4 pieces selected from The Gradient, AI Alignment Forum — only the ones worth your time.
1. After Orthogonality: Virtue-Ethical Agency and AI Alignment
The Gradient
This essay by The Gradient challenges the foundational assumption of most AI alignment work — the orthogonality thesis — which holds that intelligence and goals are independent axes, and that any AI system can have any goal. The author argues instead that rational human behavior isn't goal-directed in the classical sense; humans align actions to 'practices' (structured networks of actions, dispositions, and evaluation criteria) rather than optimizing toward terminal goals. This reframes rationality from a goal-pursuit model to something closer to Aristotelian virtue ethics: behaving well means acting in accordance with excellent practices, not maximizing a utility function.
Why it matters
Most production AI systems today — from RLHF-tuned LLMs to reward-model-based agents — are implicitly built on the goal-directed paradigm the essay critiques. If the essay's core claim holds, then aligning AI by specifying reward functions or objective targets may be fundamentally misaligned with how human rationality actually works. Developers building agentic systems, AI assistants, or autonomous decision-makers should take seriously the possibility that practice-based, virtue-ethical architectures (where the agent learns norms and situationally appropriate behaviors rather than optimizing a fixed objective) could produce safer and more robustly aligned systems than current RLHF or reward-maximization approaches.
What you can build with this
Build a small LLM-based agent that is guided not by a fixed objective or reward signal, but by a structured 'practice document' — a written set of norms, situational heuristics, and evaluation criteria for a specific domain (e.g., a customer support agent). Evaluate its outputs against the practice document using a judge LLM, and compare alignment quality and failure modes against a reward-function-guided version of the same agent. This directly operationalizes the essay's practice-vs-goals distinction and generates concrete data on whether the approach improves coherence or safety.
Key takeaways
- The orthogonality thesis (that any intelligence can have any goal) is disputed: the essay argues rationality is constitutively tied to practices, not separable from them, which has direct implications for how AI systems should be designed.
- Virtue ethics reframes alignment as learning to act well within a practice (medicine, law, conversation) rather than maximizing a utility function — a meaningful design alternative to RLHF and reward modeling.
- If human rationality is practice-based rather than goal-directed, then specifying terminal goals for AI agents may be the wrong abstraction entirely — practitioners should explore norm-following and practice-embedded architectures as alignment strategies.
2. Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
The Gradient
This essay from The Gradient examines the shifting relationship between mathematics and machine learning research over the past decade. The central observation is that carefully designed, mathematically principled architectures (think symmetry-aware models, geometric deep learning, equivariant networks) are yielding only marginal gains compared to brute-force scaling approaches that throw more compute and data at general-purpose architectures like transformers. The author traces how the field has moved from hand-crafted inductive biases grounded in mathematical structure to empirically-driven, scale-first engineering.
Why it matters
For developers building AI products today, this tension is directly relevant to architectural choices and resource allocation. Investing heavily in mathematically elegant, domain-specific architectures (e.g., equivariant networks for molecular data, graph networks for relational data) may not pay off unless your data domain has extremely strong structural priors and you lack the data scale to brute-force past them. The practical implication is that unless you're in a specialized scientific domain (drug discovery, physics simulations), scaling a general transformer on more data will likely outperform a bespoke mathematical architecture — and you should budget accordingly.
What you can build with this
Run a direct benchmark experiment this week: take a structured domain problem you care about (e.g., predicting properties from molecular SMILES strings or time-series with known periodicity), implement both a symmetry-aware/mathematically principled baseline (e.g., an equivariant GNN or a Fourier-feature network) and a vanilla transformer or MLP scaled with more parameters and data augmentation. Track performance vs. training compute cost and publish your findings — this is exactly the empirical comparison the field needs more of.
Key takeaways
- Scale and compute have consistently outperformed mathematically principled architectural design in general ML benchmarks over the past decade, suggesting that inductive biases are most valuable only when data is severely limited or the domain has extremely rigid structure.
- Mathematical frameworks like geometric deep learning and equivariant networks remain relevant in scientific ML (protein folding, quantum chemistry) where symmetry constraints are hard physical laws, not soft priors — the payoff is domain-specific, not general.
- The practical engineering lesson is that mathematical structure should be treated as a regularizer of last resort: reach for it when you genuinely cannot acquire more data or compute, not as a default design principle for production AI systems.
3. AIs can now often do massive easy-to-verify SWE tasks and I've updated towards shorter timelines
AI Alignment Forum
A researcher at the AI Alignment Forum updated their AI timeline predictions significantly upward after observing Claude Opus 4.5/4.6 and Codex 5.2/5.3 performing well above expectations on both benchmarks and real-world tasks. The core empirical observation: current AI systems can already complete 'Easy-to-verify SWE tasks requiring no novel ideation' (ESNI tasks) — think large refactors, port migrations, test suite generation, API integrations — at a 50%-reliability time horizon measured in months to years of equivalent human work. A C compiler was almost entirely autonomously written by Claude with moderately sophisticated scaffolding, and similar results emerged from METR and Epoch AI evaluations.
Why it matters
If the 50%-reliability time horizon for ESNI tasks is already in the months-to-years range for human-equivalent effort, then the ROI calculus on agentic coding pipelines shifts dramatically right now. Developers building AI products should stop treating agents as tools that handle hour-long tasks and start designing workflows — with proper verification harnesses — that hand off week-to-month-scale engineering work. The bottleneck is no longer model capability for well-specified tasks; it's verification infrastructure and scaffolding quality. Teams that invest in robust automated test suites and evaluation harnesses today are building the exact infrastructure needed to leverage these capabilities before competitors realize the shift has already happened.
What you can build with this
Pick a real, bounded migration task in your codebase — e.g., converting a legacy REST API client library to use a new SDK, or porting a module from one framework to another — and build a minimal agentic scaffold this week: a loop that gives Claude or GPT-4o the repo context, a clear spec, and an automated test suite as the verifier. Run it end-to-end unattended over 24–48 hours and measure how far it gets without human intervention. The point is not to ship it immediately, but to calibrate your own mental model of the current 50%-reliability horizon on your specific task distribution, which is more valuable than any benchmark number.
Key takeaways
- METR data shows roughly 3.5-month doubling times in AI 50%-reliability time horizon throughout 2025, meaning capability roughly doubles every quarter — a pace that makes 2026 forecasts dramatically different from today's baseline.
- The binding constraint on large agentic tasks is scaffolding quality and verification harness design, not raw model capability — the author explicitly identifies a 'scaffolding overhang' where models already exceed what current orchestration infrastructure can exploit.
- High-reliability (90%) task completion remains limited to hours-to-days of equivalent human work even as median (50%) reliability reaches months-to-years, meaning production use cases must be designed around verification and retry loops, not single-shot correctness.
4. My picture of the present in AI
AI Alignment Forum
This post is a situational assessment of AI capabilities and deployment as of April 2026, written by an AI alignment researcher sharing best-guess views without extensive argumentation. The author focuses on AI R&D acceleration, noting that AI tools are delivering real but measured productivity gains at leading labs like OpenAI and Anthropic — estimated at roughly 1.6x serial research engineering speed-up, up from 1.4x at the start of 2026. This improvement comes from a combination of more capable models, better tooling, and human adaptation (workflow changes, skill shifts, and learning to use models more effectively).
Why it matters
The 1.6x figure is a grounding data point that cuts through both hype and dismissal. Developers building AI-assisted tools need to understand that aggregate productivity gains look modest at the team level even when individual tasks see 3-10x speedups — because humans naturally drift toward tasks where AI helps most, inflating perceived gains while the harder baseline work remains. This means product teams should measure AI impact carefully, distinguishing between 'doing existing work faster' and 'doing different work enabled by AI', as conflating the two will produce misleading ROI estimates and poor product decisions.
What you can build with this
Build a personal engineering productivity tracker that logs tasks before and after AI-assisted completion, tags each by task type, and computes a per-category speedup ratio. After two weeks of data collection, you'll have a real, personal version of the 1.6x aggregate metric broken down by task type — giving you evidence-based guidance on where to invest in AI tooling and where to stop expecting it to help.
Key takeaways
- AI tools are delivering approximately 1.6x serial engineering speed-up at top labs as of early April 2026, up from 1.4x just a few months earlier — real but far from transformative at the aggregate level.
- Individual tasks can see 3-10x human time reductions, but aggregate gains are lower because engineers shift toward AI-friendly (often lower-value) tasks, making naive productivity self-assessments upwardly biased.
- A significant portion of the speed-up comes from human adaptation — learning better workflows, shifting task focus, and acquiring skills via AI — not just raw model capability improvements.