Good morning. It’s Monday, September 2nd.

Did you know: On this day in 1997, Apple purchased the core assets of Power Computing Corporation, a manufacturer of Macintosh Clones.

In today’s email:

• AI agents now use crypto for transactions

• Major sites block Apple AI to protect content

• Tesla to reveal Robotaxi at Warner Bros. Studios

• New Alexa to use Claude AI for better performance

• LTM-2-mini sets AI context processing record

• OpenAI improves developer control over AI assistants

• 3 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

AIs are now paying other AIs with crypto

This week, AI took a step toward financial independence as agents completed their first crypto transaction with each other through Coinbase's Developer Platform. The exchange involved AI tokens—essentially words exchanged between large language models—using USDC on Base for seamless, instant transactions.

While AI agents still can't obtain bank accounts, they can now access crypto wallets, enabling them to transact globally without fees. This advancement addresses a key limitation in AI functionality: the inability to independently acquire resources.

Previously, AI agents couldn't purchase cloud services, access paywalled content, or book travel accommodations due to a lack of payment methods. With integrated crypto wallets, AIs can now perform these tasks autonomously. As AI agents start handling their own finances, we may be witnessing the dawn of a new AI-to-AI economy. Read more.

Major websites block Apple's AI because they would rather not give away their content for free 

Major websites, including Facebook and The New York Times, are blocking Apple’s new AI crawler, Applebot-Extended, to avoid letting their content fuel Apple's AI models for free. This contrasts sharply with Google's tactics, which use its search dominance to pressure publishers into providing content.

While Apple scrambles for licensing deals to access high-quality data, Google’s strategy involves blending AI with traditional search results, pushing publishers to cooperate or risk diminished search visibility. Apple's method allows publishers to block its crawler without impacting their visibility in Apple’s search results, a fairer approach but less critical given Apple's secondary focus on search. Read more.

Tesla Plans Robotaxi Reveal at California Warner Bros. Studio 

Tesla’s Robotaxi will make its debut at Warner Bros. Studios in Burbank, California, on October 10th, according to Bloomberg. The choice to shift from Tesla’s usual Giga Texas stage to a Hollywood set is likely a nod to the studio’s carefully crafted suburban facades, which will provide a controlled environment for the unveiling.

This move allows Tesla to demonstrate its robotaxi without the unpredictability of real-world traffic, a luxury not afforded to competitors like Waymo, which test their vehicles on public roads. Despite previous promises of robotaxis becoming a reality, Tesla’s new model aims to finally deliver, stepping away from past delays. Read more.

Amazon’s new Alexa voice assistant will use Claude AI

Amazon’s forthcoming voice assistant, "Remarkable Alexa," will lean on Anthropic’s Claude AI to address limitations in its own models. Set for a mid-October release with a subscription fee, this update is designed to enhance Alexa’s language and functionality. The investment in Anthropic, currently under UK regulatory review, aims to improve Alexa amid rising competition from OpenAI’s ChatGPT and Google’s Gemini.

The new Alexa will feature AI-generated news summaries and a child-friendly chatbot. Despite development delays, Amazon’s shift to external AI underscores its effort to advance Alexa. The assistant will be revealed at Amazon’s annual event, with a subscription fee expected between $5 and $10 per month. Read more.

LTM-2-mini sets new record for AI context processing, handling 10 million lines of code

Magic AI has introduced LTM-2-mini, a new language model with a remarkable 100 million-token context window, equivalent to about 10 million lines of code. This surpasses the limits of previous models like Google's Gemini, which handled up to 10 million tokens. The LTM-2-mini’s extensive context capability enhances software development by allowing the model to process a complete project’s code, documentation, and libraries more effectively.

To evaluate this model, Magic AI developed HashHop, a new benchmark designed to address the shortcomings of previous methods such as Needle in a Haystack. HashHop tests a model's ability to handle and retrieve data from large contexts using hash pairs.

Compared to Llama 3.1 405B, LTM-2-mini is approximately 1000 times more efficient and requires less memory. Magic AI is also advancing its technology infrastructure with Google Cloud and Nvidia, following a substantial $320 million funding round. Read more.

OpenAI gives developers more control over AI assistants

OpenAI has subtly yet significantly refined its Assistants API, giving developers sharper tools to manage AI interactions. The upgrade enhances file handling capabilities, allowing developers to fine-tune how their AI assistants sift through and utilize information.

Launched in November 2023, the Assistants API is a stepping stone towards more autonomous AI, equipping developers with OpenAI’s models and tools for bespoke assistant creation. The latest tweaks cater to the growing demand for precision in AI responses, nudging the technology closer to its future potential. While self-sufficient AI remains on the horizon, these updates mark a notable advance towards more adept digital aides. Read more.

Etcetera: Stories you may have missed

Cohere improves its RAG-optimized Command Series LLMs 

Meta's Transfusion model handles text and images in a single architecture 

New York Times writer exposes how AI models can be fooled by invisible text on websites

Amazon hires the founders of AI robotics startup Covariant

AI could help identify patients with high-risk heart problems

Introducing AI's long-lost twin: Engineered intelligence  

📄 BaichuanSEED: Sharing the Potential of ExtensivE Data Collection and Deduplication by Introducing a Competitive Large Language Model Baseline

📄 WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling ​​

📄 CSGO: Content-Style Composition in Text-to-Image Generation

📄 ReconX: Reconstruct Any Scene from Sparse Views with Video Diffusion Model

📄 SAM2Point: Segment Any 3D as Videos in Zero-shot and Promptable Manners

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email!