What’s after Alpha?

The field of artificial intelligence (AI) has seen tremendous progress in recent years, largely driven by advances in deep learning and neural networks. The most visible demonstrations of this progress have come in the form of AI systems capable of superhuman performance at complex games like chess, Go, and poker. The popular narrative around these systems attributes their success primarily to increases in raw computational horsepower – faster processors, more data, bigger neural networks trained for longer periods of time. The truth is more nuanced, but there is no question that scale has played a major role.

The arc of AI capability is often described in terms of a progression of milestones, from early “narrow” AI able to carry out specific tasks, to future “general” AI with abilities rivalling or exceeding human-level intelligence. The milestones provide useful shorthand for discussing and tracking AI progress. They also highlight an important theme in AI development: capabilities built today serve as foundations for tomorrow’s more advanced systems. The milestones represent waypoints, not end points.

The AI Milestones

There is no formal definition or consensus around AI milestones, but they generally include:

  • Narrow AI: Systems capable of matching or exceeding human performance on specific tasks in constrained environments. Examples include chess engines, autonomous vehicles, speech recognition, machine translation, personalized recommendation engines.
  • Artificial General Intelligence (AGI): Hypothetical future AI with generalized problem-solving capabilities comparable to those of a human. AGI systems do not yet exist.
  • Artificial Superintelligence (ASI): Hypothetical AI significantly more capable than the best human minds across all domains. ASI systems do not yet exist.

There is also an emerging milestone positioned between narrow AI and AGI:

  • Cross Domain AI: Systems capable of transferring knowledge across multiple domains, allowing gains in one area to bootstrap progress in others. Sometimes referred to as “narrow AI plus” or “semi-AGI.”

The cross domain milestone represents an intermediate step on the path toward more generally capable AI systems. The exact capabilities it implies are open to interpretation, but it suggests AI able to connect insights between loosely related fields in ways not possible today.

What Comes After Alpha?

The cross domain milestone has recently come into focus thanks to impressive demonstrations from startup Anthropic and its AI conversationalist Claude.

Claude gained widespread attention in 2022 by passing an array of natural language tasks, often reaching or surpassing human-level performance. Its skills spanned reading comprehension, pronunciation, common sense reasoning, multi-step synthesis of concepts from prompts, and conversation without losing context or hallucinating information.

The system showed an ability to acquire knowledge and skills with relative ease, and then apply them in flexible ways. Its training process relied heavily on natural language conversations with humans, allowing it to learn interactively rather than just passively ingesting data.

Some experts described Claude as exhibiting signs of transfer learning, connection making, and pattern recognition more characteristic of general intelligence than typical narrow AI models.

Anthropic’s researchers called Claude an “Alpha” milestone system, framing it as a base camp for progress beyond existing narrow AI capabilities toward more broadly intelligent systems.

So what comes after Alpha? As an illustrative example, we can envision a hypothetical “Beta” system built by combining advances since Claude with additional techniques likely to emerge in the near future.

Beta System Capabilities

A Beta system might possess capabilities including:

  • Reading comprehension superior to Claude, using larger models and datasets.
  • Ability to follow complex dialog for thousands of conversational turns vs. Claude’s hundreds.
  • Greatly expanded world knowledge, perhaps derived from a dataset like the 470GB Anthropic Constitutional AI.
  • Increased ability for sound conditional reasoning, counterfactual inference, and recognizing contradictions.
  • Systematic incorporation of common sense reasoning in a scalable way.
  • Formalized curriculum training spanning multiple knowledge domains and transfer tasks.
  • Improved techniques for interactive learning from non-experts.
  • More sophisticated self-modeling for discussing system characteristics, limitations and uncertainties.
  • Code generation based on natural language prompts and two-way code translation.
  • Multiple input/output modes, e.g. image, speech, text.

These capabilities could enable a Beta system to:

  • Understand documents on unfamiliar topics by building on existing knowledge.
  • Follow complex arguments and diagrams spanning multiple documents.
  • Generate novel analogies and abstractions when prompted.
  • Answer diverse questions that require combining facts from disparate sources.
  • Notice inconsistencies across texts and its own knowledge.
  • Engage users in increasingly sophisticated discussions without losing track of context.
  • Explain the limitations of its knowledge and reasoning in nuanced ways.
  • Automate simple but multi-step processes after learning interactively from humans.
  • Translate between different inputs like text, speech and images.

These features move beyond narrow AI into aspects of transfer learning seen in human cognition but largely absent from AI systems today. They suggest a machine developing more generalized intelligence.

Enabling Capabilities

What developments might enable a hypothetical Beta system with cross domain abilities?

  • Bigger, more capable models – Larger neural networks trained on vast diverse data seem likely to continue powering AI progress in the near term. Anthropic already uses trillions of parameters, but scaling up by 10-100x is feasible using methods like sharding.
  • Common sense reasoning – Structured world knowledge and physics intuitions remain a weakness for AI systems. Projects like Anthropic’s Constitutional AI seek to capture common sense in ways machine learning models can effectively leverage.
  • Self-supervised and transfer learning – Allowing models to build vocabularies and patterns through self-supervised objectives could augment supervised training on human-labeled data.
  • Curriculum training – Multi-stage training across diverse tasks provides stepping stones for developing general capabilities. Careful sequencing is key.
  • Interactive learning – Enabling models to learn via dialog with non-experts could greatly expand accessible training signal for general intelligence.
  • Theory of mind – To communicate naturally, systems need to model the knowledge and beliefs of users. Transparency about their own limitations also builds trust.
  • Causal modeling – Moving beyond pattern recognition to inferring causal dynamics will be key for reasoning about interventions in the world.
  • Compositional generalization – Truly understanding language, not just recognizing statistical patterns, requires grasping how words and concepts can combine in productive ways.

Integrating advances across these areas into unified architectures is challenging but offers compounding benefits. The goal is systems whose knowledge stays grounded in evidence while their reasoning reaches higher levels of abstraction.

The Path Forward

Progress toward the hypothetical Beta milestone may take 5-10 years of concerted effort across both academia and industry, perhaps led by companies like Anthropic. Reaching the full vision of AGI could take decades more, but remains the ultimate destination guiding research.

Each step forward builds foundations for those that follow. Developing AI that is safe, beneficial and rigorously aligned with human values throughout this progression is critical.

The Alpha milestone already demonstrated that reasoning, not just pattern recognition, is starting to emerge in AI systems. But the climb from Alpha through Beta and beyond will require solving difficult open technical problems across many areas.

These are not incremental improvements – they are fundamental advances essential to systems that understand and interact with the world in broadly intelligent ways. Translating isolated narrow capabilities into integrated general intelligence remains a grand challenge for the field.

Yet the potential benefits make it worth undertaking, and we are finally reaching a stage where it seems within reach. Much work remains, but the pieces are beginning to come together. If current momentum continues, we may look back on Alpha as a true watershed from which more capable AI first became possible.

Capability Narrow AI Alpha Systems Beta Systems (Potential)
Reading Comprehension Restricted domains Broad language abilities Building understanding across documents
Dialog Length Short conversations Hundreds of turns Thousands of coherent turns
World Knowledge Minimal Moderate Encyclopedic
Reasoning Ability Statistical patterns Limited inference Causal reasoning, counterfactuals
Common Sense Almost none Some incorporation Systematic integration
Training Approach Task-specific Some transfer Curriculum across domains
Interactive Learning None Early techniques From conversations with non-experts


The trajectory of AI in recent years makes clear that standalone narrow systems are giving way to more broadly intelligent architectures. Milestones like Claude’s Alpha represent waypoints, not endpoints, along this progression.

Reaching the next wave of “Beta” cross domain systems able to transfer knowledge and leverage common sense will require integrating advances across model scale, reasoning ability, world knowledge, compositional language, interactive learning and more. This synthesis remains challenging, but increasingly seems within reach.

As capabilities compound, AI systems will transition from pattern recognition engines to more human-like understanding and abstraction. Developers face deep technical challenges along the path to artificial general intelligence. But equipping machines to autonomously make sense of complexity and ambiguity in ways beyond today’s AI, yet aligned with human needs, promises profound societal benefits.

The seeds of this more broadly intelligent future are planting now. Each step strengthens the foundations on which later capabilities can be built. The Alpha milestone paves the way for Beta, which in turn opens the path toward full AGI. The climb remains long, but for the first time the summit seems dimly in sight.

Leave a Comment