Neuromorphic Chips for OpenAI, Cultural Transmission of AI Knowledge, Analogues of ChatGPT: Top AI News of the Week
Our latest AI Digest covers the biggest breaking AI news of the week. Anywhere Club community leader, Aliaksei Kartynnik, comments on the key stories.
In this article
#1 — Popular: OpenAI Acquires Neuromorphic Chips
— It seems that OpenAI intends to purchase processors worth $51 million from the startup Rain AI by the end of 2024.
Why is this interesting? These are not ordinary processors. They are neuromorphic processing units (NPUs), a relatively new processor architecture that partially imitates the functions of the human brain. Notably, data storage occurs directly in the processor, which advantageously distinguishes this architecture from the classical von Neumann architecture.
Such a move by OpenAI could bring NPUs to the mass market and significantly accelerate, and reduce the cost of, modern AI systems.
The connections between some Rain AI investors and Saudi Arabia, however, have drawn the attention of a U.S. government organization due to national security concerns. This could delay the release of the technology and complicate the deal with OpenAI.
#2 — Scientific: Cultural Transmission of AI Knowledge
— A team of researchers from Google DeepMind has introduced a new method of training AI. A system of multiple AI agents is shown a complex 3D simulation of human behavior, enabling "cultural transmission" of knowledge without the need for vast amounts of training data (which is the current practice). The AI agents use deep reinforcement learning in combination with memory and attention mechanisms, as well as automatic process planning.
Tests show that artificial intelligence can generalize tasks, memorize demonstrations, and accurately reproduce them in the future.
Cultural transmission opens up new paths for the development of AI in various practical applications and contributes to our understanding of the processes of cultural evolution and the interaction between humans and AI. This is the first step towards creating systems capable of accumulating knowledge from generation to generation, as humans do.
#3 — Practical: Analogues of ChatGPT
— Due to increased workloads, ChatGPT has not been registering new paid users for several weeks. As a result, many people are unable to easily and simply use the most progressive language model — GPT-4. I propose two ways to bypass this limitation:
- By subscribing to the poe.com service (it costs around $20), you can communicate with GPT-4, GPT-4-32k, and ChatGPT models.
- By registering an account on the OpenAI API and linking a card, you can use GPT-4 through the API. Ready-made solutions can be used as clients for the API. BetterChatGPT, for example, is easy to install and quite functional.
And, of course, don't forget that in addition to ChatGPT, there are other services that may be less functional, but that are free and equally interesting:
Want to learn more effectively with AI?