What are Microsoft, Google, Apple, Meta and Salesforce each doing with AI in 2023?

Six months after ChatGPT arrived, every large technology company has had to show its hand. Their strategies differ more than the headlines suggest. Some build their own models, some rent them from partners, some give them away to researchers, and one keeps its work almost entirely inside your phone. Knowing which is which helps you predict where AI will turn up in the software you already use.

Here is the picture as of this week.

The snapshot

Microsoft and Google: the head-to-head

Microsoft relies on its partnership with OpenAI. In February it launched a new Bing and Edge in preview, with a chat mode built on an OpenAI model tuned for search that cites its sources. In March it announced Microsoft 365 Copilot, which combines large language models, including GPT-4, with your own documents, email and calendar. Examples include a first draft in Word, a presentation built from a document, or a summary of a Teams meeting. It is being tested with a small group of customers, and pricing has not been announced.

Google builds its own models. At its I/O conference on May 10 it introduced PaLM 2, which Google says improves on reasoning, coding and multilingual work, with training text in over 100 languages. PaLM 2 comes in four sizes, named Gecko, Otter, Bison and Unicorn; the smallest is light enough to run on a phone, even offline. Google says over 25 products now use it, including Bard, which the same day dropped most of its waitlist. There are also specialized versions: Med-PaLM 2 for medical questions and Sec-PaLM for security analysis. Developers can sign up for the PaLM API.

Meta and Salesforce: research models and business tools

Meta took a different route in February with LLaMA, a family of four models from 7 billion to 65 billion parameters. Rather than a chatbot, it released them to researchers under a noncommercial license, arguing that smaller models need less computing power and so let more people study how these systems work. Access is granted case by case.

Salesforce focuses on business software. Its Einstein GPT, announced on March 7, adds generative AI to its customer relationship management (CRM) products, the tools companies use to track customers and sales. It combines Salesforce's own models, partner models starting with OpenAI's, and a company's own customer data. A salesperson might get a draft follow-up email; a support agent, a suggested reply; a developer, help writing code. It is in a closed pilot. Salesforce and OpenAI also announced a ChatGPT app for Slack that summarizes conversations and helps draft messages.

Apple: machine learning on the device

Apple has announced no chatbot. On May 16 it previewed accessibility features, coming later this year, that rely on on-device machine learning, meaning the processing happens on your iPhone, iPad or Mac rather than on a company server. Apple presents that as a privacy protection. Personal Voice lets someone at risk of losing their speech, for example after an ALS diagnosis, record about 15 minutes of reading on an iPhone or iPad and then type messages that are spoken in a voice that sounds like their own. Point and Speak uses the camera and on-device machine learning to read aloud the labels on buttons, such as a microwave's, as a user points at them.

Apple's choice shows that AI doesn't have to mean a chat window. Much of it will appear as features that simply work better, without being labeled as AI at all.

Which of these will reach you first?

For most people, Google and Microsoft are closest. Bard is open in English in more than 180 countries and territories, and Bing's chat is in preview, so you can try both this week with a free account. If you use Gmail or Docs, Google has said writing help powered by PaLM 2 is on its way, and Microsoft 365 users can expect Copilot once testing ends, though neither has given a firm date for everyone.

If you work in sales or customer service, Salesforce's Einstein GPT may arrive through your employer, once it leaves its pilot, rather than through anything you choose yourself. The same goes for the ChatGPT app in a Slack workspace your company runs.

Apple's features are announced for later this year, and they are specialized tools rather than general assistants. Meta's LLaMA is the furthest from everyday users today: it is a research release, so you are more likely to meet it indirectly, through projects built by researchers and hobbyists, than as a Meta product. Whichever arrives first for you, the habits that work in ChatGPT carry over: be specific in what you ask, and check what comes back.