How to pick your perfect AI

Welcome to AI Communicator!

Today’s issue is the first in a three-part series on getting started with AI for communications:

• Part one - How to choose the right AI tools.

• Part two - How to approach learning about AI

• Part three - Putting it all together with a personal AI audit

How to pick your perfect AI

So many AI apps!

Even after nine months learning about and experimenting with artificial intelligence, I still get bewildered by the range of options.

The leading models regularly get new capabilities, features and names. 

It makes the AI landscape even harder to navigate - which is the reason most people default to ChatGPT. 

It's currently the most popular Large Language Model (LLM) with 180 million users globally.

And more than 99 per cent of them are estimated to use the free version ChatGPT-3.5.

Most likely, this is not the best option for you.

So what is?

First, you want a state-of-the-art (SOTA) model.

Not using something at or around the level of GPT-4 might be the reason you’ve been disappointed with the quality of answer, the level of hallucinations, or its inability to do something seemingly simple. 

Admittedly, the other reason is that LLMs are odd that way.

They do some things exceptionally well, which can make it jarring when they fail on the basics - like simple arithmetic, or writing to precise word counts.

Using a SOTA model doesn’t solve these issues, but it does reduce them.

A second consideration is thinking about which additional features you would find most helpful. Here are a few:

• Vision - the ability to recognise and describe images. 

• Voice - enabling you to talk to the LLM, an experience that is a lot closer to talking to a real person than chatting with Siri or Alexa.

• Image generation - creating images from text prompts

• Large context window - how much text you can give it at one one time

• Web browsing - LLMs all have a training cut off date. Sometimes this is more than a year ago which means if you want the latest information you will need it to access the web.

A third key consideration is data security. 

Does the LLM keep your inputs and outputs secure or use them as training data? If this is of prime importance ensure you study the privacy policy of your chosen model.

With these factors in mind, let’s look at the leading contenders. 

Microsoft Copilot

There’s a free for all version of Microsoft Copilot available alongside Bing search.

It is based on the same underlying model as ChatGPT and gives you access to GPT-4 Turbo in Creative and Precise modes. The Balanced mode uses GPT-3.5. 

It is currently the best free option for accessing a SOTA GPT-4 class LLM.

Copilot comes with a nice notebook feature that lets you quickly test the effect of altering your prompt.

You also get daily credits for image creation using DALL-E 3.

And you can chat with it via voice instead of typing.

There is a Copilot Pro version that has some added benefits, such as the ability to create GPTs.

At the most basic level, these GPTs can act as shortcuts for any set of instructions you regularly use.

Given that it was first introduced as Bing Chat it’s no surprise that Copilot can browse the web for up-to-date results. Precise is the best setting when using it for factually accurate browsing.

If you already have Microsoft 365, there is also a paid-for Copilot option which brings AI into Outlook, PowerPoint, Word and Excel. 

Anyone using Copilot at work who is signed in as an Enterprise user, can have greater confidence that any sensitive information they enter is secure and won’t be retained or used for training.

Ensure you use the right settings and version if data privacy is important to you.

Claude 3

Claude 3 was launched on 4 March 2024. It comes in three flavours - Opus, Sonnet and Haiku.

Opus is the most powerful and is generally considered to outperform ChatGPT on many tasks .

It writes and chats more naturally (i.e. human-like) than ChatGPT.

A plus, or minus, of this is that it sometimes leads to an uncanny feeling that Claude 3 has self-awareness.

All versions of Claude 3 have a larger context window of 200,000 tokens (a token is roughly 0.75 of a word).

It has vision, with some reports that its image recognition is superior to  ChatGPT’s for extracting information from pictures.

Access to Opus requires a Pro subscription.

Sonnet comes for free and is a better free option than ChatGPT-3.5 for many use cases - definitely for writing.

Claude 3 says it only uses your data for training in certain circumstances. Overall, it provides less controls and guarantees for anyone concerned about entering sensitive or confidential information.

ChatGPT

For many people, ChatGPT-4 still wears the LLM crown.

Accessed through a paid-for Plus subscription it gives you a SOTA model, the ability to create custom GPTs and share these with other Plus users, vision, voice for conversations through the mobile apps, web browsing, and the added bonus of DALL-E 3 image creation.

Access to GPT-4 and many of these features only come with a Plus subscription, but it is a worthwhile upgrade.

Despite having my head turned every time a competitor upgrades, I find myself coming back to ChatGPT as my default LLM.

The context window for ChatGPT is a lot smaller than Claude 3 at 32,000 tokens.

ChatGPT gives you some control over how your data is used as you can turn off Chat History and Training in the settings.

Google Gemini 

Gemini Advanced, which requires a paid-for subscription, is a SOTA GPT-4 class LLM.

Its most impressive feature isn’t yet available to the public - a one million token context window. This allows it to take in books of text and recall it with high accuracy.

Until that rolls out more widely, most users will be working with a 32,000 context window.

Overall, Gemini Advanced is a solid alternative to ChatGPT-4.

I find that its answers are well structured and often more helpful and constructive than ChatGPT’s. 

Many people prefer it for creative and copywriting tasks.

It can also search your Gmail, Google Drive and Docs and provide answers based on these.

It has image recognition and a few other nice touches:

• Pin chats in the sidebar 

• Three drafts for  every answer

• Fine tune responses (shorter, longer, simpler, etc)

• Edit and rerun prompts

• Button to double-check responses

Gemini comes in a free version using Gemini Pro and the paid-for Advanced using Gemini Ultra, which is comparable to GPT-4. You can try Advanced for free for two month

As for data security, Google advises: “Please don't enter confidential information in your conversations or any data you wouldn't want a reviewer to see or Google to use to improve our products, services, and machine-learning technologies.”

Pi

Finally, an honourable mention for Pi from Inflection. It received an update on 7 March 2024 that puts it just behind GPT-4 on a number of benchmarks. 

The pitch from Inflection is that Pi is “a personal AI, designed to be empathetic, helpful, and safe.” 

It’s free and works best on voice mode where its conversational skills come to the fore. 

I’ve used it for brainstorming, problem solving and downloading whatever is on my mind.

It is supportive and constructive - which makes it enjoyable to use.

Again, beware that sensitive data you enter can be used for training purposes.

So which one?

As you might’ve guessed, there’s no simple answer.

It all depends on your use cases, budget, priorities and data privacy preferences.

There’s no single universally agreed benchmark for the best LLM right now and much of the opinion out there, including my own, is based on personal experience and general feelings.

And this article is far from comprehensive. For example, I haven’t mentioned the models’ coding skills.

But hopefully it gives you a starting point for your own assessment.

Finally, there’s a lot to be said for being familiar with at least two.

Feature lists and performance change with each new model update..

You don’t want to get stuck with yesterday’s AI. 

Having a backup is also important as these services can and do experience downtime.

Good luck. I hope you pick a winner.

I’d love to hear what you think about the newsletter - or anything you’d like to see featured in future issues. Reply to this email with feedback or questions.

Thanks for reading.

Calvin

P.S. I have a favour to ask. If you are enjoying the newsletter, please share it with colleagues who might be interested. They can sign up for the newsletter here.