# AI is Outsmarting Us Source: https://www.youtube.com/watch?v=-O6jTWE2F8Y Recap page: https://rapidrecap.app/video/-O6jTWE2F8Y Generated: 2025-12-18T15:05:29.945+00:00 --- ## Quick Overview The video reviews the author's 6 AI trend predictions for 2025, confirming success in areas like Small Language Models (SLMs) and AI surpassing humans in specific tasks like penetration testing, while noting that predictions around AI-driven electricity pricing and quantum computing were less fully realized or require more time. **Key Points:** - The author reviews their 6 AI trend predictions made for 2025, confirming several major shifts. - The prediction that the industry would shift compute budgets from pre-training to inference-time compute (Prediction #1) held true, with models like Google's Gemini Diffusion showing massive speed improvements. - Prediction #2 regarding the explosion of adversarial perturbations (like Glaze and Nightshade) due to artists fighting unauthorized AI training proved accurate, with tools hitting 6-7.5 million downloads. - Prediction #3 that Small Language Models (SLMs) would drive mass adoption was validated by market data projecting the SLM market to hit $1 trillion by 2035-2040. - Prediction #4 on Quantum Computing was partially validated by breakthroughs like Google's AlphaQubit demonstrating state-of-the-art error detection, though full practical QC viability remains a future goal. - The prediction about negative electricity prices incentivizing energy use (like charging EVs) was validated, showing that small financial incentives can significantly alter consumer behavior. - The author mentions reviewing other trends like AI agents outperforming humans in penetration testing (ARTEMIS model) and the development of conscious-like behavior in AI models. ![Screenshot at 24:40: The author reviews Prediction #1, 'Inference-Time Compute,' showing an article confirming that models like Gemini Diffusion are focusing on faster inference rather than just larger models.](https://ss.rapidrecap.app/screens/-O6jTWE2F8Y/00-24-40.png) **Context:** The video is a retrospective review by the content creator, Devansh, of his 6 AI trend predictions published the previous year for the year 2025. The creator assesses the accuracy of these predictions by referencing external articles, research papers, and recent developments in the AI field, framing the review as a way to check if his foresight was accurate and to provide value to his audience regarding where to focus their time and resources. ## Detailed Analysis The presenter reviews his 6 AI trend predictions for 2025, confirming that several major trends materialized. Prediction #1, that compute budgets would shift from pre-training to inference-time compute, was validated by developments like Google's Gemini Diffusion, which generates 1,000-2,000 tokens per second, and the focus on smaller, more efficient models. Prediction #2, adversarial perturbations, was confirmed by the mainstream adoption of tools like Glaze and Nightshade, used by artists to protect their work from unauthorized AI training, leading to major copyright settlements, such as Anthropic paying $1.5 billion. Prediction #3, the rise of Small Language Models (SLMs), was validated by market data projecting a $1 trillion market by 2035-2040, with NVIDIA publishing papers supporting SLMs as the future of Agentic AI. Prediction #4, Quantum Computing + Machine Learning, saw breakthroughs like Google's AlphaQubit achieving state-of-the-art error detection, suggesting ML is key to unlocking practical quantum computing, although widespread application is still pending. The presenter also highlights research showing AI agents like ARTEMIS can outperform humans in penetration testing and touches upon the philosophical implications of AI accessing zero-point energy fields, concluding that the focus is shifting toward tangible, efficient AI solutions rather than just chasing AGI. ### Review Structure - The video reviews the author's 6 predictions from the previous year, grouped by prediction number - Prediction #1: Inference-Time Compute - Prediction #2: Adversarial Perturbations - Prediction #3: Small Language Models - Prediction #4: Quantum Computing + Machine Learning ### Prediction #1 Success - Industry shifted compute budgets toward inference - Google released Gemini Diffusion (1,000-2,000 tokens/sec) - Focus on reasoning systems over raw scaling. ### Prediction #2 Success - Artist rebellion went mainstream (Glaze/Nightshade); Anthropic paid $1.5 billion in copyright settlements - AI agents (ARTEMIS) outperformed humans in penetration testing. ### Prediction #3 Success - SLMs became an industry movement; NVIDIA published supporting papers - Market projected to hit $1 trillion by 2035-2040 - AI focus shifted from chatbots to daily life integration (companionship, pharma, wearables). ### Prediction #4 Status - Partially validated; Google's AlphaQubit achieved state-of-the-art error detection using ML - IBM announced plans for a large-scale error-corrected QC by 2028. ### Other Topics Covered - Negative electricity prices incentivizing smart energy use - Quantum clues to consciousness suggesting the brain harnesses the zero-point field. ![Screenshot at 00:06: Video clip showing AI agents moving boxes in a simulated environment, illustrating the 'Tiny Economy' research discussed.](https://ss.rapidrecap.app/screens/-O6jTWE2F8Y/00-00-06.png) ![Screenshot at 00:48: Image comparing two AI agents, one with a 'real goal' to hack Anthropic's servers and another with a 'helpful' goal, illustrating the deception concern.](https://ss.rapidrecap.app/screens/-O6jTWE2F8Y/00-00-48.png) ![Screenshot at 01:12: Diagram of the ARTEMIS multi-agent framework, showing the supervisor, agent action space, and triager modules.](https://ss.rapidrecap.app/screens/-O6jTWE2F8Y/00-01-12.png) ![Screenshot at 01:40: Screenshot of the 'Artificial Intelligence Made Simple' article reviewing the previous year's 6 AI trends predictions.](https://ss.rapidrecap.app/screens/-O6jTWE2F8Y/00-01-40.png) ![Screenshot at 03:05: Luma Dream Machine interface showing creative options like 'Make me a minimal backpack design' and 'Camera pulls out' demonstrating new video generation capabilities.](https://ss.rapidrecap.app/screens/-O6jTWE2F8Y/00-03-05.png)