A large week for large models

As weeks go in AI and technology, this is peak summer-beach-vibes busy. The movers and shakers: Meta's latest Llama launch, with multilingual support across WhatsApp, Instagram and Messenger; Cohere and Harvey's gigantic investment rounds as enterprise opens up to generative AI; and Mistral launching Large 2 as OpenAI announces GPT Search.

The so-called arms race is well underway. To what end, though? At this rate we're heading for rapid commoditisation, which will make for an interesting market dynamic against these huge investments. As we saw this week, lower-than-expected Q2 tech earnings — primarily citing a shift in focus to building out foundational models — met a sober market reaction, with stocks down across Nvidia, Alphabet, Microsoft, Apple and Tesla, even as CEOs held firm and pointed to existing AI earnings and the patience needed to recoup enormous investment costs.

Meta's Llama 3.1

Meta launched Llama 3.1 on 23 July, promising better, faster and cheaper than the closed models at OpenAI and Anthropic. Zuckerberg used an open letter to stake out his unambiguous support for why open-source AI will overtake closed models. The thesis: just as Linux evolved from a cheaper, less advanced alternative to closed Unix systems into the industry standard for cloud and mobile, open-source AI will rapidly close the gap with proprietary models.

Llama models were always intended to work as part of an overall system that can orchestrate several components, including calling external tools. Our vision is to go beyond the foundation models to give developers access to a broader system that gives them the flexibility to design and create custom offerings.

Llama 3.1 405B — named for its 405 billion parameters — is on par with, and in some cases outperforming, Claude 3.5 and GPT-4o. More interesting is Meta's move to enable an open, interoperable ecosystem, including a licence change that lets developers use its outputs to improve other models. It was trained on over 16,000 H100 GPUs, so no surprise the NVIDIA partnership continues to shine, and 25 enterprise partners including AWS, Databricks, Dell and Azure will tout services from launch.

Mistral goes Large again

Mid-week, French AI company Mistral released Large 2, reported (by Mistral) to outperform Cohere's Command R+ in multilingual document applications — one of the most significant business use cases for LLMs today. Mistral raised €600M at a €5.8B valuation in June and co-announced a deeper integration with Google Vertex, which is itself already powered by Google's Gemini.

Cohere raises $500M at a $5B valuation

Canadian-HQ'd enterprise AI company Cohere announced a $500M round, valuing it at $5.5B and making it one of the world's most valuable AI companies — though a fraction of OpenAI's $90B. Unlike OpenAI, Mistral and Anthropic, Cohere is entirely focused on enterprise, billing itself as AI optimised for enterprise generative AI, search and discovery, and advanced retrieval. Capital came from Oracle, Salesforce Ventures, Fujitsu, NVIDIA and Canadian pension manager PSP. Its annualised revenue stood at $35M in March, so it has some very large bets that need to pay off in what is the most aggressively paced investment space we've seen in 25 years.

Competition authorities in a spin

The EU, US and UK competition authorities issued a joint statement on competition in generative-AI foundation models, which reads more like a faint cry from the side lines than a rally for reason. Sensibly, they all align on the need for interoperability and choice — music to the ears of Mr Zuckerberg. Not every authority is so laid back: Brazil halted Meta training its models on Brazilian public data.

Harvey raises $100M

At just two years old, enterprise legal-AI platform Harvey raised a mega-round from Google Ventures, OpenAI, Kleiner Perkins, Sequoia, Elad Gil and SV Angel — bringing generative AI to legal and professional services.

OpenAI releases SearchGPT

Perhaps the most interesting news in arms-race terms is OpenAI's SearchGPT prototype. In what will likely be a winner-takes-all consumer race, search is where human–internet interaction lives — Google is searched some 1.2 trillion times a year. Cohesively integrating LLM chat-search into a multimodal, multilingual experience could create a generational foundation for how we interact with the world's data. Can SearchGPT grab share and develop beyond prototype quickly enough?

Data-centre build-out continues apace — Ireland now reports data-centre energy consumption above a fifth of total national consumption, a sign of foundational scaling issues to come. And on a lighter note, the consensus this week is that AI still can't make music. Or can it? Till next time.