Hey Synapticians,

The weekend is here, knocking at our door. On today’s menu: Mr. Biden wants to boost AI, a rumor about GPT-5 supposedly already being released, MIT calculating the energy costs of model training, and finally, a new architecture that seems promising ; TITANS. We also invite you to take advantage of the weekend to test out Perplexity (this week’s theme).

Lastly, if you enjoy this news, feel free to share it.

Have a great weekend!

Top AI news

1. Biden Signs Order to Boost AI Infrastructure
On January 14, 2025, President Joe Biden signed an executive order to strengthen U.S. leadership in artificial intelligence (AI). The order instructs the Departments of Defense and Energy to lease federal sites to private companies for constructing advanced AI infrastructure, including large-scale data centers and clean energy facilities. This initiative aims to accelerate AI development while adhering to environmental standards and preventing additional costs for American families. President Biden emphasized that this effort is crucial for enhancing economic competitiveness and national security, as well as promoting a clean energy transition. The administration seeks to ensure that the U.S. remains at the forefront of AI technology and infrastructure development.

2. Is OpenAI Quietly Using GPT-5 for Secret Projects?
The article speculates that OpenAI has developed GPT-5 but has chosen to keep it internal, utilizing its capabilities for in-house applications rather than releasing it publicly. This strategy is hypothesized to provide a greater return on investment by leveraging GPT-5's advanced features internally. The author emphasizes that this is pure speculation, with no concrete evidence or insider information to confirm the theory. Additionally, the article references the unexplained absence of Anthropic's Claude Opus 3.5, suggesting a trend where advanced AI models are retained for internal use to generate synthetic data or enhance other models. Readers are encouraged to view this as a thought experiment pending official updates on GPT-5's development.

3. Generative AI's Environmental Impact Unveiled
The rapid development of generative AI models, exemplified by OpenAI's GPT-4, entails significant environmental costs. Training these models requires immense computational power, resulting in high electricity consumption and elevated carbon dioxide emissions. Beyond the initial training phase, deploying and fine-tuning these models for widespread use further intensifies energy demands, placing additional stress on electrical grids. Moreover, the cooling systems essential for maintaining the hardware's optimal performance consume large volumes of water, potentially straining local water supplies and affecting ecosystems. The surge in demand for advanced computing hardware also introduces indirect environmental impacts through manufacturing and transportation processes.

Bonus. New 'Titans' Model Challenges Transformer Dominance in AI
The article reviews the paper "Titans: Learning to Memorize at Test Time," which addresses the challenges faced by traditional AI models, particularly Transformers, in processing extremely long sequences due to high computational demands. The proposed "Titans" architecture introduces a Long-Term Memory (LTM) module to store essential information from earlier in the sequence and a Short-Term Memory (STM) module for immediate data processing. This dynamic memory system enables Titans to adapt and learn during testing, processing sequences exceeding 2 million tokens while maintaining high accuracy, thus outperforming standard Transformers. The approach enhances flexibility and adaptability in AI models when encountering new data during testing.

Video of the Day

2025 AI trends from IBM.

Theme of the Week

Conversational Search Engines - AI Startup Review
Perplexity AI combines advanced language models with real-time search to deliver fast, accurate answers, supported by citations. Backed by $915M, it challenges tech giants.

Stay Connected

Feel free to contact us with any feedback or suggestions—we’d love to hear from you !