Good morning. I’m up in Albany today for something cool you’ll be hearing about shortly. That was a tease. Keep your eye out!

In other news, a coalition of more than 100 AI researchers published an open letter last week affirming the continued existence and threat of bias and discrimination in AI.

The group of scientists is urging “policymakers to continue to develop public policy that is rooted in and builds on this scientific consensus, rather than discarding the bipartisan and global progress made thus far.”

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

In today’s newsletter:

• 🩺 AI for Good: Dealing with tuberculosis 

• 🏛️ Apple faces proposed class action over its lag in Apple Intelligence

• 💰 Meta has been quietly earning revenue from Llama, court filing shows

• ⚕️ The challenges of making AI for mental health care work

AI for Good: Dealing with tuberculosis

India, home to more than a quarter of the world’s tuberculosis infections, is at the center of a global battle against the disease. In an effort to meet its target of total TB elimination by 2025, the country has turned to AI-powered technological innovations. 

The details: Prominent among these is something called DeepCXR, an AI tool developed by the Institute for Plasma Research that is designed to detect early traces of TB in chest X-rays. 

• India’s Ministry of Health combined the software with handheld X-ray devices, which are being used as part of its 100-day intensive campaign to combat TB. 

• The handheld devices offer easy maneuverability, enabling the quick and early detection of TB among vulnerable populations. 

“AI tools are expected to be a game-changer in detecting presumptive TB patients and quick initiation of treatment,” Minister of State for Health and Family Welfare, Anupriya Patel, said in a statement. 

While Patel highlighted India’s progress — aided and abetted by technological innovations — in reducing TB infections and deaths, doctors as recently as August said that India’s goal of eliminating TB by 2025 was “unachievable.” 

200+ hours of research on AI tools & hacks packed in 3 hours

This free 3-hour Mini Course on AI & ChatGPT (worth $399) will help you become a master of 20+ AI tools & prompting techniques and save 16 hours/week.

Get it now for absolutely free! (for first 100 users only) 🎁

This course will teach you how to:

• 👉 Automate Tasks and Save Time: Leverage AI tools to automate repetitive tasks, freeing up your time to focus on strategy and leadership.

• 👉 Accelerate Decision-Making: Use AI-driven insights to make smarter, faster decisions that drive your business forward.

• 👉 Scale Your Personal Brand: Learn how to use AI to craft personalized content that enhances your visibility and authority on LinkedIn.

• 👉 Master AI Communication: Learn how to use ChatGPT for writing emails, reports, and presentations in minutes—saving hours each week.

Register & save your seat now! (100 free seats only)

Apple faces proposed class action over its lag in Apple Intelligence

Apple, already moving slowly out of the gate on generative AI, has been dealing with a number of roadblocks and mounting delays in its effort to bring a truly AI-enabled Siri to market. The problem, or, one of the problems, is that Apple used these same AI features to heavily promote its latest iPhone, which, as it says on its website, was “built for Apple Intelligence.” 

Now, the tech giant has been accused of false advertising in a proposed class action lawsuit that argues that Apple’s “pervasive” marketing campaign was “built on a lie.” 

The details: Apple has — if reluctantly — acknowledged delays on a more advanced Siri, pulling one of the ads that demonstrated the product and adding a disclaimer to its iPhone 16 product page that the feature is “in development and will be available with a future software update.”

• But that, to the plaintiffs, isn’t good enough. Apple, according to the complaint, has “deceived millions of consumers into purchasing new phones they did not need based on features that do not exist, in violation of multiple false advertising and consumer protection laws.” 

• Apple “enriched itself by saving the costs they reasonably should have spent on ensuring that the (iPhones) had the technical capabilities advertised,” according to the complaint. 

Apple did not respond to a request for comment. 

The lawsuit was first reported by Axios, and can be read here. 

This all comes amid an executive shuffling that just took place over at Apple HQ, which put Vision Pro creator Mike Rockwell in charge of the Siri overhaul, according to Bloomberg. 

Still, shares of Apple rallied to close the day up around 2%, though the stock is still down 12% for the year. 

Transform Your Business with Advanced AI and Automation

Discover how to use AI beyond basic copywriting and research. Today, it's possible to use AI to drive more sales, generate quality leads, and optimize efficiency. This exclusive webinar reveals proven strategies that enable AI to work for you — turning every facet of your business into a growth engine.

Reserve your spot now and start monetizing your business.

• Nvidia’s way forward: Nvidia’s GTC conference this year was, in true Nvidia fashion, record-breaking, drawing some 25,000 people to San Jose. Throughout the conference, the firm and its CEO tried hard to project confidence that the boom will continue. But last year, shares of Nvidia spiked 7% during the week it held GTC; this year, shares of Nvidia closed the week down 3.6%. Not the boost it needed.

• Europe’s semiconductor push: A coalition of nine European countries is banding together to set up plans to boost the Union’s semiconductor industry.

• Inside Google’s two-year frenzy to catch up with OpenAI (Wired).

• Perplexity wants to buy TikTok and open-source its algorithm (The Verge).

• Microsoft is exploring a way to credit contributors to AI training data (TechCrunch).

• More work for teachers? The ironies of GenAI as a labour-saving technology (Professor Neil Selwyn).

• Apple is working on turning its watches into AI devices (Bloomberg).

Meta has been quietly earning revenue from Llama, court filing shows

Meta CEO Mark Zuckerberg, a staunch defender of so-called open-source technology, wrote an impassioned defense of open AI last July, saying at the time that “a key difference between Meta and closed model providers is that selling access to AI models isn’t our business model.” 

Openly releasing Llama, according to Zuckerberg, “doesn’t undercut our revenue, sustainability, or ability to invest in research like it does for closed providers.” 

The news: But, beyond the fact that Llama isn’t really open-source by any acceptable definition, newly unredacted filings in the class action lawsuit Kadrey V. Meta revealed that Meta is actually deriving some level of revenue from its Llama models.

• Beyond the commercial opportunity around integrating Llama into its other products, the filing says that Meta has earned money from agreements with the companies that host Meta’s models. This is paid out as a percentage of the revenue those companies generate from Llama users. 

• Both the amount that Meta has earned from this arrangement, and the companies who have signed those checks, have been redacted from the court filing. 

The finding is significant to the case, which argues that Meta didn’t just knowingly train its model on datasets of pirated books, but that the company engaged in something called “seeding,” in which a user shares stolen data with other users.

The case has revealed internal communications from Meta’s AI team that show that, in an effort to catch up to the competition, Meta began pursuing the licensing of high-quality content in the form of books. But the effort was too expensive and time-consuming, according to the documents, so Meta received approval from Zuckerberg to use datasets of pirated books instead. 

The challenges of making AI for mental health care work 

Benchmarks have become something of a holy grail in the AI field. 

When we’re dealing with systems designed to automate high-level knowledge work, the need for some method of measurement is clear. But the problem, as we’ve often discussed in this newsletter, is that benchmarks are rarely a good indicator of real-world performance, especially when it comes to generative AI. 

• A simple lack of transparency crossed with the consistent black-box nature of large language models (LLMs) means that researchers don’t really know if a model is exhibiting genuine performance, or if it was just trained on the information in the benchmark test. 

• It’s the difference between, for example, a student studying for and passing a test, and a student who was given the test to study, then memorized the answers and recalled them later. 

In some realms, and for some people, the difference might not matter. 

But it’s a piece of nuance that gets more important as these models get integrated into higher-stakes environments, as it relates heavily to levels of user trust.

This is especially true of GenAI integrations in healthcare fields, an integration that is happening in full force, today. Many GenAI systems in healthcare are back-end systems, employed by researchers to speed up drug development, for instance. But increasingly, we’re seeing the rise of AI-powered clinical assistants and notetakers designed to help out nurses and reduce the administrative burden faced by doctors. 

Ethics and challenges aside, it's an integration that requires a robust reliability calculus; an effective benchmark.  

Researchers at Stanford just proposed one for mental health, a field that has not been spared the unrelenting push of AI integration. 

The details: Current benchmarks, according to the paper, are built to mimic exams, and come complete with multiple-choice answer options. The problem with this is that, “even for humans … success in these standardized tests only weakly correlates with clinicians’ real-world performance, a disconnect that can be especially problematic in psychiatry, where diagnosis and management hinge on subjective judgments and interpersonal nuances.” 

This is even worse with the benchmarks that assess LLMs, according to the researchers, which “over-simplify the complexities of day-to-day clinical practice tasks.” 

• The proposed benchmark, which the researchers made openly available on GitHub under an MIT license, was curated by a diverse group of experts across five major domains: diagnosis, treatment, monitoring, triage and documentation.

• It focuses on real-world ambiguity, assessing open-ended clinical responses, and was built without the use of an LLM.  

The researchers assessed a number of off-the-shelf models from developers including OpenAI, Anthropic and Meta. The models, which aren’t designed or intended for clinical applications, performed well in the diagnosis category, collectively achieving an 80% accuracy rate. 

But they averaged accuracy rates of less than 50% for triage and documentation, clocking only a 67% for monitoring and a 76% for treatment. 

According to the researchers, the models “perform well on structured tasks,” but “struggle significantly with ambiguous real-world tasks … underscoring the limitations of current AI models in handling uncertainty.” 

Under the impression, laid out so elegantly by Psychology Today, that “like it or not, (AI) is here to stay,” we’re going to need many more benchmarks of this nature — expert-curated and as ambiguous and nuanced as humanly possible. They’re harder to train for, and they are absolute necessities when clinicians are considering the adoption of these tools. 

Experts in domains beyond computer science absolutely need to be made aware of just how reliable, and just how trustworthy, generative AI models are (or are not). Even 80% accuracy rates ought to be pretty unacceptable in a bunch of fields. 

Of course, as the paper mentions, this doesn’t address a couple of key problems that surround this integration, which include data security and privacy, consent, reliability and algorithmic bias. 

Taken together, it adds up to the risk that “AI systems could be prematurely deployed in psychiatric care, potentially leading to harmful, biased or unreliable clinical decisions.” 

But I guess you can’t have everything. 

This is the start of a start. 

🤔 Your thought process:

Selected Image 1 (Left):

• “My instinct said that the spaced-out houses were more correct. Upon inspection it seems the perspective, lines, and windows are more consistent.”

Selected Image 2 (Right):

• “Very hard to tell now!!! 🙂”

💭 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 Claude’s web search:

A third of you think it’s a bit of a weird rollout; you already use other search-enabled systems.

21% think it comes too late and 14% think it puts Anthropic back in the game.

Back in the game:

• “I wouldn't count anybody (out) in this; if we believe this technology has lasting power, it's a marathon, not a race.”

Weird rollout:

• “Well, after half a century here’s hoping that your wish comes true and the NY Knicks once again attain greatness, winning the NBA championship.”

Yes.

If you want to get in front of an audience of 450,000+ developers, business leaders and tech enthusiasts, get in touch with us here.