A week in AI: Generative AI is spamming the web

Image credit: ETH Zurich

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 what’s been going on in the machine learning world and notable research and experiments we couldn’t cover alone.

This week, SpeedyBrand, a company that uses generative AI to create SEO-optimized content, emerged from stealth with the help of Y Combinator. It has not yet raised a significant amount of capital ($2.5 million) and has a relatively small customer base (around 50 brands). But it got me thinking about how generative AI is starting to change the fabric of the web.

As James Vincent of The Verge wrote in a recent article, generative AI models are making it cheaper and easier to generate low-quality content. Newsguard, a company that provides tools to scrutinize news sources, has exposed hundreds of ad-supported sites with generic-sounding names featuring misinformation produced by generative AI.

That’s causing problems for advertisers. Many of the sites Newsguard looks at appear to be built solely for the purpose of exploiting programmatic advertising, an automated system for placing ads on pages. A Newsguard report found nearly 400 ads for 141 major brands on 55 junk news sites.

Advertisers aren’t the only ones to worry. As Gizmodo’s Kyle Barr points out, a single AI-generated article could drive mountains of engagement. And even if he only makes a few bucks from the AI-generated article, it’s less than the cost of generating the text in the first place, and it’s possible that no ad dollars go to legitimate sites.

So what is the solution? do you have? These two questions keep me up at night more and more. Barr suggests that search engines and advertising platforms have a duty to crack down and punish bad actors who use generative AI. But given how fast the field is changing, and the infinitely scalable nature of generative AI, I’m not sure they can keep up.

Of course, spam content is not a new phenomenon, and there have been waves of it before. The web has also adapted. The difference this time is that the barriers to entry are dramatically lower, both in terms of cost and time to invest.

Vincent is optimistic, saying if the web teeth Ultimately it will be flooded with AI junk, but it could encourage the development of more well-funded platforms. I’m a little unsure. But there is no doubt that we have reached an inflection point and the decisions that are being made today about generative AI and its output will impact how the web functions for some time to come. .

Other notable AI stories from the last few days include:

OpenAI officially announces GPT-4: OpenAI announced this week that its latest text generation model, GPT-4, is generally available through a paid API. GPT-4 can generate text (including code) and accept image and text input. It is an improvement over its predecessor, GPT-3.5, which only accepted text, and performs “human-level” on a variety of professional and academic benchmarks. But as I said in my last post, it’s not perfect. (On the other hand, ChatGPT adoption is reportedly declining, and we get it.)

Putting “ultra-intelligent” AI under control: In other OpenAI news, the company has formed a new team led by chief scientist and one of OpenAI’s co-founders, Ilya Sutskever, to develop ways to operate and control “ultra-intelligent” AI systems. are doing.

New York City anti-stigma laws: After months of delays, New York City this week passed legislation requiring employers who use algorithms to recruit, hire, and promote employees to submit their algorithms to an independent audit and publish the results. Enforcement started.

Valve has implicitly given the go-ahead to AI-generated games: Valve has issued an unusual statement after it claims it rejects games containing AI-generated assets from the Steam game store. The notoriously tight-lipped developer said his company’s policy is evolving and not against AI.

Humane Announces Ai Pin: Humane, a startup founded by former Apple design and engineering duo Imran Chaudhry and Bethany Bongiorno, unveiled details of its first product, The Ai Pin, this week. After all, Humane’s products are wearable gadgets with projected displays and AI-powered features. It’s like the smartphone of the future, but in a very different form factor.

EU AI Regulation Warning: Big tech founders, CEOs, venture capitalists and industry giants across Europe signed an open letter to the EU Commission this week saying Europe could miss out on a generative AI revolution if the EU passes laws stifling innovation warned of potential

Deepfake scams are prevalent: check out this clip UK consumer finance champion Martin Lewis appears to have shilled an investment opportunity backed by Elon Musk. Sounds normal, right? not exactly. This is an AI-generated deepfake, and we may be able to catch a glimpse of how AI-generated misery is rapidly accelerating on our screens.

AI-Powered Sex Toys: Lovense — perhaps best known for its remote-controlled sex toys — announced the ChatGPT Pleasure Companion this week. Released in beta on the company’s remote control app, “Advanced Lovense ChatGPT Pleasure Companion” invites you to indulge in juicy, erotic stories that Companion creates based on your chosen topic.

Other machine learning

Our research review begins with two completely different projects by the ETH Zurich. The first is aiEndoscope, a spin-off of smart intubation. Intubation is necessary for patient survival in many situations, but is usually a cumbersome manual procedure performed by specialists. The intuBot uses computer vision to recognize and respond to live feeds from the mouth and throat to guide and correct the position of the endoscope. This could save lives by allowing people to safely intubate when needed without having to wait for a specialist consultation.

They explain it in a little more detail.

In a completely different field, researchers at ETH Zurich have pioneered the techniques needed to animate smoke and fire without falling prey to the fractal complexity of fluid dynamics, making them secondary to Pixar films. contributed to Their approach was noticed by Disney and Pixar and built for the movie Elemental. Interestingly, this is more of a style transfer solution than a simulation solution. This is a smart and obviously very valuable shortcut. (Click here for the image above.)

AI in nature is always interesting, but natural AI applied to archeology is even more interesting. A study led by Yamagata University aimed to identify the new Nazca Lines, a giant “ground line” located in Peru. Visible from orbit, you might think they’re very obvious, but the thousands of years of erosion and tree cover since these mysterious formations were created have made them invisible to the eye. You can see that there are unknown numbers hidden in plain sight. After being trained on aerial imagery of known obscured geoglyphs, the deep learning model was released on other views and, surprisingly, detected at least four new geoglyphs, as shown below. Pretty exciting!

In a more directly related sense, AI-adjacent technologies are constantly finding new jobs in natural disaster detection and prediction. Engineers at Stanford University are compiling data for training future wildfire prediction models by running simulations of heated air above a forest canopy in a 30-foot tank. Modeling the physics of flames and embers that travel outside the range of wildfires requires a deeper understanding of them, and this team is doing everything they can to approximate it.

UCLA is researching ways to predict landslides that are more likely to occur with changes in fires and other environmental factors. But while AI is already being used to predict them with some success, it doesn’t “show its effectiveness.” That is, projections cannot explain whether it is due to erosion, water table changes, or crustal activity. In the new “superimposable neural network” approach, each layer of the network uses different data, but they all run in parallel rather than simultaneously, making it more specific which variables led to increased risk. can be output as desired. It’s also much more efficient.

Google is working on an interesting challenge. It’s a way to let machine learning systems learn dangerous knowledge, but not propagate it. For example, if my training set contains a napalm recipe, I don’t want it to repeat, but in order to know not to repeat it, I need to know what won’t repeat. Contradiction! So tech giants are looking for “non-machine learning” ways to perform this kind of balancing task safely and reliably.

If you want to learn more about why people seem to trust AI models without good reason, look no further than this Science editorial by Celeste Kidd (UC Berkeley) and Abeba Birhane (Mozilla). The piece explores the psychological underpinnings of trust and authority, and shows how her current AI agents can basically use them as springboards to increase their self-worth. A very interesting article if you want to feel smart this weekend.

You’ve heard a lot about the infamous fake chess machine, the “Mechanical Turk,” and its guise actually inspired people to create imitations of it. IEEE Spectrum has an interesting story about Torres Quevedo, a Spanish physicist and engineer who created a real mechanical chess player for him. Its functionality was limited, but that’s what makes it so authentic. Some even claim that his chess machine was the first “computer game”. food for thought.