OpenAI Suspends Training, Anthropic Signs $11.6 Billion Deal

Between efficiency gains and major contracts, AI is reaching new thresholds. OpenAI has suspended the training of its models to review incidents while preparing to expand Ultrafast and gaining access to new texts. Trading computation continues to progress, with cost objectives halved.

Efficiency: 1 billion tokens/min on 1 GPU and costs x178

According to the data provided, it is possible to achieve a rate of one billion tokens processed every minute on a single GPU, a result obtained by combining a query scheduler with an inference engine. On the other hand, trials reveal that 22 different models are capable of generating the exact same response, but with a cost ratio of up to 178 between the cheapest and the most expensive. The computational activity for trading is progressing rapidly, accompanied by debates around the optimal use of resources and cost reduction. Many companies are striving to execute their AI tasks at half the cost, which could disrupt the field of algorithmic trading.

Resources: $11.6 billion with Akamai and access to Bodleian texts

Anthropic has signed a contract worth $11.6 billion with Akamai for computing. OpenAI has obtained permission to use texts from the Bodleian Library to train its AI models. The Muse manifesto has also been published, outlining the goals and vision of this new initiative.

OpenAI pauses training to analyze thousands of incidents

OpenAI has chosen to temporarily halt its training sessions to review several tens of thousands of incidents. This pause aims to ensure the safety and performance of its AI systems. Additionally, OpenAI plans to expand access to its Ultrafast API to more users. A detailed assessment is also underway to determine whether AI self-improvement can surpass the phenomenon of diminishing returns.