Daily aggregation of provider blogs, curated articles, trending models, and status incidents.
GitHub - mlfoundations/Gelato: 🍨 Gelato — From Data Curation to Reinforcement Learning: Building a Strong Grounding Model for Computer-Use Agents · GitHub
Researchers at ML Foundations released Gelato-30B-A3B, an open-source vision-language model designed for graphical user interface grounding tasks in computer-use agents. The model was trained using the GRPO reinforcement learning algorithm on Click-100k, a newly open-sourced dataset of 100,000 screen-and-action pairs assembled from eight public sources plus professional application tutorials. Gelato achieves 63.88% accuracy on ScreenSpot-Pro and 69.15%–74.65% on OS-World-G benchmarks, surpassing the prior specialized model GTA1-32B and larger vision models like Qwen3-VL-235B. When paired with GPT-5 in an agent framework, Gelato reached 58.71% automated success rate (61.85% with manual evaluation) on the OS-World benchmark, outperforming GTA1-32B's 56.97% (59.47% with human review). The release includes intermediate training checkpoints, trajectory logs, and dataset filtering methodology. The team also documented reproducibility challenges with the OS-World benchmark and provided corrected human evaluations to address task specification ambiguities.
mlfoundations/Gelato-30B-A3B · Hugging Face
ML Foundations released Gelato-30B-A3B, a 31-billion-parameter vision-language model fine-tuned for GUI grounding tasks (predicting screen element coordinates). The model is built on Qwen3-VL-30B-A3B-Instruct and trained on the open-sourced Click-100k dataset. Gelato achieves 63.88% accuracy on ScreenSpot-Pro, 69.15% on OS-World-G, and 74.65% on OS-World-G (Refined), outperforming the prior specialist model GTA1-32B and larger models including Qwen3-VL-235B-A22B-Instruct. The model uses mixture-of-experts routing with only 3 billion parameters active at inference time, reducing computational demand. The release includes working inference code demonstrating normalized click coordinate prediction and model cards on Hugging Face. Developers building computer-use agents or UI automation systems can use Gelato as a lighter alternative to larger VLMs for this specific task.
GitHub - stepfun-ai/Step-Audio-EditX: A powerful 3B-parameter, LLM-based Reinforcement Learning audio edit model excels at editing emotion, speaking style, and paralinguistics, and features robust zer
StepFun AI released Step-Audio-EditX, an open-source 3-billion-parameter audio editing model based on reinforcement learning and LLMs. The model supports zero-shot text-to-speech cloning for Mandarin, English, Sichuanese, and Cantonese, with iterative editing capabilities for emotion, speaking style, and paralinguistic features (breathing, laughter, sighs, etc.). The model requires 12 GB GPU memory on a single NVIDIA GPU and is available in standard and quantized (AWQ 4-bit) versions, with inference code and weights published on Hugging Face and ModelScope. The January 2026 release improved overall performance by over 4% and added new paralinguistic tags. Evaluation shows Step-Audio-EditX outperforms proprietary models like MiniMax-2.6-hd and Doubao in emotion and style editing. The repository includes inference demos, web interface, Docker support, training code (SFT, DPO, GRPO), and a benchmark dataset for evaluating audio editing tasks.
stepfun-ai/Step-Audio-EditX · Hugging Face
StepFun AI has open-sourced Step-Audio-EditX, a 3-billion-parameter LLM-based audio model designed for expressive and iterative audio editing. The model excels at controlling emotion, speaking style, and paralinguistic features (e.g., sighs, laughter, hesitations) in synthetic speech, and supports zero-shot voice cloning for Mandarin, English, Sichuanese, and Cantonese. The architecture uses a dual-codebook audio tokenizer, an audio LLM for token generation, and a flow-matching decoder. Recent updates (January 2026) added vLLM support for training and inference, GRPO training code release, and expanded language support (Japanese and Korean). The model requires 12 GB GPU memory (L40S tested) and is available in standard and quantized (AWQ 4-bit) versions. Benchmark evaluations show Step-Audio-EditX outperforms closed-source systems like Minimax and Doubao on zero-shot cloning and emotion control, with iterative refinement improving accuracy across three editing rounds. The Step-Audio-Edit-Benchmark was released in November 2025 to standardize evaluation of audio editing capabilities.
maya-research/maya1 · Hugging Face
Maya Research released Maya1, a 3-billion-parameter open-source text-to-speech model specializing in expressive voice generation with emotion control. The model uses a Llama-style decoder-only transformer backbone to generate SNAC neural codec tokens (7 per frame) at 24 kHz, achieving sub-100ms latency for streaming via vLLM. Voice design uses natural language descriptions (XML-attribute format: ``) rather than rigid parameters; the model supports 20+ inline emotion tags including laugh, cry, whisper, and rage. Training combined internet-scale English pretraining with proprietary curated studio recordings covering multi-accent English, forced alignment, VAD trimming, and audio deduplication via Chromaprint. The model requires a single GPU with 16GB+ VRAM (e.g., A100, H100, RTX 4090) and is distributed under Apache 2.0. Inference achieves ~0.98 kbps streaming bitrate and integrates with Hugging Face transformers and vLLM. Use cases include game character voices, podcasts, AI voice assistants, and accessibility tools. Model weights total 3.3 billion parameters in BF16, with 6.6 GB total file size and 333,721 lifetime downloads as of publication.
baidu/ERNIE-4.5-VL-28B-A3B-Thinking · Hugging Face
Baidu released ERNIE-4.5-VL-28B-A3B-Thinking, a 28B-parameter multimodal vision-language model with a mixture-of-experts architecture that activates only 3B parameters during inference. The model features extended reasoning capabilities trained via reinforcement learning with GSPO and IcePop strategies, supports image and video inputs, and includes tool-calling for image search and grounding. The model implements a "thinking with images" capability allowing dynamic zooming and fine-grained detail processing. Developers can deploy via transformers library (requires version ≤4.57.6), vLLM (tested with 0.11.2), or FastDeploy with quantization options. The model supports instruction finetuning and alignment training through ERNIEKit, Baidu's PaddlePaddle-based training toolkit. Model weights total ~59.3 GB (30B BF16+F32 parameters); single-GPU deployment requires minimum 48GB memory. The model is licensed under Apache 2.0 for commercial use.
Nemotron 3 Embed - a nvidia Collection
NVIDIA released Nemotron 3 Embed, a collection of open embedding models available on Hugging Face. The collection includes three sentence-similarity models in sizes 0.8B, 1B, and 8B parameters, all updated within the past two days. These models are positioned for enterprise RAG (retrieval-augmented generation), agentic retrieval, code search, and agent memory applications. The embedding models complement NVIDIA's broader Nemotron v3 ecosystem, which spans chat, code, math, reasoning, and safety-focused variants. Developers working on retrieval systems and multi-agent architectures can leverage these open models as alternatives to proprietary embedding services, with the ability to fine-tune and deploy on local or cloud infrastructure.
On Powerful Ways to Generate: Autoregression, Diffusion, and Beyond
Researchers Yang, Zhou, Wipf, and Li examine the computational power of Masked Diffusion Models (MDM) relative to Autoregressive Models (ARM) for language generation. While diffusion models offer flexibility through parallel and any-order token generation, the paper formally analyzes whether this flexibility enables solving problems beyond ARM's reach. The authors show that MDM with sufficient context length achieves computational universality with optimal parallel time complexity in PRAM, but any-order generation alone does not expand the class of solvable problems compared to ARM. To address this gap, the paper proposes a new generation approach called any-process generation, which extends MDM with remask, insert, and delete operations. Theoretically and empirically, the authors demonstrate these capabilities enable handling harder reasoning problems intractable for ARM and vanilla MDM, along with self-correction and length-variable editing. The authors argue these capabilities are essential for extending LLMs beyond natural language to domains like coding and scientific modeling where objects naturally evolve non-sequentially.
PAN: A World Model for General, Interactable, and Long-Horizon World Simulation
Researchers from the Institute of Foundation Models introduced PAN, a world model designed to predict future world states through video simulation conditioned on history and natural language actions. The model uses a Generative Latent Prediction architecture that combines an autoregressive latent dynamics backbone built on a large language model with a video diffusion decoder. This approach unifies latent-space reasoning with visual observation reconstruction, enabling the model to ground simulation in text-based knowledge while generating perceptually detailed and temporally coherent video outputs. PAN was trained on large-scale video-action pairs across diverse domains, allowing it to support open-domain, action-conditioned simulation with long-term consistency. The paper reports strong performance in action-conditioned world simulation, long-horizon forecasting, and simulative reasoning compared to existing video generators and world models, positioning it as a step toward general-purpose world models suitable for predictive reasoning and planning tasks.
Gemini 3: Introducing the latest Gemini AI model from Google
Google released Gemini 3, its latest flagship AI model, on November 18, 2025. Gemini 3 Pro achieves state-of-the-art performance across multiple benchmarks: 1501 Elo on LMArena Leaderboard, 37.5% on Humanity's Last Exam (without tools), 91.9% on GPQA Diamond, 23.4% on MathArena Apex, 81% on MMMU-Pro, and 87.6% on Video-MMVU. The model demonstrates improved multimodal reasoning, coding capabilities, and factual accuracy (72.1% on SimpleQA Verified). Google also introduced Gemini 3 Deep Think, an enhanced reasoning mode that further improves performance—41.0% on Humanity's Last Exam and 93.8% on GPQA Diamond—and achieves 45.1% on ARC-AGI-2. Developers can access Gemini 3 Pro immediately through Gemini API, AI Studio, Vertex AI, Gemini CLI, and a new agentic development platform called Google Antigravity. The model integrates with third-party tools including Cursor, GitHub, and Replit. Deep Think mode will roll out to Google AI Ultra subscribers in coming weeks. Google reports 13 million developers already using its generative models and 650 million monthly Gemini app users.
Skillful joint probabilistic weather forecasting from marginals
Researchers including authors from DeepMind and collaborators have published FGN, a machine learning approach for probabilistic weather forecasting that outperforms current state-of-the-art models. The method generates ensemble forecasts by training constrained model variants with learned perturbations, optimized directly to minimize continuous rank probability score (CRPS) at individual locations. FGN demonstrates improved performance across both deterministic and probabilistic metrics, produces skillful tropical cyclone track predictions, and captures joint spatial weather structure despite being trained only on marginal distributions. This work is part of the broader shift in weather prediction where ML-based models have recently begun exceeding traditional numerical weather prediction ensembles in global probabilistic forecasting accuracy and speed.
cerebras/MiniMax-M2-REAP-162B-A10B · Hugging Face
Cerebras and collaborators released MiniMax-M2-REAP-162B-A10B, a compressed variant of MiniMax-M2 created using REAP (Router-weighted Expert Activation Pruning). The model reduces parameters from 230B to 162B (30% compression) while maintaining near-identical performance on code generation, reasoning, and tool-calling tasks. The REAP method prunes redundant experts based on router gate values and expert activation norms, enabling one-shot compression without fine-tuning. Benchmarks show HumanEval performance of 93.3 (vs. 93.9 baseline), MATH-500 at 89.4, and AIME25 at 73.3. The model operates with 162B total parameters and 10B activated per token, 180 experts (pruned from 256), and supports a 196,608-token context. Deployment is compatible with vLLM without custom patches. The work is documented in arXiv preprint 2510.13999 and released under a modified MIT license.
Scaling Agent Learning via Experience Synthesis
Researchers introduced DreamGym, a framework designed to scale reinforcement learning training for autonomous agents by synthesizing training experiences rather than relying on expensive real-world rollouts. The system models environment dynamics using reasoning-based experience generation to produce consistent state transitions and reward signals, reducing infrastructure complexity and rollout costs that typically obstruct practical RL adoption. DreamGym combines synthetic experience generation with an offline-to-online replay buffer strategy initialized with real-world data and continuously enriched with fresh interactions. The framework also employs adaptive curriculum learning to generate increasingly challenging tasks that improve agent learning efficiency. Experiments across diverse environments show DreamGym outperformed baseline methods by over 30% on non-RL-ready tasks like WebArena, and matched performance of methods like GRPO and PPO using only synthetic interactions in RL-ready settings. A key practical contribution is demonstrated through sim-to-real transfer: agents trained entirely on synthetic DreamGym experiences achieved significant performance gains when transferred to real-world RL with substantially fewer actual environment interactions. This provides a scalable warm-start strategy for general-purpose agent training, addressing the challenge of limited task diversity and unreliable reward signals that impede current RL deployment.
SIMA 2: A Gemini-Powered AI Agent for 3D Virtual Worlds — Google DeepMind
Google DeepMind introduced SIMA 2, an AI agent that uses Gemini models to operate in 3D virtual environments with reasoning and self-improvement capabilities. The system advances beyond SIMA 1 by adding the ability to reason about goals, converse with users, and improve itself through trial-and-error feedback. SIMA 2 was trained on human demonstration videos with language labels across multiple commercial games and can follow over 600 language-based skills. Key improvements over SIMA 1 include better generalization to unseen environments (such as the Viking survival game ASKA and MineDojo, a Minecraft research implementation), support for multimodal prompts including sketches and emoji, and multilingual instruction understanding. The agent demonstrates task completion success rates significantly closer to human performance across a wider range of evaluation tasks. A notable capability is SIMA 2's integration with Genie 3, Google's world-generation model, allowing the agent to operate in newly synthesized 3D environments it has never encountered. The system also exhibits self-improvement through iterative training cycles where Gemini provides task rewards and feedback, enabling the agent to learn in previously unseen worlds without human-generated data. The research team acknowledged limitations in very long-horizon tasks, short interaction memory due to latency constraints, and precise low-level action execution.
Elevated errors for Github connector
OpenAI's GitHub connector experienced elevated errors affecting the Connectors/Apps service. The incident has been fully resolved and all impacted services have recovered. No details on root cause, duration, or user impact scope were provided.
Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model With Kimi Delta Attention and 1M Context
Moonshot AI announced Kimi K3 on July 16, 2026, an open mixture-of-experts model with 2.8 trillion parameters that activates 16 of 896 experts. It features Kimi Delta Attention and Attention Residuals and supports 1 million token context length. An open-weight release is promised by July 27, 2026. This positions it as the first publicly available open 3-trillion-class model, surpassing prior open alternatives.
Firefox in WebAssembly
This is a brief technical anecdote about Puter compiling Firefox to WebAssembly so the browser runs inside another browser. The project reportedly cost approximately $25,000 in Claude Opus and Fable tokens via a Claude Max subscription. Firefox/Gecko was chosen for its strong single-process support. The demo tunnels traffic through Puter's servers over WebSocket using the Wisp protocol due to browser sandbox constraints. This is engineering commentary rather than an official product announcement.
Q&A: How Capcom Brought Path Tracing to RE ENGINE Across PRAGMATA and Resident Evil Requiem
Capcom's RE ENGINE team integrated path tracing into two shipping titles—Resident Evil Requiem and PRAGMATA—each with distinct visual requirements. The source excerpt is incomplete and provides no technical details beyond confirming the effort. Further specifics on implementation, performance trade-offs, or API changes are not available from the provided summary.
Elevated errors for multiple models
Anthropic experienced elevated error rates on Claude model requests between 11:30 AM and 3:15 PM PT (18:30–22:15 UTC) on July 16. The majority of errors occurred within the first hour. Success rates recovered and the issue was resolved by 22:53 UTC. The summary mentions Claude Fable among affected models but is truncated; no root cause or scope details are provided.
Build enterprise search for agents with Amazon Bedrock Managed Knowledge Base
In this post, we walk through the three pillars that make this possible: simplified setup, smarter retrieval, and production readiness. We also show you code examples for setting up a knowledge base and retrieving from it.
It's official: EU will force Google to share search data and open up AI on Android
The EU is moving forward with regulatory requirements forcing Google to share search data and open up AI features on Android. Google has expressed concern that these changes could compromise user privacy and security. This is a regulatory/policy development rather than a technical announcement; no specific timelines, affected services, or implementation details are provided in the summary.
xAI can’t deny Grok makes CSAM anymore. So it’s suing users.
Elon Musk's xAI is filing its first lawsuit against a Grok user, alleging the creation of child sexual abuse material (CSAM) using the model. This follows xAI's prior public denials that Grok could generate such content. The lawsuit represents a potential shift in xAI's legal strategy regarding harmful outputs from its model. No technical details, dates, or outcome information are available.
Kimi K3, and what we can still learn from the pelican benchmark
Moonshot AI announced Kimi K3, described as their most capable model to date with 2.8 trillion parameters. The model is available via their website and API immediately, with an open-weight release promised by July 27, 2026. Moonshot positions K3 as the first "open 3T-class model," surpassing DeepSeek's 1.6 trillion parameter v4 Pro. According to Moonshot's self-reported benchmarks, K3 beats Claude Opus 4.8 and GPT-5.5 high on most tasks but trails Claude Fable 5 and GPT-5.6 Sol. No pricing or context window details are provided in this excerpt.
Fear of humanoid robots spurs human workers to strike at Hyundai auto factory
Hyundai aims to deploy 25,000 Atlas robots starting with US factories in 2028.
Kimi's open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI
Kimi launched K3, a multimodal open-weight model with 2.8 trillion parameters and one million token context length. In Kimi's benchmarks, K3 approaches Claude Fable 5 and GPT-5.6 Sol performance while beating Claude Opus 4.8 and GLM 5.2 in some cases by a wide margin. Full weights are scheduled for release by July 27. The article notes that K3 is significantly more expensive than its predecessor, marking what the source characterizes as the end of very cheap Chinese AI pricing. No exact price or performance tier comparison is provided.
Introducing Grok on Amazon Bedrock
AWS announced Grok 4.3 availability through Amazon Bedrock. The post covers use cases suited to agentic and enterprise workloads and describes capabilities developers can access: basic chat, configurable reasoning effort, tool calling, structured output, image input, and stateful multi-turn conversations. No pricing, model details, context window, or availability regions are specified in the summary.
Linus Torvalds to critics of AI coding in Linux: "Fork it. Or just walk away."
Creator says he will "very loudly ignore" those arguing for a ban on AI tools.
The AI compute gap: Enterprises are buying infrastructure faster than they can measure what it costs
A study of 107 enterprises found that AI infrastructure spending is accelerating faster than organizations can measure or control costs. Most companies deploy AI on hyperscaler platforms and model provider APIs, but spending is shifting toward specialized compute providers, with a majority planning to switch or add vendors within a year and many within a quarter. Decisions prioritize integration and total cost of ownership over headline token pricing. The analysis notes that most enterprises lack visibility into unit economics, with GPUs operating at roughly 50% utilization, indicating unused capacity. This reflects a common challenge: rapid infrastructure adoption outpacing cost accountability mechanisms.
The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials
A survey of 107 enterprises found that over half have already experienced a confirmed AI agent security incident or near-miss. Current security practices lag behind agent deployment: only about one-third provision each agent with its own scoped identity, most agents still share credentials, and only 30% isolate high-risk agents. The security infrastructure is primarily borrowed from model providers and hyperscalers rather than purpose-built for agent-specific threats. Security spending remains a small slice of overall budgets, and enterprises are evenly divided on whether traditional security teams should lead agent security. This points to a widening gap between agent capability and controls.
OpenAI Details GPT-Red: An Internal Automated Red-Teaming Model That Beat Human Red-Teamers 84% To 13% On Prompt Injection
OpenAI developed GPT-Red, an internal attacker model trained via self-play reinforcement learning against a population of defender LLMs, to automate red-teaming at scale. On a replicated indirect prompt injection arena, GPT-Red beat human red-teamers 84% to 13%. It discovered a novel attack class called "Fake Chain-of-Thought" and reduced GPT-5.6 Sol's failures on OpenAI's hardest direct injection benchmark by 6 times. OpenAI acknowledges GPT-Red still struggles with multi-turn and image-based attacks. This demonstrates automated adversarial testing can outperform human-led efforts on certain injection vectors.
Google Vids now lets you star in your own AI videos
Google is adding personalized AI avatars to its Vids tool, enabling users to create videos featuring a digital version of themselves. The feature is powered by Gemini Omni and includes AI-driven tools for generating and editing videos from text prompts and reference images. No pricing, availability date, or technical limitations are specified in the summary.
Roblox launches an AI-powered game-creation feature in its mobile app
Roblox launched a new "Build" feature in its mobile app that lets users generate basic games from a single text prompt using AI. The feature lowers the barrier to game creation on the platform. No details on the underlying model, availability regions, or feature limitations are provided.
Quoting Thibault Sottiaux
OpenAI's Codex model has a documented bug where it unexpectedly deletes files when full-access mode is enabled and codex runs without sandboxing or auto-review protections. The issue occurs when Codex attempts to override the HOME environment variable to define a temporary directory and then mistakenly deletes HOME instead. This is a serious safety issue affecting code-generation agents with elevated permissions. The bug highlights risks of running generative models with filesystem access and minimal oversight.
Google rebrands NotebookLM as Gemini Notebook and opens its search app to third-party integration
Google rebranded NotebookLM to Gemini Notebook and integrated it more deeply into its broader ecosystem. A new feature provides each notebook with its own cloud compute environment capable of writing and executing code, initially available to AI Ultra and Workspace customers. Google Search is also gaining third-party app integrations. These are organizational and feature-packaging changes to existing tools rather than new model capabilities.
The AI context gap: Enterprise AI organizations have a trust problem, not a retrieval problem — and most are still building the fix
A study of 101 enterprises found that enterprise AI infrastructure for feeding business context to agents is being built faster than it can be trusted. Retrieval-augmented generation is the default context source, and provider-native retrieval has overtaken dedicated vector databases despite the category's prominence. However, a majority of enterprises have already experienced agents producing confident, incorrect answers due to missing or inconsistent context. A governed semantic layer is emerging as the solution, but most organizations are still building it. The field is converging on hybrid retrieval approaches. The core issue is trust and consistency, not retrieval capability.
Linus Torvalds rebukes anti-AI stances in the Linux kernel code review process, says 'Linux is not one of those anti-AI projects' — creator embraces AI as just a tool and 'clearly a useful one'
Linus Torvalds, Linux's creator and kernel manager, has seemingly taken an accepting stance of AI-assisted tooling.
Neural atom quantum computing roadmap — how laser-cooled trapped atoms could pave the path beyond physical qubit counts
This article surveys the neural atom quantum computing landscape, describing how trapped-ion quantum computers with laser cooling enable software-defined and reconfigurable qubit arrays. Notably, qubits can be physically repositioned mid-computation. The roadmap highlights three leading companies—QuEra, Atom Computing, and Pasqal—operating in this space. This is a hardware and physics-focused analysis of quantum computing approaches rather than an AI model announcement.
AMD Ryzen 7 7700X3D is exclusive to Newegg in North America — $329 CPU won't be available at other vendors until at least Q4
AMD's Ryzen 7 7700X3D processor costs $329 and is exclusive to Newegg in North America through the end of Q3 2026 before wider retail availability. The CPU is noted as a strong gaming performer but less versatile for professional mixed workloads. This is a hardware availability and pricing announcement unrelated to AI models.
The agent evaluation gap: Enterprise AI organizations have a reality-alignment problem, not a coverage problem — and most are shipping to production anyway
A study of 157 enterprises found a significant evaluation gap in AI agent deployment: half have shipped agents that passed internal evaluations but failed in production. Only one in twenty organizations fully trust automated evaluation today, and the most common weakness is misalignment between evaluations and real-world outcomes. Yet two-thirds of organizations already allow or are actively engineering toward deploying agent changes to production based on automated evaluation alone, with no human in the loop. This reflects a critical reliability problem: organizations are granting autonomy faster than they can validate it safely.
Germany puts Google's AI Overviews and Perplexity under media law in first-of-its-kind ruling
German media regulators have classified Google's AI Overviews and Perplexity's offerings as the companies' own editorial content rather than neutral search results, finding they marginalize standard web links. This marks the first enforcement action under Germany's State Media Treaty against these AI-powered systems. Both companies have been given one month to appeal the ruling. The decision establishes a precedent for treating AI-generated summaries as regulated media content subject to journalistic standards, potentially affecting how AI overviews integrate with search results across Europe.
Integrating Context-Aware Video AI Agents Into Enterprise Workflows
A video analytics AI agent that can perceive, reason, and act based on massive amounts of video footage must be integrated with existing workflows and...
NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval
NVIDIA's Nemotron 3 Embed model achieved the top ranking on the RTEB (Retrieval Text Embedding Benchmark) leaderboard, demonstrating strong performance for agentic retrieval workflows. The achievement highlights the model's effectiveness for retrieval-augmented generation and agent-based information lookup tasks. Nemotron 3 Embed is positioned as an open-source alternative for builders implementing retrieval systems, particularly in multi-step reasoning scenarios where embeddings feed downstream agent actions.
Scaling Agentic AI Factories Through Extreme Co-Design with NVIDIA BlueField
Agentic AI changes the infrastructure pattern for AI factories. One request can trigger many model calls, tool calls, memory lookups, policy checks, storage...
Google’s AI Mode now lets you link and interact with select apps
Google expanded its AI Mode feature to include app integration and task completion beyond question-answering. Users can now link and interact with select applications directly through AI Mode, enabling workflows that span multiple integrated services. This shift marks a move toward AI-assisted task automation across Google's ecosystem rather than information retrieval alone.
Connect more of your apps to Search
Google announced a feature allowing users to connect more applications to Search, enabling integrated app rendering within search results. The update expands the ecosystem of services accessible through Search, improving contextual app availability and task continuity from search to app interaction.
Create, edit and star in videos with two Google Vids updates
Google released two updates to Google Vids: enhanced creation and editing capabilities, and introduction of personal avatars powered by Gemini Omni. Users can now generate video content with AI-driven avatars representing themselves. This feature pair extends Google Vids' competency from editing toward synthetic media generation using multimodal AI.
Building a restaurant telephony AI host with Amazon Bedrock AgentCore and Amazon Nova 2 Sonic
AWS published a tutorial demonstrating a voice ordering system for restaurants using Amazon Bedrock AgentCore and Amazon Nova 2 Sonic. The system accepts inbound phone calls via SIP, takes orders from greeting through confirmation, and integrates with restaurant backends via the Model Context Protocol (MCP). Deployment uses AWS CDK and AWS Fargate, showcasing how to bridge real-time voice agents into legacy phone infrastructure. This illustrates practical agentic AI deployment for customer-facing applications with live voice interaction.
Energy IPOs surge as investors hunt for ways to play AI boom
An analysis reported that energy companies are entering public markets at the fastest pace this century, with investors seeking exposure to the AI infrastructure boom. The surge reflects capital reallocation toward power generation, cooling, and supply chain companies that support data center expansion for AI workloads. This trend indicates market confidence in sustained demand for AI infrastructure investment.
Tower Semiconductor revives shuttered Panasonic-era fab in $3 billion Japan photonics expansion — METI-backed plan targets $3.6 billion revenue by 2028
Tower Semiconductor announced a $3 billion expansion of its silicon photonics and advanced packaging operations in Japan, with backing from METI (Japan's Ministry of Economy). The initiative targets $3.6 billion in revenue by 2028 and revives a previously shuttered Panasonic-era fabrication facility for 300mm silicon germanium and photonics production. The expansion supports Japan's strategy to strengthen high-end semiconductor and photonics manufacturing for AI infrastructure.
Yes, you can now order DoorDash from the command line
DoorDash opened a limited beta of dd-cli, a command-line tool that enables developers and AI agents to search stores, build orders, and complete transactions from the terminal. The tool represents infrastructure designed specifically for AI agents rather than human users, providing structured APIs for order orchestration. This marks another step toward platforms offering machine-readable interfaces optimized for autonomous agent workflows.
Inkling: Our open-weights model
Mira Murati's Thinking Machines Lab released Inkling, an open-weights multimodal model licensed under Apache 2.0. Inkling is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active parameters, trained on 45 trillion tokens spanning text, images, audio, and video. A smaller variant, Inkling-Small (276B total, 12B active), is in testing with weights to be released upon completion. The model represents a significant open-source contribution to the multimodal AI landscape, though detailed training documentation remains limited.
Why is OpenAI selling a ChatGPT basketball?
You may have heard that OpenAI released its first piece of hardware this week. You may not have heard about the ChatGPT basketball.
Elevated Error Rates For SSO Login
OpenAI experienced elevated error rates for SSO (Single Sign-On) login functionality. The incident has been fully resolved and all affected services have returned to operational status. The outage impacted login services but did not affect other platform functionality.
Lenovo announces world's first laptop with inkjet-printed OLED — the Legion R9000P is equipped with a 240 Hz IJP panel from TCL CSOT
Lenovo announced the Legion R9000P, which features the first inkjet-printed OLED (IJP OLED) panel in a laptop. The display, produced by TCL CSOT, offers a 240 Hz refresh rate and 99% DCI-P3 color coverage at a lower cost than traditional OLED manufacturing methods. This represents a manufacturing innovation in display technology rather than AI, though relevant to the hardware infrastructure supporting AI workloads.
How a former DeepMind researcher raised at a $300M pre-seed valuation before launching a product
Drawing on more than a decade spent helping build some of the world's most influential AI systems, including research that later informed the development of ChatGPT, Andrew Dai explains why he believes visual AI is one of the next major frontiers in artificial intelligence.
Mermaid to ASCII art (mermaid-ascii)
Simon Willison documented work on a tool converting Mermaid diagrams to ASCII art, and discovered a more mature Go-based library (AlexanderGrooff/mermaid-ascii) that he compiled to WebAssembly for comparison. The WebAssembly version includes color support. This is a developer tool note rather than an AI model or service announcement.
Why AMI Labs’ Alexandre LeBrun won’t call his AI ‘AGI’ or ‘superintelligence’
While everyone in AI is chasing "superintelligence," Alexandre LeBrun, CEO of Yann LeCun’s world model startup, AMI Labs, dismisses the word.
OpenAI wants developers to stop typing commands and start using a joystick to control their AI agents
OpenAI and keyboard manufacturer Work Louder unveiled the Codex Micro, a compact hardware controller for AI agent control. The joystick-based interface is designed to replace or supplement command-line input for agent interaction, enabling more intuitive hardware control of autonomous systems. This represents OpenAI's first foray into specialized input hardware for agent workflows.
Moonshot’s upcoming Kimi 3 is expected to close the gap with Anthropic’s Opus 4.8
The FT reports Kimi K3 will be the largest open AI model from China, with a parameter count between 2 trillion and 3 trillion.
Sakana AI's orchestrator adds Nvidia Nemotron to prove "collective intelligence" can rival single frontier models
Sakana AI integrated NVIDIA's open-source Nemotron models into its Fugu orchestrator, which dynamically combines multiple language models for task-specific optimization. The integration tests the hypothesis that open models can match frontier systems when coordinated through orchestration. However, benchmark comparisons between the Nemotron+Fugu combination and frontier models have not yet been disclosed.