AI Will Ship a AAA Game AUTONOMOUSLY by 2030!

When Will AI Deliver AAA Games?

Predictions for AI in Game Development

• The speaker discusses the timeline for AI to autonomously deliver a AAA game with minimal human interaction, predicting this capability by around 2030.

• By 2030, advancements in AI (potentially named differently) are expected to allow users to create polished games from a single prompt.

• In 2026, the expectation is that AI will be able to set up a Shopify storefront autonomously, requiring only five minutes of human input for the initial prompt.

• By 2027, it is anticipated that AI will write an entire technical book independently, including educational content and slides.

• The speaker forecasts that by 2028, AI could develop feature-rich mobile banking applications.

Future Developments in Gaming

• In 2029, the prediction includes the creation of an entire indie game solely from prompts; by 2030, this could extend to double A or AAA open-world titles like "No Man's Sky" or "Cyberpunk 2077."

Understanding Acceleration in AI Progress

The Concept of Jerk in Technology Growth

• The speaker references Jensen Huang's comment on Moore's Law squared and explains how current growth resembles Moore's Law cubed due to accelerating advancements.

• Definitions are provided: position (location), velocity (rate of change of position), acceleration (rate of change of velocity), and jerk (change in acceleration).

• An example illustrates how jerk occurs when acceleration itself changes over time—highlighting why predictions about AI have often been conservative.

Implications of Accelerated Growth

• The discussion emphasizes that multiple factors contribute to accelerated progress beyond just Moore’s Law; thus leading to more optimistic projections about future capabilities.

• While there may be potential for further breakthroughs ("snap"), such as photonic or quantum computing, these are not currently seen as likely developments.

Data Insights on Autonomous Tasks

Trends in Task Automation

• A graph based on MER data shows exponential growth in the duration tasks can be performed autonomously by AI without human intervention.

AI Acceleration and Computational Substrates

The Nature of AI Progress

• The advancement in AI is not slowing down; it continues to grow exponentially, driven by the fundamental limits of physics and thermodynamics.

• Training FLOPS (floating-point operations per second) are increasing due to financial investment rather than a fundamental shift in technology.

• Historical context shows that advancements like Moore's Law have been pivotal, but future changes may involve new computational substrates beyond silicon.

Financial Influence on AI Development

• Current trajectories suggest that unless there’s a significant change in computation methods, we will continue along existing trends for decades.

• The cost-effectiveness of calculations per dollar spent is crucial; more money leads to more intelligence generated per dollar invested.

• Scaling AI requires algorithmic improvements alongside financial input, highlighting the importance of funding as a primary constraint.

Benchmarking AI Performance

• Recent benchmarks indicate that many AI systems are reaching or surpassing human-level performance across various tasks such as image recognition and code generation.

• There is observable acceleration in progress rates, with some tasks achieving saturation much faster than before—e.g., reading comprehension surpassed human levels within two years after 2016.

Autonomy and Success Rates

• Data indicates that AI systems are becoming more autonomous approximately every four months since 2024, showing real-time evidence of acceleration in capabilities.

• A median success rate of 50% suggests variability; while AIs can perform tasks faster than humans, they do not always achieve full autonomy consistently.

Future Projections for Generative AI Adoption

• Predictions show that Fortune 500 companies' adoption of generative AI is currently at 70%, expected to reach about 90% by 2029 due to gradual acceptance despite initial hesitance towards new technologies.

The Future of AI: A Tipping Point by 2028

The Evolution of AI and Ethernet

• By 2028, generative AI is expected to become as compelling as Ethernet was for local area networking, marking a significant tipping point in AI capabilities and business deployment.

Historical Data on Deep Neural Networks

• Analysis of historical data indicates that the "jerk coefficient" (the rate of acceleration in AI development) has been declining since the inception of deep neural networks, suggesting a slowdown in rapid advancements.

Understanding Jerk Coefficient

• The jerk coefficient reflects that while acceleration is still occurring, the rate at which it accelerates is decreasing. This implies a transition from super-exponential growth to ordinary exponential growth around 2026.

Implications of Declining Jerk

• The decline in jerk suggests that the era of rapid AI acceleration may be nearing its end within one or two years unless new breakthroughs occur. However, this does not imply a halt in value proposition growth.

Cost and Utility Dynamics

• The cost per token for AI models has decreased significantly while their utility has increased exponentially. For instance, GPT-3 tokens were initially less intelligent compared to newer models like ChatGPT.

Token Cost Trends

• Early tokens such as those from GPT-2 and GPT-3 had limited utility; however, advancements have led to current models being significantly more capable with lower costs per token.

Reasoning Models Impact

• Introduction of reasoning models may temporarily dilute utility per token due to higher token usage but is expected to integrate seamlessly into future model architectures.

Projected Capabilities of Future AI Models

• Current projections indicate substantial improvements in AI capabilities over time based on benchmarks and task complexity. This includes an analysis comparing earlier models like GPT-2 and GPT-3 regarding their utility per dollar spent.

Transition from GPT-2 to GPT-3

The Future of AI Utility and Predictions

Evolution of AI Utility Over Time

• The utility of AI has significantly increased over the past five years, with advancements leading to million-token windows. ChatGPT demonstrated a substantial leap in usefulness compared to GPT-3, marking an order of magnitude increase.

• By 2025, integrations like RAG tools are expected to enhance utility per dollar spent on AI, representing one of the most cost-effective ways to boost performance.

Scenarios for AI Development by 2035

• Five scenarios were outlined regarding future AI capabilities:

• Worst Case: If all technological progress halts (less than 0.1% chance), AI will only be 30 times better than today by 2035.

• Bad Case: Even under pessimistic assumptions, AI could still achieve a power increase of up to 300 times current capabilities by 2035.

• Middle Road Scenario: Predicting a growth factor around 3,000 times more powerful than today’s technology within the next decade.

Optimistic Projections for Future AI

• In an optimistic scenario, advancements could lead to improvements exceeding tenfold increases in capability every decade. This would mean that typical AIs could be as much better than GPT-3 as ChatGPT was two years ago.

• The ultra optimistic projection suggests potential growth reaching up to five million times more powerful than current models if trends continue positively.

Key Developments Expected in Upcoming Years

Mainstream Adoption and Performance Metrics

• By 2026, agents using advanced token windows are anticipated to become mainstream across various platforms. Companies like OpenAI and Anthropic are already making strides in this area.

• Expectations include achieving a success rate of about 50% on autonomous tasks that typically take four hours for humans, indicating significant efficiency gains.

Market Saturation and Cost Trends

• By the end of 2026, it is projected that approximately 85% of Fortune 500 companies will utilize at least one generative AI function.

• Pricing for outcomes is expected to decrease roughly every nine months while performance metrics may improve even faster due to accelerated learning curves observed historically.

Commoditization Year Insights

AI Software Evolution and Future Predictions

The Shift to AI-First Software

• New software will default to being AI-first, creating pressure for existing software to become AI compatible.

• By 2028, a cognitive plateau is anticipated, aligning with predictions of achieving artificial superintelligence (ASI).

• ASI is defined by the saturation of benchmarks; however, task complexity and size remain critical metrics for evaluation.

Understanding Task Complexity and Intelligence Metrics

• The time horizon serves as a proxy for task size or complexity rather than just performance on single tasks.

• Even if benchmarks are saturated by 2028, there will still be unmeasured aspects of intelligence that indicate underlying utility per token.

Advancements in Agentic Workflows

• By 2028, agents will manage workflows significantly faster than humans; tool use will be integrated into all SDKs.

• This period marks the realization of Sam Altman's concept of "intelligence too cheap to meter," with models becoming cheaper and smarter.

Focus Areas for Vendors Post-2028

• After achieving intelligence capabilities by 2028, vendors like OpenAI and Google must prioritize autonomy, speed, privacy, and memory integration.

• While capable of developing simpler applications autonomously (e.g., mobile banking apps), complex projects like AAA games require more human input due to their intricacy.

Projections for 2029 and Beyond

• By 2029, enterprise software is expected to adopt an agentic-first approach as traditional computing paradigms fade away.

• Task length continues to double despite accelerating advancements; meta orchestration layers will differentiate future systems from simple agent supervision.

Autonomy in Project Development

• A hierarchy of agents will emerge where independent decision-making layers oversee strategy and resource allocation.

• By late 2029 or early 2030, it may be possible for AI to autonomously develop indie games using platforms like Unreal Engine.

Achieving Full Autonomy by 2030

• In 2030, autonomous projects could achieve a success rate of around 20% with zero-shot prompts—indicating significant advancements in AI capabilities.

Acceleration Trends in AI Development

Effective Acceleration Rate

• The acceleration of AI capabilities is observed to double approximately every three months, indicating rapid advancements in technology.

• Notable examples include the release of ChatGPT and its subsequent versions, which exceeded initial expectations for development speed.

Deceleration of Jerk Coefficient

• Analysis across different epochs (2012-2018, 2018-2022, 2022-present) reveals a deceleration in the rate of acceleration over time.

• The fastest doubling time was recorded from 2012 to 2018 at 3.4 months; this increased to 6.4 months by the latest epoch, suggesting a slowing trend in advancements.

Implications of Slowing Acceleration

• Despite the deceleration, overall acceleration continues; however, it indicates that while progress is ongoing, it may not be as rapid as previously anticipated.

• The concept of "jerk" helps explain discrepancies in predicting AI capabilities and suggests that understanding this factor can improve future forecasts.

Current State and Future Predictions

• There remains a "jerk regime," but it's noted that the intensity has lessened compared to previous years; stakeholders should adjust their expectations accordingly.

• Upcoming releases from Frontier Labs could significantly alter current trends; however, recent disappointments (e.g., Llama 4) may indicate further slowing down in advancements.

Measuring Intelligence Per Dollar

• A focus on measuring intelligence relative to cost shows significant improvements: GPT models have become smarter while also cheaper per token over time.

• For instance, GPT3 had a performance score of 43% on MMLU at a cost of $0.06 per thousand tokens; by comparison, GPT4 achieved an 86% score with lower costs per token three years later.

Conclusion on Capability Growth

Comparison of AI Models: Gemini vs. 03

Introduction to Model Comparisons

• Discussion begins on the comparison between different AI models, specifically focusing on intelligence saturation and the release of Gemini 2.5.

• The conversation highlights the differentiation between models, noting that while many consider Gemini Pro superior, it serves a different market niche compared to 03.

Performance Metrics

• Gemini is noted for its strength in handling long conversations and large contexts, being faster and cheaper than 03.

• Cost analysis reveals that Gemini is significantly cheaper at approximately $1.25 per million tokens for inputs compared to 03's $10, making it eight times more expensive on input costs.

Utility and Functionality

• Despite its cost advantages, 03 offers more mature tool calling capabilities which may provide greater utility depending on specific tasks.

• The empirical data suggests that Gemini has drastically reduced input pricing over recent months while maintaining competitive performance metrics against 03.

Factors Influencing Model Development

• The discussion shifts towards various components influencing model performance including Moore's Law, algorithmic efficiency, and economies of scale.

• Positive feedback loops are identified as critical in reducing costs associated with perplexity every nine months due to advancements in hardware and algorithms.

Future Predictions and Trends

• A new utility function is proposed encompassing hardware algorithms, margin compression, and tooling as essential for evaluating AI platforms' total utility.

• Concerns are raised about Meta’s approach to releasing basic models without investing in additional functionalities which could hinder their competitiveness against OpenAI.

Diminishing Returns in AI Development

• Predictions indicate a potential tapering off of development progress around mid-2026 when super-exponential growth may cease.

The Future of AI Tools and Utility

The Potential Growth of AI Tools

• The current landscape features around 600 tools, but the speaker suggests a future where this could expand to 600,000 tools, indicating significant growth potential in AI development.

• Multimodality has yet to be fully explored within the context of existing LLM (Large Language Models), hinting at untapped capabilities that could enhance tool functionality.

Understanding Utility in AI Investment

• Total utility is derived from a formula with four components, showing an upward trend. Continued investment in AI is contingent on this utility increasing; stagnation would deter further funding.

• While there may be expectations for diminishing returns on total utility per token, the speaker expresses skepticism about this notion and believes compounding returns will persist for an extended period.

Projections and Realistic Expectations

• Graphical data presented indicates anticipated trends in total factor utility per marginal dollar on a logarithmic scale. The speaker considers these projections more realistic than previous optimistic estimates.

• Acknowledgment that earlier calculations were rough estimates ("back of the napkin math") suggests room for adjustment as new data emerges.

Conclusion and Key Takeaways