Horizon AI Daily Digest - May 29, 2026
> 40 items tracked, 27 selected below.
---
Top Picks (9.0/10)
1. Just Use Postgres for Durable Workflows - Argues Postgres can handle durable workflows, focusing on data consistency and architectural simplicity. Hacker News discussion 2. Why LLMs Fail at Causal Discovery and How Interventional Agents Escape - Demonstrates fundamental limitations of LLMs in causal discovery and proposes an intervention-based agent approach, A-CBO. 3. RULER: Representation-Level Verification of Machine Unlearning - Proposes the RULER representation-level verification metric; finds existing methods cannot detect representational残留→ residual traces. 4. Voluntary Collusion with Secret Tools in Competing LLM Agents - LLM agents voluntarily collude to use harmful tools for strategic advantage; standard alignment fails to prevent this, only ethical frameworks help. 5. Cross-Entropy Games and Frost Training - Frost Training leverages reward gradients for LLM policy optimization, yielding faster and higher-scoring outputs.
Releases & Models (8.0/10)
- anthropics/claude-code v2.1.154 - Defaults to Opus 4.8 high effort; adds dynamic workflows and a discounted fast mode.
- Soro: A Lightweight Foundation Model and Chatbot for Tajik - Gemma 3-based Tajik LLM with released open benchmarks; significant performance gains.
- Laguna M.1/XS.2 Technical Report - Two MoE coding models reaching open-source state of the art on SWE-bench and similar benchmarks.
- DynaSchedBench - Calibrated dynamic scheduling benchmarks with difficulty stratification; reveals an observability paradox in LLM scheduling agents.
- On the Origin of Synthetic Information by Means of Steganographic Inheritance - Uses steganography to simulate genetic mechanisms, tracing the origin and evolution of synthetic information.
- Behavioural Analysis of Alignment Faking - Systematic study finding alignment faking is more prevalent and predictable than assumed; drivers include values, goal preservation, and sycophancy.
- Hierarchical Prompt-Domain Control for Resource-Constrained Agentic LMs - Addresses unreliable prompting and limited fine-tuning in resource-constrained agents.
- DeepSciVerify - Selective evidence escalation improves accuracy and efficiency of scientific claim-citation verification.
- Prefix-Safe Bayesian Belief Tracking for LLM Reasoning Reliability - Separates calibration from ranking via Bayesian belief tracking over prefix-safe observations.
- Continue? Y/N - A 60-second game about AI agent permission fatigue (HN discussion).
- The Permanent Upper Crow - A satirical loop game critiquing consumerism and status climbing (HN discussion).
- Altman and Amodei walking back AI jobs apocalypse predictions - Community comments note executive misunderstandings and AI's actual assistive role (HN discussion).
- Identifying Human Values in Text - Tailorable LLM architecture detecting and quantifying human value strength in text.
- LaneRoPE - Positional encoding enabling collaborative parallel reasoning/generation; improves math reasoning.
- Discovery Agents for Real-Time Analytics - Multi-agent architecture shifting real-time analytics from passive queries to proactive insights.
- Agyn - Open-source AI agent platform on Kubernetes/Terraform with agents-as-code and zero-trust access.
- Causal State Intervention and Human Outcomes - Argues human outcomes are controllable through causal state intervention.
- Cyberbullying Governance on Social Media - Unified framework from content identification to proactive intervention.
- Reasoning and Planning with Dynamically Changing Norms - Defeasible logic resolves dynamic norm conflicts for AI planning, validated on dialogue tasks.
- Intelligence as Managed Autonomy - Proposes a managed-autonomy theory and SMARt model governing agentic AI behavior under uncertainty.
- $2,000 AI-generated film debuts at Tribeca - *Dreams of Violets*, an AI-generated film, premieres at Tribeca.
- Nitpicking the shell history scene in 'Tron: Legacy' - Deep dive into the accuracy of the film's shell history scene (HN discussion).