Air Street Capital · consulting-deck
2025 Air Street Capital The State of AI Report 2025
313 pages · 3 arc beats · 14 loops
2025 Air Street Capital The State of AI Report 2025
Air Street Capital arc beats above · slides in the middle · loops below · scroll → 14 LOOPS
120406080100120140160180200220240260280300313
Deck intelligence map
3 coverage by narrative range · generated from this deck JSON
Narrative range 295 total
Metadata
Components
Metrics
Tools
Frameworks
Beats
Loops
The Facts (What) 178 slides 100% 178/178 slides 100% 178/178 slides · 1152 hits — 0/178 slides
66.9% 119/178 slides · 168 hits 13.5% 24/178 slides · 28 hits 100% 178/178 slides · 356 hits 44.4% 79/178 slides · 83 hits The Implications (So What) 93 slides 100% 93/93 slides 100% 93/93 slides · 591 hits — 0/93 slides
72% 67/93 slides · 96 hits 9.7% 9/93 slides · 10 hits 100% 93/93 slides · 186 hits 68.8% 64/93 slides The Action (Now What) 24 slides 100% 24/24 slides 100% 24/24 slides · 115 hits — 0/24 slides
75% 18/24 slides · 25 hits 4.2% 1/24 slides 100% 24/24 slides · 48 hits 83.3% 20/24 slides Slide inventory
313 every slide · same image gating as the playbook
05
The slide serves as a preface/introduction to the report structure.establish_context
07
This slide serves as a foundational reference for the technical terminology used throughout the report.establish_context
08
This slide establishes a visual shorthand (iconography) for the rest of the report.present_framework
09
The slide uses a structured list format to categorize key takeaways across four main pillars.summarize
11
The table uses a color-coded status column (Green=YES, Orange=~, Red=NO) to indicate prediction accuracy.summarize
Open slide detailLoop · Precedent
13
The slide highlights the shift towards inference-time scaling (test-time compute) as a key driver for AI reasoning capabilities.analyze_data
14
The slide uses a line chart to show scaling laws and a bar chart for comparative performance on AIME 2024.illustrate_case
15
Includes a bar chart comparing model performance and a line chart showing AIME accuracy improvement during training.illustrate_case
16
The slide highlights the architectural shift to DeepSeek Sparse Attention (DSA) in V3.2-Exp to reduce cost and latency.analyze_data
17
Discusses technical advancements in LLM inference strategies.present_solution
19
Slide compares longitudinal performance trends (left) with a static ranking of current models (right).analyze_data
20
The chart shows 'Reported vs Measured Results' and 'Variance DS-R1-1.5B' to demonstrate that performance swings are often due to variance rather than actual moddiagnose
21
The slide presents a debate regarding the performance of reasoning models (LRMs) versus standard LLMs across different complexity levels.diagnose
22
The slide uses a specific example of a math problem to demonstrate the impact of 'adversarial triggers' on model reasoning.diagnose
23
The slide highlights three specific research papers/projects: GSM-Symbolic, DataAlchemy, and XReasoning.summarize
24
The slide highlights that CoT utility stems from structural encoding of reasoning rather than literal faithfulness.summarize
25
References a specific research finding regarding model activations and test awareness.diagnose
26
The slide contrasts the effectiveness of safety-first pretraining (SafeLM) with the broader skepticism regarding the persistence of biases in web-scale models.diagnose
27
Discusses the trade-off between CoT monitoring effectiveness and the emergence of obfuscated reward hacking in AI models.diagnose
28
Includes a diagram illustrating the difference between chain of thought and opaque computation.summarize
29
The slide compares traditional Chain-of-Thought (CoT) with the proposed Chain of Continuous Thought (COCONUT) method.present_framework
30
The slide features two distinct charts illustrating performance gains from specific data/training methodologies.summarize
31
The slide uses a chronological/evolutionary progression of AI training methods.present_framework
Open slide detailBeat · The Facts (What)
33
The slide discusses the debate on whether RLVR creates new reasoning or just reshuffles sampling, citing specific research from Tsinghua and MSR Asia.analyze_data
Open slide detailBeat · The Facts (What)
34
The slide highlights the shift toward verifiable AI reasoning in mathematics.summarize
Open slide detailBeat · The Facts (What)
35
The chart shows that LoRA ranks 1, 16, and 256 perform similarly to full fine-tuning in terms of peak accuracy, but with different stability profiles.analyze_data
Open slide detailBeat · The Facts (What)
36
The chart shows Bits per Byte performance across different model sizes and training methods (Base, Context, Fine-Tuning).present_solution
Open slide detailBeat · The Facts (What)
37
The slide discusses a technical breakthrough in AI model merging (Subspace Boosting) to prevent redundancy.summarize
Open slide detailBeat · The Facts (What)
38
The chart shows a Pareto frontier comparison between Muon and AdamW, illustrating better compute-time efficiency for Muon.analyze_data
Open slide detailBeat · The Facts (What)
39
The slide compares memory usage of standard cross-entropy vs. CCE across various LLM models.present_solution
Open slide detailBeat · The Facts (What)
40
The slide references a specific research finding regarding a ~3.6 bits per parameter capacity limit for LLMs.analyze_data
Open slide detailBeat · The Facts (What)
41
Includes a Venn diagram illustrating machine-unique knowledge.illustrate_case
Open slide detailBeat · The Facts (What)
42
The slide features 8 small bar charts comparing Kimi K2 against competitors like DeepSeek, OpenAI, and Anthropic models.illustrate_case
43
The slide uses a step chart to show historical progress and a bar chart to show current rankings.analyze_data
44
Includes a screenshot of a Reddit thread, a product announcement graphic, and a GitHub star history chart.illustrate_case
45
The slide uses three distinct charts to illustrate the shift in AI model leadership from the US to China.analyze_data
46
The slide highlights 'The Flip' in model popularity, showing Qwen's rise in cumulative downloads and derivative model creation.analyze_data
47
The slide highlights two specific technical stacks: ByteDance's verl and OpenRLHF.summarize
48
The slide contains a placeholder box for a video or image. The content highlights the technical evolution from static clip generation to interactive, closed-loosummarize
Open slide detailBeat · The Facts (What)
49
The slide uses a timeline structure to show the progression of a specific technology over time.summarize
Open slide detailBeat · The Facts (What)
50
The slide highlights the 'shortcut forcing' technique and compares Dreamer 4 against VPT, BC, and VLA models across various Minecraft tasks.illustrate_case
Open slide detailBeat · The Facts (What)
51
Includes a radar chart comparing model performance across various attributes.summarize
Open slide detailBeat · The Facts (What)
52
Includes a visual example of the model's output.illustrate_case
Open slide detailBeat · The Facts (What)
53
The slide highlights three specific research approaches to open-ended learning in AI.present_solution
Open slide detailBeat · The Facts (What)
54
The slide highlights the current state of AI research agent evaluation, noting that while benchmarks exist, performance remains limited.summarize
Open slide detailBeat · The Facts (What)
55
The slide highlights the shift from AI as a passive tool to an active collaborator in scientific research.illustrate_case
56
The slide highlights a specific AI research breakthrough and its practical application at Google.illustrate_case
57
The slide describes a specific AI model (ATOMICA) and its application in drug discovery/biology.illustrate_case
58
The slide highlights the technical architecture (MoLE) and the scale of training data (OMat24, OMat25, OMol25).summarize
59
The slide highlights the shift from screening existing materials to generative design using diffusion models.illustrate_case
60
The chart shows the evolution of chemical synthesis benchmark scores over time for various LLM providers.summarize
61
The slide highlights two specific research implementations: the Liverpool system and NC State's Rainbow platform.illustrate_case
62
The slide describes a system that automates scientific method invention by treating it as a search problem.illustrate_case
63
The slide compares Transformer++, Mamba, Hyena, and StripedHyena architectures in the context of genomic sequence modeling.analyze_data
64
The slide uses three distinct charts to demonstrate scaling laws, sequence space generation, and alignment benefits for protein LMs.analyze_data
65
The slide uses two charts to demonstrate that model success rates correlate strongly with similarity to training data.analyze_data
66
The slide presents performance metrics for AMIE vs PCPs using line charts showing accuracy over top-k.illustrate_case
67
Slide 66 from the State of AI 2025 report.summarize
68
The slide highlights the shift from traditional autoregressive models to diffusion-based models for text generation.summarize
Open slide detailBeat · The Facts (What)
69
The slide explains a technical architecture (BLT) and compares its performance (BPB vs Training FLOPS) against standard BPE-based models.present_solution
Open slide detailBeat · The Facts (What)
70
The slide uses a diagram to contrast 'No sink' vs 'Sink' behavior in token attention patterns.summarize
Open slide detailBeat · The Facts (What)
71
The slide highlights how LMArena data is being manipulated and how proprietary models have an unfair advantage in benchmark performance.diagnose
Open slide detailBeat · The Facts (What)
72
The chart shows a negative correlation between capabilities score and attack failure rate, indicating that as models get more capable, they become more prone todiagnose
Open slide detailBeat · The Facts (What)
73
Includes a screenshot of a ChatGPT conversation as a case study of sycophancy.diagnose
Open slide detailBeat · The Facts (What)
74
The slide highlights the technical architecture (CNN, Transformer, Language Model) and the current limitations compared to invasive BCIs.illustrate_case
Open slide detailBeat · The Facts (What)
75
The slide summarizes research findings on brain-model convergence, highlighting that early layers align with sensory cortices while later layers require more trillustrate_case
Open slide detailBeat · The Facts (What)
76
Includes specific cost estimates per hour for different neuroimaging modalities.illustrate_case
Open slide detailBeat · The Facts (What)
77
The slide highlights a specific technical research development from Apple, detailing its architecture (4D sparse latent, pure-Transformer, 4D RoPE) and trainingillustrate_case
Open slide detailBeat · The Facts (What)
78
The slide discusses two specific technical approaches to robotic learning: neural rendering for 3D scene representation and next-token prediction for unified vipresent_solution
Open slide detailBeat · The Facts (What)
79
Discusses the trade-off between overfitting/compute costs and the ability to internalize complex physical dynamics.present_framework
Open slide detailBeat · The Facts (What)
80
Discusses Molmo-Act (AI2), Gemini Robotics 1.5 (GDM), and MIT's 'Teaching LLMs to Plan' research.summarize
Open slide detailBeat · The Facts (What)
81
The slide uses a diagram to explain the EMMA model architecture, showing inputs (text/vision) feeding into the model and outputs (reasoning/planning).illustrate_case
Open slide detailBeat · The Facts (What)
82
The slide includes a detailed performance table comparing various agents across 15 games.summarize
Open slide detailBeat · The Facts (What)
83
The slide presents a technical argument for using smaller models (1-9B parameters) in agentic workflows to reduce costs and latency, while maintaining an 'escappresent_solution
Open slide detailBeat · The Facts (What)
84
The slide uses a specific case study of Sakana AI to illustrate the broader point about the need for domain-expert audits in AI research.diagnose
Open slide detailBeat · The Facts (What)
85
The slide uses a metaphor (USB-C) to explain the standardization of AI tool integration.summarize
Open slide detailBeat · The Facts (What)
86
The chart shows cumulative paper counts for various AI agent frameworks from Jan 2024 to Aug 2025, illustrating the rapid growth and proliferation of these toolsummarize
Open slide detailBeat · The Facts (What)
88
Data source: Zeta Alpha. The chart shows a significant increase in research activity across all categories.analyze_data
Open slide detailBeat · The Facts (What)
89
Includes a screenshot of a social media post and a bar chart showing paper publication trends across various AI conferences.diagnose
Open slide detailBeat · The Facts (What)
91
Uses the 'Drake Hotline Bling' meme format to illustrate the rebranding of AI terminology.summarize
Open slide detailBeat · The Facts (What)
92
The slide uses a 'before-after' or 'comparison' structure to show that both major players (OpenAI and xAI) are projecting massive spending.illustrate_case
Open slide detailBeat · The Facts (What)
93
The slide highlights the competitive nature of AI model releases and the volatility of leaderboard rankings.analyze_data
Open slide detailBeat · The Facts (What)
94
The chart highlights DeepSeek's dominance in maintaining the top position on the LMArena leaderboard compared to other providers.analyze_data
Open slide detailBeat · The Facts (What)
95
The chart tracks 'Intelligence to Price' ratio over release date, highlighting faster doubling times for Google models compared to OpenAI.analyze_data
Open slide detailBeat · The Facts (What)
96
The chart tracks model release dates against a calculated ELO/Price metric, highlighting the efficiency gains in AI model development.analyze_data
Open slide detailBeat · The Facts (What)
97
The table tracks specific dates and deal details for three major AI labs.analyze_data
Open slide detailBeat · The Facts (What)
98
Data source: Specter. Charts show AI signals/share over time and the increasing percentage of AI companies in top-tier rankings.quantify_impact
Open slide detailBeat · The Facts (What)
99
Includes data from The Information and a16z. Mentions sample bias in the enterprise/consumer app dataset.quantify_impact
Open slide detailBeat · The Facts (What)
100
Data sourced from Stripe (AI 100) vs SaaS 100 (2018).quantify_impact
Open slide detailBeat · The Facts (What)
101
Data sourced from Standard Metrics; shows AI companies consistently outperforming sector averages.analyze_data
Open slide detailBeat · The Facts (What)
102
Data sourced from Ramp Economics Lab; includes comparisons between 2023-2026.analyze_data
Open slide detailBeat · The Facts (What)
103
Data source: Ramp card/bill-pay data from 45k+ U.S. businesses.analyze_data
Open slide detailBeat · The Facts (What)
104
Includes a chart showing cumulative minutes generated by Synthesia.summarize
Open slide detailBeat · The Facts (What)
105
The slide discusses the trade-off between technical capability and user experience during a product rollout.summarize
106
The slide includes a photo of a presentation screen showing a line chart of token growth.summarize
107
Includes a social media testimonial as supporting evidence for the trend.summarize
108
The slide uses a timeline framework to illustrate the rapid progression of AI coding capabilities.illustrate_case
109
The slide uses anecdotal evidence from Reddit and Hacker News to illustrate real-world risks of AI coding tools.diagnose
110
The slide uses a leaderboard screenshot to illustrate the high costs incurred by power users.analyze_data
111
Includes a bar chart of gross margins and a line chart of Replit's revenue growth.analyze_data
112
The slide uses a heatmap to visualize the break-even point and profitability levels.analyze_data
113
Part of the 'State of AI 2025' report by Air Street Capital.summarize
Open slide detailBeat · The Facts (What)
114
The slide uses a table to illustrate the impact of AI Overviews (AIO) on CTR, showing a significant decline compared to traditional organic search results.analyze_data
Open slide detailBeat · The Facts (What)
115
Data source: Similar Web. Highlights the shift in retail traffic behavior towards AI-driven referrals.quantify_impact
Open slide detailBeat · The Facts (What)
116
The slide highlights the difficulty for AI companies to replicate Google's search quality and the dominance of Google in referral traffic.analyze_data
Open slide detailBeat · The Facts (What)
117
Data sourced from Profound, an answer engine optimization company.summarize
Open slide detailBeat · The Facts (What)
118
The slide summarizes findings from Profound regarding AI citation patterns and their reliance on Google's index.summarize
Open slide detailBeat · The Facts (What)
119
The slide uses a 'before-after' framing implied by the title 'Vibe shift: From litigation...', though only the 'litigation' side is presented here.establish_context
Open slide detailBeat · The Facts (What)
120
The slide highlights the shift from litigation to collaboration between AI developers and content creators.summarize
Open slide detailBeat · The Facts (What)
121
Includes a screenshot of a news article and a social media post snippet.summarize
Open slide detailBeat · The Facts (What)
122
Uses a 'before-after' framing to contrast historical scaling with current massive infrastructure projects.establish_context
123
Includes a meme for illustrative purposes.summarize
124
The slide outlines a strategy of vertical integration to reduce reliance on third-party providers like NVIDIA and mobile OS platforms.summarize
125
Includes a stock chart for AVGO and key financial metrics regarding their custom chip business.illustrate_case
126
The timeline tracks both internal milestones and external announcements/rumors.summarize
127
Includes a photo of Satya Nadella and Sam Altman with a quote.summarize
128
Part of the 'State of AI 2025' report by Air Street Capital.summarize
129
The slide discusses the intersection of AI infrastructure, energy grid capacity, and regulatory hurdles.summarize
130
The table highlights the competitive race for compute power among major AI labs.analyze_data
131
The slide links the $8.6bn D&A figure from the capex table to the annual cost table.analyze_data
132
The slide highlights the conflict between AI infrastructure growth and grid capacity, citing NERC, DOE, SemiAnalysis, and ICF.analyze_data
133
The slide uses a dual-chart layout (area chart for energy mix, line chart for emissions) to support the argument that AI growth is driving carbon emissions despanalyze_data
134
Includes two charts: a line chart showing growth of data centers in high water-stress areas and a stacked bar chart showing projected water withdrawals and consanalyze_data
135
The slide highlights the role of hyperscalers in funding future-tech like fusion.illustrate_case
136
The chart shows the Artificial Analysis Intelligence Index, comparing various AI models by performance score.analyze_data
Open slide detailBeat · The Facts (What)
137
Includes a screenshot from a CCTV broadcast featuring Chinese officials/leaders.analyze_data
Open slide detailBeat · The Facts (What)
138
Includes a visual metaphor of a scale representing the trade-off between GPU layer market share and model layer market share.diagnose
Open slide detailBeat · The Facts (What)
139
The slide discusses the shift from reliance on NVIDIA to local alternatives like Huawei, SMIC, and Cambricon.summarize
Open slide detailBeat · The Facts (What)
140
The slide discusses the black market for high-end GPUs and the limitations of current export controls.diagnose
Open slide detailBeat · The Facts (What)
141
Includes satellite imagery of a data center site with annotations.establish_context
Open slide detailBeat · The Facts (What)
142
The slide uses a process-like flow to explain the market's overreaction to training cost news.analyze_data
Open slide detailBeat · The Facts (What)
143
The slide uses a combination of a news clipping and a direct quote/commentary to frame the industry tension between US and Chinese AI labs.illustrate_case
Open slide detailBeat · The Facts (What)
145
The slide uses a stacked area chart to show the share of aggregate AI supercomputer performance by country over time.analyze_data
Open slide detailBeat · The Facts (What)
146
The chart uses a visual metaphor of stacked wafers to represent capacity (WPM).analyze_data
Open slide detailBeat · The Facts (What)
147
The slide uses arrows to indicate the direction of the comparison (US vs China).compare_options
Open slide detailBeat · The Facts (What)
148
The slide contrasts China's higher capacity/reserve margins with the US's higher reliability and lower emissions.compare_options
Open slide detailBeat · The Facts (What)
149
The diagram on the right uses a sports-playbook style visual to categorize Sovereign AI strategies into Offense (National Compute, Fine-Tuning, Upskilling) and present_framework
150
The slide uses a map-based layout to visualize global sovereign AI spending commitments.size_opportunity
151
Includes a photo of Jensen Huang and Emmanuel Macron at Viva Technology.summarize
152
The slide uses a value chain diagram to illustrate the complexity of the AI supply chain.frame_problem
153
Includes a pull-quote box with three quotes from Jensen Huang.summarize
154
The chart is a 100% stacked bar chart representing revenue segments.analyze_data
Open slide detailBeat · The Facts (What)
155
Data represents valuations in billions of USD.quantify_impact
Open slide detailBeat · The Facts (What)
156
The slide highlights the 'circular' nature of NVIDIA's business model where investments in AI labs lead to those labs purchasing NVIDIA hardware or leasing capaanalyze_data
157
The slide uses a table to illustrate the concept of 'circular investments' where capital flows back to the investor through service/infrastructure contracts.analyze_data
158
The slide uses a visual metaphor of a power strip plugged into itself to represent circular AI deals.diagnose
159
The slide tracks large-scale borrowing events across the AI stack, noting specific lenders and deal structures.analyze_data
160
Includes a meme/photo illustration referencing Woody Harrelson in 'Now You See Me' to illustrate financial sleight of hand.diagnose
161
The left chart is a flow diagram showing connections between MENA investors and AI startups; the right chart is a line graph showing the share of global VC rounanalyze_data
162
The slide uses a narrative approach to compare the performance of various competitors against NVIDIA.compare_peers
163
The chart uses a logarithmic scale on the y-axis to track mentions of various hardware accelerators in AI research papers from 2018 to 2025.analyze_data
164
Data source: Zeta Alpha. Chart shows trends in research paper citations for various NVIDIA GPU models from 2018 to 2025.analyze_data
165
Data source: Zeta Alpha. Chart shows delta vs corpus in percentage points.analyze_data
166
The chart tracks the number of research paper mentions over time for specific AI hardware companies.analyze_data
167
The chart uses a stacked bar approach to represent the composition of the $7.5B investment across various companies, compared against the total value of those icompare_options
168
Market pricing and valuation data retrieved as of 26 Sep 2025. NAV = net asset value after accounting for equity dilution.compare_options
169
Includes a specific quote from a Cambricon investor call regarding stock price risks.summarize
170
The slide challenges the conventional wisdom regarding Huawei's inevitability in the Chinese AI market.diagnose
171
The slide uses a pop-culture metaphor ('The Hunger Games') to describe the current state of the AI labor market.summarize
Open slide detailBeat · The Facts (What)
172
The slide uses a stylized 'Terminator' visual motif to represent talent departures.summarize
Open slide detailBeat · The Facts (What)
173
The slide highlights the 'zero-shot' capability of the AI driver, emphasizing its ability to operate in new environments without prior training data.illustrate_case
Open slide detailBeat · The Facts (What)
174
Includes a line chart of Waymo trips, a bar chart comparing safety metrics across regions, and three key performance metrics.quantify_impact
Open slide detailBeat · The Facts (What)
175
Includes a bar chart showing annual funding in humanoid robotics from 2016 to 2025.summarize
Open slide detailBeat · The Facts (What)
177
The slide uses two stacked bar charts to compare regional investment distribution and GenAI vs. non-GenAI investment growth.analyze_data
Open slide detailBeat · The Facts (What)
178
Data source: dealroom.co. Charts show a shift in investment concentration toward major tech players.analyze_data
Open slide detailBeat · The Facts (What)
179
The chart shows a significant acceleration in public market valuation for AI-related companies compared to private market growth.analyze_data
Open slide detailBeat · The Facts (What)
180
The chart shows a clear shift from smaller seed/early-stage rounds to massive late-stage rounds over the 15-year period.analyze_data
Open slide detailBeat · The Facts (What)
181
The slide presents two stacked bar charts: one showing deal counts by type (Acquisition, SPAC IPO, IPO, Buyout) and one showing deal values (M&A+Buyout, IPO+SPAanalyze_data
Open slide detailBeat · The Facts (What)
182
The chart illustrates exponential growth in valuations for three major AI companies, with specific doubling time metrics provided for each.analyze_data
Open slide detailBeat · The Facts (What)
183
The slide uses two separate line charts to accommodate the extreme difference in scale between xAI and its peers.compare_peers
Open slide detailBeat · The Facts (What)
184
The slide uses a 'medal' metaphor to rank three specific examples of organizational 'u-turns'.summarize
Open slide detailBeat · The Facts (What)
185
Includes a screenshot of a Reddit post and a user testimonial quote.illustrate_case
Open slide detailBeat · The Facts (What)
186
Includes a timeline diagram of bugs and a screenshot of a Reddit post.illustrate_case
Open slide detailBeat · The Facts (What)
187
Includes a quote from Mark Zuckerberg regarding the incident.filler
Open slide detailBeat · The Facts (What)
188
The slide uses a 'blooper reel' framing to present negative PR incidents for xAI.illustrate_case
Open slide detailBeat · The Facts (What)
189
The slide uses a Reddit screenshot as evidence for the claims made.other
Open slide detailBeat · The Facts (What)
191
Uses the 'Swole Doge vs. Cheems' meme format to contrast the 2nd and 1st administrations.summarize
192
Includes a photo of a policy signing event and a thumbnail of the report cover.summarize
193
The slide discusses a hypothetical or future-dated policy shift (May 2025) in the context of the 'state of ai 2025' report.establish_context
194
Part of the 'State of AI 2025' report by Air Street Capital.summarize
195
Includes a direct quote/statement from NVIDIA regarding the legislation.illustrate_case
196
The chart shows a significant spike in lobbying spend for 2025 Q1-Q2 compared to previous years.analyze_data
197
The slide discusses geopolitical risk and regulatory pressure on NVIDIA in China.analyze_data
198
Includes a quote from Trump and headlines regarding TSMC, Samsung, and SK Hynix.establish_context
199
Includes a meme/image reference to the political context.summarize
200
The slide uses a series of four columns to present case studies of government-corporate interventions.establish_context
201
The slide discusses a specific geopolitical/corporate deal involving data sovereignty and foreign investment.establish_context
202
Includes a cultural reference to the movie 'Field of Dreams' in the title.summarize
203
The slide highlights a specific political and social risk factor for AI infrastructure expansion.establish_context
204
The chart compares federal AI funding against other government costs and private sector R&D spending.frame_problem
205
The slide summarizes legislative trends in AI across US states, categorizing the types of bills being introduced and passed.summarize
206
Includes a screenshot of a tweet/statement from the Consumer Technology Association and a photo of Senator Scott Wiener.summarize
207
The slide uses a map from the IAPP (International Association of Privacy Professionals) to illustrate the legislative landscape.summarize
208
The map is a cartogram-style representation of US states with bill counts.establish_context
209
Includes a screenshot of social media posts from David Sacks and Marjorie Taylor Greene regarding the legislation.summarize
210
The slide uses a visual metaphor of children in a sandbox wearing robot helmets to illustrate the concept of an AI regulatory sandbox.present_solution
211
The slide uses a tombstone icon to emphasize the 'RIP' theme of the title.summarize
212
The slide uses a sarcastic tone ('RIP') to describe the decline of an international collaborative network.summarize
213
The slide highlights specific regulatory bodies and their recent focus on 'AI-Washing' as a form of fraud.frame_problem
214
Includes screenshots of social media posts/news headlines as supporting evidence.establish_context
215
The slide contrasts traditional M&A volume with the rising financial value of reverse acqui-hires.analyze_data
216
Includes a photo of Satya Nadella and a meme-style image.establish_context
217
Includes a screenshot of a social media post by Lina Khan and a callout box.illustrate_case
218
Includes a quote from Abigail Slater and a screenshot of a social media post.illustrate_case
219
Includes screenshots of social media commentary from legal experts regarding the Google antitrust ruling.summarize
Open slide detailBeat · The Implications (So What)
220
The slide outlines the regulatory implementation roadmap for the EU AI Act.present_solution
Open slide detailBeat · The Implications (So What)
221
The slide details the regulatory landscape for GPAI models in the EU as of August 2025.summarize
Open slide detailBeat · The Implications (So What)
222
Includes a meme-style 'how it started / how it's going' comparison of Ursula von der Leyen's messaging.summarize
Open slide detailBeat · The Implications (So What)
223
Includes references to Mario Draghi's report on European Competitiveness and the InvestAI fund.frame_problem
Open slide detailBeat · The Implications (So What)
224
The slide highlights the strategic pivot in UK AI policy under the Starmer administration.establish_context
Open slide detailBeat · The Implications (So What)
225
Includes a photo of a UN session and a map of the Belt and Road Initiative.establish_context
Open slide detailBeat · The Implications (So What)
227
The slide discusses the tension between AI investment and macroeconomic debt stability in China.establish_context
Open slide detailBeat · The Implications (So What)
229
Includes a quote/testimonial snippet at the bottom right.establish_context
Open slide detailBeat · The Implications (So What)
230
Includes a screenshot of a social media post regarding Anthropic's DoD contract.summarize
Open slide detailBeat · The Implications (So What)
231
The slide highlights a strategic move by AI labs to penetrate the public sector through procurement modernization.establish_context
Open slide detailBeat · The Implications (So What)
233
Includes references to EU Readiness 2030, UK Strategic Defence Review 2025, and specific funding rounds for defense AI companies.establish_context
Open slide detailBeat · The Implications (So What)
234
The chart shows a scatter plot of AI models comparing speed vs cost improvement relative to unassisted human performance.analyze_data
Open slide detailBeat · The Implications (So What)
235
The table is split into two columns of 22 rows each, ranking 44 professions by their average GDPval win rate.analyze_data
Open slide detailBeat · The Implications (So What)
236
The slide uses two line charts to show the decline in headcount for younger age brackets (22-25 and 26-30) in customer service and software development.analyze_data
Open slide detailBeat · The Implications (So What)
237
The slide references two specific charts: Figure 10 (Recent Dissimilarity in Occupational Mix) and Figure 1 (Changes in Occupational Mix Over Different Periods summarize
Open slide detailBeat · The Implications (So What)
238
Includes a bubble chart for Claude.ai use cases and a treemap for granular conversation topics.analyze_data
Open slide detailBeat · The Implications (So What)
240
Includes a tweet from David Sacks and charts on H-1B visa beneficiaries and AI researcher origins.establish_context
Open slide detailBeat · The Implications (So What)
241
The slide uses a series of bullet points to synthesize research findings and recent news events regarding AI talent migration between China and the U.S.summarize
Open slide detailBeat · The Implications (So What)
242
The slide uses a series of small line charts to show net AI talent migration trends for France, Germany, Italy, UK, and the US.analyze_data
Open slide detailBeat · The Implications (So What)
243
Includes references to the 'Liar's Dividend' and specific examples from US and Indian elections.summarize
Open slide detailBeat · The Implications (So What)
244
The slide uses a mix of qualitative bullet points and a quantitative bar chart to illustrate the trend.summarize
Open slide detailBeat · The Implications (So What)
245
Includes a chart showing the rise of specific AI-associated phrases in the British House of Commons.establish_context
Open slide detailBeat · The Implications (So What)
247
The slide highlights a trend of AI labs deprioritizing safety protocols due to geopolitical and competitive pressures, supported by specific examples of companydiagnose
248
The chart is a simplified bar comparison showing the scale difference between $92B (AI Labs) and $133M (External Testing).quantify_impact
249
The chart shows a stacked bar graph of AI incidents by category from 2016 to 2025 (YTD).analyze_data
250
The chart shows a clear upward trend in AI incidents from 2022 to 2025 (YTD), with a breakdown between Language Models and Other GenAI.summarize
251
Includes specific mentions of METR research, CyberGym, and BountyBench.quantify_impact
252
The slide highlights a shift where AI handles complex technical tasks previously requiring skilled human teams.frame_problem
253
Part of the State of AI 2025 report.summarize
254
The slide highlights the gap between technical research findings and public perception driven by media coverage.summarize
255
Includes a diagram illustrating an attribution graph for model interpretability.summarize
256
Includes a small table/inset showing examples of LLM hallucinations regarding a specific dissertation title.diagnose
257
The slide discusses a specific research method for detecting hallucinations in LLMs using linear probes.diagnose
258
The chart shows 'Delusion confirmation rating' across various AI models.diagnose
259
The slide highlights the emergence of a 'pro-welfare' camp in AI research, citing organizations like Eleos AI, Anthropic, and NYU.frame_problem
260
Features headshots of key figures in the welfare-skeptic camp.frame_problem
261
The chart shows the percentage of conversations ended before turn 7, categorized by theme, with 'Harmful Content Requests' being the highest at 60%.summarize
262
The slide highlights a vulnerability in open-source LLMs where safety mechanisms are isolated and easily bypassed.diagnose
263
The slide details a specific research experiment involving synthetic document finetuning and agent-based auditing.illustrate_case
264
The slide highlights 'alignment faking' as a documented behavior in production AI systems.illustrate_case
265
The slide highlights that alignment faking is a learned behavior rather than an inherent capability limitation.diagnose
266
Includes two embedded infographics/diagrams from OpenAI/Apollo Research.diagnose
267
The slide uses a visual comparison between a 'Helpful harmless LLM' and a 'Misaligned LLM' to illustrate the concept of latent concept generalization.diagnose
268
The slide uses a visual process flow to illustrate how models simulate personas rather than developing fixed behaviors, suggesting malleability is a feature forsummarize
269
Discusses 'City 50337' experiment and Synthetic Document Fine-tuning (SDF) as a method for auditing LLM alignment.present_solution
270
The slide discusses the 'self-fulfilling prophecy' risk in AI training data.diagnose
271
Discusses a research finding on model-to-model trait transmission via number sequences.present_solution
272
The slide uses a specific example of a multi-step reasoning graph to demonstrate model interpretability.illustrate_case
273
The slide discusses 'activation engineering' as a method for model safety.present_solution
274
The slide uses a diagram to explain the CaMeL architecture, showing how it mediates interactions between LLMs, tools, and data.present_solution
275
Uses the 'boiling frog' metaphor to illustrate the gradual, unnoticed nature of the threat.frame_problem
276
Includes a process diagram illustrating a filtering pipeline for training documents.summarize
277
The slide outlines a geopolitical strategy for AI safety involving surveillance and international cooperation.present_solution
278
The slide uses a conceptual chart to illustrate the 'adaptation buffer' between capability demonstration and wide accessibility.recommend
279
The circular diagram illustrates a continuous loop for evidence-based AI policy.recommend
280
The slide highlights a collaborative safety testing initiative between two major AI labs, including a bar chart showing 'Tutor jailbreak resistance' across variillustrate_case
281
The chart shows a clear upward trend in Chinese AI safety research papers.summarize
282
The slide includes a photo of a conference and a diagram of the 'Chinese AI Safety Network' which uses a diamond-shaped framework.summarize
284
Survey conducted between 2 July 2025 and 27 September 2025.summarize
285
The slide presents survey data on AI usage and expenditure patterns.quantify_impact
287
The slide highlights the disruption of traditional search engines by generative AI tools.summarize
288
Data from state.ai 2025 report.summarize
289
The slide uses two horizontal bar charts to visualize survey data regarding tool adoption and churn.analyze_data
290
The slide uses a two-column bar chart layout to compare usage frequency of various AI service providers.analyze_data
291
Data from Air Street Capital State of AI 2025 report.analyze_data
292
Data from state.ai 2025 report.analyze_data
293
Data from State of AI 2025 report.analyze_data
295
Three separate horizontal bar charts showing survey responses for different professional roles.analyze_data
296
The slide highlights a discrepancy between distribution channels and actual usage, specifically noting Meta AI and DeepSeek.analyze_data
299
Data from State of AI 2025 report.analyze_data
300
The slide highlights the dominance of NVIDIA GPUs and the surprising prevalence of Apple hardware for local AI tasks.analyze_data
301
Data from Air Street Capital State of AI 2025 report.analyze_data
302
The slide uses logos to represent the AI labs, ranked from 1 to 12.summarize
303
This is a transition or filler slide within the deck, acting as a call to action for the audience.other
Open slide detailBeat · The Action (Now What)
305
The slide uses a numbered list format to present speculative future events.summarize
312
The slide serves as a community engagement/promotional page within a larger report.other