to compete in the arc challenge, the project submitted must be opensourced. So its weird they're usong this one as a benchmark when theres no submitted code to examine. Edit: Ok i see my misunderstanding now. I should just submit api access myself then? I can beat 85% already.
10 months ago (edited) | 7
Hi o3's energy impact specifically... best igor ## Energy Per Task One task on OpenAI's o3 (high-compute version) consumes: - 1,785 kWh of electricity - Produces 683.66 kg CO2e emissions - Costs between $41,000 to $2.5 million per hour to operate ## European Context To put this in perspective: - A European household uses about 3,500 kWh per year - One o3 task uses about 51% of a European household's annual electricity consumption - The energy used for one o3 task could power a European home for about 6 months ## Environmental Impact The carbon footprint is substantial: - 683.66 kg CO2e per task is equivalent to: - Multiple transatlantic flights - Several months of an average European's total carbon emissions - This is particularly concerning since most AI training relies on fossil fuel-powered electrical grids ## Future Implications The trend is concerning because: - By 2027, AI servers could consume as much electricity as Argentina - If integrated into common services, AI could increase energy demands dramatically - The gap between human and AI energy consumption for similar tasks is around 1000x - Current projections suggest 20-25 years before costs become economically viable This makes o3 one of the most energy-intensive AI models to date, raising serious questions about the sustainability of such high-compute AI systems as they become more prevalent. Citations: [1] https://blogs.cfainstitute.org/investor/2024/10/31/the-hidden-environmental-costs-of-tech-giants-ai-investments/ [2] https://www.theparliamentmagazine.eu/news/article/how-ai-energy-consumption-challenges-eu-climate-policy [3] https://carboncredits.com/how-big-is-the-co2-footprint-of-ai-models-chatgpts-emissions/ [4] https://community.openai.com/t/sustainable-development-and-ai/377448 [5] https://www.vox.com/climate/2024/3/28/24111721/climate-ai-tech-energy-demand-rising [6] https://community.openai.com/t/carbon-emissions-of-the-api/371484 [7] https://www.bruegel.org/comment/artificial-intelligence-and-energy-consumption [8] https://www.scientificamerican.com/article/the-ai-boom-could-use-a-shocking-amount-of-electricity/ [9] https://www.polytechnique-insights.com/en/columns/energy/generative-ai-energy-consumption-soars/ [10] https://www.newscientist.com/article/2361343-artificial-intelligence-training-is-powered-mostly-by-fossil-fuels/ [11] https://news.ycombinator.com/item?id=42473321
10 months ago (edited) | 2
Incremental compute costs for marginal improvement is not very impressive
10 months ago | 2
Thought experiment: what percentage of the world's human work involves tasks that have clear, objectively true answers, and has sufficient training data likely to be available to train/RL on? 🤔
10 months ago | 3
All About AI
very promising results from OpenAI and the o3 (!) model. 87.5% on the ARC AGI seems kinda shocking. excited to to hear more from tests and try it out in early 2025. have a nice weekend all :)
10 months ago | [YT] | 84