And some product updates.
OpenAI has made its Memory feature available to all ChatGPT Plus subscribers, allowing the chatbot to remember user-supplied insights and information to generate more contextually relevant answers. The feature saves personal facts and can be used to create responses without the need for repetitive prompts. Users can manage their memories and toggle the feature on or off in the settings. The feature is currently available to ChatGPT Plus subscribers, with plans to expand to other subscription types.
OpenAI has developed a new deepfake detector that can identify 98.8% of images created by its generator, DALL-E 3. However, the tool is not designed to detect images from other popular generators. OpenAI is also working on watermarking AI-generated sounds and joining the Coalition for Content Provenance and Authenticity to develop credentials for digital content.
Anthropic is launching a new paid plan called Team aimed at enterprises, along with a new iOS app. The Team plan provides higher-priority access to Anthropic’s generative AI models and additional admin and user management controls. It offers a larger context window and collaboration features. Anthropic aims to capture significant enterprise market share as more companies move from AI experimentation to full-scale deployment. Anthropic’s iOS app also provides access to its AI assistant, Claude 3, with real-time image analysis capabilities.
Microsoft has decided to pause the rollouts of new Copilot features in Windows 11 to refine them based on user feedback. The Copilot in Windows will continue to work as expected while new ideas are being developed with Windows Insiders. Microsoft is holding an AI event on May 20th, where more updates are expected to be showcased. New features for Windows are anticipated to include an AI Explorer app and the concept of a “Copilot for every person.”
Microsoft is deploying GPT-4 in the Azure Government Top Secret cloud for classified defense and intelligence customers. This isolated environment will give users access to AI-powered tools for drafting documents, writing code, and analyzing information. The model will be physically unconnected to the public internet, and users will not be able to train the model on new data. The deployment aims to improve productivity and provide classified users with the same AI capabilities in non-classified settings.
Why do we care?
Detectors are a step towards combating deepfakes, but there is no perfect solution to the problem. Building on the previous segment, the collection of tools grows and changes rapidly and adapts to use cases. Let’s focus on the isolated GPT-4 for Top Secret – the natural progression of “a model for each use case.”
I’m also intrigued by enterprise offerings like Anthropic, similar to what we’ve seen from OpenAI and Microsoft. Circle that area for opportunity.