metrics.diagnose
188
The slide features a 2x3 grid of scatter plots showing performance vs training loss.analyze_data
189
Includes a concrete example of a ChatGPT-4 interaction demonstrating sycophancy.diagnose
190
The slide uses a series of scatter plots with fitted curves to demonstrate the relationship between KL divergence and win rates, suggesting that DPO is susceptidiagnose
191
The slide discusses the persistence of RLHF in AI training, citing Google DeepMind research and Cohere's RLOO optimization.summarize
192
The slide describes 'Direct Alignment from AI Feedback' (OAIF), a method using an LLM as an annotator to perform online DPO.illustrate_case
193
The slide references specific research methodologies (semantic entropy and Google DeepMind's SAFE) to diagnose and mitigate LLM reliability issues.diagnose
194
Discusses CriticGPT and Cohere's research on LLM-generated critiques.summarize
195
The slide discusses uncertainty calibration in LLMs using fine-tuning techniques like LoRA.diagnose
196
The table displays percentage scores for various AI companies across different transparency criteria, with an average column and row.analyze_data
197
The slide references a specific experiment with 45 occurrences out of 32,768 trials.diagnose
198
The slide showcases a specific technical breakthrough in AI interpretability, using the 'Golden Gate Bridge' feature as a primary example.illustrate_case
199
Includes a screenshot of an SAE viewer tool and a histogram of activation density.summarize
200
The slide highlights three specific research developments: a CMU/Chicago team's work on binary concept variables, Moscow-based AI Research Institute's work on lsummarize
201
Discusses the dual-use nature of interpretability research in AI safety.diagnose
202
The slide highlights the ongoing debate and lack of consensus on whether LLMs significantly increase biological threat risks compared to standard internet accessummarize
203
The slide highlights the dual-use nature of AI tools in biology (e.g., RFDiffusion) and the emerging governance response.diagnose
204
The chart shows 'Impersonation' as the most frequent tactic.frame_problem
206
The slide is part of the 'state of ai' report series.summarize
Open slide detailBeat · The Action (Now What)