Meta’s AI Overreach, DeepMind Union Push, and AI Finds Clues to Alzheimer's Cure

Summary

AI Deep Dive: Meta’s AI Overreach, DeepMind Union Push, and AI Finds Clues to Alzheimer's Cure

Hosted by Daily Deep Dives

Episode Overview

In this compelling episode of the AI Deep Dive podcast, hosts A and B navigate through a spectrum of pressing topics in the artificial intelligence landscape. Released on April 27, 2025, this episode delves into Meta’s controversial AI integration in WhatsApp, the burgeoning union movement within Google’s DeepMind, Microsoft’s futuristic vision of AI-driven workplaces, and groundbreaking AI-assisted research pointing towards a potential Alzheimer's cure. Through insightful discussions and expert opinions, the hosts unpack the complexities and implications of these developments, offering listeners a comprehensive understanding of the current AI terrain.

---

1. Meta’s AI Overreach in WhatsApp

Integration Challenges and User Frustration

The episode opens with a critical examination of Meta's latest AI feature embedded within WhatsApp. Hosts A and B discuss the introduction of an optional AI chatbot powered by Meta's Llama 4 model, designed to assist users with queries such as "What's the weather in Glasgow?"

A [00:07]: "You know that feeling? Just information overload constantly. It's like, where do you even start?"

B [00:56]: "Yes, the Meta AI feature. It's labeled optional, but the catch is it's embedded right there in your chat screen."

Despite its intended utility, the feature has sparked user frustration due to its persistent presence and occasional inaccuracies.

A [01:09]: "I bet it feels a bit like here's something you didn't ask for and you're stuck with it. So it's a chatbot."

An illustrative example highlighted is the AI providing incorrect information by linking Glasgow's weather query to Charing Cross Station in London, underscoring the current limitations of AI in delivering precise responses.

Privacy Concerns and Ethical Implications

The conversation shifts to the broader privacy implications of Meta's AI implementation. Dr. Chris Reshak voices significant concerns over Meta's exploitation of its vast user base, likening users to unwitting test subjects. He criticizes the foundational methods used to train these AIs, such as scraping extensive web data, including potentially copyrighted material from sources like Library Genesis.

B [02:02]: "Dr. Chris Reshak has raised some pretty serious points. He's suggesting Meta is exploiting its dominant market position, kind of using its massive user base as well."

In response to these privacy issues, Meta asserts that the AI only accesses messages explicitly shared with it, maintaining that personal chats remain end-to-end encrypted. The UK's Information Commissioner’s Office (ICO) is monitoring the situation.

A [02:39]: "Their official line is that the AI can only read messages specifically shared with it."

Dr. Reshak emphasizes caution, reminding users that interacting with Meta's AI entails sharing information with the company, not just another individual.

B [02:56]: "Every time you use this feature and communicate with Meta AI, you need to remember that one of the ends is Meta, not your friend."

2. DeepMind Union Push

Employees Seek Collective Representation

Transitioning from user-facing AI applications to the internal dynamics of AI development, the hosts shed light on a significant unionization effort within Google’s DeepMind. Approximately 300 DeepMind employees in London are striving to join the Communication Workers Union, motivated by ethical concerns.

A [03:27]: "Yes, that's another significant development. Reportedly, around 300 DeepMind employees based in London are looking to unionize with the communication workers union."

Ethical Dissatisfaction and Corporate Response

The union movement is primarily fueled by dissatisfaction with Google’s recent policy shifts, including the revocation of a pledge against utilizing AI for military purposes and surveillance. A notable point of contention is Google's substantial $1.2 billion cloud contract with the Israeli military, which has ignited protests and ethical debates among staff.

B [03:38]: "It seems to stem from dissatisfaction over Google dropping a pledge against using AI for weapons or surveillance. That, and also concerns about the company's work with the Israeli military."

Some employees feel betrayed, leading to resignations and heightened tensions within the company.

A [04:05]: "Resigned? That shows real conviction. These are the people building the AI, questioning its use."

Google's official stance encourages open dialogue, though critics argue it may be an attempt to downplay employee grievances.

A [04:19]: "Sounds a bit like brushing it off."

The potential unionization of DeepMind employees could have profound implications, given DeepMind's pivotal role in Google's AI initiatives, raising pivotal questions about the control and ethical direction of AI development.

3. Microsoft's Vision for AI-Driven Workplaces

The Emergence of "AI Bosses"

Shifting focus to corporate visions of the future workplace, the hosts explore Microsoft's ambitious prediction of a world where AI agents take on managerial roles. Jarrett Spataro from Microsoft envisions a future where human workers primarily oversee autonomous AI agents, effectively becoming "the CEO of an agent-powered startup."

B [04:45]: "This is quite the prediction. Microsoft's looking ahead and seeing a future dominated by what they call frontier firms."

Phased Integration and Organizational Structure

Microsoft outlines a phased approach for this transformation. Initially, AI assistants will support individuals, gradually becoming integral digital colleagues within teams. Ultimately, humans will set overarching strategies while AI agents manage day-to-day workflows.

B [05:15]: "They see it evolving in phases. First AI assistants helping individuals, then AI agents becoming like digital colleagues on teams. And finally, humans setting the strategy."

A practical illustration includes AI managing complex logistics within supply chains, thereby liberating humans to focus on strategic decision-making and relationship management.

A [05:31]: "They use a supply chain example. AI agents handle all the complex logistics, freeing up humans for, say, strategic decisions and managing relationships."

Economic and Ethical Considerations

While the potential for increased efficiency is evident, significant concerns loom regarding job displacement. Reports from the UK Government's AI Safety Group and the International Monetary Fund (IMF) warn of widespread employment challenges. Even organizations like the Tony Blair Institute acknowledge potential job losses, albeit with some optimism about new roles emerging.

A [05:39]: "I can see the potential efficiency gains, but the obvious question is jobs, right?"

Dr. Andrew Rogoiski further cautions against the indiscriminate replacement of human workers with AI agents, emphasizing the irreplaceable value of human knowledge, intuition, and customer relationships.

B [06:08]: "Dr. Andrew Rogoiski raised a really important point. He warned that companies might be tempted to just swap humans for cheaper AI agents."

This dialogue underscores the delicate balance between embracing technological advancements and preserving the human elements that drive business success.

4. AI Finds Clues to Alzheimer's Cure

Discovery of PHGDH Gene’s Hidden Role

Concluding the episode on a hopeful note, the hosts delve into a groundbreaking AI-assisted study from UC San Diego that uncovers a potential new avenue for understanding and treating Alzheimer's disease. The research focuses on the PHGDH gene, traditionally known for its role in producing serine, an amino acid vital for astrocytes—support cells in the brain.

B [07:04]: "It comes from UC San Diego. They've been looking at a gene called PHGDH."

While elevated levels of PHGDH correlate with the severity of Alzheimer's, the new study posits that PHGDH may actively contribute to the disease's progression by disrupting gene expression in astrocytes.

B [07:24]: "But this new study suggests it might be more than that. It might actually be driving the disease process, at least in some forms."

AI’s Role in the Breakthrough

The pivotal discovery was made possible through advanced AI techniques that modeled the 3D structure of the PHGDH protein. This modeling revealed an unexpected "moonlighting" function of PHGDH, where it acts as a transcription factor, directly influencing gene activation.

B [08:21]: "This is where AI comes back in. They used AI analysis to get a super precise 3D model of the PhDH protein's shape, and that revealed how it could interact with DNA."

Professor Zhong, leading the study, highlights AI's crucial role in uncovering this hidden functionality, marking a significant leap in Alzheimer's research.

Potential Therapeutic Developments

Building on this discovery, researchers identified a small molecule drug, NCT503, which selectively inhibits PHGDH's gene-regulating activity without hindering its serine-producing function. Importantly, NCT503 can cross the blood-brain barrier, making it a viable candidate for therapeutic intervention.

A [09:13]: "What does it do?"

B [09:14]: "It seems to specifically block PHDH's gene regulating function without messing too much with its normal job of making serine. And crucially, it can get into the brain."

Preclinical trials in mouse models demonstrated that NCT503 reduced amyloid plaque accumulation, improved memory performance, and alleviated anxiety-related behaviors.

B [09:26]: "The results were promising in mouse models of Alzheimer's, giving them NCT503 led to fewer amyloid plaques."

This approach targets the disease's underlying mechanisms rather than merely addressing symptoms, offering a promising new direction for early-stage treatments and potential preventative measures.

B [09:58]: "It's targeting the process that leads to the plaques, potentially much earlier, before widespread damage occurs."

Caveats and Future Directions

While the findings are promising, the hosts acknowledge the early-stage nature of this research. Extensive further studies, including human trials and safety evaluations, are necessary before NCT503 can become a viable treatment option.

A [09:58]: "Of course, it's early days, isn't it? Mouse models aren't people."

The research suggests that PHGDH may be a critical nexus point where genetic, aging, and lifestyle factors converge in Alzheimer's pathology, opening avenues for multifaceted intervention strategies.

Conclusion: Navigating the Future of AI

In wrapping up the episode, the hosts reflect on the diverse and profound implications of AI's integration into various facets of life and research. From ethical quandaries in user-facing applications and internal corporate dynamics to transformative potential in healthcare, AI continues to reshape the world in multifaceted ways.

A [10:50]: "Which brings us to our final thought for you. Given that surprise discovery about the PHGHDH gene's hidden role, what other things in our biology, things we think we understand, might have secret functions with huge implications for else?"

B [11:12]: "Or switching back to AI as it weaves itself deeper into our lives, our work, our communication? How do we strike that right balance? How do you weigh the convenience against very real concerns about privacy, jobs and ethics?"

The episode leaves listeners pondering the delicate balance between leveraging AI for unprecedented advancements and addressing the ethical, social, and economic challenges it presents. As AI continues to evolve, so too does the imperative to navigate its integration thoughtfully and responsibly.

---

Key Takeaways:

• Meta’s AI in WhatsApp: Introduces user-facing AI with notable privacy concerns and user frustration due to intrusive integration and imperfect functionality.

• DeepMind’s Unionization Efforts: Highlights ethical disagreements among AI developers regarding corporate policies and military contracts, reflecting broader concerns about AI's societal impact.

• Microsoft’s AI-Driven Workplace Vision: Explores the potential future where AI agents manage workflows, raising significant questions about job displacement and the preservation of human-centric business practices.

• AI-Assisted Alzheimer's Research: Showcases how AI can accelerate biomedical discoveries, potentially leading to new therapeutic targets and treatments for complex diseases like Alzheimer's.

This episode of AI Deep Dive encapsulates the dynamic and often contentious evolution of artificial intelligence, emphasizing the need for ongoing dialogue and critical analysis as AI continues to permeate diverse aspects of society.

Transcript

[A]

You know that feeling? Just information overload constantly. It's like, where do you even start? You want the important stuff, the real gist, but finding it takes ages.

[B]

It's overwhelming.

[A]

Exactly. So that's why we're here. Welcome to the Deep Dive. We're doing the heavy lifting for you, giving you that shortcut.

[B]

Yeah. Today, we're aiming for quick, sharp summaries. We pulled some key updates from AI Deep Dives news feed covering AI developments and also something really intriguing on the Alzheimer's research front.

[A]

Right, so we'll look at AI popping up in apps you use every day and then shift gears to a potential breakthrough in, well, understanding a major disease.

[B]

Our mission, basically, is to extract the core insights for you. Concise summaries, so you leave feeling informed, not buried.

[A]

Okay, let's jump right in that little blue circle in WhatsApp. What's the story there?

[B]

Yes, the Meta AI feature. It's labeled optional, but the catch is it's embedded right there in your chat screen.

[A]

You can't turn it on?

[B]

Nope. It's just there, that blue icon, and it's causing some friction, as you might imagine.

[A]

I bet it feels a bit like here's something you didn't ask for and you're stuck with it. So it's a chatbot.

[B]

Yeah. Powered by Meta's Llama 4 model. It's meant to answer questions like, you could ask it about the weather in Glasgow.

[A]

Okay, useful enough.

[B]

But then in one example, it apparently gave the Glasgow weather, but also threw in a link to Charing Cross Station in London.

[A]

In London. Not exactly helpful. Shows it's not quite perfect yet, then.

[B]

Definitely not. And it highlights those AI limitations. But the user frustration, I think, goes deeper than just wonky answers. Yeah, it sort of echoes that backlash Microsoft faced with its recall feature. Remember that people are wary of AI that's just always there watching or recording.

[A]

Right. And WhatsApp trying to say, oh, it's just like channels or status updates. That doesn't really wash, does it?

[B]

It doesn't seem to be a landing rail. No. And this is where the privacy angle gets really sharp.

[A]

Okay, tell me about that.

[B]

Dr. Chris Reshak has raised some pretty serious points. He's suggesting Meta is exploiting its dominant market position, kind of using its massive user base as well.

[A]

Unwitting testers using people as test subjects. That's a strong claim.

[B]

It is. And he also points to the fundamental way these AIs are trained. You know, scraping vast amounts of web data, possibly including copyrighted stuff like pirated books. From places like Library, Genesis. There's that whole ongoing legal battle with authors about it.

[A]

Ah, right, the training data issue. So what's Meta's response to the privacy concerns?

[B]

Their official line is that the AI can only read messages specifically shared with it. They stress that your personal chats remain end to end, encrypted. And the UK's Information Commissioner's Office, the ICO, says they're keeping an eye on it, monitoring it.

[A]

Hmm. But Dr. Srishak had that really stark warning, didn't he?

[B]

He did. He said, and I think this is worth repeating for you, listening. Every time you use this feature and communicate with Meta AI, you need to remember that one of the ends is Meta, not your friend.

[A]

Wow, that really puts it bluntly. Meta isn't your buddy in that chat.

[B]

Exactly. Even Meta advises only sharing info you're comfortable with them potentially retaining and using. So caution is definitely advised.

[A]

Okay, so from AI in our chats to AI developers themselves, what's happening at Google DeepMind? There's talk of unionizing.

[B]

Yes, that's another significant development. Reportedly, around 300 DeepMind employees based in London are looking to unionize with the communication workers union.

[A]

300? That's quite a few. What's driving it?

[B]

It seems to stem from dissatisfaction over Google dropping a pledge against using AI for weapons or surveillance. That, and also concerns about the company's work with the Israeli military. Specifically that big $1.2 billion cloud contract that caused protests.

[A]

Right. I remember hearing about that contract. So the employees feel the company's going in a wrong direction ethically.

[B]

Apparently so. One source mentions staff feeling duped. And reports suggest at least five people have actually resigned over these issues.

[A]

Resigned? That shows real conviction. These are the people building the AI, questioning its use.

[B]

It really underscores the tension. Google's response is the usual corporate statement about encouraging constructive and open dialogue.

[A]

Sounds a bit like brushing it off.

[B]

Perhaps it's worth noting there was a smaller union effort at Google before. But DeepMind is so central to their AI work, this could potentially have more impact if it moves forward. It really raises questions for you about who controls AI's direction.

[A]

Absolutely. Okay, let's shift from the workforce to the workplace itself. Microsoft's vision, everyone becoming an AI boss. Give us the quick summary on this one.

[B]

Right. This is quite the prediction. Microsoft's looking ahead and seeing a future dominated by what they call frontier firms.

[A]

Frontier firms?

[B]

Yeah, companies where human workers primarily direct autonomous AI agents to do the actual tasks. Jarrett Spataro from Microsoft literally said you'd become the CEO of an agent powered startup.

[A]

Wow. So managing AI instead of people, that's a fundamental shift. They think this is coming soon.

[B]

They expect organizations to be heading this way within five years, structuring themselves around on demand intelligence.

[A]

How would that work? Practically?

[B]

They see it evolving in phases. First AI assistants helping individuals, then AI agents becoming like digital colleagues on teams. And finally, humans setting the strategy. And the AI agents managing workflows. Just checking in occasionally.

[A]

Okay.

[B]

They use a supply chain example. AI agents handle all the complex logistics, freeing up humans for, say, strategic decisions and managing relationships.

[A]

I can see the potential efficiency gains, but the obvious question is jobs, right?

[B]

Exactly. That's the big concern hanging over this. Reports from the UK Government's AI Safety Group and the IMF warn about potentially widespread job displacement. Even the Tony Blair Institute, while often pro AI, acknowledges significant potential job losses in the uk Though they also predict new roles will be created. It's a massive unknown.

[A]

And there are other risks too, aren't there? Beyond just job numbers?

[B]

Yes. Dr. Andrew Rogoiski raised a really important point. He warned that companies might be tempted to just swap humans for cheaper AI agents. Right. But in doing so, they risk losing invaluable human knowledge, experience, intuition, and those crucial customer relationships. It's not just about efficiency. It's about what makes a business actually work.

[A]

A real balancing act. Then productivity versus, well, people and the knowledge they hold. Okay, final topic. Let's pivot completely. Alzheimer's disease. It's such a huge challenge. And for most cases, the cause is still murky. Right. Especially for people without those known genetic links.

[B]

That's right. We know about certain genes for early onset Alzheimer's and risk factors like APOE4, but they don't explain the majority of cases that develop later in life. Life, which is, as you said, most cases, Nearly everyone over 65 shows some signs of the underlying pathology.

[A]

So what's this new insight?

[B]

It comes from UC San Diego. They've been looking at a gene called phgdh. Now, this gene was already known to be involved in making serine, an amino acid, important for brain cells, particularly support cells called astrocytes.

[A]

Okay.

[B]

And levels of PhGDH are higher in people with Alzheimer's. The more severe the disease, the higher the levels. It was seen as maybe a marker and indicator.

[A]

It's a signpost.

[B]

Yeah, but this new study suggests it might be more than that. It might actually be driving the disease process, at least in some forms.

[A]

How? By doing what?

[B]

By messing with how other genes are expressed in those astrocyte cells. The Researchers did experiments using mice and also human brain organoids, sort of mini brains grown in the lab that didn't have the known inherited Alzheimer's mutations. And they found that boosting PhGHDH activity made the Alzheimer's like pathology worse. Reducing its activity slowed things down.

[A]

Okay, that points toward a causal role. But how does PHGH mess with other genes? I thought its job was making serine.

[B]

That's the really fascinating bit. They discovered it has a hidden moon lighting function. It can also act as a transcription factor, meaning it can directly latch onto DNA and control whether other genes are turned on or off. And this function wasn't known before.

[A]

Wow. How did they figure that out?

[B]

This is where AI comes back in. They used AI analysis to get a super precise 3D model of the PhDH protein's shape, and that revealed how it could interact with DNA. Professor Zhong, who led the study, said AI was crucial for this discovery.

[A]

Incredible. So this hidden gene regulating job goes wrong somehow.

[B]

Exactly. In the context of Alzheimer's, this moonlighting function seems to get out of whack. It disrupts the normal balance of gene activity in astrocytes, particularly affecting the cells ability to clean up waste, a process called autophagy.

[A]

And that leads to.

[B]

That leads to more buildup of those toxic beta amyloid plaques, the classic hallmark of Alzheimer's.

[A]

It's disrupting the cleanup crew, allowing the junk to pile up.

[B]

That's a good way to put it. And based on this, they use AI tools again to find a potential drug, a small molecule called NCT503.

[A]

What does it do?

[B]

It seems to specifically block PHDH's gene regulating function without messing too much with its normal job of making serine. And crucially, it can get into the brain.

[A]

And did it work in the animal models?

[B]

The results were promising in mouse models of Alzheimer's, giving them NCT 503 led to fewer amyloid plaques. And the mice performed better on memory tests and showed less anxiety.

[A]

That sounds really significant because most existing treatments try to clear plaques after they formed, right? This sounds like it's intervening earlier in the chain reaction.

[B]

Precisely. It's targeting the process that leads to the plaques, potentially much earlier, before widespread damage occurs. Professor Zhong mentioned this could open doors to a new class of early stage treatments, maybe even pills.

[A]

That would be huge. Of course, it's early days, isn't it? Mouse models aren't people.

[B]

Absolutely. Major caveats apply. They need lots more research, safety studies, the whole regulatory process before it could ever reach patients. But it's a genuinely new angle it.

[A]

Suggests Alzheimer's, or at least a common form of it, might fundamentally be about this breakdown in gene management kicked off by this one gene PhDH.

[B]

Yeah, offering a potential new way to think about detection and maybe even prevention down the line. It fits with the idea that Alzheimer's is complex, involving genetics, aging, lifestyle. Maybe PhDH is a key player where these factors converge.

[A]

Okay, so just to wrap up this quick dive, we've gone from AI in your WhatsApp to AI ethics debates among developers to AI maybe being your future.

[B]

Boss and finishing with AI helping uncover a potential new cause and treatment paths for Alzheimer's.

[A]

Quite a range, definitely. It gives you, the listener, a brief snapshot of some really important shifts happening right now. Even short summaries like these pack a lot in.

[B]

They really do. Lots to think about.

[A]

Which brings us to our final thought for you. Given that surprise discovery about the PHGHDH gene's hidden role, what other things in our biology, things we think we understand, might have secret functions with huge implications for else?

[B]

Or switching back to AI as it weaves itself deeper into our lives, our work, our communication? How do we strike that right balance? How do you weigh the convenience against very real concerns about privacy, jobs and ethics?

[A]

Good questions, definitely things worth mulling over. We hope this quick summary has given you some food for thought.

[B]

We'll be back soon to dive deep into another topic.