Good morning. It’s Wednesday, March 15, also known as the Ides of March.

GPT-4 was announced yesterday. We’ll be covering all the details, including how to use it, what companies have already integrated it, and what new features have been added. First time reading? Sign up here.

You read. We listen. We recently changed the design of this newsletter. Share your feedback by replying to this email, or DM us on Twitter.

GPT-4

OpenAI has announced the release of GPT-4, a significant advancement in their efforts to scale up deep learning technology. As a large multimodal model, GPT-4 accepts both image and text inputs and produces text outputs.

If you’re a ChatGPT Plus member, it’s available to use now.

After our use of GPT-4 yesterday, we noticed that it is a tiny bit slower than ChatGPT, but the responses do seem more comprehensive and adhere to the prompt.

GPT-4 demonstrates human-level performance on a range of professional and academic benchmarks. For instance, it achieves a score in the top 10% on a simulated bar exam, compared to GPT-3.5, which scores in the bottom 10%.

The developers have spent six months iteratively aligning GPT-4, incorporating lessons from their adversarial testing program and ChatGPT, resulting in their best-ever outcomes in terms of factuality, steerability, and adherence to guardrails.

See this example of GPT-4’s results on benchmark testing:

Text + Images

GPT-4 can process both text and images as input, allowing users to assign tasks that involve visual or language elements.

Specifically, it generates textual outputs based on inputs containing a combination of text and images, whether it's processing documents with text and photos, diagrams, or screenshots, or working with text-only inputs.

The AI system benefits from test-time techniques designed for text-only language models, including few-shot and chain-of-thought prompting.

The image input feature is currently in the research preview stage and not yet available to the public.

Watch: GPT-4 official video, GPT-4 Livestream Demo

GPT-4’s Image Modality

Be My Eyes, a Danish startup founded in 2012, has been developing technology to support the community of over 250 million individuals who are blind or have low vision. The company connects visually impaired users with volunteers to assist with daily tasks like product identification or navigating airports.

Be My Eyes has started to leverage GPT-4's new visual input capability to develop a Virtual Volunteer™ within their app, which aims to provide the same level of context and understanding as a human volunteer. In early February, the company began beta-testing the GPT-4 powered assistant with a select group of employees.

The positive results have led to plans for a wider rollout within weeks. Buckley, a spokesperson for the company, expresses excitement for the potential impact on their community, noting that beta testers, including Lucy Edwards, have already expressed enthusiasm for the new feature.

Jesper Hvirring Henriksen, CTO of Be My Eyes, says that basic image recognition applications can only describe what is in front of the user but are unable to engage in a discussion to determine specific details, such as the ingredients of noodles or whether an object on the ground poses a tripping hazard. GPT-4's more advanced technology, however, can provide such insights, revolutionizing the support available to the visually impaired community.

Read more: Be My Eyes, OpenAI blog

Duolingo Announces Partnership with OpenAI

Duolingo, a popular language-learning platform, has announced the integration of OpenAI's GPT-4 technology to enhance its offering with two new features in a subscription tier called Duolingo Max.

The new features include Role Play, an AI conversation partner, and Explain my Answer, which clarifies language rules when users make mistakes. According to Bodge, a spokesperson for Duolingo, the goal was to incorporate AI-powered features that seamlessly blend with the app's gamified approach and appeal to learners.

Duolingo engineers previously experimented with GPT-3 to complement human-powered aspects of their earlier chat feature. However, lead engineer Bill Peterson mentioned that while it was close to being ready, it wasn't quite advanced enough to confidently manage complex automated chat elements.

Read more: OpenAI Consumer Stories: Duolingo

Google Workspace AI

The major players in Artificial Intelligence all chose yesterday to announce their latest AI projects and developments, it seems.

Google announced the implementation of AI-powered features to enhance user productivity across various applications, including Gmail, Docs, Slides, Sheets, Meet, and Chat.

They aim to streamline tasks such as drafting, replying, summarizing, and prioritizing emails in Gmail, as well as brainstorming, proofreading, writing, and rewriting in Docs.

The AI integration enables users to bring creative visions to life by auto-generating images, audio, and video in Slides, while also facilitating data analysis through auto-completion, formula generation, and contextual categorization in Sheets. The AI features can also generate new backgrounds and capture notes in Meet and enable workflows for increased efficiency in Chat.

One of the first AI-powered features focuses on making writing easier in both Docs and Gmail. Recognizing that blank pages can be challenging for users, generative AI has been embedded into these applications to assist in initiating the writing process.

The AI-driven feature caters to various user needs, from HR professionals creating customized job descriptions to parents drafting invitations for a themed birthday party. Workspace aims to save time and effort by generating an initial draft based on the user's input topic, allowing for further refinement and editing in collaboration with the AI partner.

Read more: Google Generative AI announcement

Anthropic AI’s Claude

The Google-backed San Francisco-based AI company Anthropic AI also had an announcement yesterday: the beta release of “Claude”.

Claude is a next-generation AI assistant designed to provide helpful, honest, and harmless support across various conversational and text processing tasks.

After a successful closed alpha test with partners such as Notion, Quora, and DuckDuckGo, the AI system is now available for wider adoption.

Claude offers assistance in various use cases, including summarization, search, creative and collaborative writing, Q&A, and coding. Users can interact with the AI through chat interfaces and APIs

What sets Claude apart from GPT is that the system can be customized for personality, tone, and behavior.

Two versions of Claude are available: Claude, the high-performance model, and Claude Instant, a lighter, less expensive, and faster alternative.

Partners have praised Claude's integration and capabilities. Quora users have reported Claude's responses to be detailed, easily understood, and conversational in their new Q&A app Poe. Juni Learning has leveraged Claude for their Juni Tutor Bot on Discord, providing high-quality, tutor-like experiences for students. Notion users have benefited from Claude's creative writing and summarization abilities, improving productivity within the workspace.

Claude has also been utilized by DuckDuckGo in its DuckAssist feature and by Robin AI in the legal industry to evaluate and suggest contract language.

Read more: Claude press release

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 your subscription.