Thursday, June 18, 2026

Next-Latent Prediction Transformers. Microsoft Research's NextLat adds latent-state self-prediction on top of next-token training, yielding better representations and 3.3x faster inference. Source

SAE Interventions are Unreliable: Post-Intervention Recovery of Suppressed Behavior. Shows sparse-autoencoder-based interventions on model internals are unreliable because suppressed behaviors recover afterward. Source

Diffusion-Proof: Recipe for Formal Theorem Proving Beyond Auto-Regressive Generation. Proposes a diffusion-based recipe for formal theorem proving that outperforms autoregressive approaches. Source

PAIWorld: A 3D-Consistent World Foundation Model for Robotic Manipulation. A 3D-consistent world model designed as a foundation for robotic manipulation policies. Source

Ollama expands local model support. Adds Kimi-K2.6, GLM-5.1, MiniMax, DeepSeek, gpt-oss, Qwen, and other recent frontier and open-weight models for local inference. Source

Kairos: A Native World Model Stack for Physical AI. A purpose-built world model stack for physical AI applications, released on Hugging Face. Source

Claude Code and Claude Design get two-way sync. Anthropic ships a /design-sync command enabling bidirectional state sharing between Claude Code and Claude Design. Source

Adam (YC W25), Open-Source AI CAD. YC W25 launch of an open-source AI-native CAD tool, receiving a mixed reception from engineers. Source

GLM-5.2 is the new leading open-weights model on Artificial Analysis. Zhipu's GLM-5.2 takes the top open-weights spot on the Artificial Analysis Intelligence Index. Source

Anthropic got hit by export rules nobody understands. Unclear US export regulations disrupted Anthropic, reigniting debate over chip and model export controls. Source

Midjourney goes from generating cat images to full-body ultrasound scans. Midjourney's image generation is being applied to medical ultrasound imagery, raising capability and safety questions. Source

LLMs are entering their Formula 1 era. Benchmark sweep across 8 advanced coding/reasoning tasks comparing GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro, and GLM 5.2. Source

Is foundational AI research still something that can be done without access to HPC?. ML community thread debating whether non-lab researchers can still produce foundational work without large-scale compute. Source

The White House–Anthropic fight over Fable. The Verge column unpacks the political conflict between the White House and Anthropic around a project dubbed "Fable." Source

The next humanoid robot might not look human at all. Coverage of how next-gen humanoid robots (e.g., Genesis AI's "Eno") are abandoning anthropomorphic form factors. Source