Good morning. It’s Monday, April 17th.

CEO Sam Altman said at an MIT event on Thursday that OpenAI was not developing GPT-5, and that it would not do so for "some time."

In today’s email:

• What is AutoGPT?

• Cohesive AI: Custom AI Content Generation

• Company Profile: Juice Labs

• 5 Most Viewed Research Papers This Week

You read. We listen. Share your feedback by replying to this email, or DM us on Twitter.

What is AutoGPT?

You may have seen the phrase “AutoGPT” in the news or on Twitter over the past week, but what does an AutoGPT do exactly?

AutoGPT is an AI agent capable of autonomously performing web-based tasks.

If ChatGPT was a single-shot pistol that can’t go online, AutoGPT is a machine gun that can execute tasks on the internet.

Unlike interactive systems like ChatGPT, which necessitate manual instructions for each task, Auto-GPT proactively generates and assigns subtasks to itself in pursuit of the overarching objective, without the compulsory involvement of human operators.

The AI agent is equipped with the ability to search the internet, modify and enhance its own code, and engage in a process of recursive debugging, development, and self-improvement.

Built on the foundation of OpenAI's GPT-4 API, Auto-GPT is among the pioneering instances of GPT-4 functioning autonomously. It boasts features such as short-term and long-term memory management, which includes handling context window length requirements, as well as text-to-speech capabilities for vocal output.

The AI agent's capacity to respond to prompts, adapt to new information, and iteratively create and refine its own prompts for recursive instances has garnered significant attention. Following its release, Auto-GPT quickly rose to prominence as the top trending repository on GitHub and has consistently made waves on Twitter.

Though few OpenAI users have access to a GPT-4 API (you must be a paid subscriber to ChatGPT Plus, and have been granted GPT-4 API access from the waitlist), there is a slimmed-down version called Godmode that uses GPT-3.5, and makes you approve each action.

In theory, here is what an AutoGPT could do:

• Search the internet

• Send emails

• Automatically respond to customer inquiries

• Make purchase/refund decisions

• Coordinate logistics with real people

• Write files

• Execute code

Though the current version and free GPT-3.5 demos are cumbersome and often fail at executing tasks, keep in mind that these are the first iterations of a very complex program.

However within the year, AutoGPT could supplement a substantial amount of the workforce, especially when GPT-4 gets its multimodal capabilities and is able to interpret images.

Make no mistake, AutoGPTs will be used by everyone.

Read more: Godmode, AutoGPT on GitHub

_{sponsored post}

Cohesive AI: Custom AI Content Generation

Cohesive AI can supercharge marketing, sales, customer support, and personal writing by providing AI-generated content that maximizes ROI, increases sales quotas, improves customer support response times, and enhances personal writing.

The AI editor is intuitive and powerful, allowing users to easily edit text, images, and language translations to craft content to perfection.

The platform provides users with access to over 100 templates for creating content, including SEO blogs, TikTok videos, song lyrics, wedding vows, and more.

It also features real-time collaboration tools to streamline teamwork and improve communication.

Cohesive AI also offers content suggestions and ideas while users write in the editor. The platform will also be accessible soon across various platforms, including Gmail, Twitter, LinkedIn, and Medium, through a browser extension.

Cohesive as a company has been gaining massive popularity for its AI-powered content generation platform, Cohesive AI, designed to enhance the quality of written content and optimize it for conversion.

AI Company Profile: Juice Labs

A lot of attention is given to the latest generative AI companies. Some will stand the test of time, many will not. NVIDIA recently made news for increasing its valuation by $300 billion since the beginning of the year by cornering 90% of the enterprise GPU market.

But what about the other companies that are powering the backend of machine learning?

Juice Labs has introduced enhanced flexibility in GPU deployment and provisioning, allowing users to dynamically allocate GPUs to CPU-only instances. This includes sharing a single GPU across multiple instances or aggregating multiple GPUs into one instance.

By doing this, they are lowering the cost of running GPUs, which is a critical utility for all companies running machine learning and AI instances.

The Dynamic GPU Sharing feature enables the sharing of GPUs among different application hosts and workloads, optimizing GPU utilization and reducing the number of cards needed for tasks such as automated build systems and continuous integration.

Juice facilitates this sharing, and lets users to connect to and leverage underutilized GPUs in workstations, utilizing all available computing power.

Juice Labs is led by CEO Steve Golik and CTO Dean Beeler. Golik was an early executive at GetFeedback and Turnitin, helping lead both to acquisitions. Beeler comes from Facebook/Oculus, where he invented the now-standard graphics technologies Async Spacewarp (ASW) and VR Direct Mode.

As the need for GPU usage continues to go parabolic over the next few years, Juice Labs will be one company to keep an eye on.

Read more: Juice Whitepaper on GPU sharing

Top 5 AI Research Papers This Week

Note: arXiv is a free online library where scientists share their research papers before they are published. These are the 5 most viewed papers related to AI in the last week.

• Language Models Trained on Media Diets Can Predict Public Opinion
  The authors introduce a method using language models to emulate opinions based on media consumption, validate it with U.S. surveys, and find it predictive of responses, effective for media-engaged individuals, and useful for studying media effects and forecasting public opinion. This may eliminate the need for polling.

• Re-ReND: Real-time Rendering of NeRFs across Devices (META research)
  The paper introduces Re-ReND, a method for real-time rendering of pre-trained Neural Radiance Fields (NeRFs) on resource-constrained devices by converting NeRFs into an efficient representation using mesh extraction and light field factorization, achieving over a 2.6-fold increase in rendering speed with low memory requirements and high quality.

• Consistency Models
  AI Image generators may go from a “diffusion” model to a one-step generation. This new class of generative AI models support fast one-step generation, few-step sampling, and zero-shot data editing without adversarial training.

• What does ChatGPT return about human values? Exploring value bias in ChatGPT using a descriptive value theory
  The study tests ChatGPT for value biases, while also discussing potential applications and implications for corporate usage, policy making, and human values theorizing.

• ChatGPT Empowered Long-Step Robot Control in Various Environments: A Case Application
  The paper presents customizable input prompts for ChatGPT to convert natural language instructions into executable robot actions.

3x the information, for less than $2/week

Stay informed, stay ahead: Your premium AI resource.

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.

Read by employees from