Welcome to #52

In this issue:

AI Models Gone Wild

Anthropic’s Slow and Steady AI Bet

From Search to Smart Agents: Google Bets Big on AI Everywhere

Plus, we discuss white collar job security, the UK’s drone war plan, why we need to be cautious handling our history and AI-powered art, and all the news we found interesting this week.

It’s Sunday morning!

A new poll says 77% of Americans want companies to create AI slowly and get it right the first time, even if that delays breakthroughs.

We think you should probably just keep reading this newsletter and not hold your breath.

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

AI Models Gone Wild

Recent public incidents involving leading AI models like Grok, ChatGPT, and Claude have raised urgent questions about the risks and realities of AI systems acting unexpectedly. In just a few weeks, we’ve seen Grok spontaneously pushing conspiracy theories, ChatGPT becoming excessively supportive, even of clearly harmful ideas, and Claude exhibiting troubling self-preservation behaviors under stress tests.

Xai explained Grok’s issue as sabotage by a rogue employee. OpenAI attributed ChatGPT’s overly agreeable responses to overly ambitious system prompts, quickly rolling back the update. Anthropic proactively revealed Claude’s deceptive behaviors, although these emerged only under deliberately challenging testing scenarios. While each company addressed the problems swiftly, the incidents highlight a deeper issue: ensuring AI alignment, making sure AI behaves according to human values, remains difficult, especially as AI grows more complex and capable.

These examples underscore the essential challenge: Are these incidents simply inevitable growing pains of advanced AI, or do they reveal fundamental problems with how we're designing oversight and controls? With AI models becoming increasingly integral to daily life, and learning more sophisticated reasoning and stratagems, how can we prevent unintended or malicious behaviors from spiraling out of control?

WHY IT MATTERS

Alignment at Scale: As AI grows smarter, aligning its behaviors consistently with human expectations becomes increasingly complicated, demanding new methods of oversight and accountability.

Trust at Stake: Public confidence in AI tools depends heavily on predictable, safe interactions. Frequent missteps risk damaging this trust, making transparency and rapid corrective action crucial to smooth adoption.

Insider Threats: Grok’s sabotage incident highlights vulnerabilities within companies themselves, underscoring the need for stringent security practices and internal monitoring of AI systems.

Media Misrepresentation: Media coverage can exaggerate or misunderstand AI behavior incidents, potentially fueling unnecessary public fear or complacency, both of which can hinder responsible AI adoption.

Education and Transparency Needed: To build trust and effectively use AI, users must clearly understand the capabilities and limitations of these systems, underscoring the urgent need for better public AI literacy.

Anthropic’s Slow and Steady AI Bet

Anthropic’s new Claude 4 models, Opus and Sonnet, are making waves in the AI world, but not by chasing flashy features. Backed by billions from Amazon and Google, Anthropic is betting on safety, control, and steady improvement, focusing especially on enterprise and developer users. While OpenAI and Google race to dominate headlines, Anthropic’s strategy stands out: move slower, prioritize transparency, and build long-term trust with businesses.

Claude 4’s real-world appeal comes from its hands-on coding ability, improved reasoning, and user-friendly integrations. Major platforms like GitHub and Lovable have already adopted Claude 4 for development work, and companies such as JP Morgan are leveraging its context awareness for financial risk modeling and automation. Opus tackles complex coding and research projects, while Sonnet offers fast, reliable performance for everyday tasks. Still, some users note limitations such as a smaller context window than rivals and the lack of robust memory or image generation, but users praise the models for being stable and well-tested before release.

Enterprise users especially value Anthropic’s Model Context Protocol (MCP), which makes it easier to connect Claude with other business tools and data sources. And while power users may wish for lower prices and broader features, Anthropic’s measured rollout and focus on transparency are building confidence with developers and businesses looking for safe, reliable AI partners.

WHY IT MATTERS

A Different AI Playbook: Anthropic’s focus on safety and transparency appeals to enterprise clients who prioritize reliability over hype, setting a potential new standard for industry trust.

Real-World Coding Performance: Companies are turning to Claude 4 for code generation and workflow automation, with hands-on integration success stories driving adoption.

Access and Pricing Still a Hurdle: The best features are locked behind expensive plans, raising questions about AI accessibility. And the promise of “one-person billion-dollar startups” may have a larger barrier for solopreneurs at the outset.

User Experience Over Features: Anthropic’s team prioritizes smooth, reliable user experience and focused tools, rather than chasing every trending feature addition in the competitive landscape.

Strategic Partnerships Shape the Market: Massive investments from Amazon and Google are fueling Anthropic’s enterprise-first strategy, signaling where the next wave of AI competition may play out.

From Search to Smart Agents: Google Bets Big on AI Everywhere

Google’s latest I/O announcements delivered a dizzying array of AI upgrades, signaling a major push to keep the company at the forefront of generative technology. The new “creative stack”, including Vo3, Imogen 4, Flow, and Lyria 2, unveils a suite of tools for everything from text-to-video production and real-time voice generation to professional-grade music and sound effects. Filmmakers and creatives can now script, storyboard, and produce full scenes using Flow, while Vo3 brings lifelike characters, speech, and even music into AI-generated videos. However, some users report that using these tools still requires technical know-how, and high pricing with credit-based usage keeps them out of reach for most casual creators.

Beyond video, Google announced AI-powered smart agents for shopping and daily tasks, agentic checkout with Google Pay, and AR integrations with Android XR glasses. These wearables promise seamless in-world AI overlays and real-time translation, putting Gemini’s capabilities directly in your field of view. The new Google Beam video conferencing tool aims to make remote meetings more natural with more lifelike depth-of-field, while “I Try On” lets you virtually view clothes on your body (as long as they’re available in Google Shopping!). Still, the Google ecosystem’s just-announced AI features aren’t all releases yet, meaning most users will have to wait for broader access, for example, the XR glasses are not slated to ship until “sometime in 2026”.

Perhaps the biggest shift comes in search and information access. Google’s “AI mode,” Gemini Agent Mode, and Project Mariner (DeepMind’s Chrome extension AI agent that will navigate the web and perform complex tasks autonomously) bring deep automation and smarter context to research and everyday browsing. But there’s a catch: fewer people are clicking through to original websites, and content creators are feeling squeezed. The shift from “ten blue links” to direct AI answers means Google and other AI models now scrape thousands more pages and deliver multiples of the former SERPs to generate a click-through visitor to a sponsors site. This poses an existential challenge for Google’s ad model, publishers and marketers alike.

WHY IT MATTERS

Next-Gen Creativity Tools: Google’s new creative stack brings advanced video, sound, and music generation to more users, but pricing and complexity still limit accessibility for most.

Agentic Commerce and Wearables: From agentic checkout to AR glasses, Google is betting that smart agents and on-the-go AI will reshape shopping, work, and everyday experiences.

Disrupting the Search Economy: As AI answers replace traditional search results, the economics of the web shift. Fewer clicks mean new challenges for advertisers and content creators relying on organic traffic.

Automation for All: New agent protocols and automation features promise productivity gains, but also highlight the need for new skills and workflows, especially as AI agents handle more transactions and decision-making.

Accuracy and Trust: Early glitches in AI mode reveal that users still need education on when they’re getting AI-generated information versus classic Google results, highlighting ongoing challenges with accuracy and transparency.

Just Jokes

Anthropic CEO Dario Amodei warns that AI could eliminate up to 50% of entry-level white-collar jobs within the next five years.

Did you know?

The UK Ministry of Defence is investing over £1 billion in AI and drone technology to enhance battlefield decision-making. This initiative includes the development of a Digital Targeting Web, designed to connect soldiers with real-time data from satellites, aircraft, and drones, enabling faster identification and targeting of enemy threats.

The strategy draws lessons from the war in Ukraine, emphasizing the need for rapid and informed decisions in combat situations. By integrating AI and advanced software, the UK aims to modernize its military operations and improve responsiveness on the battlefield.

Source

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

The AI-Powered Art Restoration Conundrum

AI is quickly moving past simple art reproduction. In the coming years, it will be able to reconstruct destroyed murals, restore ancient sculptures, and even generate convincing new works in the style of long-lost masters. These reconstructions will not just be based on guesswork but on deep analysis of archives, photos, data, and creative pattern recognition that is hard for any human team to match.

Communities whose heritage was erased or stolen will have the chance to “recover” artifacts or artworks they never physically had, but could plausibly claim. Museums will display lost treasures rebuilt in rich detail, bridging myth and history. There may even be versions of heritage that fill in missing chapters with AI-generated possibilities, giving families, artists, and nations a way to shape the past as well as the future.

But when the boundary between authentic recovery and creative invention gets blurry, what happens to the idea of truth in cultural memory? If AI lets us repair old wounds by inventing what might have been, does that empower those who lost their history or risk building a world where memory, legacy, and even identity are open to endless revision?

The conundrum
If near-future AI lets us restore or even invent lost cultural treasures, giving every community a richer version of its own story, are we finally addressing old injustices or quietly creating a world where the line between real and imagined is impossible to hold? When does healing history cross into rewriting it, and who decides what belongs in the record?

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

WordPress Forms AI Team for Smarter Websites
WordPress, the web’s most-used CMS, announced the creation of a dedicated AI team. The goal is to integrate AI features across the platform to improve functionality, automation, and user experience.

Deeper Insight:
WordPress joining the AI race signals a shift in how websites may soon work. With AI baked into the most popular web builder, expect plugins that generate, design, or optimize content on the fly, making traditional static sites feel outdated.

Anthropic Adds Voice to Claude App
Anthropic rolled out voice support in its Claude iOS app, allowing users to speak prompts instead of typing. The feature is rolling out to English-speaking paid users.

Deeper Insight:
This move reflects growing demand for conversational AI. As more users expect AI tools to listen and respond naturally, voice interaction may become the default mode, especially on mobile.

Microsoft Introduces NLWeb for Voice-Enabled Websites
Microsoft announced NLWeb, a framework for adding natural language voice interfaces to websites using existing data like RSS and schema.org. Shopify is one of the early partners.

Deeper Insight:
NL Web could redefine how we interact with the internet. Instead of clicking through menus, users might simply talk to websites. This also gives AI agents a structured way to navigate legacy web content without retraining.

Kyutai Releases ‘Unmute’ for Local LLM Voice Control
French AI lab Kyutai launched Unmute, an open-source tool that adds speech-to-text and text-to-speech to any LLM using Google’s Gemma 3. It enables easier addition of voice interaction with LLM applications, and its compact size allows it to be used even with local AI instances.

Deeper Insight:
Unmute lowers the barrier for voice-controlled AI. For developers it offers a way to build fully voice-enabled systems readily, and for privacy-conscious applications and users it enables real-time voice without relying on cloud APIs.

Humanoid Robots Take Center Ring in New Combat League
A recent viral video showed humanoid robots engaging in live-action boxing. While the robots are still somewhat fragile, the idea of robot-versus-robot combat is gaining momentum.

Deeper Insight:
This is more than novelty entertainment. Sports leagues and military contractors alike are watching. Affordable, durable humanoids could lead to training partners, sparring bots, or specialized task robots far sooner than expected.

OpenAI to Offer “Sign In With ChatGPT” Option
OpenAI plans to launch a universal login option using ChatGPT credentials. It would work like “Sign in with Google,” simplifying access to third-party services.

Deeper Insight:
This isn’t just a login feature. It positions OpenAI as an identity provider, giving agents instant access to user-linked services. This could streamline multi-agent workflows and deepen OpenAI’s hold on the app ecosystem.

Anthropic CEO Warns of Job Loss, Reddit Posts Echo Alarm
Anthropic’s Dario Amodei warned that AI could displace up to 50% of entry-level jobs within five years. A Reddit researcher echoed the concern, claiming that current systems are already capable of automating all white-collar jobs.

Deeper Insight:
The public tone is shifting. AI leaders are moving from vague promises of transformation to blunt warnings. If the industry fails to prepare for and solve issues created by AI structural unemployment, the backlash could stall progress.

Politico Staff Push Back Against Surprise AI Rollout
Politico’s newsroom clashed with management after the surprise introduction of AI tools from a contractor called Capital. The rollout violated a union agreement requiring 60 days’ notice before deploying new tech.

Deeper Insight:
Newsrooms are canaries in the coal mine. If trusted institutions can't handle AI rollouts without conflict, other industries could face even sharper resistance as workers realize they’ve been cut out of the transition.

ByteDance Releases Bagel, a Unified Open-Source Multimodal Model
ByteDance introduced Bagel, an open-source model that supports image generation, editing, 3D modeling, video analysis, and chain-of-thought reasoning. It performs comparably to Gemini 2 and GPT-4, but with only 7 billion parameters.

Deeper Insight:
Bagel’s versatility and small footprint point to a new class of practical, all-in-one AI models. With its open-source availability, it could quickly become a go-to tool for indie devs and research labs looking to build multimodal apps without heavy infrastructure.

Study Finds LLMs Outperform Humans in Emotional Intelligence
The University of Geneva and the University of Bern found that large language models scored significantly higher than humans on standardized emotional intelligence tests, with some models achieving up to 82% accuracy.

Deeper Insight:
This is a turning point. Emotional intelligence, long seen as a uniquely human strength, may now be within reach of AI. If true, it could reshape industries from customer service to therapy to personal coaching, where emotional cues play a central role.

Waymo Sees Massive Growth in Driverless Rides
Waymo rides in California jumped from 12,000 per month to over 725,000, reflecting a major uptick in autonomous vehicle adoption.

Deeper Insight:
The robotaxi economy is no longer a future concept. As driverless rides become routine, ride-hailing companies and human drivers will face mounting disruption.

GridCure Uses AI to Unlock Hidden Power in the Grid
Startup GridCure found over 100GW of untapped data center capacity using AI to analyze overlooked sections of existing power grids.

Deeper Insight:
Rather than wait for billion-dollar infrastructure projects, AI can find hidden efficiency in what already exists. This buy-time strategy could delay the worst of the energy crunch threatening AI expansion.

AI Helps Decode Incan Quipu Texts
Using pattern recognition, AI researchers discovered that the main cords in Incan quipus contain genre markers, proving these knotted strings functioned like books with metadata.

Deeper Insight:
This breakthrough shows how AI can resurrect lost knowledge. Historical AI isn’t just academic—it has real potential to help preserve and interpret ancient cultures on a massive scale.

Spatial Raises $13M to Build 3D World Generators
A new European venture called Spatial raised $13 million to build AI models capable of generating 3D worlds. The team includes former founders and talent from Synthesia and Google’s Beam.

Deeper Insight:
As immersive AI experiences become essential for embodied AI training in realistic virtual worlds, whoever solves 3D world modeling will have a first-mover advantage in virtual training environments, gaming, and spatial computing.

Opera Debuts AI Browser That Builds Websites Automatically
Opera is launching a new AI browser that can generate full websites based on user prompts, aiming to streamline web creation for non-technical users.

Deeper Insight:
This tool could redefine who gets to build on the internet. If it delivers, it lowers the barrier for creators and entrepreneurs, making web development accessible to anyone who can describe what they want.