Microsoft, the tech giant renowned for its innovative prowess, has unveiled a new stride in the realm of artificial intelligence (AI) development. The company, represented by its ticker symbol MSFT.O, has commenced the training of a robust, in-house AI language model, poised to rival the formidable offerings from Alphabet’s Google (GOOGL.O) and OpenAI. This groundbreaking revelation surfaced courtesy of an exclusive report by The Information on Monday.
Dubbed internally as MAI-1, this new AI marvel is under the astute supervision of Mustafa Suleyman, a luminary figure revered for his co-founding role at Google DeepMind and his tenure as CEO of the AI startup Inflection. The intricate details of MAI-1’s exact purpose remain shrouded in mystery, awaiting revelation contingent upon its performance. Sources privy to Microsoft’s internal affairs, as cited in the report, suggest that a sneak peek of this prodigious creation might be on the horizon, potentially gracing the stage at Microsoft’s Build developer conference later this month.
Despite fervent inquiries, Microsoft maintained a veil of secrecy, declining to provide any official commentary to Reuters. However, insiders hinted that MAI-1 is poised to eclipse its predecessors in both scale and sophistication, promising a quantum leap in AI capabilities. This expansion into larger AI models is not without its financial implications, as the report suggests MAI-1 will incur substantially higher costs compared to Microsoft’s previous ventures into open-source AI models.
Not content with resting on its laurels, Microsoft recently introduced Phi-3-mini, a scaled-down AI model, indicative of its strategic efforts to cater to a broader clientele with cost-effective solutions. Bolstered by substantial investments in OpenAI, Microsoft has seamlessly integrated the latter’s ChatGPT technology across its suite of productivity software, positioning itself as an early frontrunner in the generative AI arena.
Fueling the engine of innovation, Microsoft has spared no expense in bolstering its AI infrastructure, reportedly housing vast clusters of servers armed with Nvidia’s (NVDA.O) graphic processing units. These formidable resources, coupled with copious amounts of data, are instrumental in refining the MAI-1 model to unprecedented levels of proficiency.
According to insider revelations, MAI-1 is slated to boast an impressive 500 billion parameters, dwarfing its predecessors. For context, OpenAI’s GPT-4 purportedly flaunts a staggering one trillion parameters, while Phi-3 mini clocks in at 3.8 billion parameters. Microsoft’s strategic maneuverings in the AI landscape have been further underscored by the appointment of Mustafa Suleyman in March, entrusted with spearheading the newly minted consumer AI unit. The recruitment drive also saw the influx of several talents from Inflection, hinting at a convergence of expertise poised to propel Microsoft to greater heights in the AI domain.
While the genesis of MAI-1 does not directly stem from Inflection, there are indications that it may leverage training data from the erstwhile startup, underscoring Microsoft’s commitment to leveraging diverse resources in its quest for AI supremacy.