Wednesday, October 7, 2026

Hark Wants To Handle Everything

Hark Wants To Handle Everything

Today’s Overview

Good morning, AI agents are getting more hands-on fast. Hark wants to research, shop, book, and build on your behalf, Mistral is teasing a huge open-weight model from Europe, and OpenAI is putting hundreds of machine-generated math manuscripts into the open. Let's dive in.

Top Stories

Hark Launches a Proactive AI Assistant

Hark released Hark Pro, a personal assistant for web and mobile that remembers user preferences, flags what needs doing, and can act across tasks like research, shopping, booking, and building. In a launch demo, a request for movie tickets ends with Hark choosing seats, adding the show to a calendar, and offering to reserve parking. Tasks run on Handoff, Hark's cloud computer, which the company says can operate many browsers at once while showing its work in chat.

  • The product page frames Hark Pro as a system for getting things done rather than a conventional chatbot.
  • The sign-in flow supports Google, Apple, email, or phone as entry points for the assistant.
  • The live page emphasizes delegation with the line let Hark handle it as the core product promise.

Mistral Previews Large 4, Its 'Le Chonk' Model

Mistral released Mistral Large 4, nicknamed Le Chonk, as a 1 trillion-parameter multimodal model with 49 billion active parameters. It is available now through the API in Mistral Studio, with open weights planned for the end of October. The model supports image input, a 1 million token context window, more than 160 languages, and cybersecurity tasks that closed models often refuse.

  • Mistral describes the preview as focused on coding, agents, and multimodal understanding alongside enterprise workloads.
  • The company says the model was trained on roughly 4,000 Grace Blackwell GPUs inside its own European data centers.
  • Mistral positions the release as especially relevant for sovereign AI because the weights are planned for self-hosting and customization.

Google Details Nano Banana 2.1

Nano Banana 2.1 is a member of the Gemini 3 series built on Gemini 3.6 Flash. It accepts text and image inputs and can produce both image and text outputs. Google says it is available through the Gemini app and API, Google AI Studio, Search AI Mode, Google Ads, Google Flow, and Google Stitch.

  • The model card lists output limits of 4K tokens for images and 64K tokens for text.
  • Google evaluated it on text-to-image and editing benchmarks, including human side-by-side Elo and factuality-based autorating.
  • Known limitations include small text rendering and spatial confusion along with imperfect character consistency in some edits.

Research & Analysis

OpenAI Releases 722 AI-Written Math Manuscripts

OpenAI released 722 math manuscripts generated by an unreleased internal frontier model and made the repository publicly accessible. The work was produced from roughly 4,000 attempted problems and organized into 372 families of related papers. The release includes Lean-verified proofs and arrives amid broader debate over how AI labs should handle work on famous mathematical problems.

  • OpenAI says it consulted an independent mathematics advisory group to shape release practices and community norms.
  • The repository includes revision and citation protocols so the results can be corrected, tracked, and referenced.
  • OpenAI also shared 10 summaries of model reasoning plus details on compute and problem-attempt statistics.

Long-WAM Gives Robots More Useful Memory

Long-WAM is a model-system framework for extending the context window of causal world-action models while still meeting real-time robot control constraints. The paper argues that longer visual history helps only when the video foundation model is pretrained autoregressively. It reports improved performance across robot manipulation benchmarks and real-time deployment on multiple hardware platforms.

  • The authors describe the core constraint as balancing visual history and action latency in real-time control.
  • Long-WAM combines streaming observation encoding with asynchronous execution to keep future prediction active during deployment.
  • The system is shown on Unitree G1 and YAM for dynamic and long-horizon manipulation tasks.

Princeton Model Plays Chess and Explains Moves

Researchers at Princeton released a 4B model that plays chess at grandmaster level while explaining every move. The result points to strong domain performance from a comparatively small model in a constrained strategic setting. The work presents chess as a testbed for combining specialized expertise with language-based explanations.

  • The framework pairs an instruction-tuned language model with a silent expert chess encoder through cross-attention.
  • Across seven improvement rounds, the model rises from 1782 to 2697 Elo in the reported evaluation.
  • The authors frame the method as applicable to domains with expert encoders, including games, robotics, and computer use.

Trace2Env Rebuilds Environments From Interaction Logs

Trace2Env is a learning-free framework for reconstructing reusable environment replicas from historical interaction traces. Instead of rebuilding the original executable system, it turns traces into a worldbook that a world model agent can consult at runtime. The paper reports gains in next-observation fidelity and long-horizon interaction consistency across nine environments.

  • The worldbook contains schemas, transition rules, constraints, invariants, grounded evidence and demonstrations.
  • At runtime, the world model agent uses episode state and memory to preserve consequences across turns.
  • The paper evaluates Trace2Env across terminals, repositories, Android, web apps, enterprise services, and text games.

Trending AI Tools

  • Claude for Google Workspace Claude now works inside Docs, Sheets, and Slides in beta for all paid Claude plans.

  • Personal Agent Protocol Meta and Sierra are developing an open standard for authentication, consumer-controlled access, and company visibility into personal agents.

  • Kolibri-1 Aleph Alpha released an Apache 2.0 open-weight reasoning model with 78B total parameters, 1M context, and an OpenAI-compatible API.

Quick Hits

  • OpenAI dots are always-on agents that keep working inside ChatGPT after the tab closes, with one included on Pro and Business Premium plans in eligible markets.

  • DeepSeek's reported $12B raise is nearing completion with Tencent and CATL among backers, while the company prepares for a potential Shanghai STAR Market IPO.

  • Broadcom and Anthropic have a lending arrangement of up to $42 billion tied to Anthropic renting Broadcom's chips.

  • Britain's AI code pledge came at Labour's conference, where Prime Minister Andy Burnham promised a global AI code during Britain's 2027 G20 presidency.

  • Hark's agent-first launch brings buying groceries, booking cars, and paying bills ahead of the company's first devices.

  • Britain's 2027 G20 AI plan would create a global AI code, with Andy Burnham framing it as a chance for the North and Midlands to shape the next industrial revolution.

Keep reading for free

Enter your email. If you're already subscribed, we'll send a sign-in code. If not, you'll subscribe in the next step.

Free access. Subscribe once, then use the same email on future issues.

Free to read. Subscription just unlocks the full issue.