Generative artificial intelligence emerged as a central theme at a global tech conference.

The number of AI and machine learning-focused startups among the more than 70,000 delegates at Web Summit 2024 in Lisbon, Portugal, significantly increased compared to last year, indicating how the technology is driving a new wave of young companies eager to explore opportunities in this evolving field.

The role of foundation models, innovations in search technology, the emergence of AI wrappers and advancements in humanoid robotics were among the trends discussed at the event.

"This is an incredible opportunity for startups and developers," said Qualcomm CEO Cristiano Amon. All applications will provide an AI-first experience five years from now, Amon said at a press conference.

Model agnostic

Among startups developing GenAI applications, no single foundation model is dominating. Startups often select foundation models based on cost over performance, and OpenAI LLC's ChatGPT is one of the most cost-effective and rapid methods to launch a product quickly at scale. Startups looking to fine-tune their models frequently turn to Meta Platforms Inc.'s Llama for this purpose. Established companies typically provide a range of foundation models for customers to choose from.

A startup founder who spoke to S&P Global Market Intelligence uses Alphabet Inc.'s Gemini and said the performance of all leading foundation models is comparable.

The commoditization of current foundation models has prompted providers to rush to build the next, potentially better iteration in the hope. Meta Platforms and X.AI Corp. each have clusters of 100,000 GPUs to train the next generation of models despite growing skepticism that the current cohort could see a step-function improvement.

Re-thinking search

Startups such as Perplexity AI Inc. and SuSea Inc. demonstrated early success in leveraging GenAI to offer a novel search experience. During the Web Summit, European search firms Ecosia and Qwant announced a merger to develop a superior search engine to compete with Alphabet's Google.

The enterprise search market is nascent. The surge in organizational data created demand for tools that enable rapid information retrieval, while advancements in AI and machine learning have enhanced search capabilities.

Search is of increased interest to investors. Glean Technologies Inc. and Coveo Solutions Inc. are early leaders in this quickly developing field, with the former recently raising approximately $200 million, and competitors are emerging. Karandip Singh, founder of Italian enterprise search startup E-Plato Srl, told Market Intelligence that one of its key advantages over foreign peers is that large local enterprise customers trust local vendors with their data.

AI wrappers

The emergence of so-called AI wrappers was another recurring topic of discussion.

While they often present themselves as AI companies, these vendors primarily rely on third-party APIs to address specific use cases without having anything proprietary, such as data or unique software capabilities. This trend poses challenges for investors and potential acquirers seeking genuine AI capabilities.

An investment banker told Market Intelligence that many clients wish to acquire AI capabilities but struggle to differentiate between genuine AI companies and those that merely follow the trend. Venture capitalists have expressed concerns that the rapid pace of the AI sector makes it challenging to keep up.

Kevin Jiang, managing partner at VC Mangusta Capital, said during a panel discussion that he focuses on three key features when investing in the AI application layer: proprietary data, a capable management team and proprietary integrations — specifically, the ability to swiftly integrate the AI solution into the customer's software stack.

Open-source vs. closed-source

Open-source foundation models facilitate the rapid development of new use cases and contribute to the democratization of AI, but there is a theoretical concern that the availability of their source code to the public could enable the spread of harmful content. Meta's Llama is an open-source model, while like other closed-source software, ChatGPT is a proprietary model with a restricted source code.

The crowd at the Web Summit overwhelmingly supports open-source models.

"If you make a list today of all the misuse of AI — the deepfakes, the fake news — you will be pretty much surprised that all of them are done with closed-source models," Thomas Wolf, co-founder and chief science officer of open-source machine learning platform Hugging Face, said during a panel discussion.

Barriers to using closed-source can be lower, as open-source models can require additional investment and technical expertise. Some smaller startups told Market Intelligence that they are using OpenAI’s ChatGPT due to its lower initial development costs but plan to develop their own models using open-source technology as they become more mature.

Tackling the to-do list

The integration of GenAI and robotics is, for the first time, poised to deliver consumer humanoid robots capable of assisting with daily household chores. Advances in speech recognition and natural language understanding have made the dream of humanoid robots a reality.

“The difficult thing that has held up robotics has not been hinges and motors and actuators — it’s the intelligence in them,” said Max Tegmark, an AI and physics research professor at MIT.

Figure AI Inc., a developer of humanoid robots, raised $675 million this year at a post-money valuation of $2.67 billion, according to Market Intelligence data. Elon Musk's Tesla Inc. showcased its humanoid robots earlier this year, and the Optimus robot is expected to move into limited production in 2025. Musk said it will cost between $20,000 and $30,000.

Agility Robotics Inc. unveiled a robot capable of following instructions to perform simple human tasks at the Web Summit. Agility CEO Peggy Johnson stated that the physical structure of their robot, "Digit," is complete, and the current focus is on enhancing its intelligence through AI.

"With AI, we can now start to improve Digit’s semantic intelligence," Johnson said. "In actual use cases in warehouse facilities, this means it can do one job in the morning and another job in the afternoon."