LodeHQ
AI news for people who shipFree, by email

AI · Wednesday, 11 November 2026 · Issue 167 · 27 stories

GR00T 2 and Gemini Robotics 2.5 top 95% task success

AI is moving from lab demos toward repeatable work: NVIDIA’s GR00T 2 and Google’s Gemini Robotics 2.5 reached over 95% success on unfamiliar or factory tasks. But the day’s tests also show reliability still depends on checking what models actually saw and did.

Models & Releases

GR00T 2 takes robots from five demos to 95% task success

NVIDIA’s GR00T 2 lets robots learn tasks from human video or other robots, giving developers a way to reuse demonstrations instead of collecting everything on the target machine. It predicts both what the world will look like next and which action causes it, then adapts from five demonstrations per task: across 50 unfamiliar tasks, success rose from 40% zero-shot to 95% after a 30-minute fine-tune on one RTX 5090. The model also supports parallel grippers, dexterous hands, and humanoids.

research.nvidia.com ↗14 min read at source

Claude Fable 5.2 brings up to 30% faster output to technical work

Anthropic’s Claude Fable 5.2 and Mythos 5.2 target ambitious technical work with clearer communication, up to 30% faster output than their 5.1 versions, and lower pricing at $8 per million input tokens and $40 per million output tokens. The two models share an underlying model, while Fable adds cybersecurity and biology safeguards for broad availability; Mythos is reserved for qualifying research programs. An optional fast mode reaches up to 2.5 times token speed for twice the price.

anthropic.com ↗20 min read at source

Gemini Robotics 2.5 pushes robots toward reliable factory work

Gemini Robotics 2.5 lets partner robots handle a wider range of delicate, unstructured tasks, making work such as automotive parts packing more practical outside the lab. Google DeepMind tested it on 200 tasks, including serving food with tongs and pipetting liquid; on Apptronik’s Apollo 2, it reached a 96% success rate in a Mercedes-Benz production-line proof of concept. Compared with previous baselines, task success rose 30% to 100% and completion was 20% to 30% faster.

deepmind.google ↗5 min read at source

Gemini 4 Carbon turns Antigravity into a parallel coding team

Gemini 4 Carbon is now rolling out in Google Antigravity, giving developers a model for difficult, long-running software work plus a /swarm preview that splits one prompt across agents working in parallel. A coordinator creates specialized workers in isolated git worktrees, while adversarial verifiers run tests and challenge the changes, so edits do not collide. Carbon ranks #1 on CursorBench 4.0, and agents can use lighter models for routine edits.

antigravity.google ↗2 min read at source

MiniMax-H3-Image brings 16MP generation and editing to open source

MiniMax-H3-Image gives creators and developers one open-source checkpoint for generating images from text and editing one or more reference images, including up to 16 references. It produces roughly 16-megapixel images at every aspect ratio without an upscaler, and uses six sampling steps, making high-resolution output practical to run. You can use it with Diffusers or through MiniMax’s API and apps.

huggingface.co ↗9 min read at source

BotPsy brings reliable tool use to a 0.8B on-device model

Tether AI Research’s BotPsy gives on-device assistants a 0.8B-parameter model for function calling, so routine tasks can stay private and self-hosted instead of relying on a cloud model. Built on Qwen3.5-0.8B, it roughly doubles the base model’s scores on BFCL v4 (23.4 to 44.0), ACEBench (30.1 to 58.7), and τ-bench (23.2 to 51.5). It is meant for low- to medium-difficulty tasks, not high-stakes use.

huggingface.co ↗1 min read at source

Research

Counterfactual training makes a small medical model check the image

A 4B-parameter medical vision-language model cut false claims about erased lesions to 9.6%, versus 41-62% for nine frontier models, without a significant accuracy loss on real lesions. Training showed the model an image, an inpainted version with the lesion removed, and a control edit elsewhere, teaching it to answer from visual evidence instead of the question's wording. The result is promising but not deployment-ready: it missed all 48 pneumothorax films in a fresh external test.

arxiv.org ↗2 min read at source

EarthRecord shows satellite agents can analyze data, but not trust the pixels

EarthRecord gives researchers a harder test for multimodal agents: analyze four decades of raw satellite imagery across 56 sites, rather than answer questions about curated images. Its 238 questions leave cloud masking and cross-sensor harmonization to the agent, and tests of twelve agents found that half of correct answers followed deliberate changes to one satellite’s data. The benchmark measures whether an agent found real land change or merely fit the imagery.

arxiv.org ↗2 min read at source

GR00T 2 reaches 95% task success with five demonstrations

NVIDIA’s open GR00T 2 gives robotics researchers a way to teach dexterous skills with far less robot-specific data: it reached 95% average success across 50 tasks after one demonstration per task was expanded to five. The world action model learns from human and robot data, then transfers skills between different robot bodies without robot demonstrations. Agents can also sequence skills and recover from errors through its API and supporting tools.

blogs.nvidia.com ↗7 min read at source

ChipOmni turns fragmented chip design into a verified agent workflow

ChipOmni lets engineers hand a natural-language RISC-V specification to one agent that coordinates design tools, checks each stage, and produces a physical layout without human intervention after initialization. Its context-evolving system combines multimodal understanding, planning, multiple agents, and tool feedback, then repairs intermediate designs through repeated verification. The paper reports that it solves all 30 AnalogCoder-Pro tasks and reaches a micro-SoC GDSII layout that passes physical sign-off.

arxiv.org ↗1 min read at source

Probe accuracy can overstate what an LLM actually computes

A four-stage Bayesian test gives researchers a stricter way to tell whether geometric patterns in LLM activations drive arithmetic, rather than merely being readable from them. Applied to two-digit addition and multiplication in GPT-J 6B, Llama 3.1 8B, and Pythia 6.9B, it found 81% of judged shapes were inherited from related concepts, not owned by the target concept. Interventions retained about 61% of the full probe's effect using at most five dimensions.

arxiv.org ↗2 min read at source

Recursive self-updates can lose value after the next update

A model update that looks useful can stop being useful when the model updates again, so teams testing self-improvement need to score the full transition, not just the first step. In 64 fresh math and code runs across five 3-8B backbones, continuation reduced utility by 1.78 percentage points in the balanced Qwen3-4B/Llama-3.2-3B comparison, with immediate-selected utility near zero. The authors recommend independent scoring and a no-update baseline.

arxiv.org ↗1 min read at source

SetPrint fingerprints model lineage using only response changes

SetPrint lets you test whether a target LLM was derived from a source model without inspecting either model, using only their responses to perturbed queries. It compares how each model’s answers change relative to the average change in its own response set, then matches changes caused by the same perturbations. On LeaFBench, the training-free method consistently outperforms existing black-box fingerprinting approaches.

arxiv.org ↗1 min read at source

Physics2Weight generates PDE networks without gradient optimization

Physics2Weight lets researchers solve partial differential equations by having a large language model generate and edit neural-network weights and architectures directly from the governing equation and its initial and boundary conditions. The framework adds physics-constrained prompts, physics-guided architecture edits, and iterative weight refinement, avoiding gradient-based optimization. Tests on four equations found lower solution errors than conventional physics-informed neural networks, but nonlinear equations still exposed limits in satisfying the governing equations.

arxiv.org ↗1 min read at source

TopicScope tracks every generated token across seven topics

TopicScope turns sparse-autoencoder features into a generation-time monitor that can label each token, letting the topic label change mid-response instead of forcing one label on the whole answer. It trains on hidden states collected during prefill, then monitors tokens during decoding, and generalised across Gemma-3 1B and 4B. On synthetic multi-topic conversations, performance fell by at most 5 macro-F1 points when topics were removed, reordered, or interrupted with off-topic asides.

arxiv.org ↗2 min read at source

AI is reshaping freelance demand toward broader, better-funded work

OpenAI and Upwork’s research suggests freelancers are seeing demand move away from some familiar services toward AI-related projects with larger budgets and more hours. The study compares ChatGPT activity with Upwork job posts, finding that 43 of 63 matched work categories shifted in opposite directions across the two platforms. The stakes extend beyond one marketplace: Upwork estimates 38% of skilled U.S. knowledge workers freelance.

openai.com ↗4 min read at source

Products & Industry

Google adds TPUs for giant context windows and confidential inference

Google’s TPU 8i Max is built for AI workloads with massive context windows, while TPU 8tx adds specialized encryption for confidential inference, giving Cloud customers two systems aimed at different production constraints. Google describes a system as the chip plus its surrounding compute, networking, and storage, so these designs can respond faster to customer needs than a new chip cycle. The additions bring Google’s eighth-generation lineup to four TPU systems in one year, serving nearly 100 customers.

blog.google ↗8 min read at source

Claude turns brainstorming, spreadsheets, and data into shared workspaces

Paid Claude users can now keep plans, spreadsheets, and data work inside Claude instead of copying outputs into separate tools. Whiteboards turns a rough plan into a timeline with owners and dates, Sheets builds editable trackers with formulas and change outlines, and Claude Data answers questions against sources such as Redshift, BigQuery, and ClickHouse. Sheets is still basic, so users can export to XLSX or CSV when they need Excel’s full functionality.

claude.com ↗4 min read at source

Cursor lets teams share context across their coding agents

Cursor Projects now lets teammates and their coding agents plan, build, and ship in one shared workspace, so a new engineer can take over work without rebuilding its context from scratch. The project stores requirements, diagrams, specs, and decisions, while a shared task board shows ownership and pull requests. Agents can also coordinate across teammates, such as sequencing database and feature changes to avoid merge conflicts.

cursor.com ↗2 min read at source

Anthropic sets a $70.00 to $78.00 IPO price range

Anthropic has started marketing its proposed IPO, giving investors a $70.00 to $78.00 per-share range and a planned Nasdaq listing under “ANT.” The offering covers 825,000,000 Class A shares, including 801,351,354 from Anthropic and 23,648,646 from existing shareholders, so the company would not receive proceeds from the latter sales. The SEC registration statement is filed but not yet effective, meaning no shares can be sold yet.

anthropic.com ↗1 min read at source

AI data centers are colliding with a 54-megawatt power problem

AI data-center expansion is running into an electricity problem, not just a public-relations problem: the average facility in Virginia’s Data Center Alley has 54 megawatts of capacity, more than Denver International Airport draws. Operators can connect to the grid or generate power onsite, but both choices create visible costs and local backlash. The result is a physical constraint on expansion that better messaging cannot remove.

technologyreview.com ↗6 min read at source

Anthropic buys a 1.7 GW generator queue for $3.4 billion

Anthropic is taking direct control of 1.7 gigawatts of gas-fired generation for $3.4 billion, giving its data centers access to Caterpillar equipment that can arrive soon and over the next two years. It took over OnePWR’s orders, placed two years ago when generator supply was less constrained, plus the engineering and construction capabilities to install them, but did not acquire OnePWR or its staff. The deal works out to about $2,000 per kilowatt; the companies declined to comment.

cellcog.ai ↗6 min read at source

Policy & Safety

Fireworks finds customer data exposure after October credential breach

Fireworks says its October credential breach reached account and billing information, plus secrets, datasets, and incomplete model weights for a limited set of accounts. The company links the activity to criminals selling stolen LLM access, and says a later cloud-provider incident caused additional customer-data impact. Payment card numbers were not stored, and Fireworks says it found no impact to its zero data retention Inference, Training SDK, and Nexus services.

fireworks.ai ↗2 min read at source

OpenAI promises communities a say in every compute campus

OpenAI says every future compute campus it owns and operates will come with a local compact, giving residents input on development and a contact point throughout the campus’s life. The company also published a policy blueprint calling for developers to cover energy and infrastructure costs, protect families from new costs, and share the resulting jobs and business opportunities. OpenAI says it will report publicly each year on how it meets those commitments.

openai.com ↗4 min read at source

Tools & Open Source

NVIDIA puts OpenShell’s agent controls under CNCF governance

NVIDIA is contributing OpenShell, an open source runtime for securing and governing autonomous agents, to the CNCF Sandbox, giving teams a community-run foundation for controlling what agents can do. OpenShell applies deterministic policies as agents use tools, access inference services and run workloads, while connecting to existing compute, credentials and monitoring systems. NVIDIA says hundreds of organizations already use it, with tens of thousands of contributions since launch.

blogs.nvidia.com ↗4 min read at source

Modern Web Guidance gives AI coders browser-aware web advice

AI coding agents can now use Chrome’s Modern Web Guidance to build and update web interfaces with current platform practices instead of defaulting to generic patterns. Install it with the CLI or add it to supported agents, then ask for work such as replacing custom modals with <dialog>, migrating tooltips to Popover, or improving Core Web Vitals. Its guidance also lets developers set Baseline browser targets, so compatibility constraints shape the code from the start.

developer.chrome.com ↗2 min read at source

Language Explorer visualizes thousands of languages for researchers and educators

Google’s Language Explorer gives researchers, educators, and language enthusiasts an online way to explore linguistic diversity through interactive visualizations and data. It brings together information across thousands of languages, building on Google’s open-source LinguaMeta dataset, so users can examine the breadth of language diversity rather than treating language technology as a list of supported tongues.

research.google ↗2 min read at source

You’re caught up.

That was all 27 stories in the 11 November issue.