Artificial Intelligence

AMD says the MI325X, shipping Q4, will beat Nvidia’s H200. But Nvidia seems a step ahead; it’ll ship “several billion dollars” of its next-gen Blackwell B200 GPU in Q4, too. AMD says its Blackwell competitor, the MI355X, won’t arrive till 2H 2025.

AMD isn’t talking price, but told us it’ll undercut Nvidia when it comes to total cost of ownership.

OpenAI disrupted more than 20 foreign influence operations over the past year, according to its quarterly threat report. But there’s no “evidence of this leaning to meaningful breakthroughs in their ability to create substantially new malware or build viral audiences,” the report says.

AI has let foreign actors “more quickly and convincingly tailor synthetic content.” But so far, it isn’t reaching much of an audience.

With so many job applications now being reviewed by automated AI software, it’s little wonder some people have embraced using bots like AI Hawk to apply to thousands of roles on their behalf.

AI Hawk co-founder Federico Elia told 404 Media that he created the project to “balance the use of artificial intelligence in the recruitment process” and help re-level the playing field.

After announcing it last month, Google seems to have rolled out Google TV’s AI wallpapers, 9to5Google reports.

According to a Google TV support page, the option lives under Settings > System > Ambient Mode > Custom AI Art. Click on “Create new...” and then describe the image you want, use a template, or choose “Inspire me” and Google TV makes one for you.

It’ll roll out in more languages and to 21 additional countries by the end of the year, starting today with the UK, Brazil, Guatemala, Bolivia, Paraguay, and the Philippines. The AI assistant currently has almost 500 million active monthly users.

Meta AI is also coming to Meta’s Ray-Ban smart glasses in Australia and the UK, though the latter is only voice support “to start.”

Reuters reports that Bentley Hensel, a “long-shot congressional challenger” in Virginia, made an AI chatbot of incumbent Don Beyer to stand in if Beyer doesn’t show up to a debate:

DonBot, as the AI is playfully known, is being trained on Beyer’s official websites, press releases, and data from the Federal Election Commission. The text-based AI is based on an API from OpenAI, the maker of ChatGPT.

David Pierce called out the lack of timers in his April review of the AI Pin, but now they’re here.

The AI Pin is getting a bunch of other new features, too, thanks to the CosmOS 1.2 update that Humane says is starting to roll out.

Hearst, which owns a substantial portion of the media landscape, has just signed a deal with OpenAI to integrate Hearst content into its products (The Verge’s parent company Vox Media also partners with OpenAI).

Hearst owns 24 daily newspapers and 52 weekly newspapers, 175 websites and more than 200 magazine editions worldwide, making this one of OpenAI’s biggest media partnerships.

John Hopfield and Geoffrey Hinton were honored for their work on artificial neural networks, which created the foundation for AI tools that humanity can use “for good purposes” according to the Nobel Prize X account.

Hinton, a “Godfather of AI,” notably left Google’s Research team to speak freely about the risks of the technology and said he partially “regrets his life’s work.”

If you thought standard-issue ad tech was a little weird and creepy, get ready for the future: platforms letting marketers use all their data to make an infinite number of AI-generated ads specifically targeted to individual viewers. Digiday reports on TikTok’s Smart+, which competes with similar offerings from Meta and Google:

Marketers can let TikTok’s AI handle the heavy lifting — building and delivering ads to drive conversions, leads, or app downloads. […] The pitch is all about simplicity and speed — no more weeks of guesswork and endless A/B testing, according to Adolfo Fernandez, TikTok’s director, global head of product strategy and operations, commerce.

Super normal, everyone! No potential issues here!

Some turned out great, but others highlight just how messy this technology is, according to Pitchfork. From Shaad D’Souza’s review:

AI is simply another tool that will sometimes be used badly and sometimes be used well, and on Honey I think it’s used well—to complicate and expand the abilities of an artist well into his career, whose creative impulses can no longer be entirely satiated through the means previously available to him. There is one exception: The rap verse on “Campfire” is also Snaith, and it edges toward racial ambiguity in a way that feels queasy. At best, it’s a misguided experiment; at worst, outright minstrelsy.

I’m coming to you live from my hot-as-hell apartment here in San Francisco. We’re facing a historic heat wave and none of us have apartment air conditioning.

Even funnier, the AC at OpenAI HQ is broken, according to an employee. The office is based in the Mission district, which tends to get hotter than the rest of the city. (One Google DeepMind staffer flaunted his working AC in reply: “We’re hiring!”)

That’s straight from Rabbit CEO Jesse Lyu, who took great exception to our story from September 25th, which was sourced to a Fast Company article about his comments at one of their events. Jesse told me the actual daily user number was around 20,000, spiking up to 34,000 the day the company’s new LAM Playgrounds were launched, and that his actual comment was that 5,000 of those people were using the Rabbit at any given time. For context, Jesse also told me Rabbit has sold 100,000 R1s so far.

Fast Company has corrected its story, and we’ve updated our story as well. You can hear the whole conversation on Decoder.

With several media companies striking AI training deals with OpenAI, the number of websites blocking GPTBot has taken a big dip, according to data seen by Wired:

At its peak, the high was just over a third of the websites; it has now dropped down closer to a quarter. Within a smaller pool of the most prominent news outlets, the block rate is still above 50 percent, but it’s down from heights earlier this year of almost 90 percent.

As a reminder that AI image generators’ training data tends to include peoples’ regular smartphone photos, try entering an iPhone-like picture file name into the prompt field for Flux1.1 Pro, as this person did.

I got some similar results when I tried prompts like “IMG_4001.JPG” with its predecessor, Flux.1, a model that drives xAI’s Grok-2 image generation.

If you signed up for Google’s waitlisted feature that lets you ask Gemini questions to surface photos and videos, you might be getting it soon, as 9to5Google reports that the feature seems to be going live for some who’ve joined the list.

If you’re not already on the waitlist, you can jump in the queue by signing up on Google’s site.