slides · global inventory
Paragraph
Component collection for the active corpus filters, with each row opening the slide detail view.
Matching slides 118,471
Current window 59,791-59,820
Pages 3,950
Rows per page 30
Component
59,791-59,820 of 118,471
204
The chart shows 'Impersonation' as the most frequent tactic.2024 Air Street Capital The State of AI Report 2024 · Air Street Capital
Consulting deck VC research Beat · The Implications (So What)Loop · Pattern Hunter1 hit
209
2024 Air Street Capital The State of AI Report 20242024 Air Street Capital The State of AI Report 2024 · Air Street Capital
Consulting deck VC research 1 hit
210
2024 Air Street Capital The State of AI Report 20242024 Air Street Capital The State of AI Report 2024 · Air Street Capital
Consulting deck VC research 1 hit
212
Contains screenshots of the Air Street Press website.2024 Air Street Capital The State of AI Report 2024 · Air Street Capital
Consulting deck VC research 1 hit
02
This is a promotional slide at the beginning of the deck.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research 1 hit
03
2025 Air Street Capital The State of AI Report 20252025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research 1 hit
08
This slide establishes a visual shorthand (iconography) for the rest of the report.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · Surface Observation1 hit
13
The slide highlights the shift towards inference-time scaling (test-time compute) as a key driver for AI reasoning capabilities.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
14
The slide uses a line chart to show scaling laws and a bar chart for comparative performance on AIME 2024.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
15
Includes a bar chart comparing model performance and a line chart showing AIME accuracy improvement during training.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
16
The slide highlights the architectural shift to DeepSeek Sparse Attention (DSA) in V3.2-Exp to reduce cost and latency.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
17
Discusses technical advancements in LLM inference strategies.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
19
Slide compares longitudinal performance trends (left) with a static ranking of current models (right).2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
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 model improvement.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
21
The slide presents a debate regarding the performance of reasoning models (LRMs) versus standard LLMs across different complexity levels.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
22
The slide uses a specific example of a math problem to demonstrate the impact of 'adversarial triggers' on model reasoning.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
23
The slide highlights three specific research papers/projects: GSM-Symbolic, DataAlchemy, and XReasoning.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
24
The slide highlights that CoT utility stems from structural encoding of reasoning rather than literal faithfulness.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
25
References a specific research finding regarding model activations and test awareness.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
26
The slide contrasts the effectiveness of safety-first pretraining (SafeLM) with the broader skepticism regarding the persistence of biases in web-scale models.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
27
Discusses the trade-off between CoT monitoring effectiveness and the emergence of obfuscated reward hacking in AI models.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)Loop · Zoom In1 hit
31
The slide uses a chronological/evolutionary progression of AI training methods.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit
32
2025 Air Street Capital The State of AI Report 20252025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit
33
The slide discusses the debate on whether RLVR creates new reasoning or just reshuffles sampling, citing specific research from Tsinghua and MSR Asia.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit
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.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit
36
The chart shows Bits per Byte performance across different model sizes and training methods (Base, Context, Fine-Tuning).2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit
37
The slide discusses a technical breakthrough in AI model merging (Subspace Boosting) to prevent redundancy.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit
38
The chart shows a Pareto frontier comparison between Muon and AdamW, illustrating better compute-time efficiency for Muon.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit
39
The slide compares memory usage of standard cross-entropy vs. CCE across various LLM models.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit
40
The slide references a specific research finding regarding a ~3.6 bits per parameter capacity limit for LLMs.2025 Air Street Capital The State of AI Report 2025 · Air Street Capital
Consulting deck VC research Beat · The Facts (What)1 hit