Keeping up with an industry as fast-paced as AI is a huge challenge. Until an AI can do it for you, here’s a handy round-up of last week’s stories in the world of machine learning, as well as notable research and experiments we haven’t covered alone.
This week, Google dominated the AI news cycle with a slew of new products unveiled at its annual I/O developer conference. They range from a code-generating AI meant to compete with GitHub’s Copilot, to an AI music generator that turns text announcements into short songs.
Quite a few of these tools appear to be real labor savers – meaning more than just marketing stuff. I’m particularly intrigued by Project Tailwind, a note-taking app that uses AI to organize, summarize, and analyze files from a personal Google Docs folder. But they also reveal the limitations and shortcomings of even the best AI technologies of today.
Take PaLM 2, Google’s latest Large Language Model (LLM), for example. PaLM 2 will support Google’s updated Bard chat tool, the company’s competitor to OpenAI’s ChatGPT, and serve as the foundation for most of Google’s new AI features. But while PaLM 2 can write code, email, and more like comparable LLMs, it also responds to questions in a toxic and biased manner.
The music generator from Google is also quite limited in its possibilities. As I’ve written in my own hands, most of the songs I’ve created with MusicLM sound passable at best – and like a four-year-old latching on to a song at worst DAW.
Much has been written about how AI will replace jobs – potentially the equivalent of 300 million full-time jobs, according to one study report by Goldman Sachs. In one Opinion poll According to Harris, 40% of workers familiar with OpenAI’s AI-powered chatbot tool ChatGPT fear it will completely replace their jobs.
Google’s AI isn’t the be-all and end-all. The company’s, in fact probably in arrears in the AI race. But it is an undeniable fact that Google applies some of the best AI researchers in the world. And if that’s the best they can do, it’s proof that AI is far from a solved problem.
Here are the other notable AI headlines over the past few days:
- Meta brings generative AI to ads: Meta this week announced a sort of AI sandbox for advertisers to help them create alternate copies, generate backgrounds from text prompts, and crop images for Facebook or Instagram ads. The company said the features are currently available to select advertisers and will expand access to more advertisers in July.
- Added context: Anthropic has expanded the context window for Claude – its flagship text-generating AI model, which is still in preview – from 9,000 tokens to 100,000 tokens. The context window refers to the text that the model considers before generating additional text, while tokens represent raw text (e.g. the word “fantastic” would be split into the tokens “fan”, “tas” and “tic”) . In the past, and still today, poor memory has been a barrier to the usefulness of text-generating AI. But larger context windows could change that.
- Anthropic promotes “constitutional AI”: Larger context windows aren’t the only differentiator of the Anthropic models. The company this week detailed “Constitutional AI,” its internal AI training technique aimed at infusing AI systems with “values” defined by a “constitution.” Unlike other approaches, Anthropic argues that constitutional AI makes the behavior of systems both easier to understand and easier to adjust when needed.
- An LLM for research: The nonprofit Allen Institute for AI Research (AI2) announced plans to train a research-focused LLM called the Open Language Model, adding to the large and growing open source library. AI2 views the Open Language Model, or OLMo for short, as a platform and not just a model – one that allows the research community to take any component created by AI2 and either use it themselves or enhance it.
- New fund for AI: In other AI2 news, AI2 Incubator, the nonprofit’s AI startup fund, has grown back to triple its previous size — $30 million versus $10 million. Since 2017, 21 companies have gone through the incubator, attracting about $160 million in further investment and at least one major acquisition: XNOR, an AI acceleration and efficiency company that was subsequently bought by Apple for about $200 million .
- EU implementation rules for generative AI: In a series of votes in the European Parliament this week, MEPs backed a slew of amendments to the bloc’s draft AI legislation – including laying down requirements for the so-called base models underlying generative AI technologies like OpenAI’s ChatGPT. The changes will place an obligation on base model vendors to conduct security reviews, data governance measures and risk mitigation before they release their models to the market
- A universal translator: Google is testing a powerful new translation service that converts videos into a new language while lip-synching the speaker with words they’ve never spoken. It could be very useful for many reasons, but the company openly warned about the possibility of abuse and the measures taken to prevent it.
- Automated explanations: It’s often said that LLMs are a black box along the lines of OpenAI’s ChatGPT, and there’s certainly some truth to that. OpenAI strives to dissolve these layers Development a tool for automatically identifying which parts of an LLM are responsible for which of its behaviors. The engineers behind it emphasize that it is still in its early stages, but the code to run it is available as open source on GitHub as of this week.
- IBM introduces new AI services: At its annual Think conference, IBM announced IBM Watsonx, a new platform that provides tools to build AI models and access pre-trained models to generate computer code, text and more. The company states that the launch was motivated by the challenges many companies still face in deploying AI in the workplace.
Other machine learning
Photo credit: Landing AI
Andrew Ng’s new company Landing AI takes a more intuitive approach to creating computer vision training courses. But getting a model to understand what you want to identify in images is quite tedious their “visual prompting” technique. With just a few brush strokes you can see your intention. Everyone who has to create segmentation models says “My God, finally!” Probably many PhD students who currently spend hours masking organelles and household items.
Microsoft applied Diffusion models in a unique and interesting way, by essentially using them to produce an action vector instead of an image after being trained on many observed human actions. It’s very early days and diffusion isn’t the obvious solution for this, but since they’re stable and versatile, it’s interesting to see how they can be used beyond purely visual tasks. Her paper will be presented at ICLR later this year.
Photo credit: Meta
Meta is also pushing the limits of AI ImageBind, which is said to be the first model capable of processing and integrating data from six different modalities: images and video, audio, 3D depth data, thermal information, and motion or position data. This means that in its little machine learning embedding room, an image can be associated with a sound, a 3D shape, and various text descriptions, each of which can be asked about or used in decision-making. It’s a step towards “general” AI, as it takes in and connects data more like the brain does – but it’s still simple and experimental, so don’t get too excited just yet.
When these proteins touch… what happens?
Everyone was excited about AlphaFold, and with good reason, but structure is actually just a small part of the very complex science of proteomics. The way these proteins interact is both important and difficult to predict — but this is new EPFL PeSTO model trying to do just that. “It focuses on important atoms and interactions within the protein structure,” said lead developer Lucien Krapp. “This means that this method effectively captures the complex interactions within protein structures to enable accurate prediction of protein binding interfaces.” While not accurate or 100% reliable, it is very useful for researchers not to have to start from scratch .
The government puts a lot of emphasis on AI. The President even stopped by Meeting with a range of top AI CEOs to say how important it is to get this right. Maybe some companies aren’t necessarily the right ones, but they at least have some ideas worth considering. But they already have lobbyists, don’t they?
I’m more excited about it New AI research centers are being created with federal funds. Basic research is badly needed to counterbalance the product-centric work of companies like OpenAI and Google – that is, if there are AI centers with a mandate to study things like Social Sciences (at the CMU)or climate change and agriculture (at the U of Minnesota), it feels like green fields (both figuratively and literally). However, I would also like to say a word of greeting Meta-research on forestry measurement.
Making AI together on a big screen – that’s science!
There are many interesting conversations about AI. I thought This interview with UCLA scholars Jacob Foster and Danny Snelson (my alma mater, go Bruins). was interesting. Here’s a great thought on LLMs to pretend you made it up this weekend when people were talking about AI:
These systems show how formally consistent most typefaces are. The more generic the formats these predictive models simulate, the more successful they are. These developments push us to recognize the normative functions of our forms and possibly to transform them. After the introduction of photography, which is very good at capturing a representational space, Impressionism developed in the painterly milieu, a style that completely rejected accurate representation and instead limited itself to the materiality of paint itself.
I will definitely use that!
