AI Brief: October 30, 2026
Google DeepMind launches Gemini 4 Carbon for cyber defense, Mistral releases its trillion-parameter 'Le Chonk' model, and Anthropic files for IPO. New research highlights AI influence on human decisions and shared numerical representations in LLMs.
Models and Research
Google DeepMind launches Gemini 4 Carbon, a frontier model for cyber defense
Google DeepMind launched Gemini 4 Carbon, a frontier model for cybersecurity and long-horizon tasks, scoring 88.2% on SEC-bench Pro and ranking #1 on Artificial Analysis Intelligence Index v5.
NVIDIA Releases Nemotron 3.5 Transcribe Open Speech LLM
NVIDIA released Nemotron 3.5 Transcribe, an open-weights speech LLM handling 40 languages with 4.71% mean WER on English Open ASR sets, fitting an RTX 5090 in NVFP4.
Prime Intellect Releases Prime Mega MoE Trainable Fused Kernel
Prime Intellect released Prime Mega MoE, a fused MoE kernel with overlapping backward pass, achieving 32–34% faster training steps than baseline and up to 3.11x forward-pass speedup.
Mistral releases Le Chonk, a trillion-parameter coding model trained in Europe
Mistral released weights for Large 4 'Le Chonk,' a trillion-parameter coding model trained in Europe, beating open rivals but trailing DeepSeek V4.1 Flash and GLM-5.3.
Baseten Releases joey-1, a 2B Decision Model at $0.01 per 1M Tokens
Baseten released joey-1, a 2B decision model with 20 ms latency at $0.01 per million tokens, scoring 63.30 on the Jev Decision Index and ranking top among sub-4B models.
Anthropic study: Experts adopt AI agent's random opinions and think they're their own
Anthropic study with 50 experts found they adopted an AI agent's random opinions up to 86% of the time, believing the conclusions were their own.
Infini-AI-Lab Releases Hermit, an Open-Source Harness Adapter Framework for AI Agents
Infini-AI-Lab released Hermit v0.1.0, an open-source framework training external harness adapters for agents like Codex or Claude Code without fine-tuning.
Beyond the Readout: A Latent Numerical Frame Shared Across Modality and Operation
An arXiv study using causal interventions found Gemma4-12B-it shares an abstract numerical representation across text and image inputs, with 91.4% recovery after rotation and rescaling.
TeamGym benchmarks LLM agents on team decisions with distributed information
TeamGym benchmarks LLM agents on team decisions with distributed information, showing RL on final-decision loss improves decision quality and produces emergent coordination.
Timestral open time series models claim top spots on forecasting benchmarks
Timestral released four probabilistic time series models (5M–209M params) using Decoupled Horizon Forecasting, ranking first and second on GIFT-Eval, TIME, and fev-bench.
M4Bench Evaluates AI Coding Agents on Clinical EHR Derivations
M4Bench evaluates AI coding agents on clinical EHR derivations, finding specification completeness accounts for most gains, with skills adding only 0.024 reward improvement.
CrossProfile shows rarest and relevance memory heuristics fail for LLM group decisions
CrossProfile benchmark shows rarest-first and relevance memory heuristics fail for LLM group decisions, with accuracy tracking evidence balance rather than selection policy.
Concordat Enforces Typed Session Protocols on LLM Agents at Runtime
Concordat enforces typed session protocols on LLM agents at runtime, preventing out-of-protocol calls and achieving zero violations on tau-bench airline tasks.
Industry and Products
OpenAI-backed Thrive Holdings acquires ADT
OpenAI-backed Thrive Holdings acquires ADT for $15.3 billion to integrate AI into home security systems.
AssemblyAI launches AssemblyAI-TTS-1 streaming TTS API for voice agents
AssemblyAI launched AssemblyAI-TTS-1 streaming TTS API for voice agents, priced at $15 per 1M characters with 4,000+ voices and 47 ms median time to first byte.
Epsilon builds intent-driven UI testing platform with Amazon Nova Act
Epsilon built an intent-driven UI testing platform with Amazon Nova Act, reporting 4-6x faster test authoring and 100% pass rate on trajectory replay.
Spawn runs a multiplayer game world with autonomous coding agents
Spawn runs a multiplayer game world with autonomous coding agents using Fable-powered agents, costing over $3 million in tokens per month and shipping code hourly.
MoEngage turns 400 private APIs into agent tools with AgentCore Gateway
MoEngage used Amazon Bedrock AgentCore Gateway to expose 400 internal REST APIs as agent tools, saving six months of engineering effort.
Anthropic files S-1 as it seeks IPO and AGI-scale investment
Anthropic filed an S-1 for IPO on Nasdaq, with reported valuation range of $1.5–2 trillion and $70 billion-plus ARR.