Alternative AI Paradigms and ARC Benchmark: Path to AGI by 2030

Francois Cholet founded NDIA as a new AGI research lab with the explicit goal of creating a branch of machine learning that is much closer to optimal than current deep‑learning approaches. The lab’s central strategy is program synthesis, which works at a far lower level than the code‑generation agents that dominate today’s AI landscape.

Program Synthesis and Symbolic Descent

Program synthesis at NDIA is not merely another form of code generation. It seeks to rebuild the entire AI stack on different foundations by replacing deep‑learning’s parametric curves with small, symbolic models. The optimization technique called “symbolic descent” searches for the simplest symbolic representation that explains the data, rather than fitting parameters through gradient descent. The expected benefits are reduced data requirements, more efficient inference, and stronger generalization across tasks.

Critique of the Current AI Industry

The AI industry has poured billions into large language model (LLM) stacks, producing impressive results but also creating a single dominant path. While this path may eventually reach AGI, the speaker argues it will do so inefficiently. The vision is to leapfrog directly to optimal, symbolic methods that can accelerate progress without the massive resource consumption of existing LLM pipelines.

Verifiable Rewards and the Rise of Coding Agents

Coding agents have advanced rapidly because code provides a formally verifiable reward signal. When a program passes unit tests, the system receives a clear, trustworthy indication of success. Mathematics enjoys a similar natural verifiability, positioning both domains for swift breakthroughs. In contrast, areas such as essay writing rely on fuzzy human annotations, making progress slower. Environments that embed code‑based training with unit‑test rewards enable models to learn execution traces and improve with far less human supervision—costs as low as $0.3 cents per ARC task compared with $1‑$10 for typical LLMs.

Defining AGI Beyond Automation

AGI is framed not merely as the automation of economically valuable tasks but as a system that can approach any new problem, model it, and become competent with human‑level efficiency in data and compute. This definition emphasizes skill‑acquisition efficiency rather than sheer knowledge accumulation.

LLMs, Sample Efficiency, and the Push for Optimality

Although it is possible to build AGI on top of the existing LLM stack by adding new layers, such an approach is considered inefficient. The broader research trend is shifting toward methods that achieve higher sample efficiency and move closer to theoretical optimality.

Evolution of the ARC AGI Benchmark

The ARC benchmark was created to provide a reasoning‑focused analogue of ImageNet.
- ARC V1 highlighted the difficulty of gradient‑based reasoning, with base LLMs scoring below 10 % and GPT‑3 scoring zero.
- ARC V2 marked the emergence of agentic coding and verifiable‑reward paradigms, eventually being saturated (97 % performance) by Confluence Labs.
- ARC V3 measures “agentic intelligence,” requiring an agent to explore an interactive environment, set goals, plan, and execute actions with human‑level efficiency. This version resists the “harness” strategy used to saturate V2.

Future versions aim at continual and curriculum learning (ARC 4) and invention (ARC 5), keeping the benchmark a moving target that reveals residual capability gaps.

NDIA’s Foundational Vision

NDIA’s symbolic learning vision replaces parameter curves with the shortest possible symbolic models. Deep learning is used to guide program search, helping to break through combinatorial barriers. The lab seeks to build a compounding stack where each layer builds on reusable foundations, ultimately removing humans from the improvement loop. Science is described as symbolic compression, and NDIA aspires to recreate the scientific method algorithmically, turning messy human learning into a more efficient, first‑principles process.

Timeline Toward AGI

Extrapolating from current progress and investment, the speaker predicts that AGI could emerge around 2030, potentially coinciding with later ARC versions (V6 or V7). The “AGI moment” is defined as the point when measurable differences between human and AI capabilities disappear.

Alternative Approaches and Startup Opportunities

There remains ample space for startups to explore alternatives to LLMs. Scaling genetic algorithms, modifying existing layers (e.g., state‑space or recurrent models), or returning to foundational principles are all viable paths. Aspiring researchers are encouraged to revisit older, less‑invested papers from the 1970s and 1980s. Successful new approaches must be able to scale without human bottlenecks, enabling recursive self‑improvement.

Guidance for Open‑Source Projects

Open‑source initiatives should prioritize simple, intuitive APIs and overall usability, taking inspiration from libraries like scikit‑learn. Documentation should be informative enough to teach the domain, and community building—along with hiring enthusiastic users—should be a core focus.

Mindset for the Future

AI progress is seen as inevitable and empowering. Expertise combined with AI tools can turn developments into opportunities. The key question is not how to stop AI, but how to leverage its accelerating wave for personal and societal benefit.