Horizon AI Daily Digest - May 28, 2026
> 41 items collected, 30 best selected.
Key points
Top Story (10/10)
- The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence — The MiniMax-M2 series features 229.9B parameters with only 9.8B activated, using an end-to-end design to power agentic intelligence and self-evolution. Tags: LLMs, mixture-of-experts, agent frameworks, RL, self-evolution.
- YouTube to automatically label AI-generated videos (8/10) — HN discussion
- Anthropic and OpenAI may have found product-market fit (8/10) — HN discussion
- What Apple and Google are doing to your push notifications (8/10) — Privacy and UX concerns. HN discussion
- DuckDuckGo saw 28% more visits after Google insisted people love AI mode (8/10) — User backlash against forced AI search. HN discussion
- Last.fm is now independent (7/10) — HN discussion
- Tech CEOs are apparently suffering from AI psychosis (7/10) — HN discussion
- Facebook launches a 'Plus' subscription (6/10) — Meta expands monetization with paid tiers and AI subscription tests.
- anthropics/claude-code released v2.1.152 (6/10) — Improved code review and skill flexibility.
- Can LLMs Introspect? A Reality Check (8/10) — Behavioral evidence is insufficient to prove introspection; must distinguish it from pattern matching.
- OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling (8/10)
- Automatic Layer Selection for Hallucination Detection (8/10) — FEPoID criterion for automatic intermediate-layer selection.
- Reasoning, Code, or Both? (7/10) — Chain-of-thought proves more robust than code execution in LLM math reasoning.
- Which Changes Matter? Trustworthy Legal AI (8/10) — LexGuard framework using relevance-sensitive evaluation and formal reasoning.
- Is Agent Memory a Database? (8/10) — Long-term agent memory as a state-trajectory-driven data-management workload.
- Your Agents Are Aging Too (8/10) — Agent lifespan engineering and the AgingBench benchmark for post-deployment degradation.
- JobBench: Aligning Agent Work With Human Will (8/10) — Current models show limited performance.
- Anchor: Mitigating Artifact Drift in Agent Benchmark Generation (8/10)
- Personalizing Embodied MLLM Agents (7/10) — POLAR framework with multimodal knowledge graphs and episodic memory.
- ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence (8/10) — Verifiability framework exposing citation fabrication in current autonomous-research outputs.
- Experiments in Agentic AI for Science (8/10) — DeepTS/DeepCollector and DeepScribe automate scientific data curation.
- Managing Uncertainty in LLM-Generated Procedural Knowledge for Virtual Laboratory Planning (7/10)
- PolyFusionAgent: Multimodal AI for Polymer Property Prediction and Inverse Design (7/10)
- Exploiting Local Dynamics Regularity for Reusable Skills in Offline Hierarchical RL (8/10)
- Advancing Creative Physical Intelligence in Large Multimodal Models (8/10)
- From Static Context to Calibrated Interactive RL for Multi-turn Dialogue (8/10)
- Constraint acquisition needs better benchmarks (7/10) — MPMMine benchmark suite.
- BrickAnything: Geometry-Conditioned Buildable Brick Generation (7/10) — Structure-aware tree tokenization with autoregressive generation.
- Gemini, Gophers, and Fingers: Alternative Internets Beyond HTTPS (7/10) — HN discussion
- SimCity 3k in 4k (2025) (6/10) — HN discussion