4 pieces selected from AI Alignment Forum, The Gradient — only the ones worth your time.
1. Challenge: Hand coding weights for efficient sequence memorisation
AI Alignment Forum
We hand coded weights for one layer MLPs that memorises labels for input token sequences of length two. The number of facts our hand-coded models can memorise with 90% accuracy[1]scales roughly linearly with the models' parameter count[2], just like trained models for the same architecture. However, our hand-coded models' scaling prefactor still falls short of trained models' by a factor of
mjx
Key takeaways
2. After Orthogonality: Virtue-Ethical Agency and AI Alignment
The Gradient
The essay challenges the conventional goal-based model of rationality, arguing that humans act rationally not by pursuing fixed goals but by aligning actions with practices—networks of actions, dispositions, and evaluation criteria. It proposes that AI systems should similarly be designed without rigid goals, instead embodying virtue-ethical agency where actions are evaluated within the context of broader practices. This approach aims to address alignment issues by shifting focus from predefined objectives to adaptive, context-sensitive behaviors.
Why it matters
For developers building AI products, this matters because it offers an alternative framework for designing AI systems that are more adaptable and aligned with human values. Traditional goal-based AI can lead to unintended consequences when goals are misaligned or misinterpreted. A virtue-ethical approach could make AI systems more robust and context-aware, reducing risks associated with rigid goal-setting.
What you can build with this
Develop an AI agent that evaluates actions based on a set of ethical practices rather than predefined goals. For example, create a customer service chatbot that adapts its responses based on contextual cues and ethical guidelines rather than a fixed script or goal.
Key takeaways
- Rationality in humans is not about pursuing fixed goals but aligning actions with broader practices.
- AI systems designed with virtue-ethical agency may be more adaptable and context-aware.
- Goal-based AI can lead to unintended consequences; a practice-based approach could mitigate these risks.
3. The Long (Self-)Correction
AI Alignment Forum
The essay critiques the concepts of AI Pause and Long Reflection, arguing that neither addresses the core issue: humans are fundamentally flawed and unready to build or oversee powerful AIs. The author lists specific human flaws, such as lack of a workable moral framework, poor philosophical and strategic competence, susceptibility to manipulation, and over-optimism about partial solutions. These flaws make humans unreliable as builders, overseers, or alignment targets for AI, and the author suggests that a long, uncertain process of self-correction is needed before attempting to build powerful AIs.
Why it matters
Developers building AI products must recognize that human flaws can directly translate into AI flaws, especially in areas like alignment, ethics, and long-term strategy. This essay highlights the need for humility and caution in AI development, emphasizing that technical solutions alone are insufficient without addressing underlying human biases and incompetence.
What you can build with this
Create a self-assessment tool for AI development teams that evaluates their philosophical and strategic competence, identifying gaps and biases that could impact AI alignment. The tool could use a questionnaire to probe for overconfidence, positional values, and susceptibility to manipulation, providing a risk score and recommendations for improvement.
Key takeaways
- Humans lack a workable moral framework and are poor at philosophy and long-term strategy, making them unreliable overseers of AI.
- Human flaws like susceptibility to manipulation and over-optimism about partial solutions can directly impact AI safety.
- Addressing human flaws is a prerequisite for building safe, powerful AI, requiring a long process of self-correction.
4. AGI Is Not Multimodal
The Gradient
The essay argues against the notion that current multimodal AI models are on a direct path to Artificial General Intelligence (AGI). It highlights that while these models excel at generating human-like text and processing multiple data types, they lack the embodied, tacit understanding that underpins human cognition. The author emphasizes that human intelligence is deeply rooted in physical and social experiences, which current AI models do not possess or replicate.
Why it matters
Developers building AI products need to recognize the limitations of current multimodal models. Understanding that these models lack true embodied intelligence helps set realistic expectations and guides the design of AI systems that complement rather than replace human cognition. This insight is crucial for creating AI applications that are both effective and ethical.
What you can build with this
Develop an AI system that integrates multimodal inputs with real-world sensor data to create a more grounded understanding of user contexts. For example, build a home assistant that uses visual, auditory, and environmental data to provide more contextually aware responses.
Key takeaways
- Current multimodal AI models lack the embodied understanding that characterizes human intelligence.
- Human cognition is deeply influenced by physical and social experiences, which AI models do not replicate.
- Developers should design AI systems that complement human intelligence rather than aiming for AGI-like capabilities.