Good morning. The Nobel Prize in Literature (thankfully) did not go to another computer scientist. The streak is broken, though Andrew Ng thinks that someday, AI researchers will win Nobels in literature, medicine and peace ...

This year, though, it went instead to South Korean author Han Kang, for her “intense poetic prose.”

— Ian Krietzberg, Editor-in-Chief, The Deep View

In today’s newsletter:

• 💻 MBZUAI Research: Memes and multilingual persuasion attempts

• 🩺 Microsoft launches AI tools for doctors

• 💰 OpenAI’s long, costly road to profitability

MBZUAI Research: Memes and multilingual persuasion attempts

Memes, that wonderful hallmark of modern internet culture, are usually funny, or at least they try to be. But, increasingly, researchers have been noticing a darker side to memes: they can be weaponized to spread misinformation. 

Since their ingrained format is light on context and heavy on humor, memes are a perfect medium for bad actors to spread misleading or downright false information as far and wide as possible. 

This reality has created a bit of a content moderation challenge for social media networks, resulting in ever-increasing research into methods of automatically detecting hate speech, propaganda and offensive content in memes. 

What happened: New research from the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) seeks to enhance efforts in this area. 

• The study — a shared task that garnered 153 registered teams — specifically focused on methods of multi-lingual, multimodal detection of persuasion techniques in memes. The researchers assembled four datasets of memes that covered four languages: English, Arabic, Bulgarian and North Macedonian. 

• The intended approach involved three tasks: identifying if a meme exhibits a persuasion technique, identifying the persuasion technique on display in the meme’s textual content and identifying the persuasion technique on display in the meme’s textual and visual content combined. 

All proposed systems performed better than the benchmark. The winning system was a transformer-based model whose training data was “augmented following a Chain-of-Thought-based data augmentation approach using GPT-3.5.”

Importantly, the study “introduces multilinguality, covering four languages, and features the largest dataset in English, along with a new hierarchical evaluation method.”

To learn more about MBZUAI’s research visit their website.  

• Suki, the company that offers AI assistants to hospitals, raised $70 million in a Series D round.

• Enterprise AI testing platform Distributional raised $19 million in Series A funding.

• The rise of AI-powered job application bots (404 Media).

• The bill finally comes due for Elon Musk (The Verge).

• Inflation rate at 2.4% in September, topping expectations; jobless claims highest since August 2023 (CNBC).

• Florida devastated by flooding, fierce winds and tornadoes from deadly Hurricane Milton (NBC News).

• Ancient traditions meet modern technology (Rest of World).

If you want to get in front of an audience of 200,000+ developers, business leaders and tech enthusiasts, get in touch with us here.

• AI Researcher: Labelbox, San Francisco Bay Area

• Senior Applied Scientist: Microsoft, New Jersey

• NoteGPT: An AI-based content summarizer.

• Blog Ideas: A tool to create and optimize blog topics and titles.

Microsoft launches AI tools for doctors

Microsoft on Thursday launched a series of generative AI and data tools designed for use by doctors and the healthcare industry. These include automated documentation tools, healthcare imaging models and a healthcare agent service. 

The details:  Microsoft unveiled what it referred to as a “suite” of medical GenAI models — available on Azure — which it said are designed to handle a different type of medical data. 

• These models, according to Microsoft, can analyze text, image and genomic data, and can also be used by healthcare organizations to build new tools and applications. 

• Microsoft’s healthcare agent service is designed for healthcare practitioners to custom-build Copilots meant to serve specific functions as determined by the organizations in question. 

Microsoft said the tools will improve patient outcomes while reducing clinician burnout, adding that all the platforms are secure and in line with industry standards. 

The context: Concerns about data biases, model hallucinations and data security remain, and were not addressed, either at all, or in any sort of detail, in Microsoft’s press release regarding the launch. Misunderstandings about the capability of the models in use here — especially in medical settings — could cause patient harm.

There was also a recent report out of the Ada Lovelace Institute that determined that AI in healthcare needs to slow down, waiting for proper regulation to address instances along the lines of personalized insurance rates based on private, genomic data that gets parsed and processed by GenAI tools.  

OpenAI’s long, costly road to profitability

We’ve talked recently — and regularly — about the financial challenges, to put it mildly, that surround the industry of generative artificial intelligence. The GPU chips the developers need to train and operate their models are enormously costly, with a single Nvidia chip costing in the region of $30,000 to $40,000, making their clusters of tens of thousands of chips a bit of an expensive venture. 

That doesn’t account for other data center costs, licensing fees (legal fees) and the cost of retaining top talent, which has become significant. 

OpenAI serves as a microcosm of the industry as a whole, and, according to new reporting out of The Information, the startup’s investors will have to be down to lose a lot of money for a long time before they start to see a bit of a return. 

• Based on an analysis of financial documents — including financial statements and forecasts — obtained and reviewed by The Information, OpenAI doesn’t expect to turn a profit until 2029. 

• According to the documents, OpenAI expects to spend some $200 billion through 2030, not including the costs of stock compensation. The Information reported that between 60% and 80% of this expenditure would go toward training or running OpenAI’s models. 

Just to emphasize how large a number $200 billion is, you could spend $5,000 a day for roughly 109,589 years to exhaust that amount (or $1 million a day for 547 years). With $200 billion, you could buy 200,000 homes that are each worth $1 million. Or, you could spend it building generative AI models. I guess. 

Between 2023 and 2028, according to the documents, OpenAI expects to lose a total of $44 billion, again, not including stock compensation. In 2026, it expects its losses to hit $14 billion. 

OpenAI expects to then make $14 billion in profit in 2029. 

• In that golden year of profitability, OpenAI expects to rake in $55 billion in revenue from ChatGPT, $22 billion from its API and $24 billion from vague “new products,” for a grand total of $101 billion in revenues. 

• Microsoft, according to the documents, gets a 20% slice of OpenAI’s revenue as part of the terms of its multi-billion-dollar investment, adding to the firm’s costs. 

Now, this year, OpenAI is on track to lose $5 billion, despite earning between $3 billion and $4 billion in revenue, largely derived from ChatGPT subscriptions. This is, again, excluding stock compensation, which OpenAI spent $1.5 billion on in the first half of this year. 

The way the costs break down, OpenAI in 2024 is spending around $5 billion solely on compute to run and train its models. The remaining costs are coming from data, salaries, sales and general corporate infrastructure. 

OpenAI did not respond to a request for comment. 

This, of course, comes shortly after OpenAI raised $6.6 billion at a $157 billion valuation, in addition to securing a $4 billion line of credit. 

Couple of things. First, this paints a really confident picture, despite the fact that OpenAI doesn’t have a moat and that other developers — such as Meta — are offering similar products for free. 

Second, this indicates that OpenAI expects one of two things to happen: either it will massively grow its user base on ChatGPT (something, again, I find unlikely given the availability of cheaper models), or two, it will massively hike the cost of subscription (something that will likely come soon). 

Even if AI becomes ubiquitous by 2029, I don’t think OpenAI will ascend to Apple-like levels of consumer popularity and market control. It offers one product that is offered by many other companies, and that a lot of people won’t pay anything at all for. 

• Note: back in the ‘70s, Microsoft and Apple, those steady tech empires of today, became profitable almost immediately. 

People buy smartphones today because they have become necessary. If and when AI becomes necessary, it will already be there, integrated for free into the smartphones people have already purchased, something Apple is doing as we speak. This reduces OpenAI’s market to the enterprise, of which there are a number of lesser-known enterprise-specific AI companies — who are not trying to build AGI, by the way — which will make that space quite challenging for the ChatGPT makers. 

My prediction: 2030 will mark for the end for OpenAI, at which point it will be consumed and turned into a division of Microsoft.

🤔 Your thought process:

Selected Image 1 (Left):

• “These are getting really hard to tell, so I have to zoom in now! It came down to whether the lens focus/blurriness was consistent, which was only true for #1. The focus/ detail sharpness on the first image was consistent. Image #2 had some inconsistent focus - there were a few plants in more sharp focus that were both close and far away, and plants/grass next to those plants that were not in sharp focus.”

Selected Image 1 (Left):

• “Light pattern not consistent on lower image.”

💭 A poll before you go

Thanks for reading today’s edition of The Deep View!

We’ll see you in the next one.

Here’s your view on Google’s breakup:

35% are excited about the idea of a breakup; 35% think it’s over-regulation. The rest are undecided.