Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
OpenAI detailed its response measures to the TanStack "Mini Shai-Hulud" npm supply chain attack, including system and signing certificate protection plans, and required macOS users to update the OpenAI app by June 12, 2026.
Google DeepMind released Co-Scientist, a multi-agent AI system built on Gemini designed to accelerate scientific hypothesis generation; the system operates through generation and debate.
The Ethereum Foundation, in collaboration with wallet developers and security firms, has launched the Clear Signing open standard to address blind signing issues. Key components include the ERC-7730 description format, a registry hosted by the Foundation, independent verification mechanisms, and supporting tools for both existing and new applications. Ledger initiated ERC-7730, with contributions from ZKnox, Sourcify, Cyfrin, WalletConnect, and others. Tool documentation is available at clearsigning.org.
Playwright v1.60.0 released, adding tracing.startHar()/stopHar() for HAR recording, locator.drop() for cross-browser drag-and-drop simulation, page.matchAriaSnapshot() assertions with boxes option, and test.abort() for immediate test termination; extended BrowserContext lifecycle event listeners, getByRole description option, toHaveCSS pseudo option, highlight style option, etc.; removed multiple long-deprecated APIs; supports Chromium 148, Firefox 150, WebKit 26.4.
The Ethereum Foundation released its May 2026 protocol cluster update, confirming that the Glamsterdam upgrade has entered the final hardening phase: establishing a 200 million gas limit target, stable operation of ePBS across multi-client devnets, formal finalization of EIP-8037 with complete repricing; meanwhile, the FOCIL prototype for Hegotá is available, with multi-client devnets as the next step. Will Corcoran, Kev Wedderburn, and Fredrik have taken over as protocol cluster coordinators, while former coordinators Barnabé Monnot, Tim Beiko, and Alex Stokes will depart or take academic leave.
Sakana AI and NVIDIA collaborated to publish the paper "Sparser, Faster, Lighter Transformer Language Models" at ICML 2026.
Figure demonstrated two Helix-02 humanoid robots collaboratively completing a bedroom organization task in under two minutes, including actions such as opening doors, hanging clothes, storing headphones, closing books, taking out trash, pushing chairs, and making beds; both operate on a single vision-language-action policy.
In its Q1 earnings call, Amazon disclosed that nearly 20% of users who interacted with brand-sponsored prompts in Rufus continued conversations about that brand. OpenAI launched ads in ChatGPT for select US advertisers in February, abandoned Instant Checkout in March, and this week opened its Ads Manager to all US advertisers while removing the $50,000 minimum spend requirement. Shopify released connector apps for ChatGPT and Claude this week, stating that 83% of merchants are already using ChatGPT.
OpenAI has launched a new real-time voice model in its API, supporting voice reasoning, translation, and transcription to enable more natural and intelligent voice experiences.
ChatGPT has introduced a Trusted Contacts feature, an optional security function that notifies user-designated trusted contacts when severe self-harm tendencies are detected.
Google DeepMind has released AlphaEvolve, a Gemini-based coding agent that has been used to optimize algorithm design, reducing error rates by 30% in genomics (DeepConsensus).
Hugging Face has introduced private English speech datasets with multiple accents and styles, provided by Appen and DataoceanAI, into the Open ASR Leaderboard.
SGLang v0.5.11 has been officially released, upgrading the default CUDA version to 13.0 and PyTorch to 2.11; Speculative Decoding V2 (including overlap scheduling) is now set as the default.
OpenAI has released the MRC (Multipath Reliable Connection) supercomputing network protocol, opened via OCP, to enhance the resilience and performance of large-scale AI training clusters.
OpenAI rebuilt the WebRTC stack to support large-scale, low-latency real-time voice AI and seamless turn-taking.
The Ethereum Foundation hosted the Soldøgn interoperability meeting in the Svalbard archipelago, where over 100 core contributors advanced the Glamsterdam network upgrade. Three core objectives were achieved: determining a lower bound of 200M for the gas limit, implementing a stable external builder solution for ePBS, and finalizing the repricing values for EIP-8037. Additionally, prototype progress was made on forward-looking Hegotá topics such as Block-Level Access Lists (BAL), FOCIL, and native account abstraction.
Google DeepMind announced the launch of the AI co-clinician research initiative, proposing a 'triadic care' model where AI collaboratively participates in patient care under the clinical authority of physicians.
The Tongyi team released Qwen-Scope, a set of sparse autoencoder (SAE) interpretability modules trained on the Qwen3 and Qwen3.5 series models. It opens sources 14 sets of weights, covering 7 dense and Mixture-of-Experts (MoE) models; it supports scenarios such as targeted control of inference results, data classification and synthesis with low data dependency, anomaly feature localization to assist training, and redundancy analysis of evaluation sets; training used 0.5B tokens of pre-training data, and it is now available for trial on Hugging Face and ModelScope platforms.
Standard Intelligence proposes constructing general agents via full-scale pretraining on computer-use videos in pixel space and releases its first foundational model, FDM-1; trained on a dataset of 11 million hours of computer operation videos, this model can perform tasks such as Blender modeling, autonomous driving simulation, and software state-space exploration.
Shopify has introduced Agentic Storefronts (sales channels for Shopify merchants) and the Agentic Plan (a lightweight integration solution for non-Shopify merchants), enabling products to be discovered and purchased through AI platforms such as ChatGPT, Microsoft Copilot, Google AI Mode, and Gemini apps.
Google Research released a preprint introducing Empirical Research Assistance (ERA), used to generate expert-level empirical software; subsequent applications include epidemiology (influenza/COVID-19/RSV hospitalization forecasting).
Zhipu's official research blog disclosed that the GLM-5 series models exhibited anomalies such as garbled text, repetition, and rare characters in high-concurrency, long-context Coding Agent scenarios. Investigation identified these issues as KV Cache race conditions under the PD separation architecture (BugFix#1) and missing HiCache loading timing sequences (BugFix#2).
The IBM Granite team released the Granite 4.1 series of dense decoder large language models (3B, 8B, 30B), based on approximately 15T tokens of five-stage pre-training, with the context window expanded to 512K tokens.
Figure announced that it has achieved the ramp-up of Figure 03 mass production at its own factory, BotQ, increasing the production rate from 1 unit per day to 1 unit per hour, representing a 24x throughput increase; over 350 units have been delivered, with a first-pass yield exceeding 80% for complete robots and 99.3% for battery lines; simultaneously disclosed new capabilities driven by an expanded robot fleet—whole-body control under perception conditions (S0), supporting complex terrain behaviors such as stair traversal via zero-shot sim-to-real transfer.