Dense Slides: Trends Across the AI Frontier — George Cameron, ArtificialAnalysis.ai

Source Video

Trends Across the AI Frontier — George Cameron, ArtificialAnalysis.ai

Method

This deck is slide-only. The existing captured video frame set supplies candidate frames, then local OpenCV rejects sponsor/title/speaker-only frames, crops visible slide surfaces, deduplicates, and saves the cropped slide images.

Cropped Visible Slides

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.88
Text source: none.
OCR decision: ready — Mixed slide with small labels, screenshots, and dense text; OCR will be more reliable than direct transcription.

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
OCR decision: ready — Dense chart slide with many small labels and bars; OCR is appropriate.

Slide text:

their latest reasoning models, followed closely by other labs

Artificlol Analysis Intolligonce Index (incorporates MMLU-Pra, GPQA, Humanity's Last Exam, LivoCodeBench, SciCoda, AIME, MATH-500) Leading Large Language Model (LLMs), by Al lab

A Artlflctal Analysts

70 02 68 68 67 64 2 9 5 3 53 51 42 41

(high) G x sndd A\ 2358 Reasonin 9 2 AI Llama 8 P M ews.

Sourco: Artiticint Anotysis indopondont benchmarking A Artificlal Analysis

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
OCR decision: ready — Scatterplot/diagram with dense labels and small text; OCR is appropriate.

Slide text:

a helpful framework for understanding today's model landscape Reasoning models: Treating reasoning & non-reasoning models as distinct categories is

Intelligence vs. Output Tokens Used to Run Artificial Analysis Intelligence Index

Artifciel Analysis fnteligence Index (Version 2, released Feb 25),. Output Tokens Used (~5M input tokens)

Most attractive quadrant

Artificlal Analysis Intelligence Index 75 - 55 8- 35 - 45 59 60 40 Prerler Non-Reasoning Models Zora Mistrdl Large GPI-4.1- 2 (Noy 24) [Nor 24) GPT-40 GPT-4.1 mini Grok3 Gamm3 278 DeepSeeck V3:0324 Maverick Lama4 Llama 4 Roasoning Nemotron: Uiua 2538 Lema 3.1. α-mn (high) Reasoning Models DoepSeek R1 Gomini 2.3 Pro A Artlficll Arulysls Grck 3 mint Ressonlng (high) (uuonoog) Gemin 2.5 Clsude 3.7: aupurul Sonnet Fash.

4u Cutput Tokons Used In Artificlal Analysls Inteillgonce Indox (Log Scale) 2014 30H HOO!

A Artlflclal Analysls

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
OCR decision: ready — Dense benchmark chart with many labels and timing values; OCR is appropriate.

Slide text:

making them less suitable for latency sensitive tasks Reasoning model latency: Reasoning models are slower to provide their response,

End-to-End Response Time

Seconds to Output 5o0 Tokens, including reasoning model thinking' time; Lover is better Reasoning models

Input processing time 'Thinking' time (reasoning models) Outputting time NON-EXHAUSTIVE

Roasoning models typicolly take: longer to respond. A Artificiat Anatysis $20.53 106.1

19x! 36.9 42.4 43.2: 60.4 81.

3.3 4.3 4.5 4.7 6.3 7.7 10.9: 11.5 15.5 '12.1...15.6 14.2! 116.41 19.5 12.7 19.8 [20.51 28.6. 38.5 39.8

Llama 8 Scout 8 GPT- 7 b. ayrs G G AV

Sourco: Artificiat Anelyis mdopondant bonchmurking 8 A Artiflclal Analysis

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
OCR decision: ready — Two dense bar charts with many small labels; OCR is appropriate.

Slide text:

Today's open weights frontier is led by China-based Al labs, namely DeepSeek and Alibaba

Reasoning: Open Weights Language Models, by Country Non-reasoning: Open Weights Language Models, by Country

Artificial Analysis lntelligence Index, leading open welghts reasoning models USA Chlna A Artificial Analysis USA Chlna E France canada A Artificial Analysis

68

62 61 60 59

52 51 60 48 R 53 51 47 46 40 40 40 38 M 38

34

DerpSak Gwn3 Llu 3.1 DoipSt Qun32B DoopSert Llmh 33 Phl A1(Mry 235e A228 Nemolron 2024)(Raeonr0 U+ 25B Un4y Reasening '202s$ 1(lssonrg R1DhtllNemoton (resnonirg Qum ire Super 4Be Rertoning Hrto 7B PL. CttpSotk. Lm 708 A1 Dtpta Floth1 Feks V(Hr' 2S] Haverck 21S8 A228 DetaSottLemad Qm3. CUHt8Lma3uCemmnd imtruct 708 1.Tet.01278tnstruc HinMat.Gemma 3 (r2. A0H) Lonto 2 MatM

Source: Artiricidt Annlysis Intellgerce Index A Artlflclal Analysis

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/contrast.
OCR decision: ready — Dense cost comparison chart with many labels and values; OCR is appropriate.

Slide text:

Cost to Run Artificial Analysis Intelligence Index

Cost(USD) to run all evaluations in the Artificial Analysis Intelligence Index (Version 2,released Feb 25, ~5M input tokens) Reasoning models

InputCost ReasoningCost ■Output Cost NON-EXHAUSTIVE

Artificial Analysis

$1951 30X 650X

$1485 $218 $1335

$1786 $1356 S1112 $597 $627 $323 $304 $266 $319 $200 $220 $106 $76 $66 $65 $49 $17 $13 $10 $7 $3

AI G 235B? G Reasoning Nova Premier aws ® GPT-4. ® 7 x (Mar 8 Llama 4Scout 8 nano

3

Sourco:Artificial Analysis indopcndentbcnchmarking. Artificial Analysis

Classification audit: raw/sources/slide-ai-classification/dense/sRpqPgKeXNk/audit.json