Rumour: Microsoft may be developing its own, cheaper LLM

Seeking an alternative to OpenAI's expensive technology

Rumour: Microsoft may be developing its own, cheaper LLM

AI is extending its reach across business and society as Big Tech firms and their backers pursue ambitious plans.

A flurry of rumours have claimed Microsoft has a developed a dependence on artificial intelligence firm OpenAI, which it wants to shake because its revolutionary technology is too expensive to run.

Microsoft refused to clarify reports that spread after The Information claimed the software giant was developing its own large language models (LLMs), the systems behind the AI boom, despite investing billions in ChatGPT-developer OpenAI.

The website quoted an unnamed Microsoft source, and another unidentified former Microsoft employee, in its story.

The leak coincided with big AI announcements by Microsoft rivals Meta and Amazon. Meta, on Wednesday, launched AI assistants for its social media tools WhatsApp, Messenger and Instagram, as well as coming Ray Ban Smart Glasses and its Quest 3 virtual reality goggles.

Amazon Web Services, on Thursday, launched Amazon Bedrock, a system offering developers access to a choice of AI platforms that rival ChatGPT, the OpenAI system that spurred industry and attracted Microsoft's investment.

The Amazon service options include both its own LLM and Meta's Llama 2. At the same time Amazon announced an AI coding assistant to rival Microsoft's, called CodeWhisperer.

AI hardware to rival smartphones

Separately, the Financial Times has quoted unattributed sources claiming that OpenAI founder Sam Altman had been brainstorminging with former Apple iPhone designer Jony Ive over plans to produce a consumer hardware device, which would do for AI what smartphones did for communications.

The FT claims Japanese tech investor SoftBank Group is preparing to invest $1 billion in the plan. SoftBank would not confirm its story.

Also this week, OpenAI announced that it had restored ChatGPT's ability to incorporate live search results into its deliberations.

ChatGPT's answers have previously been limited to information current up until the time of its last web crawl, in September 2021. That meant that, for example, developers using it as a coding assistant were generating code that might be based on old libraries and information.

The live web feature appeared briefly in May, but was quickly dropped. It appeared instead as a feature of Microsoft's Bing search engine.

OpenAI says it listened to feedback and revised the way its bots searched the web, ensuring that they respected robots.txt files, by which websites can request that web crawling bots pass them over or limit their searches to defined areas.

The Microsoft rumours in The Information claimed that the software firm is developing LLMs that are smaller and cheaper than ChatGPT, to be used as office productivity software assistants.

High performance AIs such as ChapGPT bring high quality answers at the cost of a high demand on compute resources, and immense cost in training the models. ChatGPT's most sophisticated AI, GPT-4, is reported to have cost $50 billion to train.

Experts, meanwhile, say less sophisticated models can be employed more cheaply and effectively for simpler, everyday tasks than the advanced applications to which software developers the world over have been putting ChatGPT.