Best way to prompt ChatGPT, Midjourney's vary feature, Canva launches ChatGPT plugin

 Dario’s Picks 

• Midjourney introduces Vary. Midjourney (for those who don’t know) is a popular tool that converts natural language into images. The new feature is very similar to Photoshop’s much hyped “generative fill”, which allows you to add or change elements in preexisting images. This will definitely come in handy, as there are so many cases where you generate an image, and want to change just a certain part of it without losing it’s essence. It can also help with character consistency.

‍

‍‍

‍

• Meta releases Code Llama. ChatGPT is not the only kid on the block that knows how to code anymore. Meta just released a code-specialised version of its Llama 2 large language model – benchmark testing shows it surpasses other code-specific LLMs and is on the level with ChatGPT.

• Nvidia just made 6$ billion dollar in profits in Q2 due to unprecedented demand for its generative AI chips. Nvidia is forecasting a staggering 16$ billion in profits next quarter, again due to AI. Despite some clickbaity articles predicting doom and gloom for AI, I think that the technology, the use cases and the market are all aligned for continued, rapid expansion of AI. Perhaps not at the same speed as we’ve seen so far this year – the hype might dampen a bit – but nevertheless at an impressive pace.

‍

‍

 ChatGPT plugins 

• There’s now 943 ChatGPT plugins, of which 23 were launched in the past 7 days.

‍

• Notable launches: Canva (100 million+ users) just launched a ChatGPT plugin, allowing users to search for Canva templates directly inside ChatGPT.

• ChatGPT has already become a staple in the workflow of creators using Canva, particularly for design research (finding aesthetics, colour palettes, copywriting, and more). For example, designers use ChatGPT to help them find and verbalise a specific design aesthetic, then take the keywords and search for templates for their use case inside Canva.
• These users can now search for things like “minimalist yoga instagram post” (and other combinations of styles and publishing formats) and get relevant templates back, using the Canva plugin. I wouldn’t call this revolutionary by any means, but it does add an additional layer of convenience to the research process for designers.
• ChatGPT plugins are proving valuable in different research domains (hence the popularity of plugins for academic research, pdf conversations, and more); and it seems like major companies are slowly recognising this utility.

‍

Categories with most new launches:

• Marketing and SEO - 48 plugins (4 new)
• Data & Research - 79 plugins (3 new)
• Automation and Integration - 26 plugins (3 new)

‍

 In Focus 

The best way to prompt ChatGPT (and other chatbots)

Understanding some of the underlaying technology of how chatbot works can be extremely useful for getting you the answers you need.

Consider this example (try to guess the best way to prompt):

‍

Prompt 1: [Problem/question description] State the answer and then explain your reasoning.
 
Prompt 2: [Problem/question description] Explain your reasoning and then state the answer.

- Example featured in Andrew NG’s newsletter The Batch

‍

You might have noticed that the format of prompt number 2 often yields a better result when working in ChatGPT.

That’s because chatbots generate text based on statistical probabilities, not factual understanding. Asking for the answer first may lead the chatbot to make a quick guess and then attempt to rationalize it. On the other hand, asking for reasoning before the answer encourages the model to generate a more considered response, often resulting in a better guess.

As chatbots develop and the technology gets more mature, you will probably not depend on specific wording/sequences like this to optimize your results. Think about how clumsy new tech can often be in the beginning – that’s likely where we are now with our new robot friends GPT-4, Bard, Claude 2, and so on. So for now, understanding this concept can make your results better, as a user of chatbots, or as a developer integrating LLMs into your application.

‍

 AI Bytes & Resources 

• YouTube embraces AI music.

The company emphasises the need to strike a balance between using AI’s potential to “enhance music’s unique creative expression” and ensure protection of artists and the integrity of their work.

• Hugging Face raises $235M from investors, including Salesforce and Nvidia.

Hugging Face is a community making tools for building AI products – and now has a rockstar line up of investors.

• Meta releases multilingual speech translation model.

Translates your voice into different languages! You can test it here.

• OpenAI launches fine-tuning for GPT-3.5 Turbo + Scale AI partnership.

Fine-tuning makes it possible for devs to bring their own data to tailor GPT to specific use cases. Scale AI was brought on board to support with enterprise-level fine-tuning.

• EU started enforcing new Digital Services Act (DSA) last Friday.

Major platforms are adapting their services accordingly. For example, Meta and TikTok will start allowing users to get their feeds (Reels, Stories, Search, etc.) in chronological order rather than by the platform’s algorithm.

• Potential NYT lawsuit could force OpenAI to wipe ChatGPT and start over.

The potential lawsuit is noteworthy, although the title is a bit clickbaity. Surely there must be a more reasonable outcome than “wiping” out ChatGPT.

• From Mad Men to machines? Big advertisers shift to AI.

Massive companies like Nestlé and Unilever are increasingly using AI to cut costs and enhance productivity. "The savings can be 10 or 20 times", according to WPP CEO, Mark Read.

‍