A week in AI: Google goes all out for I/O amid heightened regulation

Image credit: Google

Keeping up with a rapidly changing industry like AI is no easy task. So until AI can do it for you, here’s a quick rundown of last week’s coverage of the world of machine learning and notable research and experiments we didn’t cover individually.

This week, Google dominated the AI ​​news cycle with a range of new products announced at its annual I/O developer conference. They run the gamut from a code-generating AI meant to compete with his Copilot on GitHub to an AI music generator that turns text prompts into short songs.

A good number of these tools appear to be legitimate labor-saving measures rather than just marketing. I’m particularly intrigued by Project Tailwind, a note-taking app that uses AI to organize, summarize, and analyze files in her personal girlfriend’s Google Docs folder. But they also expose the limitations and shortcomings of today’s best AI technologies.

Take for example PaLM 2, Google’s latest large-scale language model (LLM). PaLM 2 powers Google’s latest Bard chat tool, the company’s competitor to OpenAI’s ChatGPT, and serves as the underlying model for most of Google’s new AI capabilities. However, while PaLM 2 can code, email, etc. just like their LLM counterparts, they can also respond to questions in harmful and biased ways.

Google’s music generator is also pretty limited in what it can do. As I wrote hands-on, most of the songs I create with MusicLM sound like her 4-year-old wandering around in her DAW, at best passable, at worst.

Much has been written about how AI will replace jobs, potentially representing 300 million full-time jobs, according to a Goldman Sachs report. Harris’ survey found that 40% of employees familiar with OpenAI’s AI-powered chatbot tool, ChatGPT, feared it would replace their jobs entirely. doing.

Google’s AI is not final. In fact, the company is definitely lagging behind in the AI ​​race. But there’s no denying that Google employs some of the world’s top AI researchers. And if this is the best they can do, it is a testament to the fact that AI is far from a solved problem.

Other notable AI headlines from the past few days include:

• Meta brings generative AI to advertising. Meta this week announced an AI sandbox of sorts to help advertisers create alternate copy, generate backgrounds with text prompts, and crop images for Facebook and Instagram ads. The company said the feature is only available to select advertisers at the moment, but plans to expand it to more advertisers in July.
• Added context: Anthropic has expanded the context window of Claude, its flagship text-generating AI model (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, whereas the tokens represent the raw text (e.g. the word ‘fantastic’ corresponds to the tokens ‘fan’, ‘tas’ and ‘tic’). split). Historically and today, poor memory has hampered the usefulness of text-generating AI. However, things can change as the context window grows.
• Anthropic touts “constitutional AI”: The only differentiator of the Anthropic model is not just the extended context window. This week, the company detailed “Constitution AI,” an in-house AI training method aimed at imbuing AI systems with “values” as defined by “Constitution.” In contrast to other approaches, Anthropic argues that his AI in constitution makes it easier to understand the system’s behavior and make adjustments if necessary.
• LLMs built for research: The nonprofit Allen Institute for AI Research (AI2) has announced plans to train a research-focused LLM called the Open Language Model to add to its large and growing open source library. AI2 sees the Open Language Model (OLMo for short) as a platform, not just a model. This will allow the research community to take each component his AI2 builds and use or improve upon themselves.
• New Fund for AI: In other AI2 news, AI2 Incubator, a nonprofit AI startup fund, is back on its feet, triple its size ($30 million vs. $10 million). Since 2017, 21 companies have passed through the incubator, attracting about $160 million in additional investment and at least one major acquisition. XNOR is an AI acceleration and efficiency company later acquired by Apple for about $200 million.
• Introductory Rules for EU Generated AI: In a series of votes in the European Parliament, MEPs this week backed a number of amendments to the European Union’s AI bill, including setting requirements for the so-called foundational models underpinning generative AI technologies like OpenAI’s ChatGPT. . This amendment places the responsibility on underlying model providers to apply safety checks, data governance measures, and risk mitigation before bringing models to market.
• Universal Translator: Google is testing a powerful new translation service that re-dubbs a video in a new language while keeping the lips of the speaker in sync with words they never spoke. While this can be very useful for a number of reasons, the company has been candid about potential abuse and the steps taken to prevent it.
• Automatic description: It’s often said that LLM along with OpenAI’s ChatGPT is a black box, and there’s certainly some truth to that. To peel back that layer, OpenAI is developing tools that automatically identify which part of the LLM is responsible for its behavior. The engineer behind it stresses that it’s still in its early stages, but the code to do it is available to him 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 to build AI models and provide access to pre-trained models for generating computer code, text, and more. The company says the motivation for the launch lies in the challenges many companies still experience in deploying his AI within the workplace.

Other machine learning

Andrew Ng’s new company Landing AI takes a more intuitive approach to creating computer vision training. Getting the model to understand what you want to identify in an image can be quite a daunting task, but using the “visual prompting” technique, he only needs to make a few strokes of the brush and he understands the intent from there. can. Anyone who needs to build a segmentation model says, “I finally did it!” There are probably many graduate students who spend hours masking organelles and household items right now.

Microsoft applied the diffusion model in a unique and interesting way, essentially using the diffusion model to generate action vectors instead of images, and trained it on many observed human actions. It’s still in its early stages and ubiquity is not a clear solution to this, but it’s stable and versatile, so it will be interesting to see how it can be applied beyond purely visual tasks. Their paper will be presented at his ICLR later this year.

Meta is also pushing the cutting edge of AI with ImageBind. It claims to be the first model that can process and integrate data from six different modalities: image and video, audio, 3D depth data, thermal information, motion or position data. This means that in its small machine learning embedding space, images can be associated with audio, 3D shapes, and various textual descriptions, any of which can be questioned or used in decision making. To do. This is a step towards “general” AI in that it absorbs and associates data like the brain does. But it’s still basic and experimental, so don’t get too excited just yet.

Everyone was excited about the AlphaFold, and for good reason. But in reality, structure is just one part of the very complex science of proteomics. How these proteins interact is both important and difficult to predict, and his new PeSTo model at EPFL attempts to do just that. “It focuses on the important atoms and interactions within the protein structure,” said lead developer Lucien Krupp. “This means that the method effectively captures complex interactions within protein structures and allows for accurate predictions of protein binding interfaces.” Even if it’s not accurate, 100% Not having to start from scratch, even if it wasn’t reliable, is very convenient for researchers.

The federal government is big on AI. The president stopped by a meeting with top AI CEOs to talk about how important it is to get this right. Maybe many companies aren’t exactly the right people to ask, but at least they have some ideas worth considering. But they already have lobbyists, right?

I am even more excited about the new federally funded AI Research Center. Fundamental research is sorely needed to balance the product-focused research being done by the likes of OpenAI and Google. So if you have an AI center mandated to research things like social sciences (CMU) or climate change or agriculture (universities), Minnesota), it feels like a green field (figuratively and literally). I would like to say a little loudly about this meta-study on forestry measurement.

There are many interesting conversations about AI. I found this interview with UCLA (Go to my alma mater, Bruins) scholars Jacob Foster and Danny Snelson interesting. Here are some great ideas about LLM that are perfect for pretending you came up with this weekend when people are talking about AI.

These systems reveal how coherent most sentences are formally. The more generic the format these predictive models simulate, the higher the success rate. These developments have led us to recognize the normative function of forms and their transformative potential. After the introduction of photography, which was very good at capturing expressive space, the pictorial environment developed Impressionism. This style completely rejected an accurate representation in order to retain the materiality of the paint itself.

I definitely use it!