New Versions Galore | Issue 16

A busy week on the major product announcements this week. OpenAI had its developer day, releasing a whole suite of interesting changes. Midjourney image generator released ‘tuning’ (see premium version of the newsletter for more information on that), and X.ai released ‘Grok’.

Best regards,

Paul, Co-founder (and newsletter editor)

The ask me anything event

We settled on 10th November, 12pm (noon). Click here if you can make it, its free and will be a chance to chat about anything AI.

OpenAI Development day.

The key takeaway for me as a business user. A vastly simplified the user experience by removing the complex variations, and allowing the tool itself to decide and execute. This combines with the multi-modal (image, pdf reading, spreadsheet, code, and text) to make a very flexible powerful toolset.

The knowledge cut-off date has been updated to April 2023, with an intention to continue to move the date forwards.

Note: Good prompting, and understanding of how to use it still apply, they have not ‘fixed’ hallucination or the other limitations.

Summary of the announcement.

• GPT-4 Turbo: This new AI model can process a huge amount of information, similar to digesting over 300 pages of text at once, making it a powerful tool for complex tasks. It's also more cost-effective.

• Extended Context: 128,000 token context (which now exceeds claude.ai making it the current best in market), the equivalent of 300 book pages.

• Assistants API: Streamlines the creation of AI applications that can perform specific tasks, making it user-friendly for developers.

• Multimodal Capabilities: The AI now has abilities like seeing and creating images, and converting text to speech.

• GPT’s: Very simple agents and agent construction, like a personal or shared playbook of prompts, which can also have additional supplied data as part of the setup. ‘Programmed’ using natural language and business user prompts. Easier than an excel macro!

• Copyright Shield: which offers legal defense and coverage of costs for customers facing copyright infringement claims. This applies to the widely available features of ChatGPT Enterprise and the developer platform.

Enhancements for Efficiency:

• Multi-Function Execution: AI can now perform multiple tasks from a single request.

• Sharper Instruction Following: It better understands and executes precise instructions.

• Stable Outputs: A new feature that ensures consistent results, which is essential for reliable business applications development.

• Cheaper: (for developers and application execution) by 2.75x

GPT’s and Agents

OpenAI has introduced a new feature called GPTs, allowing customization of ChatGPT for specific tasks. These custom versions are easy to create without any coding, tailored to individual or business needs, and can be shared.

The GPT Store, launching soon, will enable users to discover and monetize their creations, though as you will see from the early release apps, I am not sure how useful the first ones will be.

These are built with privacy and safety at the forefront, allowing users maintain control over their data, and OpenAI has implemented systems to prevent misuse.

At this moment, I see GPT’s (I have NO idea why they chose that name) as more of pre-staged prompts, sort of like a playbook which can be personal, shared, or potentially sold.

This is the first step towards ‘Agents’, which are essentially autonomous process tools which will allow you to set AI to work on a sequence of tasks, without you having to stay at the application. The combination of agents, ‘low code’ solutions for process automation will be how to move AI and chatGPT out of personal productivity into a more functional workflow.

As of writing rollout of being able to create these yourselves is still in progress.

New reader Quick Starts

Here are a few back issues that will help you get up to speed.

Other things announced by openAI

Whilst the focus was on the chatGPT and API changes for developers, a couple of other things were announced. These both were around the updated speech to text (whisper V3), which is faster and cheaper. This is often used behind several of the other transcription and closed caption services, and will also be available via API calls. This will allow the creation of multilingual product descriptions, customer content.

In addition, text to speech (voice). Six available voices with ultra realistic speech. By far the best I have heard so far. Have a listen here, and here

Image prompt example

Let’s revisit MidJourney, the new feature is called /tuner

product mockup, scented candle, soft lighting

the /tune command is used to build a style

/tune prompt: autumn, warm colours, photography, depth of field

The process then builds a variation tuner for you to choose from, where you pick variations to influence the style generated. This pops up as a web link

You then add the style code to the end of the image you want to use it on —style [code]. The code is at the bottom of the web page provided by MidJourney bot

Quick things to check out

Runway ML 48 Short film competition

This competition has pulled together some incredibly creative short films made entirely using AI.

Events

We have an exciting series of events and training running in conjunction with Macildowie Recruitment and Retention.

C-Suite / Business Leader Event (early bird discount code MACS), November 28th.

Free online introduction to AI events (probably more of a share to your friends!)

And if you want to go deeper on chatGPT prompt construction, or just ask questions.