GPT-4o mini: advancing cost-efficient intelligence

OpenAI is introducing GPT-4o mini, a highly cost-efficient small model designed to make AI more accessible. With a price of 15 cents per million input tokens and 60 cents per million output tokens, GPT-4o mini is over 60% cheaper than GPT-3.5 Turbo and an order of magnitude more affordable than previous models. It supports text and vision in the API and has a context window of 128K tokens with support for up to 16K output tokens per request. GPT-4o mini surpasses GPT-3.5 Turbo in chat preferences and outperforms other small models in benchmarks like MMLU, MGSM, and HumanEval.

GPT-4o mini demonstrates strong performance in both textual intelligence and multimodal reasoning, scoring higher than Gemini Flash and Claude Haiku on various academic benchmarks. It excels in reasoning tasks, mathematical reasoning, and coding, making it suitable for applications requiring multiple model calls, large volumes of context, or real-time text responses. Companies like Ramp and Superhuman have found GPT-4o mini to be significantly better than GPT-3.5 Turbo for tasks like data extraction and email response generation.

Safety measures are integral to GPT-4o mini, with built-in mitigations similar to those in GPT-4o. The model has been rigorously tested by external experts to identify and address potential risks, and new techniques like the instruction hierarchy method improve its resistance to jailbreaks and prompt injections. GPT-4o mini is available through various APIs, and starting today, Free, Plus, and Team users in ChatGPT can access it, with Enterprise access beginning next week. OpenAI continues to reduce costs and enhance model capabilities, aiming to make powerful AI applications more affordable and integrated into daily digital experiences.

Related:GPT-4o mini is exactly what I've been looking forward to.

The biggest names in AI have teamed up to promote AI security

Google, OpenAI, Microsoft, Amazon, Nvidia, Intel, and several other major AI companies are uniting to form the Coalition for Secure AI (CoSAI), as announced on Thursday. This initiative aims to tackle the “fragmented landscape of AI security” by offering access to open-source methodologies, frameworks, and tools. CoSAI, which will operate under the Organization for the Advancement of Structured Information Standards (OASIS), seeks to provide a collaborative platform to enhance AI security across the industry.

The formation of CoSAI addresses significant concerns in the AI industry, such as the potential for leaking confidential information and automated discrimination. By developing best practices for AI security, addressing challenges, and securing AI applications, CoSAI aims to mitigate these risks. The coalition's approach involves creating standardized practices that can be adopted by organizations of all sizes, thus promoting a secure-by-design philosophy in AI development and deployment.

Heather Adkins, Google’s vice president of security, emphasized the dual nature of AI's potential for both defenders and adversaries. She highlighted that CoSAI will assist organizations in securely and responsibly integrating AI, allowing them to benefit from its advancements while managing the associated risks. With companies like IBM, PayPal, Cisco, and Anthropic also joining, CoSAI represents a significant step towards unified AI security efforts in the industry.

Proton launches ‘privacy-first’ AI writing assistant for email that runs on-device

Proton, known for its privacy-centric apps, has introduced Proton Scribe, an AI writing assistant for email that operates entirely on the user’s device. This tool, built on the Mistral 7B open source language model, assists in composing, redrafting, and proofreading emails without sending data to third-party servers. Proton Scribe is designed to alleviate privacy concerns, especially for enterprises, by ensuring that no user data is shared or learned from, aligning with Proton's commitment to data security.

Following Google's recent integration of AI into Gmail, Proton's move marks its entry into the AI-driven productivity tools market. The Swiss company's new assistant offers a privacy-first alternative to other AI tools, addressing the growing concern about sensitive data leakage. Proton Scribe's deployment at the local device level ensures that user information remains private, making it a suitable choice for businesses wary of generative AI due to data privacy issues.

Founder and CEO Andy Yen emphasized the necessity of such tools, noting that users will turn to AI regardless of privacy risks. By embedding privacy-first AI tools within Proton Mail, the company aims to prevent users from exposing sensitive communications to third-party AI services with poor privacy practices. Proton's approach aims to provide a secure solution for those needing AI assistance while maintaining stringent privacy standards.

Bye-bye bitcoin, hello AI: Texas miners leave crypto for next new wave

Houston-based Lancium and Denver-based Crusoe Energy Systems have announced a multibillion-dollar deal to construct a 200-megawatt AI data center near Abilene, Texas, marking the first phase of a larger 1.2-gigawatt build-out. Designed to meet the unique needs of AI companies, the facility will support advanced cloud computing applications like medical research and aircraft design. Lancium President Ali Fenn highlighted that this data center campus, once fully operational, will be among the largest in the world, reflecting a significant shift from bitcoin mining to AI-driven infrastructure.

This transition is driven by the high demand for AI infrastructure and the diminishing profitability of bitcoin mining. With bitcoin miners possessing extensive data centers, access to fiber lines, and substantial power, they are well-positioned to support compute-intensive AI operations. The pivot has already proven beneficial, with companies like Bit Digital and Hut 8 shifting to AI services, generating significant new revenue streams. Core Scientific, for example, emerged from bankruptcy and saw a stock upgrade following strategic partnerships to enhance its AI capabilities.

Crusoe Energy Systems, initially known for utilizing flare gas to power bitcoin mining, has long envisioned AI infrastructure as part of its strategy. The new Abilene facility will primarily use renewable energy, aligning with Crusoe’s and Lancium’s goals to integrate AI with sustainable power solutions. Lancium’s technology can quickly adjust energy consumption to balance grid demand, making AI data centers assets rather than liabilities. This shift represents a broader trend as data centers, increasingly optimized for energy availability and sustainability, are poised to meet the burgeoning needs of the AI industry.

Other stuff

• OpenAI GPT-4o voice will be alpha later this month, GA later

• A Brief History of AI: How Did We Get Here, and What's Next?

• LLM model size competition is intensifying… backward!

• What is Apple Intelligence, when is it coming and who will get it?

• TSMC boss predicts AI chip shortage through 2025, says Trump comments don't change his strategy

• Overcoming the limits of current LLM

• Everyone Is Judging AI by These Tests. But Experts Say They’re Close to Meaningless

Hope you enjoyed today's newsletter

⚡️ Join over 200,000 people using the Superpower ChatGPT extension on Chrome and Firefox.

OR