Daily AI Briefing — May 13, 2026

Thinking Machines released TML-Interaction-Small, a 276B-parameter MoE realtime voice model advancing SOTA in continuous human-AI interaction—the day's most significant frontier development. OpenAI dominates headlines across multiple fronts: Sam Altman testified in the Elon Musk trial over OpenAI's for-profit restructuring, another wrongful-death lawsuit emerged from a teen fatality linked to ChatGPT, and the company launched Daybreak, a new cybersecurity initiative built on Codex Security.

• Google announced a major Gemini Intelligence overhaul for Android, expanding agentic app automation across Pixel and Samsung devices
• AntAngelMed (103B-parameter open-source medical MoE) released by Chinese researchers as the largest open medical LLM
• Tilde Research introduced Aurora optimizer fixing neuron death in Muon, achieving SOTA on nanoGPT speedrun
• A $1.3B funding round for a compute-access startup and $790M for Nscale in Norway signal continued massive AI infrastructure investment
• A Hugging Face malware incident impersonating OpenAI accumulated 244K downloads, highlighting supply-chain risks
• Amazon employees are 'tokenmaxxing'—gaming AI usage metrics to satisfy management pressure on their MeshClaw tool adoption

A strong day for MoE architecture design and optimizer theory, alongside critical AI safety findings on monitoring reliability and emergent misalignment.

• Slicing and Dicing (Meta) delivers the first systematic MoE design study across 2,000+ pretraining runs up to 6.6B params, establishing actionable guidelines for expert count, granularity, and load balancing
• ExploitGym (Carlini, Song, Wallace et al.) benchmarks whether AI agents can convert known CVEs into working exploits, a critical dual-use capability evaluation
• Overtrained, Not Misaligned analyzes 1M+ model responses across 12 open-source models, attributing emergent misalignment to overtraining rather than genuine value drift
• Classifier Context Rot reveals that frontier safety monitors (Opus 4.6, GPT-5.4, Gemini 3.1) miss dangerous actions 2–30× more often late in long contexts

On the optimization front, SODA unifies Muon, Lion, AdEMAMix, and NAdam under a single Optimistic Dual Averaging framework, while a separate study shows Muon's spectral rescaling success extends to random and inverted spectra, challenging its geometric narrative. ProFIL combines linear probes with RL filtering to reduce reasoning theater in chain-of-thought. FragBench exposes cross-session attack blind spots where benign-looking fragments bypass all tested safety filters. Geometric Factual Recall proves logarithmic embedding dimensions suffice for factual memorization in transformers, and 20/20 VLMs demonstrates +11.7pp average improvement through data curation alone.

A landmark funding round and debates about AI's real-world impact dominated the day. Demis Hassabis announced Isomorphic Labs raised $2.1B for AI-driven drug discovery, the largest known AI-bio investment. Andrew Ng published a comprehensive argument against the AI jobpocalypse narrative, generating massive engagement.

• Google DeepMind went viral with an AI-enabled mouse pointer concept reimagining the 50-year-old interface using Gemini, drawing 909K views
• Gary Marcus and François Chollet converged on neurosymbolic AI arguments — Marcus called Claude Code the most neurosymbolic system he's seen; Chollet argued symbolic learning replaces gradient descent entirely
• Ethan Mollick made a sharp observation that AI labs forming consulting groups contradicts their own ASI claims, and criticized OpenAI for silently removing Study Mode
• Anthropic's Claude Cowork with Opus 4.7 demonstrated autonomous booking, while Perplexity published GB200 inference research on Qwen3 235B
• Allie K Miller identified AI's 'secondary intent problem' — models failing to understand subtext and soft refusals in human communication