The U.S. Department of Homeland Security issued a notice on Monday requiring all civilian government agencies to remove DeepSeek products and third-party services based on the model from their IT systems and prohibiting new purchases; the rationale is that usage could result in the transmission of sensitive data to the People's Republic of China and generate actionable intelligence against U.S. individuals or government entities. Previously, the Department of Commerce, the Navy, and the states of New York, Tennessee, and Virginia had restricted DeepSeek; the U.S. Federal Communications Commission will vote on October 29 regarding whether to ban Chinese labs from testing U.S. electronic devices.
Searchtok has released new features enabling AI agents to search TikTok and Instagram content; it supports integration with Claude, ChatGPT, Cursor, or Claude Code via a single URL; it can identify niche viral posts (ranked by the extent to which they exceed an account's typical view counts), quote opening hooks verbatim, and display creator information; targeting founders, marketers, content creators, and small agencies.
Yara Reads is an open-source reading application for children who have just mastered phonics. It supports word-level pronunciation hints, automatically validates story readability based on learned phonemes, and can integrate with AI models such as Claude or ChatGPT that support MCP to generate stories adapted to specific phonemes.
JPMorgan has deployed JPM Coin to the Ethereum mainnet for vetted institutional clients, supporting redemption into USD at any time on a 1:1 basis; the token was previously launched on the Coinbase Base network in November 2025, and plans are in place to build a 24/7 tokenized deposit network with banks such as Citigroup by the first half of 2027.
NVIDIA published a solution on its technical blog for accelerating Mamba-2 forward computation based on CUDA Tile (cuTile). The solution includes three custom GPU kernels: chunk_cumsum, fwd_pass_state, and compute_out. It utilizes Tensor Cores to implement log-space accumulation, state jump updates, and causal mask output fusion. On B300 SXM6, it achieves kernel-level speedups of 1.14x–1.82x compared to the Triton implementation.
TwelveLabs has launched the Pegasus 1.5 video language model, now available for invocation via Amazon Bedrock (us-west-2 region). The model supports end-to-end analysis of videos up to two hours long, provides structured metadata output with timestamps, offers multimodal prompting (supporting reference images), enables real-time processing without pre-indexing, and accepts input via S3 URIs or base64 strings.
Deepgram migrated its Nova-3 speech recognition model to AWS Trainium2 chips. Through TorchNeuron compilation, fused kernel replacements using NKILib (such as FFN, depthwise conv1d, and conv2d), tuning with the Neuron Explorer and NAD toolchain, and a custom Rust data-parallel runtime, it achieved an 11–70% improvement in offline batch processing throughput on trn2.3xlarge instances (measured by audio processed per dollar).
Multiple developers have used AI tools such as Claude Code to remotely clean up Android TV/Google TV devices via Wi-Fi, removing pre-installed ad apps, replacing native launchers (e.g., transforming them into Apple TV style), and disabling ACR tracking features; the projects dumbtv and cleanyourtv have been tested on devices from multiple brands including Fire TV, Xiaomi, Sony, and TCL, without requiring root access; users shared specific prompt examples and operational procedures, but there is a risk of accidentally deleting system packages or bricking the device.
AWS engineers implemented and open-sourced a push-based NIXL KV transfer mechanism in vLLM. By having prefill instances actively write KV caches to decode instances, it moves registration, connection establishment, and transfer submission out of the latency-critical path, thereby reducing Time To First Token (TTFT) and Time Per Output Token (TPOT).
AWS introduced bidirectional KV cache transfer in vLLM v0.21.0, enabling prefill nodes to reuse KV blocks retained by decode nodes during multi-turn conversations, thereby reducing redundant prefill computations; in a 1P1D deployment on Amazon EC2 P5en + EFA, TTFT for Qwen3-32B and zai-org/GLM-5-FP8 decreased by up to 1.6x and 3.3x respectively; this feature requires a KV-cache-aware router and NixlConnector, and supports configurable block TTL and kv_recompute_threshold.
AMD officially released the open-source decision optimization engine rocOpt, which supports running on MI300X and MI355X GPUs and requires ROCm 7.2.3; it provides four types of solvers: PDLP (for large-scale LP), hybrid CPU/GPU MILP, vehicle routing metaheuristics, and dual simplex, currently without support for barrier solvers; it is compatible with the cuOpt API and can be deployed via GitHub source code, Docker images, or official installation guides.
AMD officially released hipCIM 26.06.00 as an AMD ROCm port of cuCIM, supporting accelerated whole-slide image (WSI) I/O on AMD Instinct™ MI355X (gfx950) GPUs. This version is based on ROCm 10.0. The GPU path achieves approximately 10x the read speed of OpenSlide for Aperio SVS format slides, while the CPU path reaches 7–11x. Decoding accuracy matches the libjpeg-turbo reference (maximum absolute difference of 1 for 8-bit). OME-TIFF format support coverage increased from 0.8% to 90.3%, and zero-copy direct pass-through to CuPy and PyTorch-ROCm in GPU memory was implemented.
Due to insufficient power supply at some data centers, billions of dollars worth of Nvidia GB300 server racks are stuck in warehouses unable to be deployed; the company uses an 'early rack acceptance' process to secure equipment but must pay upfront, exacerbating free cash flow pressure. Capital expenditures for 2026 are expected to reach $220 billion, with free cash flow already turning negative. Coordination challenges across multiple stages—including power permits, construction labor, and grid access—create uncertainty regarding the actual commissioning of 44 GW of data center capacity currently under construction.
DeepSeek has open-sourced its internally used inference framework DeepInfer and token encoding/decoding component deepseek-recipe. It supports NVIDIA and Huawei Ascend platforms, integrating capabilities such as distributed inference, speculative decoding, SWA Bounded Replay, and structured output. The disclosed daily average throughput for DeepSeek V4.1 Flash is 155 trillion Tokens/day, with revenue per card per hour at $38; deepseek-recipe supports mainstream large model API protocols and Tokens encoding/decoding, ensuring that third-party deployments match the official API experience; KVCache management and load balancing components have not been open-sourced due to coupling with internal infrastructure.
According to sources including the Financial Times, Reuters, and RegLayer, at least 47 major AI data center projects in the US have been delayed, withdrawn, or blocked since 2026; RegLayer statistics indicate that up to 50 projects are obstructed. AI-related debt issuance has fallen by approximately 80% from its June peak, CoreWeave bond yields have risen to about 15%, and its stock price has experienced severe volatility. S&P Global estimates that capital expenditures related to AI for Amazon, Microsoft, Alphabet, Oracle, Meta, and SpaceX will exceed $7 trillion between 2025 and 2030, with free cash flows potentially remaining negative until 2029. The OpenAI 'Stargate' plan, the Anthropic and Fluidstack collaboration, and Nscale's West Virginia project have all encountered delays or financing difficulties.
This project released an open-source Python kit for electrocardiogram (ECG) signal processing, including zero-phase Butterworth bandpass filtering (0.5–45 Hz), 50/60 Hz power line notch filtering, adaptive R-wave detection, and calculation of time-domain HRV metrics such as SDNN/RMSSD; it provides a pre-configured FastAPI interface (POST /v1/process-ecg), supports one-click Docker deployment and local development, and comes with a pytest test suite and MIT license.
defiabell developed a CLI tool named "Same Table" that utilizes the open MCP interface of McDonald's China to schedule two identical orders for delivery at the same time across different cities. The tool has no website or server and runs locally; manual price verification is required before payment. After generating an invitation card, it can be forwarded to the other party, who can receive the meal without needing an account or performing any operations. The code repository is https://github.com/Defiabell/mcd-same-table.
UnitCtrl, developed by xxzhaoming, supports unified viewing and operation of multiple Android devices at home from macOS. It supports Wi-Fi or USB connections and provides real-time screen display, mouse and keyboard control, audio forwarding, app management, file transfer, screenshot and screen recording, and device status monitoring; the Android version supports mutual control between phones. All communication is limited to the local network, requiring no account, bypassing the cloud, with no background services and no ads. The Mac version is released for free on GitHub, while the Android version is in closed testing on Google Play. The project code and interface were completed through collaboration between the author and Claude.
Mitsubishi UFJ Financial Group (MUFG), Sumitomo Mitsui Banking Corporation (SMBC), and Mizuho Bank completed a proof-of-concept test for yen stablecoin cross-border payments from July 17 to 23, 2026. The test adopted the specific trust beneficiary rights structure under the Payment Services Act, with Mitsubishi UFJ Trust and Banking acting as trustee and TISI providing blockchain and wallet system support. The testing focused on fund transfers between domestic and overseas locations of Mitsubishi Corporation, verifying the entire process of issuance, transfer, and redemption. The project aims to launch actual transactions within fiscal year 2026 (by no later than the end of March 2027), but the official date, issuance scale, and fee arrangements have not yet been announced.
KiviDB is a multi-threaded in-memory database written in Rust, fully compatible with the Redis protocol (RESP2/RESP3), supporting native full-text search and vector search (FLAT/HNSW). On a 32-core AWS Graviton3 instance, it achieves 21,200,000 pipelined GETs per second, with throughput approximately 9 times that of Redis 8 and 2.3 times that of Dragonfly; p99 latency is less than half that of Dragonfly. It listens on port 6380 by default and offers Docker images, precompiled binary packages, and various Cargo build variants (including Lua and TLS support). The project is open-source under the AGPL-3.0 license, with the GitHub repository located at github.com/kividbio/kividb.
0sql is an open-source (MIT) Rust tool that deterministically compiles business language query specifications (e.g., 'country, revenue, country = Canada') into standard SQL without connecting to data warehouses or executing queries. It supports 13 dialects including Postgres, Snowflake, and BigQuery. It offers a CLI (zsql), an HTTP server, and an MCP-compatible interface. Models are defined in YAML within Git, featuring security policies, branch management, and testing capabilities. The project source code is hosted on GitHub.