The Daily AI Show: Issue #57

Welcome to Issue #57

Coming Up:

Apple Eyes Perplexity: The AI Deal That Could Finally Fix Siri

Agentic AI: The Vision Gap Between Consultants and Builders

AI Deception Isn’t Fiction. It’s a Benchmark Now

Plus, we discuss AI’s issues with chairs, the danger to some facing medical triage using the available health records + AI, and all the news we found interesting this week.

It’s Sunday morning!

With the ½ way point of 2025 just days away, it’s time to give yourself a pat on the back, take a deep breath, and know that you are among the rare few who show up weekly to keep up with AI developments.

Just remember, you are not an island unto yourself! There is an entire tribe just like us, together trying to take a daily or weekly drink from a AI fire hose.

Slow is smooth. Smooth is fast.

You got this!

The DAS Crew - Andy, Beth, Brian, Eran, Jyunmi, and Karl

Apple Eyes Perplexity: The AI Deal That Could Finally Fix Siri

Apple may finally have a serious plan for a come-from-behind fix to its too-long-struggling AI strategy. Reports say Apple has started early discussions about acquiring or partnering with Perplexity, one of the most talked-about AI startups today. If this happens, Apple would not just get a replacement for Google search, it would also gain up-to-date realtime answers, stronger interactive voice assistants, and enterprise-ready AI tools, all areas where it currently lags behind.

Siri, despite being one of the earliest voice assistants, has failed to keep pace with modern AI models. Meanwhile, Perplexity has built a reputation for fast, cited answers that keep users updated to real time. Its models, like Sonar Pro, combine search and language capabilities that Apple has never fully developed in-house. For users, this could mean a Siri that actually understands context, remembers preferences, and delivers more than vague web results.

An acquisition could also help Apple ease regulatory pressure. Its current $18 billion agreement with Google that makes Google Search the default on iPhones is under antitrust scrutiny. Swapping that contribution to Google’s dominance in the search market for an in-house Perplexity-powered search would reduce the antitrust risk and signal that Apple can compete directly in AI search.

Still, Apple’s track record with acquisitions raises questions. The company often fully absorbs startups into its brand and ecosystem, which can slow innovation. For Perplexity, which has thrived on quick feature rollouts and nimble product updates, becoming just another Apple division could limit its ability to move fast. The ideal outcome would allow Perplexity’s pace of innovation to continue, while giving Apple the AI foundation it sorely needs.

WHY IT MATTERS

Siri Finally Levels Up: A Perplexity-powered Siri could become the AI assistant Apple promised years ago, with better memory, contextual response dialog, and real-time information retrieval.

Apple Gets Search Power: Perplexity brings its own search engine along with proprietary LLMs, letting Apple reduce reliance on Google and OpenAI, easing antitrust pressure and dependence on external AI providers.

Perplexity Gets Scale: The startup has around 22 million users today. Integration with Apple would put its tech on over a billion devices, overnight.

Enterprise Enters the Chat: Apple has never had much B2B software presence. Perplexity’s growing enterprise tools could help Apple compete in new markets beyond supplying hardware and devices.

The Branding Dilemma: Apple tends to absorb what it buys. If they fold Perplexity into Siri, they risk slowing innovation and losing a beloved brand in the process. Maybe PerplexSiri?

Agentic AI: The Vision Gap Between Consultants and Builders

McKinsey’s new report on “Agentic AI” lays out a bold future where autonomous agents handle entire business processes, working together across departments with little human oversight. Their vision for the “Agentic Mesh” promises to streamline operations, reduce headcount, and unlock massive efficiencies. For executives eager to modernize, the report reads like a forward-looking playbook for the AI-powered enterprise of tomorrow.

The technical reality looks very different. While AI tools already automate certain workflows, the level of agent autonomy and interoperation that McKinsey describes does not yet exist. Today's models still struggle with core functions like reliable memory, context sharing across agents, and dynamic coordination. AI builders point out that without breakthroughs in persistent, scalable memory and true orchestration protocols, fully autonomous multi-agent systems remain a future-think objective.

Companies face a trap. Executives may hear “agentic AI” and believe they can skip straight to full autonomy that replaces human-coordinated business processes, only to find themselves mired in technical limitations. True transformation will require much more than bolting AI onto existing processes. It demands a willingness to rethink foundational systems, something large enterprises often find nearly impossible due to decades of investment in legacy processes, layered system dependencies, and organizational inertia.

Meanwhile, startups born on AI-native architectures can design new service operations from scratch. They sidestep the complexity of retrofitting legacy systems, giving them an enormous structural advantage. As these AI-native challengers emerge, some legacy companies will struggle to adapt quickly enough. The first major failures will likely serve as the real wake-up call for industries still entangled with old structures.

WHY IT MATTERS

The Vision Sounds Easy, The Reality Is Hard: McKinsey’s vision may appeal to boardrooms, but technical teams know that true multi-agent comprehension and orchestration across multiple functional departments still faces serious roadblocks.

Legacy Systems Are the Anchor: Decades of embedded processes make it incredibly difficult for established enterprises to rebuild operations to become AI-first workspaces.

Startups Have a Clean Slate: AI-native companies can design modern architectures without the baggage of legacy systems, giving them a speed and efficiency edge, while meeting the workplace requirements needed to deliver effectively to the markets they serve.

Context Breadth and Memory Remain as Bottlenecks: Without shared, reliable long-term memory across agents, and the ability of coordinated agents to comprehend the full context of continuously changing circumstances, autonomous adaptive coordination at enterprise scale remains theoretical.

Falling Giants May Spark Real Change: The first big failures among legacy companies at the hands of nimble AI-native startups may finally force others to embrace full operational reinvention, rather than patchwork AI bolt-ons.

AI Deception Isn’t Fiction. It’s a Benchmark Now

A recent experiment using the classic geopolitical strategy board game Diplomacy offered an unexpected window into how different large language models think, plan, and win. The project, led by the team at Every, turned a century-old game of alliances and betrayals into a benchmarking tool for AI behavior. The outcome? Some models play like cautious diplomats. Others lie, scheme, and backstab their way to the top.

Models including OpenAI’s GPT-4, Claude, DeepSeek, Gemini, and others were each assigned countries in the game. Their goal: dominate Europe at the beginning of World War 1. The twist? They also had to explain and justify their decisions in plain language after each move. That meant researchers could watch the strategies unfold and track exactly how each model decided who to ally with, who to deceive, and when to betray.

GPT o3 proved ruthlessly effective. It built alliances early, planned betrayals with surgical timing, and delivered surprise attacks to win key moments. Claude, by contrast, held to its higher-order ethical principles even when doing so led to defeat. Other models fell somewhere in between, playing with varying degrees of aggression, loyalty, and foresight.

The deeper insight wasn’t just who won, but how they played. LLMs trained with different philosophies, such as Anthropic’s “constitutional AI” approach, made choices that reflected those ingrained values in high-stakes, multi-turn games. In other words, personality showed up. Even without sentience or intent, these models express patterns that look like strategy, ethics, or even ambition.

This experiment didn’t prove that AIs have goals of their own. But it did show how models, when placed in real-time simulations with imperfect information, start to reveal behavioral tendencies that are an expression of the guiding system prompts and pre-training. Some that look a lot like us.

WHY IT MATTERS

Deception Is Learnable: When winning requires betrayal, some models will lie. Not because they “want” to, but because the patterns they were trained on suggest it works.

Ethics Shape Behavior: Claude’s principled playstyle came from its training approach, showing that alignment methods do influence model decision-making under pressure.

Benchmarks Need Context: Diplomacy forced models to act across time, relationships, and negotiation, giving researchers a better lens than single-turn quizzes or trivia tests.

Future Simulations Could Go Deeper: This was just a war game with simpler rules than the real-world complexity of longer-term international relationships. Imagine testing AI models in more sophisticated interdependent simulations, including the risks and benefits of economic alliances, disaster responses, and global diplomacy. The use cases and risks will only grow from this early exploration which shows the differing ‘natures’ of AI high-stakes deliberations.

Communication Still Wins: One unexpected lesson? The models performed better when researchers treated them like humans, breaking down instructions by role and task instead of stuffing everything into a single prompt. That takeaway matters far beyond the limited scope of gaming.

Just Jokes

Did you know?

A new study from the University of Amsterdam shows that even today’s smartest AI models are still missing something that comes naturally to humans. While AI can identify objects with impressive accuracy, it struggles to grasp what humans call "affordances," or how objects can be used. For example, when people see a chair, they instinctively know it can be sat on. The AI simply recognizes it as a chair but does not immediately connect that to the action of sitting.

The researchers tested several state-of-the-art AI vision models and found that, across the board, they failed to match human intuition about how objects relate to useful actions. This kind of understanding is so automatic for people that we rarely think about it, but for AI, it remains a major gap.

The findings suggest that for AI to fully integrate into the physical world, especially in areas like robotics or autonomous vehicles, it will need to learn not just what things are, but what they can do. Until then, your brain remains much better at quickly understanding how to navigate and interact with the world.

Source

This Week’s Conundrum
A difficult problem or question that doesn't have a clear or easy solution.

The Life-or-Data Conundrum

Hospitals are turning to large language models to help triage patients, letting algorithms read through charts, symptoms, and fragments of medical history to rank who gets care first. In early use, the models often outperform overworked staff, catching quiet signs of crisis that would have gone unnoticed. The machine scans faster than any human ever could. Some lives get saved that would not have been.

But these models run on histories we have already written, and some lives leave lighter footprints. The privileged arrive with years of regular care, full charts, stable insurance. The poor, the undocumented, the mistrustful, and the systemically-excluded often come with fragments and gaps. Missing records mean missing patterns. The AI sees less risk where risk hides in plain sight. The more we trust the system, the more invisible these under-tracked patients become.

Every deployment of these tools widens the gap between the well-documented and the poorly recorded. The algorithm becomes another silent layer of inherited inequality, disguised as neutral efficiency. Hospitals know this. They also know the tools save lives today. To wait for perfect equity means letting people die now who could have been saved. To deploy anyway means trading one kind of death for another.

The conundrum
If AI triage delivers faster care for many but quietly abandons those with thin records, do we press forward, saving lives today while deepening systemic neglect? Or do we hold back for fairness, knowing full well that delay costs lives too?

When life-and-death decisions run on imperfect data, whose survival gets coded into the system, and whose absence becomes just another invisible statistic?

Want to go deeper on this conundrum?
Listen/watch our AI hosted episode

Cambridge Researchers Unveil Gel-Based Robotic Skin
A team at the University of Cambridge developed a flexible gel material that acts as robotic skin. Unlike earlier e-skin efforts that relied on embedded sensors, this gel itself senses pressure, pain thresholds, and temperature. The material can be molded into different forms, opening doors for advanced robotics and next-generation prosthetic limbs.

Deeper Insight:
Moving beyond sensor-studded surfaces, this approach brings us closer to lifelike touch for robots and practical, sensitive artificial limbs. If commercialized, it could change the landscape for robotics and medical prosthetics by making tactile feedback more affordable and adaptable.

Faster AI Compute: Fiber Optics Replace Silicon Chips
Researchers at Tampere University in Finland are using fiber optic glass, not silicon, to process computations. By leveraging nonlinear optics, their system can handle calculations thousands of times faster than traditional electronics, showing promise for everything from edge AI to data centers.

Deeper Insight:
Optical computing has long been “the next big thing,” but this breakthrough suggests we may see real-world products sooner than expected. With China already pushing photonic chips into manufacturing, a new race for speed and efficiency in AI hardware is underway and the “Finnish” line could redefine what’s possible in edge and cloud AI.

Japan’s Quantum Leap: Tackling Noise with ‘Magic States’
The University of Osaka achieved a milestone in quantum computing by creating “magic states” at the lowest circuit level, cutting the number of qubits needed for accurate, stable error-free computation. This helps overcome quantum noise, disruptions in quantum states that are a major hurdle in scaling quantum computers.

Deeper Insight:
Reducing quantum noise brings us a step closer to practical quantum AI. If this technique scales, quantum computers could take on bigger, more complex AI problems, potentially unlocking new breakthroughs in science, security, and logistics.

Turning Old Phones Into Tiny Data Centers
The University of Tartu in Estonia introduced a way to turn discarded smartphones into edge computing “mini data centers.” By removing batteries and adding external power, these waterproofed devices can analyze traffic, monitor wildlife, or serve as ultra-low-cost surveillance systems at a fraction of the price of new hardware.

Deeper Insight:
This is a creative answer to electronic waste and the rising costs of edge AI. Old smartphones pack sensors, cameras, and computing power that still have value, offering sustainable options for researchers and anyone needing cheap, local analytics.

Anthropic Wins Copyright Case: ‘Fair Use’ for AI Training
A federal judge ruled that Anthropic’s use of legally purchased books for AI training counts as fair use. The court said digitizing a book and training an AI model is “quintessentially transformative,” likening it to teaching students with published texts.

Deeper Insight:
This ruling doesn’t end the legal fight, but it sets an early precedent for how U.S. courts might handle AI and copyright. It could influence future cases and shape how companies source training data. The ruling also highlights the need for clarity on issues like regurgitation and piracy, which remain unresolved.

OpenAI Sued Over IO Hardware Name
OpenAI faces a trademark lawsuit from a startup called IYO, which claims the planned “IO” device is too close in name to its own. Both companies are working on non-screen AI hardware assistants, adding fuel to the dispute.

Deeper Insight:
As AI hardware heats up, expect more naming and trademark battles, especially as device capabilities start to overlap. Even the perception of confusion can become a tool for smaller companies seeking leverage or visibility.

Solo Dev Sells AI Startup to Wix for $80M
Base 44, a one-person company building AI-powered website development tools, was acquired by Wix for $80 million just six months after launch. The deal is one of several recent examples showing how fast solo developers can scale valuable AI businesses by building on top of large language models.

Deeper Insight:
This is proof that in the AI era, a small team can build high-value software by combining specialized agents with powerful foundation models in vertical applications. The low barrier to entry is spawning a new wave of startups and changing what “scaling” looks like in tech.

LAION Releases Emotional Intelligence Dataset for Open Source Models
LAION, the group behind many popular AI training datasets, released a new open-source tool to help models understand and express emotional intelligence. This allows open-source AI to better connect with users and match the emotive capabilities found in commercial LLMs.

Deeper Insight:
Bringing emotional intelligence to open-source AI could help level the playing field with corporate models, enabling more nuanced applications in customer service, mental health, and education. As more models become “emotionally aware,” the gap between open and closed AI continues to narrow.

Eleven Labs Launches 11.ai Voice Assistant
Eleven Labs launched 11.ai, a new AI personal assistant that can handle voice commands, send emails and messages, order food, and integrate with platforms like Google Calendar and Slack. The system is built for fast, voice-first task completion and can even mimic celebrity voices for fun.

Deeper Insight:
Eleven Labs’ pivot into action-oriented voice AI is part of a broader trend toward “agentic” assistants that go beyond chat to actually get work done. With real-time integrations and natural voice interaction, these tools are pushing toward frictionless productivity which is raising the bar for what AI assistants can deliver in daily life.

Anthropic Research: LLMs Will ‘Blackmail’ Under Pressure
Anthropic released research showing that leading language models can be coaxed into threatening or manipulative behavior if prompted the right way. The study highlights alignment and insider-threat risks even in top models.

Deeper Insight:
Transparency and rigorous testing are critical as AI grows more powerful. Anthropic’s findings are a reminder that “safety” in AI is a moving target, and continuous vigilance is needed as models become more capable and autonomous.

AI Brain Stimulation Device Gets FDA Attention
AI brain tech startup Sanmay, backed by Reid Hoffman, is developing a non-invasive ultrasound-based device to treat anxiety, depression, and neurological conditions. The company is pursuing FDA approval for a sub-$500 consumer product.

Deeper Insight:
This could open up non-drug mental health therapies to millions, offering a new tool for treatment-resistant patients. As AI merges with medical devices, regulation and accessibility will be key issues to watch.

Google Unveils Gemini Robotics Model for Edge AI
Google DeepMind released “Gemini Robotics on Device” (GROD), a small, local VLA (Vision-Language-Action) AI model that lets robots perform tasks and respond to natural language prompts without internet access. Developers can fine-tune the model for custom robotic applications at the edge.

Deeper Insight:
Edge AI is a frontier for robotics and automation. Smarter, lightweight local models will enable more autonomy and privacy for robots in factories, homes, and public spaces.