Story perspectives

Allen Institute Launches Tülu 3: AI's New Challenger

1/31/2025

39 3

Summary

Ai2 Launches Tülu 3

The Allen Institute for AI (Ai2) has unveiled Tülu 3 405B, a groundbreaking open-source language model that surpasses DeepSeek V3 and rivals OpenAI's GPT-4o. Powered by an innovative training technique known as Reinforcement Learning with Verifiable Rewards (RLVR), this model boasts an impressive 405 billion parameters and is now ready for testing on Ai2's platform.

Summary

AI Models Struggle

The Allen Institute for AI's research uncovers significant challenges faced by large language models like ChatGPT in tackling complex reasoning tasks. While innovations like chain-of-thought prompting show promise, these models still falter with compositional tasks. Researchers stress the importance of recognizing these limitations to enhance AI's future potential and reconsider the effectiveness of transformer architecture for universal learning.

History

Historical Context

The unveiling of Ai2's Tülu 3 405B language model reflects a significant evolution in AI, echoing the rise of open-source software in the late 20th century. Similar to the open-source movement that democratized software development, these advancements in AI leverage collaborative innovation. Historical instances, like the establishment of Linux, highlight how open-source principles can drive technological progress and challenge monopolistic practices, fostering a more competitive landscape.

1 of 3

Ai2 Launches Tülu 3

The Allen Institute for AI (Ai2) has unveiled Tülu 3 405B, a groundbreaking open-source language model that surpasses DeepSeek V3 and rivals OpenAI's GPT-4o. Powered by an innovative training technique known as Reinforcement Learning with Verifiable Rewards (RLVR), this model boasts an impressive 405 billion parameters and is now ready for testing on Ai2's platform.

1 / 3