EN · International edition · 13/11/2026
Inferama daily · 13/11/2026
The most important AI developments since the previous newsletter, selected for editorial relevance.
News coverage period: 12/11/2026 08:30 → 13/11/2026 08:30 (Europe/Madrid).
ByteDance releases sparse autoencoders to analyze three layers of Qwen3-30B-A3B
ByteDance Seed has released sparse autoencoder weights for layers 4, 30, and 46 of Qwen3-30B-A3B. The associated paper, submitted to arXiv on November 5, 2026, studies how interpretable features are distributed among experts in a MoE model; its results apply to that specific model and analysis, not to a general evaluation of language models.
Read article →
Ai2 releases Molmo2-4B-tb and ManiFlow-tb for tabletop block stacking
Ai2 has released weights for two models built for tabletop block manipulation. The ToyBlocks system pairs a high-level planner with a low-level action policy and reports results in simulation and on a real robot.
Read article →
Inkling 2 Small: Launch and confirmed availability
Thinking Machines Lab has released Inkling 2 Small, a multimodal model with 594B total parameters and 17B active per token, available on the Tinker platform for sampling, fine-tuning, and beta serverless inference since November 5, 2026.
Read article →
ByteDance Seed releases Cross-Layer TopK CLT weights for Qwen3-8B-Base
The public repository contains 36 encoders and 36 decoders in FP32, totaling 94.2 billion parameters under an Apache 2.0 license that permits commercial use.
Read article →
Anthropic releases Claude Fable 5.2 and Mythos 5.2 with distinct access and safeguards
Anthropic announced two models sharing the same underlying base on November 9, 2026: Fable 5.2, available generally with reinforced cybersecurity and biology safeguards, and Mythos 5.2, restricted to approved organizations. Performance improvements and technical results are company claims, not independent validation.
Read article →
MIT introduces UQ-DAAAM, a system that corrects imprecise memories in robots
A team at MIT has developed UQ-DAAAM, an extension of its robotic memory system that identifies and corrects erroneous object descriptions stored in a robot's memory, improving its ability to interact with the physical environment.
Read article →
Google introduces the Gemini Decisions API for fast, calibrated multimodal decisions
The new API allows developers to evaluate up to 1,000 selection, verification, or scoring questions on multimodal inputs in a single pass, with latencies of 150 to 250 ms and a cost of $0.10 per million input tokens.
Read article →
OpenAI’s AI pipeline generated 948 Linux privilege-escalation exploits in a 30-hour run
OpenAI says an agent pipeline using an unreleased cyber-focused model reproduced 1,724 Linux vulnerabilities and generated 948 local privilege-escalation exploits in 30 hours, at a reported API cost of $1.21 million.
Read article →
Anthropic documents the viability of AI self-replication on household devices
Anthropic researchers demonstrate that small AI models can run distributed across appliances and TVs, bringing the possibility of an autonomous worm that spreads without human intervention closer to reality.
Read article →
FlashProve: Anthropic’s cryptographic proof for checking which AI model produced a response
Anthropic’s open-source research prototype FlashProve uses cryptographic proofs to check whether an output token matches a model committed to in advance, without revealing the model’s weights. Its demonstration has important limits: each proof covers one output token, and the system was tested under specific conditions.
Read article →