Image credits: Google
Keeping up with a fast-moving industry like artificial intelligence is a tough task. So in order for AI to do it for you, here’s a handy roundup of the past week’s stories in the world of machine learning, along with notable research and experiments that we wouldn’t cover alone.
Google this week dominated the AI news cycle with a handful of new products launched at its annual I/O developer conference. They run the gamut from code-generating AI meant to compete with GitHub’s Copilot to an AI music generator that transforms Text prompts to short songs.
Quite a few of these tools appear to be legitimate labor savers—more marketing fluff, that is. I’m particularly intrigued by Project Tailwind, a note-taking app that makes use of artificial intelligence to organize, summarize, and analyze files from my personal Google Docs folder. But they also reveal the limitations and shortcomings of even today’s best AI technologies.
Take PaLM 2, for example, Google’s latest large language model (LLM). PaLM 2 will power Google’s updated Bard chat tool, the company’s competitor to OpenAI’s ChatGPT, and serve as the base model for most of Google’s new AI features. But while PaLM 2 can write code, email messages, and more, like comparable LLMs, it also responds to questions in toxic and biased ways.
Google Music Generator, too, is somewhat limited in what it can accomplish. As I wrote in my own hands, most of the songs I’ve created with MusicLM sound passable at best—and at worst like a four-year-old released on a DAW.
Much has been written about how artificial intelligence will replace jobs — potentially equivalent to 300 million full-time jobs, according to a report by Goldman Sachs. In a Harris survey, 40% of workers familiar with OpenAI’s AI-powered chatbot tool, ChatGPT, worry it will completely replace their jobs.
Google’s AI isn’t everything. In fact, the company can be said to be behind in the AI race. But it is an undeniable fact that Google employs some of the best AI researchers in the world. And if that’s the best they can manage, it’s a testament to the fact that artificial intelligence is far from solving the problem.
Here are other AI headlines noticed from the past few days:
- Meta brings generative AI to ads: This week, Meta announced an AI sandbox, of sorts, for advertisers to help them create alt copies, create background through text prompts and crop images for Facebook or Instagram ads. The company said the features are available to select advertisers at this time and will expand reach to more advertisers in July.
- Added context: Anthropic has expanded the context window for Claude — its flagship text generation and AI model, which is still in preview — from 9,000 tokens to 100,000 tokens. The context window indicates which text the form considers before creating additional text, while tokens represent raw text (eg “cool” would be split into the tokens “fan”, “tas”, and “tic”). Historically and even today, poor memory has been an obstacle to the usefulness of text generation for AI. But larger context windows can change that.
- Anthropy promotes “constitutional artificial intelligence”: Larger context windows are not the only differentiating factor for anthropic models. This week, the company detailed “Constitutional AI,” its in-house AI training approach that aims to imbue “values” in AI systems with a “constitution.” In contrast to other approaches, Anthropic argues that constitutional AI makes the behavior of systems easier to understand and simpler to modify as needed.
- The LLM is designed to research: The nonprofit Allen Institute for Artificial Intelligence Research (AI2) has announced that it is planning a research-focused LLM training called Open Language Model, adding to a large and growing open source library. AI2 sees the Open Language Model, or OLMo for short, as a platform rather than just a model—one that will allow the research community to take every component AI2 creates and either use it themselves or seek to improve it.
- New AI fund: In other AI2 news, AI2 Incubator, the nonprofit AI startup fund, is back at triple its previous size — $30 million versus $10 million. Twenty-one companies have passed through the incubator since 2017, attracting about $160 million in further investment and at least one major acquisition: XNOR, an AI accelerator and efficiency device that Apple later acquired for about $200 million.
- EU Introduction Rules for Generative AI: In a series of votes in the European Parliament, MEPs this week supported a range of amendments to the bloc’s AI bill — including settling requirements for so-called foundational models that underpin generative AI technologies such as OpenAI’s ChatGPT. The amendments placed the onus on foundational model providers to implement safety checks, data governance measures and mitigate risks before bringing their models to market.
- Universal Translator: Google is testing a powerful new translation service that replays video in a new language while also lip-syncing a speaker with words they’ve never spoken. It can be very useful for a lot of reasons, but the company has been upfront about the potential for abuse and the steps it takes to prevent it.
- Instrumental explanations: It is often said that OpenAI’s ChatGPT-style LLM is a black box, and sure enough, there is some truth to that. In an effort to peel back its layers, OpenAI is developing a tool to automatically identify which parts of an LLM are responsible for which of its behaviors. The engineers behind it confirm that it is in its early stages, but the code to run it is available in open source on GitHub as of this week.
- IBM launches new AI services: At its annual Think conference, IBM announced IBM Watsonx, a new platform that provides tools for building artificial intelligence models and provides access to premade models for creating computer code, scripts, and more. The company says the launch was driven by the challenges many companies still face in deploying AI in the workplace.
other machine learning
Andrew Ng’s new company Landing AI is taking a more intuitive approach to creating computer vision training. Making a model that understands what you want to define in the images is pretty daunting, but their “visual stimulus” approach just lets you make a few brush strokes and set your intention from there. Anyone who has to build segmentation models says “Oh my God, finally!” There are probably a lot of graduate students who currently spend hours hiding organelles and household items.
Microsoft applied scattering models in a unique and interesting way, essentially using them to create a scattering vector rather than an image, having trained it on a lot of observed human actions. It’s still very early days and diffusion isn’t the obvious solution to this, but since they’re so stable and versatile, it’s interesting to see how they can be applied beyond purely visual tasks. Their paper is being presented at ICLR later this year.
Meta is also pushing the edges of AI with ImageBind, which it claims is the first model that can process and combine data from six different modalities: image, video, audio, 3D depth data, thermal information, and motion or positional data. This means that in a small machine learning embedding space, an image may be associated with sound, 3D shape, and various text descriptions, any of which can be subtracted or used to make a decision. It’s a step toward “general” AI in that it absorbs and correlates data like a brain – but it’s still basic and experimental, so don’t get too excited just yet.
Everyone got excited about the AlphaFold, and for good reason, but the structure is really only a small part of the very complex science of proteomics. How these proteins interact is important and difficult to predict – but this new PeSTo model from EPFL attempts to do just that. “It focuses on the important atoms and interactions within a protein’s structure,” said lead developer Lucien Crabbe. “This means that this method effectively captures complex interactions within protein structures to enable accurate prediction of protein-binding interfaces.” Even if it isn’t 100% accurate or reliable, not having to start from scratch is hugely beneficial for researchers.
The Feds are going AI. The president even took part in a meeting with a group of senior AI executives to point out how important it is to get this right. Maybe a group of companies won’t necessarily be the right one to ask, but they’ll at least have some ideas worth considering. But they already have lobbyists, right?
I’m most excited about the emergence of new federally funded AI research centers. Fundamental research is sorely needed to balance the product-focused work being done by the likes of OpenAI and Google – so when you have AI centers with mandates to investigate things like the social sciences (at CMU), or climate change and agriculture (at the U of Minnesota), it seems It’s like green fields (figuratively and literally). Although I also want to give a little shout-out to Meta Research on Forest Measurement.
Lots of interesting conversations about artificial intelligence. I thought this interview with academics Jacob Foster and Danny Sleeson was interesting. Here’s a great LLM idea to pretend you came this weekend when people are talking about AI:
These systems detect the consistency of most writing formally. The more general formats these predictive models simulate, the more successful they are. These developments prompt us to learn about the modular functions of our forms and their possible transformation. After the introduction of photography, which is very good at capturing a representational space, the environment of painting developed Impressionism, a technique that rejected exact representation altogether to remain with the materiality of the paint itself.
Definitely using that!
#week #Google #exits #regulations #grow