• AI Trailblazers
  • Posts
  • Forget AGI. Focus on AI That Acts, Learns, and Respects Boundaries

Forget AGI. Focus on AI That Acts, Learns, and Respects Boundaries

April 21, 2025

In this edition of the newsletter, we discuss the Bloomberg article on AGI - is it possible? We also delve into the necessary pairing of AI agents with people and the closing of a massive AI security loophole. Happy reading.

Now is the time to register for our annual AI Trailblazers Summer Summit, happening on June 5th in New York, NY. Registration is now open—don’t wait to secure your spot, as space is limited. And while you're at it, check out the AI Trailblazers Power 100, created in partnership with ADWEEK.

AGI Is a Mirage—Follow the Code, Not the Hype

The term artificial general intelligence (AGI) is capturing headlines and investor attention—but the reality is far more complex and uncertain than the hype suggests. In this thought-provoking piece by Bloomberg, Gideon Lichfield unpacks why AGI may not be the right framework for understanding the future of AI. Instead, he urges us to move beyond sci-fi metaphors and focus on what today’s and tomorrow’s AI systems are actually capable of doing.

  • The term "AGI" is murky and misleading. While tech leaders like Sam Altman and Elon Musk predict its arrival within years, definitions of AGI vary wildly — from Nobel-level cognition to physical-world autonomy. The label often serves more as hype to attract funding than as a clear technical milestone.

  • LLMs are impressive but fundamentally narrow. Though they can perform tasks like passing exams and generating text, they often fail at generalizing knowledge or reasoning creatively. Even as they become more capable, their intelligence remains bounded, inconsistent, and task-specific — far from true human-level adaptability.

  • Human intelligence isn’t the universal benchmark we think it is. Our minds evolved for our specific bodies, environments, and social needs — intelligence isn’t a linear ladder with humans on top. Expect a proliferation of non-human, task-specialized AIs, each powerful in its own domain but unlikely to resemble us or become omnipotent gods.

  • AI agents and embodied systems are the next frontier. Agentic AIs can now take actions, not just give answers, and embodied AIs may build understanding from real-world interaction — but these are still specialized systems. Instead of fixating on “AGI,” we should focus on what AI can actually do and how it behaves in practice, not how human it seems.

AI Trailblazer Takeaways: AGI is the most talked about technology today, yet most computer scientists are claiming we’re nowhere near it. But many questions abound, will AGI be smarter than a human? Will AGI replace human thinking? Do we even need AGI? The answers, like the technology, remain to be answered in the future.

Meet the Swiss Army Knife of AI Agents

AI agents are evolving fast, but they’re still not ready to run the show on their own. WIRED investigates a new model from Simular, called S2, is pushing the boundaries by switching between different AI systems depending on the task—offering a glimpse into the future of more capable, adaptable agents. This article explores how multi-model design, memory, and even human collaboration could be the keys to unlocking truly useful AI assistants.

  • Simular’s AI agent S2 introduces a multi-model approach. It combines general-purpose models like GPT-4o with smaller, specialized models to handle specific tasks like navigating GUIs. This setup allows S2 to outperform other agents in complex computer and smartphone tasks.

  • S2 is designed to learn and improve over time. It uses an external memory to track actions and feedback, refining its performance on repeated tasks. This makes it better equipped for real-world application use, although it's still far from perfect.

  • Benchmarks show progress, but also limitations. On OSWorld, S2 can complete 34.5% of 50-step tasks, surpassing OpenAI’s Operator, but still well below human performance. Even with advances, agents still get tripped up by edge cases and can behave erratically.

  • Human-AI collaboration may be the key forward. Tools like CowPilot show that pairing humans with agents yields better results than either alone—completing 95% of tasks with just 15% human effort. As agents evolve, the most practical systems may be hybrids that balance automation with human intuition.

AI Trailblazer Takeaways: Pairing humans with AI agents appear to be the most likely path forward. Even more reason for everyone to educate themselves on AI….their future may depend on it.

AI’s Biggest Security Hole May Finally Be Plugged

Prompt injection has been one of the most persistent and dangerous vulnerabilities in AI systems, allowing attackers to manipulate language models with hidden commands. ArsTechnica reports that Google DeepMind may have finally found a viable solution with a new architecture called CaMeL. By rethinking how AI systems process and act on data, CaMeL offers a fresh, security-first approach that could redefine how we build safe and reliable AI agents.

  • Prompt injection has long been AI’s most stubborn security flaw. It allows malicious actors to sneak hidden instructions into user content, leading AI models to perform unintended or harmful actions. Despite years of attempts, no consistent fix has been found—until now.

  • Google DeepMind’s CaMeL introduces a radically different approach. Instead of relying on a single AI to self-regulate, it splits tasks between a “privileged” LLM that plans and a “quarantined” LLM that parses data. This architecture, inspired by proven software security principles, enforces strict boundaries and prevents compromised data from triggering harmful actions.

  • CaMeL translates AI prompts into secure Python code. As the code runs, it tracks data origins to ensure no untrusted content can silently influence critical actions like sending emails. This method mirrors long-established security practices, such as the principle of least privilege, giving AI systems fine-grained control over what they’re allowed to do.

  • While promising, CaMeL isn’t a silver bullet—yet. It still requires users to manage security policies, and poorly designed interfaces could lead to “yes fatigue,” where users ignore safety prompts. Still, experts believe this model is a significant step toward making AI agents safer, more trustworthy, and genuinely useful in everyday applications.

AI Trailblazer Takeaways: AI is getting better all the time and it’s reliability as well. Security and accuracy are at the heart of concerns regarding the adoption of any new technology. While AI may seem like the wild west today, loopholes such as prompt injection are being mitigated every day.

Quote of the Week

“AI is in the hands of a small group of reckless, careless men who seem to have no understanding of society.” 

- Carole Cadwalladr, British investigative journalist during her TED Talk on April 20, 2025

AI Trailblazers Upcoming Forums Calendar

  • June 5th, 2025 - AI Trailblazers Summer Summit, New York, NY

  • December 3rd, 2025 - AI Trailblazers Winter Summit, New York, NY

Get in touch for speaking and sponsorship opportunities at these upcoming forums.

Partner with AI Trailblazers

Our community of business leaders is growing fast and we are planning more forums. To discuss partnership opportunities that connect our partners with AI decision makers including at our summits and this newsletter, email us.

What is AI Trailblazers?

AI Trailblazers is a vibrant platform dedicated to uniting corporate marketers, technologists, entrepreneurs, and venture capitalists at the forefront of artificial intelligence (AI). Our mission is to fuel growth, innovation, and career development among our members, who all want to be at the forefront of incorporating artificial intelligence into their businesses and their lives for strategic advantage. More information here.