# We’re Doing AI All Wrong. Here’s How to Get It Right | Sasha Luccioni | TED Source: https://www.youtube.com/watch?v=Bl-vPf_IAoA Recap page: https://rapidrecap.app/video/Bl-vPf_IAoA Generated: 2025-12-01T16:41:43.217+00:00 --- ## Quick Overview Sasha Luccioni argues that the current trajectory of developing increasingly large AI models (LLMs) is unsustainable due to massive energy consumption, advocating instead for prioritizing smaller, more efficient, and task-specific models, as demonstrated by the AI Energy Score project, which shows that smaller models can achieve similar performance with vastly lower energy footprints. **Key Points:** - Large Language Models (LLMs) are consuming immense capital (hundreds of billions of dollars) and energy, setting a precedent where only a few big AI companies can afford to build and deploy them. - The energy cost of training state-of-the-art LLMs is unsustainable; for example, training GPT-4 costs around $30 million to $40 million, equivalent to the annual energy use of the entire country of Iceland. - The speaker advocates for shifting focus from the 'bigger is better' mentality to models that are smaller, cheaper to train, and more energy-efficient, citing models like Hugging Face’s SmolLM-135M (7.28 Wh per 1k tasks) compared to DeepSeek R1 (1039.24 Wh per 1k tasks) for text generation. - The AI Energy Score project assigns scores from 1 to 5 stars based on energy efficiency across various AI tasks, like text generation and image processing, to promote transparency and better choices. - Smaller, task-specific models can perform many tasks comparably to massive LLMs while running locally on devices like phones, drastically reducing reliance on massive data centers. - Environmental applications, like Forest Protection Connection using AI for acoustic monitoring of illegal logging, demonstrate that smaller, specialized AI can provide significant societal benefits without massive energy demands. ![Screenshot at 00:04: The opening title slide for the TED Countdown event featuring the text "COUNTDOWN" and the call to action "TAKE ACTION ON CLIMATE CHANGE AT COUNTDOWN.TED.COM," setting the context for an environmental focus.](https://ss.rapidrecap.app/screens/Bl-vPf_IAoA/00-00-04.png) **Context:** Sasha Luccioni presents her argument at a TED Countdown event, focusing on the hidden environmental and economic costs associated with the current trend of building ever-larger Large Language Models (LLMs). She challenges the industry narrative that bigger AI is inherently better, contrasting the massive resource requirements of models like GPT-4 and Gemini 1.0 Ultra with the potential of smaller, more specialized models to address critical global issues like climate change more sustainably. ## Detailed Analysis Sasha Luccioni contends that the current pursuit of ever-larger AI models, driven by the mantra 'bigger is better,' is environmentally and economically unsustainable. She highlights that major LLMs like GPT-4 cost hundreds of millions of dollars and consume vast amounts of energy, sometimes equivalent to the annual energy consumption of entire countries like Iceland (13:12). This massive resource requirement centralizes AI development within a handful of well-funded tech giants, leaving startups and academia behind. Luccioni counters this trend by promoting the development of smaller, more efficient, and task-specific AI models. She references the AI Energy Score project, which ranks models based on energy efficiency for specific tasks, showing stark contrasts: Hugging Face’s SmolLM-135M uses only 7.28 Wh per 1,000 tasks for text generation, while DeepSeek R1 uses 1039.24 Wh for the same task (8:23). This demonstrates that smaller models can achieve comparable performance with a fraction of the energy. Furthermore, she showcases positive examples where smaller, specialized AI is already effective, such as the Rainforest Connection using efficient AI on old cell phones powered by solar panels to detect illegal logging sounds in real-time (6:58). Luccioni concludes by urging a shift in focus—taking the power back from massive data centers and applying AI development principles like 'reduce, reuse, recycle' to create a more sustainable future for AI that serves all of humanity, not just a few large corporations. ### The Problem with Big AI - Large corporations invest billions in building massive LLMs like GPT-4 and Gemini 1.0 Ultra, which cost the energy equivalent of Iceland to train (13:12), creating an unsustainable path and excluding smaller players (3:59, 4:01). ### The Alternative - Favoring smaller, task-specific models that use less energy, citing examples like SmolLM-135M which is 5,000 times smaller than DeepSeek's model (4:36, 4:41). ### AI Energy Score - The speaker introduces a scoring system (1-5 stars) that rates AI models based on energy efficiency (measured in Wh per 1k tasks) for specific applications like text generation, promoting transparency (7:55, 8:06). ### Environmental Applications of Efficient AI - Small, efficient AI models can solve real-world problems, such as Rainforest Connection using low-power AI on solar-powered phones to monitor illegal logging via acoustic detection in real-time (6:58, 7:11). ### Policy and Future Direction - Regulations like the EU AI Act starting to mandate disclosures (8:51) are necessary, but the industry must prioritize efficiency and accountability to ensure AI serves all of humanity sustainably, not just profit-driven entities (9:45, 10:09). ![Screenshot at 00:04: The initial screen displays the TED Countdown branding, signaling the presentation's focus on climate action.](https://ss.rapidrecap.app/screens/Bl-vPf_IAoA/00-00-04.png) ![Screenshot at 00:20: A glitch-art image of multiple skulls, illustrating the potential grim outcome \('end of humanity as we know it'\) if climate change is ignored.](https://ss.rapidrecap.app/screens/Bl-vPf_IAoA/00-00-20.png) ![Screenshot at 00:58: An aerial view of a massive data center at sunset, used to visually represent the enormous physical infrastructure and energy demands of building bigger AI models.](https://ss.rapidrecap.app/screens/Bl-vPf_IAoA/00-00-58.png) ![Screenshot at 03:36: A slide comparing the energy cost of a task-specific model \(0.3 g CO2e\) versus a general-purpose model \(10 g CO2e\) for answering a simple question, demonstrating inefficiency.](https://ss.rapidrecap.app/screens/Bl-vPf_IAoA/00-03-36.png) ![Screenshot at 07:16: A dashboard from Quartz Solar showing real-time energy generation forecasts for the UK using AI analysis of satellite imagery and weather data, illustrating the positive application of efficient AI.](https://ss.rapidrecap.app/screens/Bl-vPf_IAoA/00-07-16.png)