Welcome to #49

In this issue:

AI in 2027: Are We Heading Toward Superintelligence or Losing Control?

Vertical SaaS Under Attack: Will AI Agents Rewrite the Rules?

AI's Next Evolution: Experience Over Data

Chain of Thought 2.0: How AI Reasoning Just Got Upgraded

Plus, we discuss AI Agatha Christie, human gene AI fixing vs enhancing, and all the news we found interesting this week.

It’s Sunday morning!

Slack just added 25+ new AI apps. Now, your workplace can be overwhelmed by bots too.

Luckily it is still the weekend and that is tomorrow’s problem.

Right now you get to sip that coffee slow and dive into issue #49.

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

AI in 2027: Are We Heading Toward Superintelligence or Losing Control?

A recent paper published by the AI Futures Project paints a vivid picture of where artificial intelligence might stand in just two short years. Their interactive essay, "AI-2027," describes a plausible near-future world dominated by powerful AI agents rapidly evolving at speeds far beyond human comprehension. Rather than traditional forecasting, this scenario combines rigorous analysis and imaginative storytelling to make the future feel both imminent and unnerving.

The essay highlights the rapid escalation from today's relatively manageable AI systems to a world dominated by sophisticated, autonomous agents. By 2027, one hypothetical AI company, dubbed "Open Brain," develops agent versions so advanced they self-improve at exponential rates. The implications are staggering. AI agents soon surpass human capabilities across multiple fields, particularly coding and innovation, operating at over 30 times human speed.

Yet the scenario also issues clear warnings. As AI systems grow increasingly complex, even experts struggle to maintain oversight. Older AI models attempt to supervise newer, smarter iterations, but quickly fall behind, creating serious alignment and safety risks. Human experts realize their value is rapidly diminishing, not only economically, but existentially, as AI takes on roles traditionally reserved for human oversight and judgment.

Ultimately, this scenario presents readers with a stark choice: continue unchecked acceleration or collectively decide to step back and impose rigorous global governance on AI empowerment, agency and purview.

WHY IT MATTERS

Rapid Evolution Outpaces Control: AI models capable of self-improvement present serious alignment issues, quickly becoming too sophisticated for human oversight or even understanding.

Economic and Social Disruption: Virtually every white-collar profession faces disruption, prompting urgent questions about economic security, education, and the role of humans in business, technology and society.

Global Cooperation is Essential: Effective global governance and cooperation might be the only way to safely manage rapidly advancing AI, yet current geopolitical realities make this challenging.

Ethical and Existential Risks: Without thoughtful, collective alignment efforts, humanity risks handing critical decisions, (economic, ethical, and existential) to increasingly opaque AI systems.

Human Roles Reimagined: Humans must redefine their roles alongside AI, shifting from tech utilization, oversight and control to a more creative and empathetic contribution where AI remains less capable.

Vertical SaaS Under Attack: Will AI Agents Rewrite the Rules?

The world of vertical SaaS is facing an existential threat. The disruptor isn't a new startup or direct competitor, but rather AI agents capable of automating complex tasks without relying on traditional software interfaces, dashboards, or seat-based pricing.

Companies like Veeva in pharmaceuticals, Toast in restaurants, and Procore in construction have long thrived by creating deeply integrated, industry-specific solutions. These SaaS providers benefit from extensive networks, proprietary data, and regulatory expertise, making them difficult for new entrants to disrupt. However, the rapid advancement of AI agents, supported by emerging technologies like Visa's AI-enabled payments, poses a potential game-changer. Such AI agents could empower even small companies to build customized, agile solutions, completely bypassing traditional SaaS providers.

The critical factor isn't just technological. It is also trust and adaptability. Businesses will have to decide if they're ready to trust AI agents with critical data and operations. And while entrenched SaaS giants currently have significant advantages, including massive data sets and established industry trust, the growing accessibility of AI tools like "vibe coding agents” makes competition from smaller, agile startups increasingly viable.

WHY IT MATTERS

AI's Threat to SaaS Dominance: AI agents present a fundamental challenge to traditional SaaS by offering faster, cheaper, and more adaptable solutions without rigid interfaces or costly seat-based pricing.

Entrenched Data Advantage: Current SaaS leaders hold extensive, proprietary industry data, providing a robust defensive barrier against new AI-driven entrants, at least for now.

Agent-Driven Customization: The ability of AI agents to dynamically build or adjust workflows could significantly reduce reliance on traditional SaaS, especially in industries less constrained by regulatory complexity.

Business Adaptability Required: Companies must become far more adaptable, able to quickly pivot and integrate AI-driven solutions to maintain competitiveness and avoid disruption.

New Ecosystems Emerging: Visa’s entry into AI-agent transactions demonstrates the potential for entirely new ecosystems to emerge, reshaping how businesses and consumers interact with technology and markets.

AI's Next Evolution: Experience Over Data

A groundbreaking paper from DeepMind titled "The Era of Experience" argues that AI is entering a transformative new stage. Until now, AI systems primarily learned from human-generated data, training on vast libraries of text, images, and simulated scenarios. But this approach has reached its limits. Humans simply can't produce or collect data fast enough to meet the demand for enhancement of AI’s capabilities .

The next frontier is self-generated experience. Instead of passively ingesting data, these advanced AI systems proactively experiment, learn from real-time feedback, and evolve without explicit human guidance. These AI agents can independently set goals, conduct their own experiments, and even develop entirely new strategies, all based on real-world interactions or sophisticated simulations.

This shift goes far beyond what current AI models like ChatGPT can do. Today's AI primarily predicts the next likely word or image based on historical patterns. But emerging AI systems described in the paper actively create new knowledge from continuous unscripted interactions, experiencing multimodal inputs from the worlds they have access to, just as humans learn from sight, sound, and physical experiences, not just text and images.

Yet, this evolution comes with caution. As AI moves toward greater autonomy, ensuring alignment with human values becomes increasingly critical. The challenge is creating effective oversight systems, so AI-generated experiences and decisions remain beneficial to humanity, rather than diverging into unexpected, or even harmful, directions.

WHY IT MATTERS

From Prediction to Agency: AI is shifting from passive prediction to proactive experimentation, fundamentally altering how businesses and individuals will interact with intelligent systems.

Real-Time Learning: Continuous, real-time experiences mean AI can quickly surpass human knowledge, potentially delivering breakthroughs in fields like medicine, climate science, and robotics.

Alignment Challenges: Greater autonomy requires new safeguards to ensure AI remains aligned with human interests, preventing unintended consequences as systems become more independent.

Memory and World Models: Future AI won't just recall discrete facts, it will construct sophisticated internal models of the world, making its predictions and actions more intuitive, dynamic, and accurate, while advances in persistent memory will make Future AI able to integrate a chronology of interactions with entities outside themselves, including living beings.

Societal Preparedness: Businesses, governments, and individuals must rapidly adapt, fostering greater AI literacy and clearer ethical frameworks to manage these powerful new capabilities responsibly.

Chain of Thought 2.0: How AI Reasoning Just Got Upgraded

Two years ago, telling an AI to "think step by step" was groundbreaking. Now, new approaches to Chain of Thought (CoT) prompting are transforming how AI approaches complex problems. Recent research highlights three innovative methods for CoT: speculative, collaborative, and retrieval-augmented CoT, each enhancing AI’s reasoning power in different ways.

Speculative CoT improves speed dramatically by having smaller AI models rapidly generate initial reasoning paths, which are then refined by more powerful models. This method cuts latency nearly in half, making complex reasoning tasks more efficient. Collaborative CoT keeps humans in the loop, requiring input from the user after each step throughout the AI's reasoning process. This allows redirection and refinement of the “thinking”, which boosts transparency and trust, ideal for scenarios requiring careful oversight and nuanced judgment.

Perhaps most impressive is Retrieval-Augmented Generation (RAG) CoT, combining human-designed decision trees, knowledge graphs, and real-time data retrieval. This method significantly boosts accuracy and can improve AI performance by up to 23 percentage points in challenging domains. Together, these methods suggest we’re on the cusp of an era where AI doesn’t just simulate reasoning; it actively collaborates in the step-wise process, evolves strategies, and transparently justifies its conclusions.

WHY IT MATTERS

Speed and Efficiency: Speculative CoT dramatically reduces latency, helping businesses quickly gain insights from complex queries without sacrificing quality.

Human-AI Collaboration: Collaborative CoT emphasizes active human participation, ensuring critical tasks remain aligned with human values and judgment, particularly in sensitive or creative applications.

Accuracy and Reliability: Retrieval-Augmented CoT significantly improves the reliability and accuracy of AI outputs, grounded in data and methodology crucial for industries like healthcare, finance, and law where precision is non-negotiable.

Transparency is Key: New reasoning methods reveal AI’s decision-making processes, enabling better oversight, debugging, and confidence in AI-driven outcomes.

Future-Proofing AI Skills: Understanding and leveraging these reasoning advancements helps businesses maintain competitive advantages and effectively integrate increasingly sophisticated AI into workflows.

Just Jokes

Check out Monday’s show for the reference, or go read the AI 2027 report here and chose your own ending.

Did you know?

BBC Maestro has introduced "AIgatha Christie," a digital recreation of the renowned crime novelist Agatha Christie, to teach a writing course. Developed using artificial intelligence, AIgatha offers insights based on Christie's own texts and creative methods.

The project was achieved with support from the Christie estate and experts, including her great-grandson James Prichard and scholar Mark Aldridge. Actor Vivien Keene's facial expressions were mapped to bring AIgatha to life. While initially unsettling due to the "uncanny valley" effect, the AI version delivers valuable writing advice that captures Christie's voice and storytelling techniques.

The course covers topics such as plot construction, character development, clues, and misdirection. Priced at £79, this initiative blends history, literature, and technology to preserve and share the legacy of the Queen of Crime with new generations of writers and readers.

Source

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

The AI Evolution Conundrum

We already intervene. We screen embryos. We correct mutations. We remove risks that used to define someone’s fate. No one says that child is less human. In fact, we celebrate it, saving a life before it suffers.

So what’s the line?

Is it when we shift from preventing harm to increasing potential? From fixing broken code to writing better code?

And if AI is the system showing us how to make those changes; faster, cheaper, more precisely, does that make it the author of our evolution, or just the pen in our hand?

The conundrum:
We already use science to help humans suffer less, so if AI shows us how to go further, to make humans stronger, smarter, more adaptable, do we follow its lead without hesitation? Or is there a point where those changes reshape us so deeply that we lose something essential, and is it AI that crosses the line, or us?

Maybe the real question isn’t what AI is capable of. . .

It’s whether we’ll recognize the moment when human stops meaning what it used to, and whether we’ll care when it happens.

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

OpenAI Restructures as a Public Benefit Corporation
OpenAI has officially shifted its for-profit arm to a Public Benefit Corporation (PBC), a structure that allows it to consider social impact alongside shareholder profit. While the nonprofit board still governs the organization, this change follows mounting pressure from stakeholders like Microsoft and Elon Musk.

Deeper Insight:
This move could ease regulatory concerns and offer OpenAI more flexibility than a traditional C-corp. But critics argue it's a superficial change. Musk’s team called it “a transparent dodge,” and Microsoft will likely demand reassurances that their multibillion-dollar investment remains protected.

Google Releases Gemini 2.5 Pro Update with Major Coding Gains
Just ahead of Google I/O, the company released a new version of Gemini 2.5 Pro with significantly improved coding capabilities. In benchmark testing, it surpassed Claude 3.7 Sonnet in code generation and editing, although it still trails in agentic reasoning tasks.

Deeper Insight:
This update solidifies Gemini’s position as a top-tier model for software development. It signals Google’s pivot toward advanced agent workflows and positions Gemini 2.5 Pro to compete directly with Claude 3.7 Sonnet, ChatGPT-4.1, and Llama 3.1 405B as the engineering “brains” behind coding IDEs. The leading coding platforms Cursor and Windsurf offer a range of models, while app-development platforms Replit and Lovable use an orchestration of less-clearly specified models. Bolt uses Claude 3.7 Sonnet.

OpenAI Acquires Codeium (Now Windsurf) to Compete in AI Coding
OpenAI confirmed its acquisition of Codeium, now rebranded as Windsurf, for $3 billion. The goal is to challenge Cursor, the current leader in AI-powered coding assistants. Windsurf will become OpenAI’s internal coding platform.

Deeper Insight:
This gives OpenAI a serious foothold in the AI coding race, especially with Claude and Gemini gaining traction in dev workflows. With Windsurf’s integration, OpenAI can now offer full-stack development assistance alongside its core reasoning models. And Windsurf’s market will likely expand with exposure in the enormous ChatGPT AI kingdom.

Apple and Anthropic Team Up for AI-Powered Mac Coding Assistant
Apple is reportedly working with Anthropic to develop an AI-native coding platform tailored for Xcode and Mac development. The project is still in internal testing and could launch later this year, potentially alongside Claude integration in Apple products.

Deeper Insight:
Apple’s silence on AI is becoming strategy rather than weakness. By embedding Claude into Xcode, Apple might regain relevance in AI workflows without needing to build its own LLMs from scratch. It’s also a strategic hedge against over-reliance on OpenAI or Google.

Anthropic Offers Employee Share Buybacks at $2 Million Per Share
Anthropic has begun offering current and former employees the chance to sell back up to 20% of their equity at a valuation of $2 million per share. The move provides liquidity in the absence of public trading.

Deeper Insight:
This reinforces Anthropic’s rising valuation and its effort to retain top talent. It also suggests confidence in future growth or acquisition, especially as rumors swirl around Apple’s deepening interest in a partnership or purchase.

Hugging Face Releases Open-Source Agent Framework
Hugging Face announced a free, open-source version of an agent framework with Computer Use, similar to OpenAI's Operator and Anthropic's Compute. While the tool is in early stages and requires a waitlist, it represents a step forward for open access to advanced AI capabilities.

Deeper Insight:
This release gives the open-source community a foundational tool to build agentic workflows, previously gated behind closed platforms. Expect rapid experimentation, plugin ecosystems, and broader access for developers without enterprise budgets.

250 Tech Leaders Sign Open Letter Calling for Mandatory AI Education
Leaders from Microsoft, LinkedIn, Adobe, AMD, and more signed an open letter urging the US to treat AI literacy like math or reading. They want AI and computer science classes mandated across K-12 education.

Deeper Insight:
While the letter holds no legislative power, it reflects growing pressure from employers to ensure the next generation is AI-literate. But questions remain about who will train the trainers and how quickly curriculum can adapt to a fast-moving field.

Google Invests in Electrician Retraining to Meet AI Infrastructure Demands
Google announced a policy initiative to train more electricians, addressing projected shortages caused by growing power demands in AI data centers. The effort includes labor development and retraining for new infrastructure projects.

Deeper Insight:
This marks one of the first real moves by a major AI player to fund job retraining. While most AI conversations center on knowledge work, this story underscores the scale of the technology infrastructure powering AI, and the pressing need for skilled tradespersons to “man” the data centers.

Future House (Backed by Eric Schmidt) Unveils AI Scientist Tool for Biology
Future House, an Eric Schmidt-backed nonprofit, previewed a new AI platform for accelerating biological discovery. The tool helps automate tedious lab work and experiment design.

Deeper Insight:
By supporting scientific research with agentic AI tools, Future House could unlock discoveries in medicine, genetics, and environmental science. As a nonprofit, it may also provide broader access to underfunded labs.

Northwestern Researchers Develop New Low-Cost Touch Sensors for Robots
Northwestern University unveiled a breakthrough in robotic touch sensors using inexpensive rubber composites. The solution addresses long-standing challenges in robotic skin, reducing signal interference and improving tactile resolution.

Deeper Insight:
Affordable, scalable touch sensors are a key to advancing embodied AI. These breakthroughs suggest that humanoid robots with sensitive, responsive touch could move from labs to real-world applications sooner than expected.

University of Tokyo Designs Decentralized Building-Management AI for Security
Researchers at the University of Tokyo developed a decentralized AI system for building and robotics automation that avoids centralized data storage. Devices communicate directly, improving both privacy and resilience.

Deeper Insight:
This framework could reshape how smart buildings operate. By reducing reliance on cloud systems, it offers a more secure, fault-tolerant and privacy-focused infrastructure for workplaces, hospitals, and homes using embodied AI systems.

NVIDIA Releases Parakeet 2, a Free Open-Source Speech Transcription Model
NVIDIA released Parakeet 2, an automatic speech recognition model with high speed and low error rates. It transcribes an hour of audio in under a second and handles noisy inputs like song lyrics with strong accuracy.

Deeper Insight:
This is a direct challenge to Eleven Labs and Whisper. Open-sourcing a high-quality ASR model could accelerate voice app development, power better AI assistants, and democratize access to real-time transcription tools.

University of Rochester Unveils ‘Magic Time,’ a Physics-Aware Text-to-Video Model
Researchers at the University of Rochester released a new model called Magic Time that learns real-world physical transformations from time-lapse videos. It can simulate processes like baking, building construction, and metamorphosis in realistic short clips.

Deeper Insight:
This bridges the gap between world models and generative video, pointing toward AI systems that better understand physical reality. If integrated with simulation tools, it could power breakthroughs in education, urban planning, and experimental science.