Horizon AI Daily Digest - May 30, 2026

A curated selection of 35 standout items from 47 tracked stories, sourced from Hacker News, arXiv, GitHub, and tech media.

Key highlights

• Liquid AI reveals 8B-A1B MoE trained on 38T tokens — a new sparse-activation mixture-of-experts model with strong performance (9.0/10). Source · HN discussion
• COLAGUARD: Robust guardrails with latent reasoning — compresses safety reasoning into a continuous latent space for near 13x speedup at parity (arXiv:2605.29068).
• Trace-answer dissociation in reasoning models — under multi-turn adversarial pressure, chains of thought stay correct while final answers flip, exposing evaluation blind spots (arXiv:2605.29087).
• Out-of-band metadata for safe autonomous agents — the Redpanda agentic data plane enforces security policy and auditing outside the agent read/write path (arXiv:2605.29082).
• California passes the 'Protect Our Games Act' — digital games must remain playable after service shutdown or face sales bans (Source).
• Mistral AI Now Summit notes — Mistral trails Chinese and US rivals technically, but its on-prem strategy appeals to regulated industries (Source).
• GTA 6 developers unionize — Rockstar developers announce a union seeking pay transparency and an end to crunch (Source).
• Is AI causing a repeat of frontend's lost decade? — AI may erode accidental complexity reduction and professional depth in web development (Source).

Research (arXiv)

• Behavior-Induced Mirror-Prox TD Learning — accelerates off-policy linear prediction using a behavior-policy Bellman matrix (arXiv:2605.28849).
• URIEL — ultra-reduced-impact selective logging of tropical forests via airborne robotics with AI and post-harvest silvicultural treatment (arXiv:2605.28883).
• Review Arcade — LLM paper reviews show limited alignment with human reviewers and are gameable by targeted revisions (arXiv:2605.28897).
• Orthogonal Concept Erasure for Diffusion Models — precise concept removal via orthogonal updates without harming generation quality (arXiv:2605.28902).
• Frontier LLM agents for ontology curation — automated mapping of phenotype text to ontology terms (arXiv:2605.28965).
• Adopt ≠ Adapt — longitudinal analysis of in-the-wild LLM conversations shows sticky user behavior and growing task complexity among active users (arXiv:2605.29018).
• When Models Disagree — multi-model disagreement as a diagnostic for LLM evaluation of public comments (arXiv:2605.29025).
• Differentiable Belief-based Opponent Shaping in multi-agent RL (arXiv:2605.29042).
• Cognitive Categorical Transformer — category-theoretic inductive biases cut GPT-2 Small perplexity by 12%, driven by simplex message passing (arXiv:2605.28864).
• VFEAgent — multimodal agents generating executable finite element analysis code from images (arXiv:2605.28978).
• BEAMS — a benchmarking initiative for AI in modeling and simulation (arXiv:2605.28994).
• Mind Your Tone — prompt tone significantly and model-dependently affects LLM accuracy (arXiv:2605.29027).
• Hallucination mitigation via nested learning plus semantic caching in agentic pipelines (arXiv:2605.29055).
• Sim-to-real for RL-based industrial dispatching through explicit execution semantics (arXiv:2605.29078).
• AI in clinical trials — a hybrid human-AI trend analysis finding US/China dominance with multi-country growth (arXiv:2605.29096).
• Behavior-aware auxiliary corrections for stable off-policy TD prediction (arXiv:2605.28855).
• AI-enhanced education survey — 72 higher-education practitioners via the DOT framework (arXiv:2605.29041).

Engineering & industry

• SQLite is all you need for durable workflows — simple and effective, with concurrency caveats (Source).
• The dead economy theory — efficiency gains from technology shrinking employment, requiring resource redistribution (Source).
• On Rendering Diffs — how CodeView renders large diffs in the browser (Source).
• Bijou64 — a variable-length integer encoding with SIMD compatibility tradeoffs (Source).
• Framework 12 hard to justify — weak value-for-money, though repairability and Linux support still attract some buyers (Source).
• We should be more tired than the model — in AI coding, developer taste matters more than raw skill (Source).
• Claude Code v2.1.157 — plugin auto-loading/init and agent field support (Release notes).
• TV Explorer — an advanced UI for free online TV (tvexplorer.live).
• CAPTCHAs can still detect AI agents — though their main role is user tracking, raising privacy and accessibility concerns (Source).
• Tech companies want to film you doing chores — free housekeeping in exchange for robot-training video data (Source).