Model releases and updates
MiniMax Releases Open-Source H3-Image With Native 4K Image Generation
QbitAI reports that MiniMax has released H3-Image, an open-source image generation model that supports native 4K output at 4096×4096 pixels and up to 16 reference images. The author describes tests involving CAD views, exploded diagrams, academic text, and design assets, and says 4K images took about 10–20 seconds in the web interface. The article also reports API pricing at 50% to 70% of major competitors, with a further 50% discount on 16MP requests during the first week.
IBM Releases LLMON Preview Models Based on Granite 4.1 3B and 8B
IBM is releasing two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, trained to consume and generate text marked with LLMON, an LLM-native format that separates instructions from data. On the OPI benchmark, the post reports that the previews avoided 47% more attacks than the best tested frontier model and 95% more than the best tested open-source model. The post says the models are for exploration and evaluation, not production, and that adaptive attacks remain under development.
Product releases and updates
AWS Machine Learning BlogFirst-party
AWS Announces Faster, Larger Amazon Bedrock AgentCore Runtime
AWS announced updates to the Amazon Bedrock AgentCore runtime, including sessions that start in under 2 seconds, scale up to 8 vCPUs and 32 GB, and support container images up to 10 GB. The runtime can reclaim unused memory and let idle agents scale to zero, while sessions retain hardware isolation. The updated runtime is available today in AWS Regions across North America, South America, Europe, and Asia Pacific.
Anthropic Introduces Claude Meetings for Google Meet
Anthropic is rolling out Claude Meetings in beta for Team and Enterprise plans, adding Claude as a participant in Google Meet calls. It can produce meeting notes, answer questions, and create documents and whiteboards during the call; notes, transcripts, and ordinary questions are included during the beta, while connected tool use counts toward the user's allowance. Anthropic says other meeting platforms are coming soon.
Industry
TechCrunch · AIJensen Huang Resigns as Nvidia CEO Amid Scandal and Investigations
Nvidia confirmed in an SEC filing that its board accepted Jensen Huang’s resignation, and CFO Colette Kress will serve as interim CEO while continuing her CFO duties. The timeline also recounts reports of Huang’s relationships with Megaspeed executive Alice Huang and Nvidia executive Kimberly Powell, as well as investigations into Megaspeed over suspected chip diversion to China. Nvidia, Huang, and Powell have not confirmed the report about Powell.
Grok Build Apps Exposed Sensitive Data Through Public Links, Report Says
QbitAI reports that security firm Claroc found more than 35,000 Grok Build apps accessible without login, with an estimated 650–800 containing private data. Examples included SpaceXSI and X sales information, Tesla employee performance and bonus details, ICU patient records, and plaintext passwords. Claroc said passive DNS databases could reveal app URLs; Ars Technica reported that 24 apps created by SpaceXSI and X employees had been taken down by publication time.
TechCrunch · AI
Nebius Tests the Neocloud Playbook With $40 Billion in Planned Spending
Nebius outlined $40 billion in planned spending for 2027, funded partly by customer prepayments and loans backed by customer contracts. Its New Jersey data center, operated by DataOne, is months behind schedule, pushing more than half of Microsoft’s capacity delivery to Q2 2027. Nebius said third-party site delays contributed to a $200 million cut to its 2026 revenue forecast, now $3 billion.
Research
TCSExam Evaluates AI Agents on 100 Theoretical Computer Science Research Tasks
A joint research team introduced TCSExam, a benchmark of 100 theoretical computer science tasks that tests AI agents on theorem proving, solver development, mathematical constructions, and open conjectures. The framework uses Lean 4, program checkers, and checkpoints to verify results and distinguish partial progress from completed proofs. The article also reports that GPT-6 Astra extended a result for the Alon–Jaeger–Tarsi conjecture from 8 to 9 dimensions over a five-element finite field, while the general conjecture remains unresolved.
The Decoder
OpenAI Analysis Links Frontier Models’ Hidden Reasoning to Training Choices
OpenAI’s safety team analyzed models from GPT-5 through GPT-6 Astra and identified four factors linked to reduced chain-of-thought transparency: base-model capability, architecture, reinforcement learning, and transfer from teacher models. The analysis finds that some architectures gain more on silent multi-step reasoning, while more reinforcement learning increases controllability in Astra-class models. It also describes a tradeoff: longer, more visible reasoning can be easier to monitor but raises deployment compute costs.
Tips and perspectives
Nathan Lambert · XNathan Lambert Reacts to Claims About Chinese Models’ Training Chips
Nathan Lambert asks, “I thought people already thought/knew this?” while quoting Will Depue’s post about a forensic analysis of open model weights. Depue claims the analysis matched six Chinese models to NVIDIA Blackwell systems and says the findings were independently verified by senior researchers at OpenAI and Anthropic.
A Guide to Fine-Tuning TRELLIS.2 for 3D Footwear Generation
Hugging Face provides scripts and Docker containers for fine-tuning TRELLIS.2 on custom 3D assets locally or with Hugging Face Jobs. In an experiment using 188 footwear objects from Google Scanned Objects, the authors added LoRA adapters to the sparse-structure and shape DiTs while freezing the texture DiT. On nine held-out objects, the adapters improved the mean F-score from 0.276 to 0.803 and reduced Chamfer distance from 0.174 to 0.033.
Tomorrow at 08:00, a briefing like this arrives in your inbox.
The same kind of briefing as the one above: a few stories worth reading. Free. Mail starts after you confirm, and every letter has a one-click unsubscribe. Read a past issue · 2026-11-12