Introduction There’s a lot of buzz surrounding OpenAI’s Strawberry. Some claim it could be the...
Counter Intuitive 2025 AI Predictions -- AGI, Evals, Agents, Enterprise, GPT-5/Orion
Introduction
The world of AI is buzzing with anticipation for 2025. However, amidst the excitement and hype, it’s important to separate what’s realistic from what’s overly optimistic. In this blog post, I’ll break down some lesser-discussed predictions for the future of AI, including AGI, GPT-5/Orion, AI agents, and enterprise adoption.
AGI in 2025: Why Claims May Fall Short
The talk around Artificial General Intelligence (AGI) is everywhere, but we’re still a long way from achieving it. My prediction for 2025 is that there will be bold claims of AGI, but they’ll likely be misleading. The definition of AGI is vague, and some companies may “move the goalposts” to claim success.
For instance, if an AI system performs exceptionally well on an evaluation metric like the EV Val (Evaluation Metrics for AGI), it’s still not AGI. Passing an evaluation like this doesn’t necessarily indicate that the AI has reached true human-like general intelligence.
The missing components for AGI are still significant, and current models—such as large language models (LLMs)—cannot meet the requirements for AGI. So, when you hear claims of AGI, take them with a grain of salt. The real breakthrough is still years away.
GPT-5/Orion: A Step Forward, But Not a Game-Changer
The next iteration of GPT (GPT-5 or Orion) will be more than just an LLM—it will integrate a "chain of thought" process, expanding its functionality beyond simple text input and output. However, this will not be the massive leap forward that some might expect. Despite all the excitement, GPT-5 will likely only be a marginal improvement over GPT-4 in practical use cases.
One key issue is the EV Val scores. While GPT-5 will likely achieve better results than GPT-4 on these evaluations, EV Val metrics are limited in scope. These evaluations are important, but they don’t capture the full picture of AI capabilities. A higher score doesn’t necessarily equate to true AI progress.
EV Val Metrics: Their Limitations and Why They Aren’t the Full Story
As AI models improve, we’ll continue to see marginal increases in EV Val scores. However, these metrics, while useful, are narrow in their application. A model that scores 90% on an EV Val might not be significantly more capable than one that scores 70%, especially when comparing multiple models that are within a fraction of a percentage point of each other.
The problem with celebrating these incremental improvements is that they don’t always indicate real-world advancements. It’s like judging the progress of an elephant by taking a picture of just its foot. Sure, it’s part of the elephant, but it doesn’t give you the whole picture. So, while we’ll see plenty of media coverage about top models, don’t be fooled into thinking these are major breakthroughs.
AI Agents: Why They Still Aren’t Ready for the Masses
AI agents have been generating a lot of buzz, but there's a fundamental misunderstanding about how they work. While there will be progress in AI agent development, it’s unlikely that they will live up to the hype in 2025. Many AI agents will still be limited in their capabilities and will struggle to meet the expectations of everyday users.
One common misconception is that AI agents will be accessible to everyone, even those without coding knowledge. While this sounds appealing, it’s akin to saying, “Anyone can build a car, even without understanding engineering or physics.” The reality is, building reliable AI agents requires professional knowledge in software development and a deep understanding of how to integrate LLMs into larger software systems. Expect AI agents to remain a niche product for professionals who know how to build them properly.
Enterprise AI Adoption: Slow But Steady Progress
AI adoption in enterprises will continue to move at a slower pace than many anticipate. Despite all the advancements, it may take 5 to 10 years for widespread AI adoption in the enterprise space. The reason for this is that most AI solutions today are LLM-centric, which leads to reliability issues, particularly in fields like healthcare and finance where dependability is crucial.
The solution lies in using LLMs to address specific, narrow challenges within larger software systems rather than relying on them to handle all aspects of a complex solution. This approach is more reliable and sustainable in the long term, but it requires a shift in mindset from LLM-focused solutions to more software-centric ones.
The Road Ahead: Balancing Innovation with Realism
While AI continues to evolve, it’s essential to cautiously approach bold claims. From AGI hype to GPT-5's incremental improvements, 2025 will showcase meaningful advances but not the seismic shifts some may expect. The true potential lies in practical applications that enhance real-world systems rather than chasing the mirage of AGI or overestimating AI agents' capabilities.
Ready to Leverage AI for Your Business?
At 42robotsAI, we specialize in building custom AI solutions that focus on real-world results, not just trends. Whether you’re exploring AI adoption or looking to refine your operations, we’re here to guide you every step of the way.
Contact us today to discover how we can help you integrate AI effectively and drive measurable success.
