OpenAI reveals 'Strawberry' AI to the feds, powering its new 'Orion' project.

In today’s email:

🔥 Google Meet’s automatic AI note-taking is here
😭 ‘Make your health insurance company cry’: One woman’s fight to turn the tables on insurers
📚 Why AI can’t spell ‘strawberry’
🧰 12 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

OpenAI reveals 'Strawberry' AI to the feds, powering its new 'Orion' project.

OpenAI is making significant strides with a new AI model code-named "Strawberry," which is designed to tackle complex problems without the hallucinations that often plague current AI systems. Although Strawberry is reportedly slower and more expensive at inference time, it excels at solving intricate challenges on the first try. The model's main role is to generate synthetic data for "Orion," OpenAI's upcoming flagship large language model (LLM). This focus on creating accurate synthetic data could help Orion reduce errors and improve its reasoning capabilities, making it a significant advancement in AI development.

OpenAI is also pushing to incorporate a distilled version of Strawberry into ChatGPT, potentially as early as this fall. This smaller, simplified version would allow ChatGPT to benefit from Strawberry's improved reasoning abilities, offering more accurate responses across a range of tasks, including math problems and programming challenges. While this integration could slow down response times, the trade-off would be worth it for users seeking more reliable and thoughtful answers.

In addition to enhancing ChatGPT, OpenAI has demonstrated Strawberry's capabilities to U.S. national security officials, underscoring its potential broader applications. The model's ability to generate synthetic data could address limitations in obtaining high-quality training data from real-world sources, a critical issue for the development of future AI systems. Strawberry's origins trace back to research led by Ilya Sutskever, a former chief scientist at OpenAI, and continued by Jakub Pachocki and Szymon Sidor.

The AI Gold Rush: Are You In or Out?

While others scramble, savvy entrepreneurs are quietly transforming their businesses with AI.

Here's the kicker: You don't need a Ph.D. to join them.

Our no-BS guide cuts through the hype, giving you:

A jargon-free breakdown of large language models (the engine behind AI's magic)
Concrete steps to leverage AI for immediate business growth
Real-world examples that turn "prompt engineering" from gibberish into your secret weapon

Don't let the tech heads have all the fun. Grab your share of the AI pie – no coding required.

Get Your Free AI Playbook

Google Meet’s automatic AI note-taking is here

Google Meet has introduced a new AI-powered feature called "take notes for me," which aims to automate note-taking during meetings. This feature, now rolling out to select Google Workspace customers, can summarize key points from discussions rather than just transcribe the entire conversation. Initially, the feature supports only spoken English and is available to customers with the Gemini Enterprise, Gemini Education Premium, or AI Meetings & Messaging add-ons. Once the meeting concludes, the notes are saved in a Google Doc, which is automatically attached to the calendar event and shared with the meeting organizer and anyone who enabled the feature.

In addition to summarizing conversations, the tool is designed to provide a meeting recap for latecomers, allowing them to catch up quickly without disrupting the flow. If users also opt to record and transcribe the meeting, those files will be linked in the same document as the notes. This integration aims to streamline post-meeting organization and accessibility, particularly benefiting those who may find it difficult to process spoken language while taking notes simultaneously.

While this feature holds promise for improving efficiency and accessibility, there are still concerns about its accuracy. Users of Google's transcription tool have reported issues with incorrect transcriptions, often needing to review and edit the generated content. As a result, there is some skepticism about whether the note-taking tool will offer a significant improvement in capturing the essence of conversations accurately. Despite these concerns, Google is hopeful that the rollout, expected to reach all eligible customers by September 10th, 2024, will meet users' needs.

Cerebras Launches the World’s Fastest AI Inference

Cerebras has unveiled its new AI inference solution, which claims to be the fastest in the world. The Cerebras inference system delivers an impressive 1,800 tokens per second for Llama3.1 8B and 450 tokens per second for Llama3.1 70B—20 times faster than traditional GPU-based cloud solutions. The offering also stands out with industry-leading pricing, starting at just 10 cents per million tokens for Llama3.1 8B. Developers can now access Cerebras' inference capabilities via an open API, making it easier to integrate into various applications.

Powered by Cerebras' third-generation Wafer Scale Engine (WSE-3), the system breaks through the memory bandwidth limitations that often slow down traditional LLM inference. By storing the entire model on a single chip, Cerebras can achieve speeds that far exceed GPU-based solutions. With an aggregate memory bandwidth of 21 petabytes per second, the WSE-3 offers 7,000 times more memory bandwidth than leading GPUs like the H100, eliminating bottlenecks and enabling real-time inference for even the largest models.

Unlike some other platforms that compromise accuracy by reducing model precision, Cerebras inference uses Meta's original 16-bit weights to maintain the highest accuracy possible. This dedication to precision ensures that the Llama3.1 models perform optimally in complex tasks, such as multi-turn conversations and reasoning challenges. Cerebras is also planning to add larger models, including Llama3 405B and Mistral Large, in the near future, further expanding the platform's capabilities.

The launch of Cerebras inference signifies a major leap forward in AI technology, offering both unmatched speed and cost-effectiveness. The platform’s high-speed processing opens up new possibilities for AI applications, from real-time decision-making to advanced code generation. By solving long-standing challenges in AI inference, Cerebras is positioning itself as a leader in the future of AI development.