Hello and welcome to Eye on AI.
It’s been a busy week for OpenAI. Coming off its product-packed Dev Day conference, the company followed up with an announcement that offers a window into the next stage of its data collection plans while simultaneously battling a DDoS attack that caused major disruptions to ChatGPT and other OpenAI tools. We also saw the very beginnings of an ecosystem for GPTs, custom ChatGPT-like models users can create for their own specific purposes.
OpenAI on Thursday announced Data Partnerships, a program for collaborating with third parties to build datasets for training AI models. A blog post introducing the program says the company is “interested in large-scale datasets that reflect human society and that are not already easily accessible online to the public today.”
“To ultimately make AGI that is safe and beneficial to all of humanity, we’d like AI models to deeply understand all subject matters, industries, cultures, and languages, which requires as broad a training dataset as possible. Including your content can make AI models more helpful to you by increasing their understanding of your domain,” reads the blog post, adding that the company is “particularly looking for data that expresses human intention.”
It goes on to state that OpenAI is seeking partners to help it create an open-source dataset for training LLMs as well as private datasets for training proprietary AI models, including its own foundation, fine-tuned, and custom models. It’s not exactly clear what participating in this program looks like (there’s no mention of compensation, data ownership, etc.). But it’s clear the private dataset efforts will go beyond empowering AGI in theory and will directly enrich OpenAI’s own models, and therefore its bottom line. It’s also impossible to ignore the drumbeat of the call for data that isn’t “already easily accessible online to the public today,” seeing as OpenAI already scraped the internet as it pleased, and is now facing an onslaught of lawsuits as a result.
While this announcement was trickling into the world, Open AI was simultaneously fighting off a targeted DDoS attack that struck just two days after its first-ever in-person event, interrupting service for ChatGPT and other OpenAI tools for more than 24 hours. But despite this, users started gaining access to GPTs, building their own models, and sharing them with the world.
Kristina Martin, a marketer and educator, has already made three GPTs: one to aid K-12 educators in writing progress report comments, an assistant for grading MLA English composition papers, and a brand-building guide for solopreneurs. She told Eye on AI that she found the builder “relatively easy to use,” though she did have to actively test and refine the GPTs to get an outcome close to what she envisioned, and she never quite achieved the type of decision-tree line of questioning she was hoping to achieve with the brand-building model.
“They are fun to build, and I do think they can help with certain tasks,” she told Eye on AI. “However, those tasks can’t be too complex. As a personal branding and content strategist, I would typically ask my clients a series of questions in a 1:1 interview and ask additional questions based on their responses. I think the GPT builder needs to be more sophisticated to handle those types of tasks.”
While OpenAI has yet to release the GPT Store, which it says will roll out later this month, users jumped at the opportunity to compile GPTs themselves and have already created various public lists and searchable GPT repositories. I discovered and experimented with a wide range of single-function GPTs doing everything from creating playlists (like PlaylistAI:Spotify) to emulating specific people (like Paul Graham and Elon Musk).
Most notably, I saw a large swath of purely playful GPTs that reminded me of the earlier days of the internet. For example, there’s FortuneTeller GPT (described as a “cryptic and mystical AI agent”), OppositeDay GPT (which answers inquiries with the opposite of the truth), RoastMeGPT (self-explanatory), and PessimismGPT (which had some interesting things to say about OpenAI). There’s even PirateSpeakGPT, which communicates exclusively in the style of a stereotypical pirate and elicits memories of when everyone discovered how to change their language to “Pirate” on Facebook.
These are far from Earth-shattering use cases, but they do show how GPTs are opening up new lanes for people to get creative on an internet that often feels as if it’s lost its magic. Not to mention, it is pretty amazing that people with zero technical expertise can create these models in minutes using only natural language—not a single line of code is required.
Now we just have to wait for the killer app, or in this case, the killer GPTs.
And with that, here’s the rest of this week’s AI news.
Reminder: Brainstorm AI is just weeks away!
Gain vital insights on how the most powerful and far-reaching technology of our time is changing businesses, transforming society, and impacting our future. Join us in San Francisco on Dec. 11–12 for Fortune’s third annual Brainstorm AI conference. Confirmed speakers include such A.I. luminaries as Google Assistant and Google Bard GM Sissie Hsiao, IBM Chief Privacy and Trust Officer Christina Montgomery, Walmart International SVP and CTO Sravana Karnati, Pfizer Chief Digital and Technology Officer Lidia Fonseca, and many more. Apply to attend today!
And with that, here’s the rest of this week’s AI news.
Sage Lazzaro
sage.lazzaro@consultant.fortune.com
sagelazzaro.com
AI IN THE NEWS
SAG-AFTRA approves new contract after holding out to win AI protections, ending Hollywood’s longest-ever actor’s strike. In her remarks, SAG-AFTRA president Fran Drescher framed AI as a key issue and said the union held out to win AI protections. The tentative agreement stipulates that companies must get consent before making digital replicas of actors (including deceased actors), disclose what the replica will be used for, and compensate actors for digital replicas. It also includes guidelines around “synthetic fakes,” or fake performers based on the likeness of actors, that can be used to train generative AI, according to Rolling Stone. Some actors have criticized that the AI protections don’t go far enough.
Nvidia is gearing up to release three new AI chips for China after U.S. restrictions blocked the company from selling three of its current chips in the Chinese market. That’s according to the Guardian. The news comes just weeks after the Biden Administration announced the updated export restrictions, and the company could announce the chips—the HGX H20, L20 PCIe, and L2 PCIe—as soon as Nov. 16.
Meta will require campaigns to disclose if political ads are altered by AI. The company also announced that labels noting the use of AI will appear when users click on ads altered by AI, according to the Wall Street Journal. The rules will be applied worldwide, although a date for them to go into effect has not been set. The announcement follows the news that Meta is banning political campaigns and groups from using its new generative AI advertising products.
Microsoft shares its AI election integrity plan, including “Content Credentials as a Service.” With this new Microsoft service, electoral campaigns and other users can prove the authenticity of their images and videos with metadata—such as who created the content, when it was created, and if AI is involved. Microsoft also announced it has formed a team to help campaigns “navigate the world of AI” and will launch what it’s calling an Election Communications Hub, where election authorities will be able to tap Microsoft’s security and support teams in the weeks leading up to their election.
OECD updates its definition of AI as finalization of the EU AI Act nears. Defining “artificial intelligence” has been one of the key challenges of regulating the technology. The Organisation for Economic Co-operation and Development, an intergovernmental organization comprised of 38 countries, has settled on a definition that is expected to be used in the final EU AI Act, according to Euractiv. The definition: "A machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that [can] influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment."
GM’s Cruise fallout continues as the carmaker recalls 950 driverless vehicles. After having its permits suspended in California and then taking its cars off the road nationwide, Cruise has now recalled nearly 1,000 of its autonomous vehicles and will likely issue more recalls, according to Reuters. The response comes after an incident last month in which a pedestrian was struck by a hit-and-run driver, thrown into an adjacent lane, and then hit a second time and further dragged by a Cruise robo-taxi that wasn’t able to stop in time.
Samsung unveils a family of LLMs, including a ChatGPT competitor. Called Samsung Gauss, the family of models consists of three tools: Samsung Gauss Language, Samsung Gauss Image, and Samsung Gauss Code. The company is currently using the models internally and said they’ll be available to the public “in the near future,” according to TechCrunch.
EYE ON AI RESEARCH
It’s raining AI. Weather prediction is becoming more important as the rate of extreme weather events increases—plus, it’s just helpful to know when to grab an umbrella and what type of jacket to wear.
In a paper published today in Science, Google DeepMind researchers detail a weather prediction model trained using machine learning that “significantly outperforms” the current industry standard. Called GraphCast, the model predicts hundreds of weather variables to deliver 10-day weather predictions more accurately and much faster (in less than one minute) than the industry gold-standard weather simulation system: the High Resolution Forecast (HRES), produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). They write that GraphCast also supports improved prediction for severe weather events, including tropical cyclones tracking, atmospheric rivers, and extreme temperatures.
“We believe this marks a turning point in weather forecasting, which helps open new avenues to strengthen the breadth of weather-dependent decision-making by individuals and industries, by making cheap prediction more accurate, more accessible, and suitable for specific applications,” reads the paper.
The authors are open-sourcing the model and write that GraphCast can open new directions for other important geo-spatiotemporal forecasting problems around climate, ecology, energy, agriculture, and more. For more information, you can check out the paper and accompanying blog post.
FORTUNE ON AI
AI is going to ‘turbocharge winners’ faster than anyone is expecting, predicts an MIT research scientist —Sheryl Estrada
Sending an AI bot to your Zoom meetings is the latest office power move, but it’s also a management mistake —Lila MacLellan
AI was not even in the top 20 business risks in a ‘shocking’ survey of nearly 3,000 corporate leaders —Rachyl Jones
BRAINFOOD
Demo day. This past week also delivered the first consumer hardware launch of the generative AI era, the Humane AI Pin. The $700 screen-less, app-less device is essentially a wearable for interacting with LLMs. Measuring about 1.7 by 1.8 inches, it clips onto the user’s clothing and acts as an AI assistant they can speak with as they go about their day. Instead of typing a question into ChatGPT, imagine speaking your question out loud into a pin on your shirt and it answering you out loud, too. Ars Technica called it “a bizarre cross between Google Glass and a pager.”
The entire demo video is worth a watch. The ability to translate an in-person conversation in real time is undeniably powerful. And the “catch me up” feature, which surfaces the most important items that need attention from across a user’s various messages, certainly sounds like a time-saver. These capabilities, however, have everything to do with LLM technology and almost nothing to do with the AI Pin itself. Humane’s press release cites collaborations with Microsoft and OpenAI, but doesn’t specify exactly which models it’s using.
Unlike previous voice AI assistants like Siri and Alexa (not to mention the in-development Rewind Pendant AI device designed to always be recording), Humane says its AI Pin isn’t always listening and doesn’t use a wake word. You have to physically touch it to activate it. But the device isn’t without privacy concerns.
One frame from the demo depicting the device’s video capture feature—which shows the PoV of someone riding a bike, recording the person riding in front of them—gave me an instant flashback to “The Entire History of You” episode of Black Mirror, which haunts me the most out of all the episodes I’ve seen (apologies for the cliché Black Mirror reference). The company touts its “Trust Light” feature, which blinks to let the user and supposedly everyone else know it’s collecting data, but there are so many instances where this would not be effective—for example, the very situation depicted in the bike-riding scene where the AI Pin-wearer is recording someone in front of them who may not even know someone is behind them, let alone be able to see a tiny blinking light on a pin on the person’s shirt.
It’s interesting that the first hardware device created specifically for interacting with LLMs runs almost entirely on voice. In Humane’s world, we’d all be walking around talking into our shirts everywhere we go. It feels like a tough sell. Overall, the reviews have skewed mixed to negative. The Verge wrote that “it’s not yet entirely clear what you’re supposed to use it for.” At this point, the AI Pin seems best suited for analyzing and imagining what the next big device will actually be like. I can’t imagine we’ll all adopt the AI Pin in its current form, but a world after smartphones? Something screen-less and centered around LLMs? It feels very plausible.
This is the online version of Eye on AI, Fortune's weekly newsletter on how AI is shaping the future of business. Sign up for free.
