Tech Brew Ride Home - Mon. 10/30 – AI Guardrails

Episode Date: October 30, 2023

Governments around the world announce guardrails for AI… or, at least, suggest some. OpenAI takes steps to keep you from going to plugins. Meta seems serious about offering a subscription option to Europeans. Your earbuds are about to get smarter. And how AI might finally give us useful robots. Sponsors: Collective.com/ride NPR's Planet Money Links: Biden releases AI executive order directing agencies to develop safety guidelines (The Verge) Exclusive: G7 to agree AI code of conduct for companies (Reuters) New Version Of ChatGPT Gives Access To All GPT-4 Tools At Once (Search Engine Journal) Google Commits $2 Billion in Funding to AI Startup Anthropic (WSJ) Meta to Offer Ad-Free Facebook, Instagram Subscriptions in Europe (Bloomberg) Google can turn ANC earbuds into a heart rate monitor with no extra hardware (9to5Google) Stacking Boxes? Treating Cancer? AI Needs to Learn Physics First (WSJ) Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to the TechMe right home for Monday, October 30th, 2023. I'm Brian McCalla today. Governments around the world announce guardrails for AI, or at least suggest some. Open AI takes steps to keep you from going to plugins.

Meta seems serious about offering a subscription option to Europeans. Your earbuds are about to get smarter and how AI might finally give us a useful robot. Here's what you miss today in the world of tech. Global authorities are at least making a show of putting up some sort of AI regulatory framework today. For example, President Joe Biden has signed an executive order on generative AI directing the NIST, DHS, and other agencies to create new safety standards, protect privacy, support workers, and more. Quoting the verge, the National Institute of Standards and Safety, NIST, will be responsible for developing standards to red team AI models before

public release, while the Department of Energy and Department of Homeland Security are directed to address the potential threat of AI to infrastructure and the chemical, biological, radiological, nuclear and cybersecurity risks. Developers of large AI models like OpenAIs GPT and Metazlama 2 are required to share safety test results. The Biden administration would only provide a briefing sourced to a senior official who said, quote, we're not going to recall publicly available models that are out there, the official said. Existing models are still subject to the anti-discrimination rules already in place, end quote. To protect users' privacy, the White House called on Congress to pass data privacy regulations.

The order also seeks federal support for the development of privacy preserving techniques and technologies. Part of the order plans to prevent the use of AI to discriminate, including addressing algorithmic discrimination and ensuring fairness when utilizing the technology for sentencing, parole, and surveillance. It also orders government agencies to provide guidelines for landlords, federal benefits programs, and contracts on how to prevent AI from exacerbating. discrimination. Agencies are directed to address job displacement and produce a report on the impact of AI on the labor market. The White House also wants to encourage more workers to work in the AI ecosystem and ordered the launch of a national AI research resource to provide key information to students and AI researchers and access to technical assistance for small businesses. It also directed the rapid hiring of AI professionals for the government. The Biden

administration first released an AI Bill of Rights outlining a set of principles. Developers of AI models should follow. These were later turned into a series of agreements between the White House and several AI players, including Meta, Google, OpenAI, Nvidia, and Adobe. But an executive order is not a permanent law and generally only lasts the length of Biden's administration. Lawmakers are still discussing how to regulate AI, though some politicians said they want to pass laws around AI before the end of the year, end quote. Meanwhile, Reuters says that the group of seven industrial countries will today announce a voluntary 11-point code of conduct for companies developing advanced AI systems seeking to mitigate risks. Quote, the voluntary code of conduct

will set a landmark for how major countries govern AI amid privacy concerns and security risks. The document seen by Reuters showed, leaders of the group of seven or G7 economies made up of Canada, France, Germany, Italy, Japan, Britain, and the United States, as well as the European and Union kicked off the process in May at a ministerial forum dubbed the Hiroshima AI process. The 11-point code, quote, aims to promote safe, secure, and trustworthy AI worldwide, and will provide voluntary guidance for actions by organizations developing the most advanced AI systems, including the most advanced foundation models and generative AI systems. The G7 document said, the code urges companies to take appropriate measures to identify,

evaluate, and mitigate risks across the AI lifecycle, as well as tackle incidents and patterns of misuse after AI products have been placed on the market. Companies should post public reports on the capabilities, limitations, and the use and misuse of AI systems and also invest in robust security controls, end quote. ChatGPT plus users say that OpenAI has added in all tools feature and capabilities for analyzing docs like PDFs, potentially sidelining third-party chat GPT plugins. Quoting search engine journal. OpenAI plans to unveil a new way to use multimodal GPT4 with access to all tools without switching and more document analysis capabilities. Screenshots shared by numerous chat GPT plus users on X show new capabilities for PDF and document

analysis and an all tools feature. All tools gives users access to all GPT4 features without having to switch between one over the other. The move has been hailed as a significant leap, pushing the boundaries of generative AI capabilities as it goes beyond text-based queries. Now users can upload an image and ask Dolly 3 to create a response, adding a new dimension to their workflow. It could be a calculated move to streamline user experience and sideline third-party additions that have historically offered similar functionalities. The ability to analyze PDFs and other files directly within the system effectively negates the need for third-party chat GPT plugins that have been filling these gaps until now. By consolidating such

features in the latest version of ChatGPT, OpenAI responded to user feedback to create a more powerful tool that does not rely on external functionality. In addition to users gaining access to GPT for all tools, others have noticed a more recent knowledge cutoff date of roughly September 2023 or less than 30 days ago, end quote. Google has agreed to invest $2 billion in Anthropic. It would be $500 million up front and then $1.5 billion more over time. Now, they previously invested $550 million in the generative AI startup, actually earlier this year, quoting the journal. Anthropic has also signed a multi-year deal with Google Cloud worth more than $3 billion,

said one person familiar with the matter. The contract was signed a few months before the new investment, the person said. The three major providers of on-demand computing, Amazon, Google, and Microsoft have increasingly allied themselves with either Anthropic or Open AI, the two startups with the grandest ambitions for future artificial intelligence models. Anthropics founders led by the siblings Dario and Daniela Amodai left OpenAI two years ago after a dispute with Sam Altman over how to safely develop artificial intelligence. Since then, Anthropic has been in an intensifying battle with OpenAI to secure the training resources and deep-pocketed backers

needed to become the technology's leaders. Anthropic offers an AI assistant called Claude that competes with ChatGPT as well as a similar tool it sells to business. businesses. Anthropics' two mega deals with Amazon and Google mean it has raised almost $7 billion in the past year alone. Anthropic has also struck large deals with Amazon and Google to train and run its models. The AI company was largely bankrolled by Sam Bankman-Freed's FTX until the crypto exchange collapsed last year, pushing Anthropic to find other backers. It turned to the cloud giants as well as more traditional Silicon Valley investors like Spark Capital and Menlo Ventures. Earlier this year, it achieved a valuation of $4 billion.

Anthropic has told investors that the leaders of the AI race could be cemented as soon as next year and has painted a rosy picture of the future where AI tools could lead to virtual assistance, more intelligent search engines, and more advanced content generation for things like movie scripts and video games, end quote. Bloomberg says that Meta plans to offer Instagram and Facebook users in the European Union and European Economic Area, as well as Switzerland. as early as next month, an ad-free option for 999 euros or 1299 euros per month, alongside the usual free-with-adds version. Again, this is not yet official, but if this does

happen next month, it's going to be a big seismic shift historically, even if few people actually end up signing up for it. Quote, the subscriptions are a reaction to increasing regulations around how user data is collected and used in Europe, and a blog post explaining the decision. Meta cited the EU Court of Justice's decision that went against the company in July. The court said at the time that companies should consider offering an alternative service for customers who don't want their data collected and sold to advertisers, if necessary for an appropriate fee. We'll continue to advocate for an ad-supported internet, even with our new subscription offering in the EU, EEA and Switzerland, the company said in a statement, but we respect the spirit and purpose of these evolving European regulations

and are committed to complying with them, end quote. The company said that the new European subscription model has been, quote, factored into its most recent business outlook and guidance, end quote. A couple of trends that I want to put on your radar today. First up, I've been referring to this obliquely recently, but you should expect your headphones and earbuds to get a ton smarter over the next few years. Because, number one, if everyone wants to put an AI assistant on your phone, the dream of the Star Trek computer, as I said recently, would probably see voice interaction as the

key user interface. And that's before we even consider how smart earbuds could help in an AR-V-R future. But also, it turns out, you can put a ton of health tracking sensors into these devices. I've seen this from a slew of startups that have crossed my transom for the ride home fund recently. But here's one particular example from 9 to 5 Google. quote, Google today detailed its research into audio plus semography or APG that adds heart rate sensing capabilities to active noise-cancelling headphones and earbuds with a simple software upgrade. Google says the ear canal is an ideal location for health sensing, given that the deep ear artery, quote, forms an intricate network of smaller vessels that extensively permeate the auditory canal.

This APG approach works by sending a low-intensity ultrasound probing signal through an ANC headphones speakers. This signal triggers echoes, which are received via onboard feedback microphones. They observe that the tiny ear canal skin displacement and heartbeat vibrations modulate these ultrasound echoes. A model that Google created works to process that feedback into a heart rate reading, as well as heart rate variability or HRV measurement. this technique works even with music playing and bad earbuds seals. However, it was impacted by body motion and Google countered with a multi-tone approach that serves as a calibration tool to, quote, find the best frequency that measures heart rate and use only the best frequency to get

high-quality pulse waveform. Google performed two sets of studies with 153 people that found APG, quote, achieves consistently accurate heart rates of 3.21 percent median error across participants in all activity scenarios and heart rate variability, 2.7% median error in interbeat interval measurements, end quote. Compared to existing HR sensors, this system is not impacted by skin tones. Ear canal size and suboptimal seal conditions also do not impact accuracy. Google believes this is a better approach than putting traditional PPG and electrocardiograms or ECG sensors into the devices, as well as a microcontroller in headphones and earbuds, end quote. And finally today, friend of the show Christopher Mims has a piece up that talks about something that I've been hearing a lot about as well recently.

Turns out, quietly, behind the scenes, AI might end up being the key innovation that brings functionally useful real-world robots to IRL. It's all about using machine learning to master basic physics, quoting the journal. There are a handful of names for this approach, including physics-informed neural networks and scientific machine learning. but they all have one thing in common. They give AI someplace to start. That starting point is what we already know about a system, be it a bridge or a battery from decades or even centuries of hard-earned knowledge. This framework helps limit the universe of solutions and AI has to experiment with before it can make useful predictions. It's sort of like if someone's trying to solve a maze and there are certain paths you've already blocked off for them, says Karyann Bergen, who leads a machine learning research

group at Brown University. Let's say you're trying to teach an AI how to direct a robot to walk. The early stages of this learning process have to happen in a simulation because it's much faster than doing it with a real robot and less costly, says Jonathan Hurst, chief robot officer of agility robotics. Agility makes a bipedal robot that Amazon recently announced it will test in one of its warehouses. If the basic laws of physics are already built in, a machine learning algorithm has to explore far fewer possible combinations of limb and body movements when figuring out how to direct a robot to walk. If the simulation didn't have those laws, the AI might come up with correct solutions that aren't plausible, like passing through solid objects

or misunderstanding gravity. The idea that AI works best when the problem it's tackling is as narrowly defined as possible is a common theme in AI systems that generate measurable value for people and companies. Only in this case, the constraints are the laws of the natural world. Physics-informed machine learning systems can make predictions using far less data than AIs that are naive about the real world, says Karen Wilcox, director of the Odin Institute for Computational Engineering and Sciences at the University of Texas, Austin. A classic example of this approach is the digital twin of a jet engine, she says. Companies like General Electric have long used such models which incorporate machine learning to predict when an engine needs maintenance.

What's changing now is that with the growth of computing power and the spread of new kinds of machine learning algorithms, this physics-informed approach is becoming much more widespread. Wilcox's TED Talk on the subject is especially illuminating. Using the same approach, combining knowledge of the natural world continuously gather data and machine learning, it's in theory possible to create a digital twin of a cancer patient to direct their care, she adds. This is something that Wilcox's research group is studying in collaboration with the MD Anderson Cancer Center. So far, the team has only tested the approach in silico, that is, on a pool of data drawn from representative patients, but they are discussing a possible

clinical trial. In Formula E-Racing, which is the fully electrified version of Formula One, managing the amount of energy left in your battery is the difference between winning and losing. WAE Technologies, which makes the batteries for Formula E race cars recently created a division Eliza to commercialize its battery management software that uses physics-informed neural networks. Eliza's systems can determine the status of a battery with far less data than would normally be required because these AIs already, quote, know a great deal about how batteries work. This allows engineers to push batteries closer to their limits, eking out more power without damaging them. The result could be more range from existing EV batteries, including the one in your

driveway, says Eliza Technical Director Tim Engstrom. Dexterity, a robotics company is combining machine learning with models of how boxes behave in the real world to create robots that can finally load trucks. Taking boxes out of trucks, an easier problem was solved first. What made stacking box is nearly impossible without these models was that objects in the real world don't always behave in an idealized fashion, says Samir Menon, chief executive of the company, They might weigh more or less than a robot expects. Their contents might shift, and they might settle after they are dropped into place. Handling all of the weirdness of the real world requires a pretty good model of it, he adds.

It's still early days for physics-informed approaches to machine learning, say the experts I interviewed for this piece. Scientists are wary of the hype that comes with other forms of AI, such as the chatbots and art-generating models that are currently all the rage, says Bergen. But they're also excited by the potential of scientific machine learning, which at its core can be a way to gain new insights about systems, especially when we don't yet fully understand them, she adds, end quote. Reminder that there is that Apple event tonight. We'll cover that tomorrow for obvious reasons. The Blank Check podcast did the social network on their podcast this weekend as part of their David Fincher miniseries. And I think they had the best analysis of that movie as a movie that I've maybe ever heard.

Gave me a deeper appreciation for that movie than I had before. and my appreciation of it was pretty high. And then this morning, various outlets are reporting that David Fincher says he and Aaron Sorkin have discussed doing a sequel to the social network. Now, Fincher attaches himself to lots of ideas that never actually become movies. But that would be interesting. Talk to you tomorrow.