Open models are great but they have a big catch; Synthesia releases Expressive Avatars; generative AI arrives in healthcare; AI assistants and busywork.

The future of AI gadgets is phones; Saudi Arabia and the UAE race to become global AI superpowers; Cohere looks towards the enterprise for success; most global companies are unprepared for AI.

Apr 26, 2024

∙ Paid

The generative AI race welcomed another important contender last week when Meta released two pre-trained and fine-tuned Llama 3 language models with 8B and 70B parameters, respectively. Just in the past three months, Microsoft, Google, Databricks, Cohere and Mistral have all made foundational models available for broad use, creating a growing trend among technology companies.

First, let’s make one thing clear: most of these models were released in the form of open weights coupled with inference code, not open source code and datasets. The difference is important because code represents the well-structured, human-readable text written in a programming language that describes computer systems. Weights are just the output of training runs of a model on data and are almost impossible for a person to understand (you can think of them as very similar to the binaries of traditional software packages). There’s a separate debate between free software and open source software, but we won’t go there today.

Llama 3 has the best models existing today at the 8B and 70B parameter scale.

On its face, releasing the weights of AI models seems like an unequivocally good thing: researchers get access to state-of-the-art technology and can extend the models to create new variants of them. This allows smaller companies, researchers, and hackers to build new applications that wouldn’t have otherwise existed, as foundational models are incredibly data and compute intensive. It also creates an open ecosystem that spurs innovation in unexpected directions as different groups experiment with the models. For example, MBZUAI (the world’s first AI university) built BiMediX, a bilingual model for medical applications, by changing and fine-tuning Mistral’s Mixtral 8x7B mixture of experts model. The Yi-34B model from Chinese startup 01.ai uses Llama's architecture, reducing the effort required to build it from scratch and enabling the utilization of the same tools within the Llama ecosystem.

Open software also has a track record of producing more secure technologies and products, as more people can better stress test the code and discover bugs that wouldn’t have been found in closed environments.

However, the details of how large language models are created and the scale of resources required to build them make AI different to traditional software, raising questions about whether truly decentralized development is possible, even with an open source codebase.

There are three uncomfortable truths that advocates of openness in AI typically avoid.

First, despite opening their models, large tech companies are still likely to wield outsized influence over the trajectory of AI progress. The challenge stems from “the dirty secret of generative AI”: the code of foundational models is becoming more and more irrelevant. What’s more important is the immense computing power and data required to train the foundation models that can then be adapted for downstream applications. Only a handful of companies have the specialized AI supercomputers and training data measured in the trillions of words to create these large models.

So while a company may release the code or weights for an AI model under a permissive license, the ability to materially improve, update, or build new models at that scale will soon be limited only to major tech players. The open source license simply makes it easier for others to repurpose and adapt what has already been created by the company that developed the original foundation.

In addition, the comprehensive training datasets required to create large language models often include web crawled data, books, and databases to which the tech companies have preferential access or have spent years gathering and curating. This data disadvantage for those outside big tech puts up another barrier to entry.

As a result, we could see an AI development ecosystem where a small group of big tech companies open up their models, but still largely control the rate and direction of fundamental progress due to their unmatched resources for retraining and updating. Smaller players and researchers could build innovative applications atop the open sourced foundations, but would remain reliant on the tech giants to advance the technology.

Theoretically, big tech companies could embrace more of an authentic open source approach, working with academics and smaller teams to federate some of the training data and compute resources. But the fiercely competitive nature of the market and the immense economies of scale at play make it unlikely that these companies will cede control over the core model development pipeline.

Secondly, an open weights approach is a way to rapidly commoditize foundational models. If the market is crowded, there’s no faster way to eliminate competition than making a great product free. And given the high costs of training large models, very few startups (and even medium to large sized companies) can justify the price tag of trying to go head-to-head at the model level, unless they’re building for very targeted applications.

And thirdly, AI is a powerful technology unlike anything else we’ve seen in traditional software. As such, it gives bad faith actors unprecedented power to create harmful content at scale. While some debate abstract existential risks, real people are already experiencing the ramifications of unchecked AI today, with explicit deepfakes affecting everyone from high profile celebrities to primary school children. Given the lack of worldwide legislation on AI-related harms and the struggles of internet platforms to moderate AI-generated content, is it wise to continue making this technology openly available to everyone, including potential abusers?

Therefore, as AI capabilities rapidly advance, surface-level openness is unlikely to be enough. Instead, we should encourage real decentralization of power if we are serious about mitigating the risks around the control of such a powerfully transformative technology. For all the real benefits to opening up large AI models, we shouldn't be lulled into thinking the step of opening up alone will prevent the biggest tech firms from controlling the overarching direction and pace of AI progress.

And now, here are this week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

MIT Technology Review: An AI startup made a hyperrealistic deepfake of me that’s so good it’s scary
FT: AI service providers mine a rich seam of demand for efficiencies
New York Times: Generative A.I. Arrives in the Gene Editing World of CRISPR
The Verge: The future of AI gadgets is just phones
CNBC: Less burnout for doctors, better clinical trials, among the benefits of AI in health care
New York Times: ‘To the Future’: Saudi Arabia Spends Big to Become an A.I. Superpower
Bloomberg: Most Global Tech Leaders See Their Companies Unprepared for AI
Fortune: The CEO of OpenAI rival Cohere shakes off the haters: ‘We’re still sort of the underdog’
Bloomberg: Caught Between the US and China, a Powerful AI Upstart Chooses Sides
Axios: AI shakes up corporate boards
WSJ: At Moderna, OpenAI’s GPTs Are Changing Almost Everything
New York Times: How a Virtual Assistant Taught Me to Appreciate Busywork
Washington Post: The AI hype bubble is deflating. Now comes the hard part.

⚙️Computer does

AI in the wild: how artificial intelligence is used across industry, from the internet, social media, and retail to transportation, healthcare, banking, and more

Fortune: Cargill leans on regenerative agriculture and generative AI to feed the planet
TechCrunch: Rabbit’s R1 is a little AI gadget that grows on you
MIT Technology Review: This creamy vegan cheese was made with AI
BBC: How robots are taking over warehouse work
The Verge: The Ray-Ban Meta Smart Glasses have multimodal AI now
TechCrunch: Sanctuary’s new humanoid robot learns faster and costs less
Business Insider: AI is helping Amazon send fewer small items in comically large boxes
TechCrunch: How United Airlines uses AI to make flying the friendly skies a bit easier
Fortune: AI is shaking up how sports like rugby, soccer and cricket are played—and could mint big money for sports clubs
AP: A coffee roastery in Finland has launched an AI-generated blend. The results were surprising

🧑‍🎓Computer learns

Interesting trends and developments from various AI fields, companies and people

Keep reading with a 7-day free trial

Subscribe to Computerspeak by Alexandru Voica to keep reading this post and get 7 days of free access to the full post archives.