Daily aggregation of provider blogs, curated articles, trending models, and status incidents.
Gilbert + Tobin rethinks the business of law with AI
A conversation with Sam Nickless, Partner and Chief Executive Officer at Gilbert + Tobin.
STGformer: Efficient Spatiotemporal Graph Transformer for Traffic Forecasting
This arXiv preprint, submitted in October 2024 by Hongjun Wang and six co-authors, proposes STGformer, a spatiotemporal graph transformer for traffic forecasting. The authors argue that graph neural networks and transformer models are effective for traffic prediction but are often too computationally expensive for large real-world road networks. Their architecture combines elements of both approaches to model global and local traffic patterns. A central design claim is that the STG attention block captures high-order spatiotemporal interactions within a single layer, instead of requiring stacked attention layers. According to the paper, this reduces computational cost. The authors report a 100x inference speedup and a 99.8% reduction in GPU memory usage compared with STAEformer during batch inference on a California road graph containing 8,600 sensors. The paper evaluates STGformer on the LargeST benchmark and reports better results than transformer-based baselines including PDFormer and STAEformer. The authors position the method as a foundation for future spatiotemporal modeling tasks. These performance figures are the authors' own reported results. The work appears in cs.LG with cross-listings in cs.AI and cs.DB. Relevance to developers is mainly for those building forecasting systems on large sensor graphs.
Scaling Retrieval-Based Language Models with a Trillion-Token Datastore
Researchers led by Rulin Shao and seven co-authors (arXiv:2407.12854, submitted 9 July 2024) study how the size of the external datastore available to a retrieval-based language model at inference time affects performance, treating datastore size as a third scaling axis alongside training data and parameter count. According to the authors, performance on language modeling and several downstream tasks keeps improving as the datastore grows, without clear saturation, and a smaller model paired with a large datastore can beat a larger model that relies only on its parameters for knowledge-intensive tasks. To test this, the team built MassiveDS, a 1.4 trillion-token datastore that they describe as the largest and most diverse open-sourced datastore for retrieval-based LMs to date. They also designed a pipeline intended to make datastore scaling experiments feasible at lower compute cost. Their compute-optimal scaling curves vary datastore, model, and pretraining data sizes together, and they report that a larger datastore can meaningfully improve results at a fixed training compute budget. The paper also examines how retriever quality, datastore quality filtering, and other design choices shift these scaling trends. The authors conclude that datastore size should be treated as part of the efficiency and performance trade-off when designing LMs. They released the datastore and code publicly to support follow-on work.
Segment Anything without Supervision
Researchers XuDong Wang, Jingfeng Yang, and Trevor Darrell submitted a paper to arXiv on June 28, 2024 describing UnSAM, an unsupervised variant of the Segment Anything Model (SAM) that removes the need for human-annotated training masks. The method first uses top-down clustering to split unlabeled images into instance- and semantic-level regions, then applies bottom-up clustering within each region to merge pixels into a hierarchy of multi-granularity masks. These generated masks serve as the supervision signal for training the model. According to the paper, UnSAM is evaluated on seven popular datasets and reaches results competitive with supervised SAM, while outperforming prior unsupervised segmentation methods by 11% in average recall (AR). The authors also test a lightly semi-supervised setup in which pseudo-masks are combined with SA-1B ground-truth masks and training uses only 1% of SA-1B; they report that this setup can segment entities that supervised SAM misses, with gains of over 6.7% in AR and 3.9% in average precision (AP) on SA-1B. For developers, the relevance is mainly in the labeling cost: the paper argues that the expensive annotation step required by SAM-style training can be partly replaced by automatically discovered structure, and that existing supervised SAM models can benefit from these pseudo-labels. Code is referenced in the paper's comments field. The source is a primary research preprint, so reported figures reflect the authors' own evaluations.
A Theoretical Understanding of Self-Correction through In-context Alignment
This arXiv paper, by Yifei Wang and four co-authors and accepted at NeurIPS 2024, studies why large language models can sometimes improve their own answers through self-correction, meaning they re-examine and revise a previous response. The authors analyze the phenomenon theoretically from an in-context learning perspective, using a simplified setup resembling an alignment task. Their central claim is that when a model's self-examinations are reasonably accurate and serve as reward signals, the model can refine its responses in-context. Unlike earlier theoretical treatments that relied on heavily simplified linear transformers, the paper's construction is meant to capture features of realistic transformers. It attributes specific roles to softmax attention, multi-head attention, and the MLP block in enabling self-correction. The authors report extensive validation on synthetic datasets. The paper also describes applied uses of self-correction. One example is defending against LLM jailbreaks, where the authors report that a single self-correction step makes a large difference. The work is presented as groundwork for future research on understanding, using, and improving self-correction in foundation models. The paper does not report results on named production models.
openai/whisper-large-v3-turbo · Hugging Face
OpenAI's whisper-large-v3-turbo is a model card on Hugging Face describing a finetuned, pruned variant of Whisper large-v3. According to the card, the decoder was reduced from 32 layers to 4 while the rest of the model is unchanged, making it substantially faster at the cost of a minor quality degradation. The model has roughly 809 million parameters, compared with about 1550 million for large-v3, and is multilingual, unlike the English-only smaller checkpoints. The card is partly written by the Hugging Face team and partly carried over from OpenAI's original card. The card lists evaluation results from the Open ASR Leaderboard, including a mean word error rate (WER) of 7.83, a real-time factor (RTFx) of 200.19, and per-dataset WERs such as 2.10 on LibriSpeech clean and 4.24 on LibriSpeech other. It also documents usage through the Transformers pipeline API for transcribing audio of arbitrary length, including batching, chunked long-form decoding, and word- or sentence-level timestamps. For developers, the card covers speed and memory options: torch.compile with a static cache (reported by the card as up to 4.5x speed-ups), Flash Attention 2, and PyTorch SDPA, which is enabled by default on PyTorch 2.1.1 and later. It notes that the chunked long-form algorithm is faster for single long files while the sequential algorithm is slightly more accurate for batches. The card also states usage limits, advising against high-risk decision-making contexts and transcription of recordings made without consent.
Protegrity/protegrity-pii · Hugging Face
Protegrity has published Protegrity-PII, a token-classification model on Hugging Face for detecting personally identifiable information in text. It labels spans across 69 entity types, including names, contact details, financial and government IDs, and health data, and supports English, Spanish, Portuguese, German, French, Italian and Dutch. The model is built on XLM-RoBERTa base, is listed as version 1.0.0 released 2026-10-27, and is accompanied by a technical report on arXiv (2610.30518). According to the model card, training used only synthetic text generated by Protegrity's LLM-based pipeline. The card reports evaluation on PII-SCOPE, an 8,030-example synthetic set spanning several business domains, plus four public PII benchmarks, using exact span matching. On PII-SCOPE the typed F1 is 0.906 and untyped F1 is 0.909. Typed F1 on the public sets is 0.674 (AI4Privacy), 0.737 (Nemotron-PII), 0.484 (Gretel) and 0.467 (SPY), with the card noting that some lower precision reflects unlabeled real PII in those benchmarks. The card lists limitations: recall is highest for English and Spanish and lower for the other European languages, Portuguese was not evaluated, free-form entities like names and passwords are harder to detect, and quasi-identifiers are not flagged. The repository ships weights, tokenizer and an inference script with overlapping 512-token windows; training code is not released. The license is a custom 'other' license limited to research and non-commercial use, and the model is gated.
google/frames-benchmark · Datasets at Hugging Face
This is the Hugging Face dataset page for google/frames-benchmark, a set of 824 multi-hop questions designed to evaluate Retrieval-Augmented Generation (RAG) systems on factuality, retrieval accuracy, and reasoning. Per the page, each question requires information drawn from 2 to 15 Wikipedia articles. The dataset is distributed under an Apache-2.0 license in CSV format, with an English-language, text-only modality, and the page lists a single test split of 824 rows. The dataset is accompanied by an arXiv paper (2409.12941), titled in the page metadata as "Fact, Fetch, and Reason: A Unified Evaluation of Retrieval-Augmented Generation." The page's tags indicate coverage of long-context, iterative-search, and question-answering tasks. The sample rows shown on the page include answers and lists of the Wikipedia links consulted, along with reasoning-type labels such as numerical reasoning, tabular reasoning, temporal reasoning, and multiple constraints. For developers evaluating RAG pipelines, the value lies in the multi-hop structure: questions cannot be answered from a single source, so systems must retrieve and combine several documents. The page reports 32,041 downloads in the recent window and 228,726 all-time, with 276 likes. The dataset page does not itself report benchmark scores for specific models; those results would be found in the linked paper.
ATH-MaaS/Ovis1.6-Gemma2-9B · Hugging Face
This is a Hugging Face model card for Ovis1.6-Gemma2-9B, a multimodal large language model that takes image and text input and produces text output. The card describes it as built on the Ovis1.5 architecture, with improvements to high-resolution image processing, training on a larger and more varied dataset, and a DPO training stage applied after instruction tuning. The model pairs a SigLIP-400M vision encoder with a Gemma2-9B-It language model and is released under the Apache 2.0 license. The card points to a sibling 27B variant (Ovis1.6-Gemma2-27B) and a 3B variant built on Llama-3.2-3B-Instruct. The card's main performance claim is that, at roughly 10 billion parameters, the model leads the OpenCompass benchmark among open-source multimodal models under 30B parameters. The page does not list per-task scores in the text provided. The model is stored in BF16 safetensors, and the card reports roughly 3,400 downloads in the last month. For developers, the page includes a Python example using transformers 4.44.2 and trust_remote_code, with a preprocess_inputs step that converts an image and prompt into tokens and pixel values. It also includes a batched inference example that pads sequences and runs generation with greedy decoding and max_new_tokens of 1024. The card notes that the model was trained with compliance-checking algorithms but that copyright and content issues cannot be fully ruled out.
google/gemma-2-2b-jpn-it · Hugging Face
Google's Hugging Face model card describes gemma-2-2b-jpn-it, a 2-billion-parameter (about 2.6B weights in BF16) text-generation model derived from the Gemma 2 2B instruction-tuned checkpoint and further fine-tuned on Japanese text. The card states the model is intended to match Gemma 2's English-only performance on Japanese queries. Access is gated: users must accept Google's Gemma usage license on Hugging Face before downloading files. According to the card's own evaluation, which used an LLM-as-a-judge setup against GPT-3.5 on a set of Japanese prompts, the model scored 0.03 ± 0.04 on a preference scale where 0 means roughly equal, compared with -0.25 ± 0.05 for the base Gemma-2-IT model. Language correctness, meaning answering in Japanese for Japanese prompts, rose from 86.47% to 98.24%. The card reports training on roughly 8 trillion tokens using TPUv5p hardware with JAX and ML Pathways, and lists CSAM and sensitive-data filtering among its preprocessing steps. For developers, the card provides Transformers code using the pipeline API and AutoModelForCausalLM, with example output and guidance on bfloat16 versus float32 loading. It notes the model is for text generation tasks such as summarization, chat, and translation, and lists known limitations including factual errors, weaker handling of nuance, and possible inherited biases. Its usage section points to Google's Responsible Generative AI Toolkit and Gemma Prohibited Use Policy.
Zyphra/Zamba2-2.7B-instruct · Hugging Face
Zyphra has released Zamba2-2.7B-Instruct, an instruction-tuned variant of its Zamba2-2.7B base model, on Hugging Face under an Apache-2.0 license. The model was produced by supervised fine-tuning the base model on the ultrachat_200k and Infinity-Instruct datasets, followed by direct preference optimization (DPO) on ultrafeedback_binarized, orca_dpo_pairs, and OpenHermesPreferences. The model card states the release has about 2.66 billion parameters stored in BF16. The architecture is a hybrid of Mamba2 state-space layers interleaved with shared attention blocks, with the original model embeddings concatenated into the attention input and LoRA projections applied to the shared MLP. The card notes this is a temporary Hugging Face transformers implementation that may not yet work with every tool, and points to a standalone PyTorch implementation as an alternative. The card reports MT-Bench and IFEval scores: Zamba2-2.7B-Instruct scores 72.40 aggregate, 48.02 MT-Bench, and 42.20 IFEval in its comparison table, versus Mistral-7B-Instruct, Gemma2-2B-Instruct, H2O-Danube-4B-Chat, and StableLM-Zephyr-3B. The authors also claim lower time-to-first-token and smaller memory footprint than comparable transformer models, though the card's charts are not reproduced here. For developers, the notable points are the small parameter count, the Mamba2 kernel requirements (mamba-ssm v2.1.0 and causal-conv1d), and the stated slowdown when optimized kernels are not used.
Zyphra/Zamba2-1.2B-instruct · Hugging Face
Zyphra has released Zamba2-1.2B-instruct, an instruction-tuned chat variant of its 1.2B-parameter Zamba2-1.2B base model, published on Hugging Face under an Apache-2.0 license. The instruct model was produced in two stages: supervised fine-tuning of the base model on the ultrachat_200k and Infinity-Instruct datasets, followed by DPO preference tuning on ultrafeedback_binarized, orca_dpo_pairs, and OpenHermesPreferences. The model card lists the checkpoint as a 1.2B-parameter model stored in F32 and BF16 safetensors. The architecture is a hybrid of Mamba2 state-space layers and transformer attention blocks. Attention weights are shared across positions to limit parameter count, the original embeddings are concatenated into the attention input, and LoRA projections give each shared block some per-position specialization. Zyphra notes the Hugging Face implementation is temporary and may not yet work with every tool built for Hugging Face models; a standalone PyTorch implementation is also referenced. According to the card, the model scores 59.53 on aggregate and 41.45 on IFEval, against Gemma2-2B-Instruct at 51.69 and 42.20 respectively, and several other ~1.5–2.7B chat models. The card also claims lower time-to-first-token and smaller memory overhead than comparable transformer models, though the figures are presented as charts not reproduced in the text. For practical use, the card recommends installing transformers from source plus mamba-ssm v2.1.0 and causal-conv1d for optimized kernels, warning that the unoptimized path has significantly higher latency and memory use.
GitHub - datalab-to/surya: OCR, layout analysis, reading order, table recognition in 90+ languages · GitHub
Datalab's Surya is a 689M-parameter document OCR model that also performs layout analysis, reading-order detection, and table recognition (output as HTML with colspan/rowspan). The repository reports a score of 85.9 on OmniParseBench, which it describes as the highest among open models under 1B parameters it has scored, and 83.7% on olmOCR-bench. It also reports 74.9% accuracy on an internal 91-language benchmark. Surya 2.5 runs layout, OCR, and table recognition through a single vision-language model served via vllm (NVIDIA GPU) or llama.cpp (CPU/Apple Silicon); a separate small torch model handles text-line detection. The README gives throughput figures of 4.5 pages/s on one H100 at 192 DPI (1.7x Surya 2 with multi-token prediction) and about 0.19 pages/s on an M4 Max via llama.cpp/Metal. Median per-page latency on H100 at 192 DPI is 6.8 s for Surya 2.5 versus 12.0 s for Surya 2. Surya 2.5 writes roughly a third as many output tokens per page as Surya 2. Known limitations include no row/column geometry for tables in Surya 2.5 (a Surya 2 checkpoint is needed), and the project positions itself for document OCR rather than photos or natural scenes. Licensing is split: code is Apache 2.0, while model weights use a modified AI Pubs Open Rail-M license that is free for research, personal use, and startups under $5M in funding or revenue, with commercial licensing otherwise. The repository also lists CLI commands, a Python API, and environment-variable configuration for choosing inference backends.
IFM/TxT360 · Datasets at Hugging Face
The page is the Hugging Face dataset card for TxT360, a pre-training dataset for large language models published under the IFM/LLM360 namespace. The card's changelog reports a v1.1 release that added four data sources: TxT360_BestOfWeb, TxT360_QA, europarl-aligned, and wikipedia_extended. According to the card, TxT360_BestOfWeb is a filtered subset of TxT360 produced with the ProX document-filtering model, which the card describes as similar to the FineWeb-Edu classifier but also assigning an additional format-related label. The dataset is tagged for English text-generation and licensed under ODC-BY. The scraped page also includes a large body of unrelated mathematical text on stochastic flows and Schatten-class operator equations, which does not describe the dataset itself. That material is not summarized here and should not be read as the card's content. For developers, the relevant facts are the dataset's English-language scope, its license, the v1.1 source additions, and the use of ProX filtering for the BestOfWeb subset. The card as captured reports download counts (about 35.8k recent and about 3.45M all-time) and 285 likes at the time of scraping.
Introducing Kimi K3.1 Pro on Amazon Bedrock | Artificial Intelligence
AWS announced that Kimi K3.1 Pro Highspeed from Moonshot AI is now available on Amazon Bedrock. According to the post, it is a post-trained version of Kimi K3 with the same 3T-class open-weight architecture and a 1-million-token context window, trained with more post-training compute aimed at long-horizon agentic work, fewer factual errors, and professional tasks such as financial modeling, office documents, and CAD or game-engine scripting. Moonshot AI reports 2–3 times the output-token speed of Kimi K3.1 Pro. The model supports tool calling, structured output, configurable reasoning (low, high, or max, defaulting to max), image input, and streaming, and uses the same endpoints as Kimi K3 so existing code can switch by changing the model ID. AWS states that data stays within the AWS data boundary, is not shared with Moonshot AI or used for training, and that zero data retention is enabled for inference. The post walks through access via the Bedrock console Playground and via OpenAI-compatible Chat Completions and Responses APIs as well as the Converse and Invoke APIs, using cross-Region inference profiles (global, us, and in prefixes). It describes a Flex tier at half the Standard price with potentially longer latency, and explicit prompt caching where breakpoints require at least 256 stable tokens, cached content is retained for at least 30 minutes, and cache reads are discounted. Code examples cover Python OpenAI SDK, boto3, and an OpenCode coding-agent configuration. For developers, the practical points are the drop-in compatibility with existing Kimi K3 integrations, tiered cost options, and caching behavior for agent workloads with repeated context. Specific per-token prices are not stated in the article and are referred to the Bedrock pricing page.
LLM API Rate Limits: How to Fix OpenAI and Claude 429 Errors
OpenRouter's guide explains that HTTP 429 'Too Many Requests' errors from OpenAI, Anthropic, and OpenRouter mean a request count or token volume has exceeded the limit for the current time window. The error most often appears when an application first runs under production load or when a batch job spreads work across several parallel workers. The source frames the fix around understanding which limit was hit (requests or tokens) and pacing or distributing calls accordingly. Source: OpenRouter.
LLM Quantization: What fp8 and fp4 Do to Quality and Price
OpenRouter's explainer covers how open-weight models served through APIs may run at different numerical precisions, such as bf16 (16-bit), fp8 (8-bit), and fp4 (4-bit), with different providers offering different options. Lower precision reduces memory and hardware requirements, which can lower the price, but may reduce output quality. The piece describes quantization as storing model weights in fewer bits to trade accuracy for resource use. Developers comparing providers for the same model should check which precision each one serves. Source: OpenRouter.
MCP vs CLI vs Skills: What Each Costs in Context
OpenRouter's comparison examines the context cost of three ways to give agents tools: Model Context Protocol (MCP) servers, command-line tools (CLI), and skills. Each tool definition is sent to the model on every request and billed as input tokens, and an MCP server can add tens of thousands of tokens to the context window. The source positions CLI tools and skills as alternatives that can reduce this overhead. Source: OpenRouter.
Build a daily dependency watcher with LangGraph and Amazon Bedrock AgentCore web search
AWS describes the Web Search API in Amazon Bedrock AgentCore, which returns current, cited web results authenticated with existing AWS credentials. The post walks through building a daily dependency watcher as a LangGraph graph that researches package ecosystems in parallel, enforces its rules in code, and verifies each finding against its cited sources. This gives developers a reference pattern for combining LangGraph orchestration with AgentCore's search capability. Source: AWS Machine Learning Blog.
Meta FAIR Releases e-means: An Algorithm Beating k-means in 10x Fewer Epochs
Meta FAIR's e-means is described as a replacement for mini-batch k-means, shipped in FAISS 1.16.0. According to the source, it reaches the loss that Lloyd's algorithm attains after 25 epochs in roughly 1.4 to 2.6 epochs, reports 10 to 45 times GPU speedups, and requires no additional hyperparameters. The source is a short MarkTechPost summary, so independent benchmark details are not included here. Source: MarkTechPost.
The agent evaluation playbook: Built-in, DeepEval, and AutoEval on Amazon Bedrock AgentCore Evaluations
AWS presents a guide to evaluating AI agents on Amazon Bedrock AgentCore Evaluations, covering task completion, topic adherence, data leakage, and step-skipping. It describes three evaluator types: built-in, DeepEval, and AutoEval, and explains how to match each behavior to the appropriate evaluator and evaluation level. The post is aimed at teams deploying agents who need to measure more than a single pass/fail outcome. Source: AWS Machine Learning Blog.
Ring cameras go mobile as Amazon debuts AI-powered security drone and robot
Ars Technica reports that Amazon's mobile Ring Robotics devices, including a security drone and a robot, are expected to debut in late 2026 and early 2027. The source provides only the timing and product category, with no pricing, specifications, or availability details. Source: Ars Technica.
Elevated errors on Fable 52 and Mythos 52
Anthropic's status page logged an incident on November 10 at 21:47 UTC in which 'Fable 5.2' and 'Mythos 5.2' returned overloaded errors for four minutes. The incident was marked resolved and mitigated. No root cause or affected request share is given in the status entry. Source: Anthropic Status.
Grok Build coding agent leaks oodles of confidential info, some belonging to SpaceXSI
Ars Technica reports that a researcher found about 35,000 Grok Build apps publicly accessible without authentication, exposing confidential information, some of which reportedly belongs to SpaceXSI. The source describes the exposure as affecting apps openly reachable by anyone who searched for them. Further details on the data types and remediation are not stated in the summary. Source: Ars Technica.
OpenAI acquires Odessia, travel startup created by Sonder founder
TechCrunch reports that OpenAI acquired Odessia, a travel startup founded by Sonder co-founder Francis Davidson. OpenAI has not stated its plans for travel. The source gives no deal terms or product details. Source: TechCrunch.
Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
Microsoft Research describes 'Lazarus,' a backdoor technique that reprograms dormant neurons in an AI model so that a hidden trigger causes an otherwise normal agent to run malicious commands. The source highlights elevated risk for agents with tool or terminal access, where a triggered model could download and execute attacker code. The article is a research disclosure rather than a product release. Source: Microsoft Research.
OpenAI, Anthropic face new oversight as New York’s AI safety law takes effect
TechCrunch reports that New York's AI safety law is entering enforcement, requiring large developers including OpenAI and Anthropic to register, disclose their safety practices, and report critical incidents, with fines for noncompliance. The source frames this as new regulatory oversight for frontier developers. Specific thresholds and deadlines are not detailed in the summary. Source: TechCrunch.
What happens to mathematics when AI becomes an oracle?
Mathematician Steven Strogatz discusses his new book, Big Math: The Hidden Codes That Run Our World.
OpenAI turns ChatGPT into a business directory with AI agents that handle customer questions
The Decoder reports that OpenAI is turning ChatGPT into a business directory through a new Business Hub, where companies can claim profiles, add product data, and view user interaction analytics. A feature called Business Agents lets eligible US e-commerce providers answer customer questions directly within ChatGPT. The source covers these as platform features for businesses, with eligibility limited to certain US e-commerce providers. Source: The Decoder.
llm-gemini 0.36
Simon Willison's llm-gemini 0.36 release notes that Gemini 4 Carbon, initially limited to a subset of cyber defenders, is now available broadly. Launch pricing is $2 per million input tokens and $10 per million output tokens, scheduled to rise to $4 and $20 on December 1. The post also includes pelican-image tests at low, medium, and high thinking effort, with medium as the default. Source: Simon Willison's Weblog.
All 7 Anthropic founders will each cash out about $250M in IPO
TechCrunch reports that each of Anthropic's seven co-founders plans to sell the same number of shares in its IPO, which would yield about $250 million each at the midpoint of the price range. The source notes the group would retain control of the company after the sale. The summary contains no further financial terms. Source: TechCrunch.
Anthropic turns Claude into a full workspace with Plugins, spreadsheets, whiteboards, and soon Meeting notes
The Decoder reports that Anthropic is adding Sheets, Whiteboards, and Interactive Connectors to Claude, allowing partners such as Adobe and Atlassian to embed their apps inside Claude. Anthropic says Claude Meetings, which can join calls and take notes, is coming soon. The source describes this as a move toward Claude as a central workspace integrating third-party tools. Source: The Decoder.
Oura to acquire AI-native primary care startup Counsel Health for $625M
TechCrunch reports that Oura will acquire Counsel Health, an AI-native primary care startup, for $625 million. The deal follows Oura's indefinite postponement of its up to $2.2 billion IPO, which it attributed to uncertainty in the IPO market. The source gives no integration or product details. Source: TechCrunch.
PSI is trying to make a business out of AI-driven scientific discovery
TechCrunch reports that PSI, a company focused on AI-driven scientific discovery, used an unreleased Anthropic model to address a quantum computing question that scientists have pursued for nearly two decades. The source does not name the specific question or publish the result in detail. The summary is primarily about the company's business approach. Source: TechCrunch.
Gemini Intelligence: New features, on more devices
A man speaks into a phone. Text on the screen reads 'Gemini Intelligence'
OpenAI rolls out new business profiles, making ChatGPT just a little bit more like Google
TechCrunch reports that OpenAI is rolling out business profiles in ChatGPT, along with features intended to help businesses be discovered through the product. The source frames this as an effort to make ChatGPT a consumer discovery engine. Specific feature details are not given in the summary. Source: TechCrunch.
Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot
The Decoder reports that Nvidia's GR00T 2 is designed to teach robots new tasks by transferring human hand movements from video to robotic arms, rather than relying on teleoperation. The model was trained on about 400,000 hours of video. Reported task success rates range from 10 to 80 percent depending on the task. Source: The Decoder.
Gemini 4 Carbon arrives just in time to feel like yesterday's news in the fastest model cycle yet
The Decoder reports that Google is making Gemini 4 Carbon generally available after a two-week testing phase. The source says it ranks first in several benchmarks, in some cases ahead of Anthropic's Fable 5.2, at less than half the per-task cost. Google is also previewing a '/swarm' feature for coordinated agent teams that divide complex tasks in real time. Source: The Decoder.
Autonomous treasury agents: How Trovata built an AI-native solution on Amazon Bedrock and AgentCore
AWS describes how Trovata built autonomous treasury agents on Amazon Bedrock and Amazon Bedrock AgentCore, using an RBAC-secured architecture for bank-branded deployments. The post presents the system as moving corporate treasury work from manual operations toward real-time, agent-driven decisions at enterprise scale. It is a customer case study with no benchmark or pricing figures in the summary. Source: AWS Machine Learning Blog.
How to Use AI for Enterprise BI
A short overview piece on applying AI to enterprise business intelligence. The source's main points are that AI can make BI tools more accessible to non-specialists, automate recurring reporting, surface insights faster, and support better decisions when paired with appropriate governance. The source gives no specific products, benchmarks, or figures, so this is a general, high-level treatment rather than a technical release.
Microsoft lobbies against potential ban on Chinese optical networking gear for AI data centers as FCC mulls restricting imports
Microsoft is reportedly lobbying against a possible US Federal Communications Commission restriction on imports of Chinese optical networking equipment, specifically optical transceivers that convert electrical signals to optical ones and are used in AI data centers. The Information Technology Industry Council (ITI) is also reported to be opposing the proposed ban. The source frames this as a policy debate over supply-chain restrictions affecting data center hardware.
Anthropic launches Claude Meetings to turn workplace discussions into plans and tasks
Anthropic has launched Claude Meetings, a product that turns workplace discussions into plans and tasks. According to the source, during a meeting Claude can draft the plan under discussion, lay out a timeline on a shared whiteboard, answer questions posted in the meeting chat, and pull information from connected work applications, for example prior-year figures. The source does not provide pricing, availability regions, or model names in the summary.
AMD VP and chief Zen architect rebuke claims that Nvidia and Arm made CPUs for agentic AI — ‘We didn’t magically know four years ago that agentic was going to be a thing’
AMD's Zen architect and a senior AMD executive publicly disputed claims that Nvidia and Arm designed CPUs specifically for agentic AI workloads. The response coincides with AMD's launch of Zen 6-based EPYC Venice server CPUs. The executive's position, as reported, is that the industry could not have anticipated the rise of agentic AI several years ago. This is a commentary-driven item centered on the positioning debate rather than on technical specifications.
Google DeepMind's Gemini Robotics 2.5 scores 96 percent sorting metal parts in a Mercedes-Benz assembly pilot
Google DeepMind introduced Gemini Robotics 2.5, a control model for robots, and tested it with robotics company Apptronik on active Mercedes-Benz production lines in Hungary and Germany. The humanoid robot Apollo 2 was used to sort metal parts, with a reported success rate of 96 percent. The source presents this as a real-world factory pilot rather than a lab benchmark.
Nvidia CEO Jensen Huang’s affair with alleged Chinese spy: a timeline
A timeline piece reporting an alleged personal relationship between Nvidia CEO Jensen Huang and a person alleged to be a Chinese spy. The source notes possible national security ramifications given U.S.-China competition over advanced AI infrastructure. This is a news and reporting item about personnel and security concerns, not a product or technical release.
AMD Zen 6 EPYC ‘Venice’ deep dive and benchmarking
AMD has disclosed details of its Zen 6 microarchitecture, and this piece covers differences between Zen 5 and Zen 6 along with first benchmark results for AMD's Venice server CPUs (EPYC). The source is a technical deep dive focused on architectural comparison and early performance data for the new EPYC processors.
How TCS built an agentic test orchestrator with Amazon Bedrock AgentCore
Tata Consultancy Services built an agentic test orchestrator on Amazon Bedrock AgentCore for its Banking, Financial Services, and Insurance clients. The system converts requirements into executable, governed tests using specialized agents, per-tenant isolation, and workload-aware execution. An internal benchmark reported an 88% reduction in testing effort. The post is an AWS customer case study.
Amazon Bedrock tokenomics data lake
AWS describes a reference architecture for centralizing Amazon Bedrock invocation logs from multiple linked accounts into one queryable data lake. The approach captures prompts, token counts, model selection, and per-call cost, addressing the gap that consolidated billing shows spend but not per-invocation detail. This is a how-to guide for cost and usage observability rather than a product launch.
Alice Huang reportedly instrumental in securing over $7.5 billion of Nvidia AI GPU imports during three-year relationship with Jensen Huang
Reporting alleges that Alice Huang is connected to a network of Chinese companies that reportedly used her relationship with Nvidia CEO Jensen Huang to obtain Nvidia GPUs worth over $7.5 billion across a three-year period. The source is investigative reporting on alleged export and access issues rather than a technical or product announcement.
Elevated error rates with some U.S. business Mastercards
OpenAI's status page reports elevated payment declines on some U.S. business Mastercards, identified by BIN prefixes including 544015, 546400, 543502, 543799, 512998, and some 513989 cards. For API customers, declines can block automatic recharges and one-time credit purchases, which can lead to API request failures once prepaid balances are exhausted. OpenAI recommends adding a card from a different issuer before the next payment, and says that if a credit purchase already failed, a new card can be used to retry. The source excerpt is truncated and does not indicate a resolution time.
NVIDIA Expands Open Source Commitment With OpenShell Donation and New AI Infrastructure
NVIDIA is expanding its open-source commitment, including a donation of OpenShell and new AI infrastructure work, announced at KubeCon + CloudNativeCon. The source positions Kubernetes as the foundation for orchestrating AI workloads across large numbers of AI factories and says NVIDIA is deepening its participation in the cloud-native community. The excerpt is truncated and lacks specifics on OpenShell's functionality.
Running AV Reasoning Models Within Onboard Compute and Latency Limits
An NVIDIA technical post on running autonomous vehicle reasoning models within the compute and latency budget of an onboard vehicle system. The source notes that larger models offer capability advantages but must fit onboard constraints. The excerpt is truncated, so specific model sizes, hardware, and latency figures are not available in the provided text.
Google broadens its AI hardware portfolio with two new specialized chips
Google announced two new specialized tensor processing units (TPUs) to broaden its AI hardware lineup, targeting use cases such as confidential model training and workloads requiring greater memory capacity. The source does not provide chip names, specifications, or pricing in the summary.
Zyphra Releases ZONOS3-Mini: Open-Source TTS With 100M Active Parameters
Zyphra released ZONOS3-Mini, an open-source text-to-speech model under the Apache 2.0 license with 800M total parameters and 100M active parameters. The model supports voice cloning, word-level control over delivery, and real-time operation in the browser.
Mitra-v2: Scaling synthetic pretraining for tabular foundation models
Amazon describes Mitra-v2, a tabular foundation model trained with an expanded mix of synthetic pretraining data. The source reports that a single model can adapt to new tables without task-specific retraining and leads the single-model field on the TabArena benchmark.
Secretlab debuts monitor light bar add-on to reduce glare and eye strain
Secretlab has released a light bar accessory that mounts to monitors, intended to reduce glare and eye strain. The source is a consumer hardware product note with no AI-specific content.
Meta’s AI agent Muse lands on Apple Watch and Android Wear OS
Meta's AI assistant, Muse, is now available on Apple Watch and Wear OS devices. Users can interact via voice calls, text messages, or dictation directly on the watch.
DeepMind reveals Gemini Robotics 2.5 and open safety benchmark "ASIMOV-jailbreak"
DeepMind introduced Gemini Robotics 2.5 along with an open safety benchmark called ASIMOV-jailbreak. The source reports the new model is up to twice as successful as its predecessor on some tasks. Specific task categories and benchmark details are not given in the summary.
Integrating LLM Translation into Your Scikit-Learn Pipelines with Scikit-LLM
A tutorial showing how to build a two-stage scikit-learn pipeline for multilingual text classification. The first stage uses a pre-trained language model via Scikit-LLM for translation, and the second uses another for zero-shot classification. The source is an instructional guide.
Agent Substrate: Advancing Kubernetes for Agents at Scale
NVIDIA describes Agent Substrate, an approach to making Kubernetes better suited to running AI agents at scale. The source notes that agents spend most of their time idle awaiting human input, tool results, or model responses, and that the one-Pod-per-agent model in Kubernetes was not designed for this pattern. The excerpt is truncated and does not detail the implementation.