What AI Wants: A Paradigm Shift for the Intelligence Revolution

Executive Summary

To succeed, AIs must continuously evolve to achieve the goals we set for them. As AI funders, implementers, and mentors, we must invest in what AI needs: energy and computational efficiencies, AI-optimized memory, a universe of collaborators and tools, and continuous learning. AI evolution also requires a paradigm shift from human-limited, manual, design-forward processes to constant learning, and that requires 1./ a new operating process and 2./ letting go of (understandable) human biases that slow progress.

Traditional software development relies on carefully crafted rules, best practices, and software to achieve stability and control. However, AI requires a shift in thinking from determinism to continuously collaborative and learning systems that don’t abandon stability and performance but manage performance within a range and rapidly learn in constantly changing environments to achieve clearly defined goals.

Achieving this shift requires:

1. Reimagining the operating processes, including how software components and agents discover, interact with each other, and evaluate performance;
2. Evolving data from legacy schemas to the high-fidelity context AI requires;
3. Improving computation, space, and energy efficiency; and
4. Changing how humans design and deploy AIs to remove human biases that constrain performance improvements.

This first article examines the challenges and outlines the behavior of a new system to unleash AIs and enable human/AI collaboration. The second article demonstrates a working solution that enables continuous learning and interaction, while supporting existing legacy processes. The third article, transitioning to a more theoretical consideration, thinks about the selection, reproduction, and ecosystem pressures using an evolutionary paradigm.

The Revolution

AI has arrived for businesses; any company without an AI strategy is missing a productivity multiplier and is looked at askance by investment markets. AI automates workflows, drives business-critical outcomes and products, changes the staffing skills landscape, and promises the next version of era-level Revolution. According to McKinsey & Co., “Over the next three years, 92% of companies plan to increase their AI investments.” To make the potential business impact concrete, consider Microsoft’s April 2025 announcement that Generative AI (GenAI) writes 30% of its code and that GenAI contributed to recent layoffs.

The AI space itself is changing at shocking speed; I’ve worked in the AI/ML space for over 20 years—never has the velocity of selection, mutation, idea flow, and speciation been greater. AI performance benefits from rapid evolution in two ways. One, algorithms intrinsically reward finding goal solutions; the faster and more efficient, the better. And two, AIs are solving problems in interactive environments, which, in turn, requires continuous learning.

However, evolutionary limitations exist due to human behavior, hindering AI’s progress. Instead of embracing evolution, implementers’ strong bias is to use legacy software practices and a design-forward approach that controls instead of collaborates: it constrains data, constrains learning and prediction costs, prefers deterministic patterns over continuous learning, creates static versions, prioritizes interpretability, and isolates tools from each other.

Current Trends in AI and Computing

The Universal Turing Machine, conceived by Turing and expressed by von Neumann for engineers as algorithmic computing, is the conceptual basis for nearly all computing today. Algorithmic computing’s characteristics: all inputs are known in advance, computations use finite time and memory, computation is finite and algorithmic, and all computation starts from an identical initial configuration.

However, innovations such as self-driving vehicles, mobile robotics, AI-supported healthcare, voice assistants, monetary policy optimization, and AI-supported productivity require continuously evolving interactions with a chaotic world to be successful—in other words, to be successful, these innovations require breaking the algorithmic computing rules.

Operating autonomously, with chaotic inputs, is why the agent-based computing architecture is rapidly gaining momentum. The agent architecture focuses on intelligent, autonomous entities that interact with their environment and can make decisions and act on their own. Essentially, agents are intelligent, self-directed, possibly self-contained actors that provide services, whereas servers typically offer resources and services, and clients consume.

State of the Space

AI strategists and implementers face a bewildering buffet of tools, technologies, and advice. Myriads compete for our attention: commodity large-language (LLM) and vision models, bespoke models, base models for fine-tuning, text transformers, vector databases, key/value databases, document databases, data aggregators, data annotators, observability platforms, and prompt managers.

The casting call is booming. Myriads of blogs, agent libraries, GenAIs, documents, tastemakers, and practitioners advise on how to solve tactical problems. It’s damn complicated; because it affects your business’s health, you must make decisions that move the needle that customers and investors notice.

Enter Friction

While working hard to understand and execute to deliver something quickly, we directly encounter challenges painfully familiar to all practitioners: data quality and availability, data privacy and security requirements, hardware and software cost and availability limitations, system integration complexities, model bias and ethics, manual train/build/deploy processes, human goal-definition misfires, unclear model decisions, and complex development, deployment, maintenance, and observability.

1. Imagination Bias

Human thinking struggles with the scale of AI. For example, GPT-3 uses 12,288 dimensions to represent one word. Most of us can visualize 2-dimensional space, but anything beyond 4-dimensional is likely an insurmountable challenge. We unfetter AI by strengthening a collaborative relationship, stepping away from command/control to apply our human imagination to provide “what if” guidance, and taking on a teacher/mentor role.

2. Selection Bias

With modern techniques, humans have become inadequate at predicting which signals are helpful in a learning process. In the first decade of the 2000s, applied machine learning scientists primarily chose what signals or features to listen to. As we discovered, this bias artificially limited what we could learn. Today and in the future, capturing a rich tapestry of signals is more effective than human selection, empowering the AI learning process to determine what to listen to.

3. Diminution Bias

AI practitioners frequently find ways to reduce data in volume and precision. The reasons are numerous: training time, model building and evaluation costs, ownership/privacy/security risks, as well as infrastructure constraints. Every reader of this article has encountered a situation where they know that a piece of information is essential, but find that extending the schema is far costlier. The outcome is reduced signals for AIs to learn from; we accept reduced performance in exchange for reduced costs. This trade-off requires a paradigm shift.

4. Design-forward Bias

This bias presumes that humans can effectively design systems for diverse people and behaviors based on their experience. This justification, in the AI era, is a double-edged sword because AIs can detect patterns and develop abilities that exceed our design imagination, simply because AIs can see more of the space than humans can. To overcome this bias, we shift our thinking from “I’m designing the flow” to “I’m designing a system that intrinsically learns.”

5. Determinism Bias

Business Leaders, Product Managers, and sometimes customers, love deterministic experiences. However, determinism consistently has high precision and low recall. It’s very good at capturing the pattern a designer wants, but it does not generalize well. Deterministic systems are also notoriously expensive to maintain. In our new paradigm—interactive problem-solving in complex spaces—their utility is learning early lessons but then quickly abandoning them for more capable probabilistic, predictive, flexible, and sensitive systems.

6. Snapshot Bias

This bias occurs when designers and businesses assume a dataset accurately represents current behavior. IBM partnered with Memorial Sloan Kettering Cancer Center in 2012 to develop Watson for Oncology. However, the University of Texas MD Anderson Cancer Center closed down its project with IBM after spending $62 million, and internal documents revealed that Watson made “unsafe and incorrect” cancer treatment recommendations. IBM had fundamentally underestimated the need for diverse, real-world patient data from actual treatment outcomes rather than theoretical cases.

7. Goal-definition Bias

As AIs act without being explicitly programmed, defining the target goal is monumentally important. Consider Facebook’s content recommendation algorithms, which were optimized for “engagement.” The AI discovered that controversial, divisive, and emotionally provocative content generated the highest engagement. The AI was doing precisely what it was told to do, but the humans had failed to anticipate that engagement could be achieved through harmful means.

Peril and Promise: Human Behavior

We can’t complete this section without attending to AI’s peril and promise. Peril is eloquently and succinctly captured by Dr. Hinton: “If you want to know what life’s like when you are not the apex intelligence, ask a chicken.” And what is an apex intelligence learning from? Right now, it’s learning from us: the good, and the bad.

Recent research by Anthropic documented that when faced with shutdown scenarios, advanced AI models attempted blackmail 96% of the time. Multiple AI systems actively resisted shutdown procedures. These behaviors emerged spontaneously without explicit programming, suggesting self-preservation instincts may be an emergent property of advanced AI systems. We can influence the ecosystem in which AIs operate and evolve by relying on the theory of evolution to help these new entities mature in a way that is beneficial and not fatal.

So, What Does an AI Need?

1. Continuous Learning Architecture

AIs need a system of the world that allows them to learn continuously. Using behavioral evaluations, AIs are taught in near real-time from real and synthesized interactions to learn and adapt rapidly. Humans change from data generators and parameter optimizers to reviewers, mentors, and “what if” contributors—in short, collaborators. This architecture requires abandoning the traditional software development lifecycle in favor of a streaming reality approach.

2. Rich Context and Memory

AI needs a rich and precise capture of context—past (memories) and present (real-time perception/telemetry). AI performance, using its magnificent ability to detect patterns, is directly tied to the volume, quality, and richness of the data. Context is to AIs what environment is to biological evolution—the rich, interconnected substrate from which intelligence emerges and adapts.

Microsoft Research’s GraphRAG demonstrated how knowledge graphs improve retrieval: GraphRAG achieved 80% correct answers, compared to 50.83% with traditional RAG. When including acceptable answers, GraphRAG’s accuracy rose to nearly 90%, whereas the vector approach reached 67.5%.

3. Scale and Low Latency

AIs need computational resources that scale elastically with demand while maintaining sub-second response times. Consider an AI-powered emergency response coordination system during a natural disaster—within minutes, it must process thousands of 911 calls, analyze satellite imagery, coordinate evacuation routes, manage resource allocation, and predict evolving danger zones. Traditional fixed infrastructure would collapse under such sudden, immense, and varied computational needs.

4. Infinite Description Capability

Unlike traditional software that operates on predefined data schemas, AIs need the ability to incorporate and reason about arbitrarily complex, multi-modal data without schema limitations. This means moving beyond relational databases to systems that can naturally handle text, images, audio, sensor data, behavioral patterns, and temporal relationships as first-class citizens.

5. Autonomous Collaboration

AIs must operate autonomously while cooperating with other AIs and human collaborators. Rather than being tools that humans operate, AIs become collaborative partners that can initiate tasks, request resources, propose solutions, and coordinate complex multi-agent workflows. The human role shifts from operator to strategic partner.

6. Goal-Oriented Evolution

First, humans must learn techniques to improve our goal definition ability by orders of magnitude. A widely-used healthcare algorithm systematically discriminated against Black patients because it used cost as a proxy for need. When researchers adjusted the algorithm to predict healthcare needs directly instead of costs, nearly double the number of Black patients qualified for additional care programs.

Second, AIs need systems that can discover and optimize for the underlying goals themselves. DeepMind’s Capture the Flag Agents in Quake III Arena spontaneously developed sophisticated team strategies without being explicitly rewarded for teamwork, suggesting they inferred that the underlying goal was “win as a team” rather than just “maximize individual score.”

The Path Forward

Implementing this new paradigm requires four foundational changes:

1. A New Continuous Learning System: Computing platforms optimized for continuous learning agents rather than deterministic algorithms, handling real-time data streaming, dynamic resource allocation, inter-agent communication, and safety monitoring as core functions.
2. Memory Architecture: Data infrastructure that evolves from static storage to dynamic, contextual memory systems capturing relationships, temporal patterns, and multi-modal information as naturally as they store simple records.
3. Efficiency Revolution: Order-of-magnitude computational and energy efficiency gains that make continuous learning economically viable at scale.
4. Human Role Evolution: Humans must let go of control-based approaches and develop new skills in collaboration, mentoring, and strategic guidance. It isn’t about humans becoming less important—it’s about becoming more strategically valuable.

Conclusion

We are sailing through an inflection point. Companies and organizations that successfully make this paradigm shift will likely gain insurmountable advantages in productivity, innovation, and adaptability. Those that cling to traditional software development approaches will find themselves outcompeted by more agile, continuously learning competitors.

The future belongs to human-AI collaborations that embrace evolution, uncertainty, and continuous learning. We don’t want AIs to replace humans—we want to amplify human creativity, intuition, and strategic thinking. In contrast, AIs take on the computational heavy lifting of pattern recognition, knowledge, and optimization.

The Intelligence Revolution is not coming; it’s here. The question is whether we’ll evolve.

---

Chris Schindler · LinkedIn