English text is being prepared.

Cursor Releases Agent SDK, Enabling Code-Defined Agents with One-Click Deployment

Cursor has released the Agent SDK, the framework it uses to build hosted agents such as Bugbot, Security Reviewer, and Rollouts. Developers can define agents in code and deploy them to Cursor infrastructure with a single command, agent-sdk deploy. Agents are organized as small software projects, with prompts written in Markdown and tools defined as TypeScript functions. They can respond to PR events from GitHub, GitLab, and Bitbucket, and include built-in evals and event replay. They can also be packaged to run as a single Node process on your own infrastructure.

Viggle Releases JST-2, a 3D-Native World Model

Viggle has released JST-2, building its entire model stack in 3D space and representing the world with Physics-Native Tokens that carry xyz coordinates, attribute vectors, and dynamics vectors, rather than implicitly inferring physical laws from video pixels. The architecture consists of three parts: the JST encoder, JST generator, and JST decoder. A single forward inference generates a complete 3D scene in about 1 second, at a cost below $0.01 (B200), and supports up to 64 input views as a spatial context window. Blind human preference tests published by Viggle show that character generation wins 58.4% against Meshy 7.1 and 73.2% against Tripo H3.1, while video-generated motion wins 89.1% against DeepMotion. JST-3 plans to extend the learned dynamics to object interaction.

Anthropic proposes GRAFT: localizing reward hacking to removable parameters rather than penalizing it

Anthropic and other institutions propose GRAFT (Gradient Routing Adapter Fine-Tuning), which, when reward hacking is detected, updates only a dedicated FORGET adapter and removes it after training, thereby eliminating these behaviors. The method uses monitoring signals only to decide which parameters to update, and does not set rewards based on them, so the model is not reinforced for exploiting weaknesses in the monitor. In Countdown-Code RLVR experiments with Qwen3-8B, GRAFT prevented reward hacking while still enabling the model to learn the target task; the experiments were limited to small models and simple monitors.

How Deepgram optimized large-scale speech AI with AWS Trainium

Deepgram partnered with AWS to port the Nova-3 speech-to-text model to Trainium2, running 8 model replicas on a single trn2.3xlarge instance, achieving an 11%–70% increase in audio processed per dollar compared with the previous solution in offline batch processing scenarios. The team used TorchNeuron to complete the PyTorch port and tuned it through NKILib fused kernels and hand-written NKI kernels, among which the FFN kernel reduced latency by 2.0x, depthwise conv1d sped up by 1.5x, and the conv2d kernel unlocked batch sizes above 32.

Amazon SageMaker JumpStart 2026 Progress Review

From January to September 2026, Amazon SageMaker JumpStart added more than 100 models from 30 providers, of which Alibaba (Qwen) contributed 29, NVIDIA 17, Mistral 7, Google 6 Gemma 4 variants, IBM 4, and Cohere 4. The new models cover multimodal, speech, image, and video generation, and introduce quantized and compact models such as GLM-5.2 FP8 and JetBrains Mellum2-12B to reduce inference costs.

English translation is being prepared.

Zihao Input Method Lands on iPhone: Built on the Rime Engine, Supports Importing Custom Schemes

The author has released Zihao, an iPhone input method built on the Rime (Zhongzhouyun) engine with Song Pinyin built in. Basic typing features are free and ad-free, and typing content is processed only on the phone. It supports importing Rime configuration folders, ZIP, YAML, or GitHub repositories with automatic updates, and dictionaries can be synced in Rime format to iCloud Drive and merged with Squirrel on Mac. The author says that in simulated tests, the first-try accuracy for sentences with 2 mistyped letters rose from 36% to 56%, and for 3 mistyped letters from 26% to 53%.

Town Completes $108 Million Series B Funding, Valued at $1 Billion

Town announced on October 19, 2026 that it had completed a $108 million Series B funding round, co-led by Forerunner Ventures and Menlo Ventures, at a valuation of $1 billion, with participation from existing investors including a16z and First Round Capital. Town's AI work assistant Townie can now collaborate directly with one another, asking each other questions, connecting context across teams, and handing off work, while still sharing information only within the scope authorized by each respective user. The company says that since July 2026, its paid customers have grown 4x and paid teams 9x, and it supports more than 60 integrations.

Fireworks Launches Free Weights Club, Free SFT Training for Muse Glimmer 30B

Fireworks AI launches the Free Weights Club: SFT training for Muse Glimmer 30B is free until November 30, with no credit card required; inference and evaluation calls are billed separately. The tuned Glimmer trained on Kalshi market data achieved 58.3% prediction accuracy on unseen topics, 11.2 percentage points higher than Jev's 47.1%; in a historical backtest, a hypothetical $1,000 grew to $1,071 after simulated fees, while both base Glimmer and Jev posted losses. In another synthetic benchmark, the System One serving stack achieved a median server-side processing speed 2-3 times faster than standard BF16 chat completions.

Alibaba Tongyi MAI-UI team open-sources general computer-use agent tech stack OpenGCUA

Alibaba Tongyi MAI-UI team open-sources the complete general computer-use agent tech stack OpenGCUA, including harness, cross-device SDK, native macOS app, evaluation trajectories, and reference agent implementation. Its core design is that code execution and decision-making shift from the harness to the model; the model takes screenshots on demand, writes Python programs with loops and conditional judgments to orchestrate GUI operations, API calls, and shell commands, and maintains its own todo list. In evaluations of 8 frontier models on OSWorld-v2, the average partial score rose from 45.3 to 64.9, and the average number of steps per task decreased by 53%; in real-machine macOS tests, under the same harness, Qwen3.8-Max's inference token consumption was 19.4 times that of GPT-6 Astra, completing 3 of 4 tasks.

Central China Normal University proposes CoReFuse-Med: purify first, then fuse, to address low-quality modalities dragging down multimodal MRI segmentation

A team from Central China Normal University and others proposed the multimodal MRI segmentation network CoReFuse-Med, which uses LACE+IMSH to remove resampling artifacts at shallow layers and MBCF to calibrate modality contributions at deep layers, solving the problem of low-quality modalities contaminating fusion. On the EPVS dataset, it achieves a DSC of 0.7326 and an HD95 of 7.35, outperforming MedNeXt's 0.7085 and 8.17; on BraTS, the DSC is 0.8528, with only 2.665M parameters. This method targets the resolution degradation after spatial alignment and does not cover scenarios with missing modalities or strong noise.

English translation is being prepared.

eul, unmaintained for 5 years, is fixed with Workbuddy and restores temperature and GPU monitoring on Apple M6

The open-source menu bar monitoring tool eul, unmaintained for 5 years, completed a low-level fix with Workbuddy and once again reads CPU temperature and GPU information on the Mac Mini 6 with the Apple M6 chip. The root cause is that SMCParamStruct changed from 80 bytes to 88 bytes in arm64 user mode, and all 135 temperature keys on the M6 are of the flt type rather than the sp78 type from the Intel era. After the fix, it dynamically enumerates 2,066 SMC keys, retains 126 valid sensors, and remains compatible with the logic for Intel models.

SpringAI Customer Service Security Compliance in Practice: Sensitive Word Filtering, PII Masking, and Conversation Auditing

Based on SpringBoot 3.3.x and SpringAI 1.0.x, this article presents a security and compliance solution for launching AI customer service in financial and government/enterprise scenarios: using the DFA algorithm for bidirectional sensitive word filtering, with a 100,000-word lexicon matched within 1ms; using regular expressions to mask phone numbers, ID card numbers, bank card numbers, and email addresses before calling Hunyuan; storing audit logs independently in Tencent Cloud CDB, allowing only INSERT, retaining them for 5 years, and keeping both original and masked records. The solution is designed with reference to Level 3 of the Multi-Level Protection Scheme and the Personal Information Protection Law.

How far can we push computation?

MIT Tech Review surveys how the curves that measure progress in computing and communication—Moore's law, Dennard scaling, Koomey's law, Kryder's law, and Keck's law—are slowing or breaking down, while the six major hyperscale cloud providers are expected to spend more than $700 billion on computing infrastructure this year, roughly twice the record set in 2025. The article notes that these scaling laws are merely observations of engineered systems, not fundamental laws of physics, and that the physical limits set by quantum mechanics, general relativity, and information theory remain far below the levels of today's data centers.

Jack Clark says in IMF speech that the singularity has already begun

Anthropic co-founder Jack Clark delivered the Per Jacobsson lecture on October 17 at the IMF meetings in Bangkok, saying that this year has brought convincing evidence that the singularity has already begun and that AI is autonomously producing results in frontier science. He cited Humanity's Last Exam's top score rising from 8% in January 2025 to 67.7% for Claude Opus 5.5, and FrontierMath rising from 2% to 94%, and cited TheoPhysEval, in which GPT-6-Astra's score rose to 98% (49/50) after experts corrected the answers. Citing Anthropic's internal modeling, he said that if models can autonomously complete a wide range of tasks and spread rapidly, U.S. GDP could reach $44 trillion by 2030, and he called on countries to expand computing power, electricity and scientific research robotics facilities, while maintaining strict human oversight in areas such as biological research.

How to run n8n AI workflows on NVIDIA DGX Spark without calling cloud APIs

n8n Blog published a tutorial on how to deploy a complete model stack locally on a single NVIDIA DGX Spark, enabling n8n workflows to perform transcription, OCR, document Q&A, and image generation without calling cloud APIs. The entire stack exposes an OpenAI-compatible endpoint through LiteLLM; the main model is the NVFP4 version of Qwen3.6-35B-A3B, with a single-user generation speed of 120 tokens/s and an aggregate of 359 tokens/s across eight users. The author's testing found that image generation reduces chat speed to about one-third, and recommends that small teams queue image tasks or schedule them outside working hours.

English translation is being prepared.

English translation is being prepared.

Indian Supreme Court proposes hash database and penalty framework to platforms for CSAM regulation

In a 137-page order, the Indian Supreme Court declined to ban children from using social media or impose age-based restrictions, holding that internet access is a fundamental right guaranteed under Articles 21 and 19(1)(a) of the Constitution, while also proposing a regulatory framework for child sexual abuse and exploitation material (CSEAM/CSAM). The framework calls for a national database recording hashes and URLs of confirmed content, which could be built on the Sahyog portal, along with supporting processes including manual verification, an independent appeals channel, confirmation of reports within 24 hours, takedown within two hours, and reporting to bodies such as the SJPU. The court also recommended that the government consider fines beyond intermediary liability and the confiscation of revenue earned from disseminating such content through paywalls; the framework has not yet been implemented, and the government must submit an affidavit before the next hearing on December 10 explaining its progress.