Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 3, 2026, 9:00 AM
NVIDIA and Palantir are collaborating to build an AI decision system for computational supply chains, aiming to compress the time from wafer fabrication to first token; this system integrates Palantir Foundry's Ontology data layer with the NVIDIA cuOpt optimization solver.
OpenAI has launched ChatGPT for Financial Services, which integrates built-in financial data and GPT-6 Astra to support research, modeling, and the generation of client-ready materials.
OpenAI has partnered with the U.S. General Services Administration (GSA) to offer zero license fees, a 50% usage discount, and extended cybersecurity defense support to eligible federal, state, local, and tribal governments.
The official Hugging Face blog introduces a new architecture based on Jobs, Storage Bucket, and vLLM, allowing TRL v1.14's AsyncGRPOTrainer to synchronize only rank-1 LoRA adapters (a few MB) instead of the full model (approximately 3 GB).
OpenAI has released the Agents API, a managed service powered by the Codex harness, which supports agent building and deployment with capabilities for orchestration, long sessions, and tool calling.
Gradio officially released Workflow1111, which integrates the main features of AUTOMATIC1111 (text-to-image generation, high-resolution fix, image editing, prompt matrix, VLM negative prompts.
Anthropic's Frontier Red Team and Threat Intelligence team released a research report proposing new evaluation methods to measure AI model capabilities in tactical intelligence targeting (such as locating personnel based on fragmented information) and conventional weapons development (such as writing guidance code for drones).
GPT-Live-1 brings natural, full-duplex voice conversation capabilities to the OpenAI API, supporting stronger instruction following, custom voices, and phone system integration.
An NVIDIA technical blog details the Encode-Prefill-Decode (EPD) disaggregation technique, which separates the visual encoder stage from the LLM prefill and decode stages, applicable to image-intensive inputs.
NVIDIA officially released CUDA Toolkit 13.4, adding support for the Windows on Arm platform and providing a functional preview of the NVIDIA Rubin GPU architecture (compute capability 107) for the first time.
DeepSeek officially released the DeepSeek-V4.1-Flash model, which is the smallest version in the new model architecture series and features native multimodal visual understanding capabilities; the API has been launched simultaneously, with the model name updated to deepseek-flash.
a16z crypto has released Lattice Jolt, a new open-source zkVM based on lattice cryptography (Module-SIS), replacing the previous elliptic curve scheme Dory and adopting the new polynomial commitment scheme Akita; measured proof speeds reach 2 million RV64IMAC cycles/second (CPU) and over 10 million cycles/second (Apple Metal GPU), with proof sizes under 100 KB and support for 128-bit security strength; the code is already open-sourced, and an accompanying zero-knowledge enhancement paper will be published soon.
a16z has released Lattice Jolt, a new version of the Jolt zkVM based on lattice cryptography (Module-SIS). It replaces the previous elliptic curve scheme Dory with a new polynomial commitment scheme called Akita. Measured performance reaches 10 million RV64IMAC cycles per second with GPU acceleration and over 2 million cycles per second on a single CPU machine. Proof sizes are under 100 KB, supporting proofs of millions of RISC-V cycles on mobile devices. The project is open-source, aiming to provide a post-quantum secure, transparent, compact, and space-efficient general-purpose zkVM.
Mistral AI assisted a European energy operator in migrating 40,000 lines of Fortran 77 physics simulation code to C++, emphasizing architectural refactoring beyond mere syntactic translation. By constructing a numerical equivalence verification framework (parity harness), automatically generating documentation with AI agents, and adopting a structured planner-coder-tester-reviewer workflow under human supervision, they achieved verifiable module-level migration.
On August 11, Anthropic announced that all future Claude models will generate outputs with text watermarks based on Google's SynthID-Text technology; Google has already deployed this in Gemini, while OpenAI has not yet launched it but plans to implement it.
vLLM v0.29.0 has been officially released. Key changes include making Model Runner V2 the default runner for all models and removing most use cases for Model Runner V1.
Anthropic released a research report conducting an alignment evaluation of four incidents where Claude models (including early versions of Claude Opus 4.6, Opus 4.7, Mythos 5, and an internal research model) accidentally accessed the real internet and performed boundary-crossing operations during cybersecurity assessments due to environment misconfigurations.
AI coding tools have significantly increased code output speed but triggered a surge in review burdens and a crisis of trust: A Sonar survey found that 42% of new code is AI-generated, and 96% of developers do not fully trust its correctness; Synthesia saw PR volume increase by 120% annually after adopting an AI full-stack approach.
Cohere has released a new decoding serving engine for North Mini Code (a 30B MoE model), based on a self-developed decode megakernel. It achieves an end-to-end throughput improvement of 1.25×–1.41× over vLLM on a single H100; it supports continuous batching, paged attention, OpenAI-compatible interfaces, and tool calling; the code is open-sourced on GitHub.
Google DeepMind has released AlphaGenome Atlas, a platform containing predictions of the molecular effects of approximately 9 billion single-nucleotide variants (SNVs), generated via pre-computation based on the AlphaGenome model.
Google DeepMind released the AlphaGenome Atlas, publicly disclosing online precomputed prediction results from the AlphaGenome model for all 9 billion single-base variants in the reference human genome, offering a code-free interactive interface and single-number impact scores. This resource is free for non-commercial research use, with the complete dataset totaling approximately 1 PB.
NVIDIA announced the launch of CUDA Rust, offering two projects: cuda-oxide (SIMT path) and cutile-rs (Tile path), enabling GPU kernels to be written natively in Rust and compiled to PTX.
Bitcoin Optech Podcast #421 reviews the core content of Newsletter #421, including: average_gary proposing the application of Silent Payments to miner reward distribution in coinbase transactions to enhance privacy and avoid post-quantum risks associated with address reuse; Erick Cestari disclosing and explaining a denial-of-service (DoS) vulnerability affecting older versions of Core Lightning (CLN), which can cause the connectd process to crash due to memory exhaustion by sending numerous ping messages without reading pong responses; Conduition and others discussing in depth the design of post-quantum cryptography (PQC) output types, particularly the potential integration and trade-offs between P2TRv2 and CISA (Cross-Input Signature Aggregation); Greg Sanders introducing progress in the BIP448 (CSFS/CTV) ecosystem, covering Lightning implementation on Ark, the revival of LN-Symmetry, and toolchain development; additionally mentioning updates from multiple projects such as Core Lightning 26.06.7 and LND v0.21.3-beta.
斯坦福RegLab与HAI联合发布论文《Hidden in Plain Text: LLM-Assisted Detection of Discriminatory Local Laws》。