more models and more Open Source


OpenAI remains the dominant player. But for Andreessen Horowitz, companies are now investing in other LLMs that will reshape the generative AI landscape. (Photo: Jonathan Kemper / Unsplash)

According to the Andreessen Horowitz Foundation, large companies increasingly rely on a diversity of LLMs and favor Open Source models. Two factors undermining OpenAI’s dominance.

AdvertisingGrowing budgets and a variety of technologies that are developing significantly. This is the observation made by the American foundation Andreessen Horowitz (or a16z), founded in 2009 by Marc Andreessen and Ben Horowitz. Based on a survey of 70 of the 500 largest US companies – accompanied by in-depth interviews with leaders of the same organizations – the foundation’s study estimates that, based on encouraging initial results during experiments with GenAI, companies will increase investments in technology. Almost all companies we spoke to planned to increase their spending by a factor of 2 to 5 in 2024 to support the spread of more applications in production, writes a16z, which explains that the companies’ reserves regarding technology (hallucinations, handling of confidential data) is not enough to slow this growth.

Better, according to this study, US companies are devoting budget envelopes to GenAI, although the phenomenon is still in the minority: only 19% of organizations have created a specific budget for LLMs, 37% will further draw from budget innovation or in other IT – budgets. “On a much smaller scale, we’re also starting to see some managers implement GenAI budgets based on staff savings, especially in customer service,” Andreessen Horowitz points out, citing the case of an unidentified company that would save $6 per employee. handled calls by its customer service equipped with an LLM, which would push this organization to multiply its investment in technology by 8.

The challenge of calculating ROI

Note that this growth in investment is due more to perceptions than solidly established KPIs. 56% of companies believe the ROI of technology is positive… without measuring it. In the short term, managers are still implementing this technology and determining the best metrics to quantify results, but over the next two to three years, return on investment will become increasingly important. While managers are still searching for the answer to this question, many are relying on the words of their employees, who claim to make better use of their time thanks to technology, a16z analyzes.

The major language models most widely used among the 70 major US companies surveyed by a16z.

The investigation of the investment fund also zooms in on the companies’ technological choices. First observation: the undivided dominance of OpenAI is crumbling and we are moving towards a more diversified technological landscape. Today, when we talk to business leaders, they all test – and in some cases even use in production – several models that allow them to adapt use cases based on performance, size and cost, to avoid technology lock-in and too fast to take advantage of advances in a rapidly evolving field, the authors write The studio. Thus, 57% of companies already use, in production or for their prototypes, 4 or more LLM models, often with flexible architectures that allow them to move from one to the other. If OpenAI is still at the forefront – especially when we focus on applications in production – Google and Meta are reaching significant levels of implementation. And the French Mistral rears its ugly head, in 5th place (with 17% of companies using its models, mostly for testing).

AdvertisingOpen source to strengthen control over models

And if we trust the trends outlined by Andreessen Horowitz, this landscape should continue to evolve. Because more than 8 out of 10 surveyed decision-makers expressed their desire to strengthen the use of Open Source models, such as Llama or Mistral. A reversal of the trend according to the foundation, which estimates the market share of proprietary models at 80 to 90% in 2023, with the majority of this share going to OpenAI. And this choice of open source code is not dictated by costs, but primarily by the control options that these models offer (safety of confidential data and understanding of the results) and by their customization options. With the advent of high-level Open Source models, most organizations choose not to train their own LLM from scratch and instead use Retrieval Augmented Generation (RAG) or fine-tune a model for their specific needs,” emphasizes a16z.

Overall, the investment fund estimates that companies’ annual spending on generative artificial intelligence—excluding features embedded in third-party software—will rise from $1.5 to $2 billion by the end of 2023 to $5 billion by the end of this year.

Share this article

Source link

Leave a Comment