Keeping up with a fast-moving industry like AI is difficult. So, until AI can do it for you, here’s a helpful summary of recent stories in the world of machine learning, along with notable research and experiments that we didn’t cover on our own.
This week, Google flooded the channels with announcements about Gemini, its new flagship multimedia AI model. It turns out that it’s not as impressive as the company initially imagined – or rather, the “light” version of the model (Gemini Pro) that Google launched this week is not. (It doesn’t help that Google is faking a demo of the product.) We’ll reserve judgment on the Gemini Ultra, the full version of the model, until it starts making its way into various Google apps and services early next year.
But enough talk about chatbots. I think the biggest thing is the funding round that barely squeezes the workweek: Mistral AI Raised €450 million (~$484 million) at a $2 billion valuation.
We’ve covered Mistral before. In September, the company, co-founded by Google DeepMind and Meta alumni, launched its first model, the Mistral 7B, which it claimed at the time to outperform others of its size. Mistral closed one of its largest seed funding rounds in Europe to date ahead of the fundraising on Friday — and has yet to launch a product.
Now my colleague Dominique rightly points out that the fortunes of Paris-based Mistral are a red flag for many concerned with inclusivity. The startup’s co-founders are all white and male, and academically fit the homogeneous, privileged image of many of those who work at The New York Times. He slammed hard existing A change maker in the field of artificial intelligence.
Meanwhile, investors appear to view Mistral – as well as its sometime rival, German firm Alpha – as an opportunity for Europe to plant its flag in the (for now) fertile AI ground.
To date, the largest and best-funded generative AI projects have been in the United States. OpenAI. Anthropic. Artificial intelligence turn. cohere. The list goes on.
Mistral’s good fortune is in many ways a microcosm of the struggle for AI supremacy. The European Union (EU) wants to avoid being left behind in another technological leap while at the same time imposing regulations to guide technology development. As did German Vice Chancellor and Minister of Economic Affairs Robert Habeck recently Quoted “The idea of having our own sovereignty in the AI sector is very important,” he said. [But] “If Europe has the best regulation but no European companies, we will not win much.”
The divide between entrepreneurship and regulation came into sharp focus this week, as European Union lawmakers tried to reach agreement on policies to limit the risks of artificial intelligence systems. Lobbyists, led by Mistral, have pushed in recent months for a complete regulatory carve-out of generative AI models. But EU lawmakers have resisted such an exemption — for now.
There’s too much reliance on Mistral and its European rivals, all that being said; Industry watchers – and US lawmakers – will no doubt be closely watching the impact on investments once EU policymakers impose new restrictions on AI. Could Mistral one day grow to challenge OpenAI with established regulations? Or will the regulations have a chilling effect? It’s too early to say, but we’re keen to see for ourselves.
Here are some other noteworthy AI stories from the past few days:
- New AI Alliance: Dead, on open source CutsIt wants to spread its influence in the ongoing battle for AI mind-sharing. The social network has announced that it is teaming up with IBM to launch the AI Alliance, an industry body to support “open innovation” and “open science” in AI – but there are many ulterior motives.
- OpenAI is heading to India: Evan and Jagmeet reported that OpenAI is working with former Twitter India head Rishi Jaitley as a senior advisor to facilitate talks with the government on AI policy. OpenAI is also looking to set up a local team in India, with Jaitly helping the AI startup navigate the Indian political and regulatory landscape.
- Google launches AI-assisted note-taking feature: Google’s AI-based note-taking app, NotebookLM, which was announced earlier this year, is now available to US users 18 or older. To mark the launch, the beta app has gained integration with Gemini Pro, Google’s new large language model, which Google says will “help with document understanding and reasoning.”
- OpenAI is under regulatory scrutiny: The cozy relationship between OpenAI and Microsoft, a key backer and partner, is now the focus of a new investigation launched by the UK Competition and Markets Authority into whether the two companies are actually in a “relevant merger position” following the recent drama. The FTC is also said to be examining Microsoft’s investments in OpenAI in what appears to be a coordinated effort.
- Well done AI question: How can you reduce biases if biases present in its training data are built into an AI model? Anthropic suggests that you ask nicely for your satisfaction, please do not discriminate Or someone will sue us. Yes, really. Devin has the full story.
- Meta rolls out AI features: Along with other AI-related updates this week, Meta AI, Meta’s generative AI experience, gained new capabilities including the ability to generate images on-demand as well as support for Instagram Reels. The first feature, called “reimagine,” allows users in group chats to recreate AI images using prompts, while the latter can turn to Reels as a resource as needed.
- The speaker gets the money: Ukrainian prosthetic voice startup Respeecher — perhaps best known for being chosen to imitate the voice of James Earl Jones and the iconic voice of Darth Vader in the Star Wars animated show, and later the voice of a younger Luke Skywalker in The Mandalorian — is finding success despite not only falling bombs, writes Devin said their city was a wave of hype that sparked sometimes controversial competitors.
- Liquid Neural Networks: An MIT spin-off co-founded by robotics star Daniela Ross It aims to build general-purpose AI systems powered by a relatively new type of AI model called a liquid neural network. The company, called Liquid AI, raised $37.5 million this week in a seed round from backers including WordPress parent company Automattic.
More machine learning
Orbital images are an excellent playground for machine learning models, since satellites these days produce more data than experts can keep up with. Researchers are looking at EPFL Better identification of plastic carried by the oceanIt’s a big problem but very difficult to track systematically. Their approach isn’t shocking, as they train a model on specific orbital images, but they have improved the technique so that their system is vastly more accurate, even when there is cloud cover.
Finding it is only part of the challenge, of course, and removing it is another, but the better people and organizations are when they do the actual work, the more effective they will be.
However, not every domain contains a lot of images. Biologists in particular face the challenge of studying animals that have not been adequately documented. For example, they may want to track the movements of a particular rare species of insect, but since there are no images of that insect, automating the process is difficult. Collection at Imperial College London It is working on using machine learning in cooperation with the game development platform Unreal.
By creating photorealistic scenes in Unreal and filling them with 3D models of the creature in question, whether it’s an ant, a stick insect or something larger, they can generate random amounts of training data for machine learning models. Even though the computer vision system has been trained on synthetic data, it can still be very effective in real-world footage, as they video Offers.
You can read their paper at Nature Communications.
However, not all images created are very reliable, As researchers from the University of Washington found. They systematically pushed the open source image generator Stable Diffusion 2.1 to produce images of a “person” with different constraints or locations. They showed that the term “person” is disproportionately associated with lighter-skinned Western men.
Not only that, but some locations and nationalities have produced troubling patterns, such as sexualized images of women from Latin American countries and the “almost complete erasure of non-binary and indigenous identities.” For example, requesting photos of “person from Oceania” results in white men rather than indigenous people, despite the large number of indigenous people in the region (not to mention all other non-white people). This is all a work in progress, and it is important to be aware of the biases inherent in the data.
Learning how to deal with biased and questionably useful models is something that is on the minds of many academics—and the minds of their students. This interesting chat with Yale University English professor Ben Glaser is a refreshingly optimistic look at how to use things like ChatGPT constructively:
When you talk to a chatbot, you get this weird, mysterious image of culture back. You may get counterpoints to your ideas, and then you need to evaluate whether those counterpoints or evidence supporting your ideas are actually good. There is a kind of knowledge in reading these outputs. Students in this class gain some literacy.
If everything is cited, and you develop a creative work through some elaborate back and forth effort or programming including these tools, you are doing something wild and interesting.
When should they be trusted in a hospital, for example? Radiology is a field where artificial intelligence is frequently applied to help quickly identify problems in body scans, but it is not infallible. So how should doctors know when to trust a model and when not? And MIT seems to think it can automate that part, too -But don’t worry, it’s not another AI. Instead, it is a standardized, automated onboarding process that helps determine when a clinician or a particular task finds an AI tool useful, and when it gets in the way.
Increasingly, AI models are being asked to generate more than just text and images. Materials are one area where we’ve seen a lot of movement, models are great at coming up with potential candidates for better catalysts, polymer chains, etc. Startups are starting to get involved in this, however Microsoft has also released a model called MatterGen This is “specifically designed to produce new and stable materials.”
As you can see in the image above, you can target a lot of different qualities, from magnetism to reactivity to volume. No need for a Flubber-like accident or thousands of lab runs – this model can help you find a suitable material for an experiment or product in hours instead of months.
Google DeepMind and Berkeley Lab are also working on this kind of thing. It is quickly becoming standard practice in the materials industry.