Horizon Lens — 11 September 2026
Slack puts small AI-built tools inside the conversation
Slack is bringing AI-generated dashboards, reports and other interactive tools into chat. The Verge reports that its new Slackforce Surfaces feature lets people describe what they need to Slackbot, which draws on relevant conversations and connected services such as Google Drive or Salesforce. Colleagues can then use, comment on and pin the resulting tool in a channel.
There are two practical qualifications. According to the report, Surfaces is available across customer plans, including free accounts, where Slackbot is enabled; using it with live data starts in October. Slack also says it only draws on information its AI tools have permission to access. A shared dashboard today should therefore not automatically be read as a continuously refreshed operational view.
Analysis: The useful change is that a team can turn a discussion into something it can work with together. For a first experiment, choose a narrow report with a known answer, check it against the underlying records, and make its refresh status obvious. Permission to read a source and a correct interpretation of that source are separate checks; an attractive chart does not remove the second one.
A visual AI workflow can also become a reusable tool
A 10 September Hugging Face tutorial demonstrates Workflow1111, a Gradio canvas combining image generation, editing, background removal, image-to-video and related pipelines. The authors describe rebuilding much of the familiar AUTOMATIC1111 feature set as connected operations. Users can try the example with a Hugging Face account or access token, but the model calls consume the user’s own quota.
The more useful engineering detail is how the same workflow can be reused. The authors say output nodes become REST endpoints, allowing other software to call them. Enabling the MCP server also exposes those outputs as tools for compatible AI assistants. Some operations are local Python functions; the demonstrated model calls use external providers or hosted Spaces, although functions can be connected to locally loaded models.
The tutorial also shows why retaining generation details matters. Its image pipeline writes settings into PNG metadata and provides a way to read them back, including prompts, seed, image size and model. This is a concrete aid to revisiting an experiment: the saved image can carry information about how it was made, alongside the visual result itself.
Analysis: This offers a route from a visual experiment to a repeatable service without separately rebuilding its interface for an agent. Before adopting a workflow, map which steps send data away, whose quota pays for them, and which outputs need human inspection. Being able to run the canvas without owning a GPU does not make its remote computation free or its output automatically dependable.
Two different ways to help robots learn from human video
In a 10 September account of Skild AI’s S1, NVIDIA describes teaching a robot a new task by giving it a demonstration video, without updating the model’s weights for that task. S1 itself launched the previous week. NVIDIA says demonstrated activities can last up to ten minutes, including potting plants and preparing coffee. These are the supplier’s descriptions of its partner’s system, rather than an independent deployment assessment.
One number deserves careful reading: NVIDIA reports roughly 66% success at each step in Skild’s tests on new multistep tasks, against 9% for a comparison system. That is a step-level result. It should not be presented as the proportion of complete jobs finished successfully, or as proof that a robot can be left unattended in an arbitrary workplace.
Related research context, submitted on 9 September: HuRo tackles video learning earlier, during model pretraining. Its authors describe converting human videos into robot-aligned observations and actions, producing about 630,000 episodes. Across four real-world manipulation tasks, they report overall completion rising from 51.5% to 80.3% as pretraining scale increases. Those results describe their experimental setup, not S1’s performance or a comparison against it.
HuRo also reports a separate test under changed spatial and visual conditions, where completion rose from 34.9% to 72.2%. Keeping that measure separate from overall completion is valuable: it asks how well a trained system copes when the scene changes. The study’s four tasks still set the boundary of what those reported improvements demonstrate.
Analysis: Both point towards making human demonstrations more useful, but at different stages: building a model’s skills and specifying a task to an existing model. For anyone evaluating a robot, ask separately about complete-task success, recovery after mistakes and the range of environments tested. A strong learning result is a reason to investigate those operational questions, not to skip them.
Huang’s growth outlook puts AI demand under the spotlight
NVIDIA chief executive Jensen Huang reiterated an outlook of roughly 70% revenue growth next year at Thursday’s Goldman Sachs technology conference, according to TechCrunch. He argued that NVIDIA’s relationships across model developers, infrastructure providers and data-centre projects give it unusually broad visibility into demand. This is an executive’s forward-looking assessment, not growth that has already happened.
The report also describes questions about NVIDIA investing in companies that buy its technology. Huang’s defence was that prospective recipients already have revenue-generating customer contracts. TechCrunch notes that much of the current expansion comes from AI-native startups raising substantial funding and spending heavily on AI themselves. That makes the source of demand part of the story alongside the scale of hardware orders.
Analysis: The question to follow is whether spending on AI infrastructure is supported by customers getting enough lasting value to keep paying. Huang’s confidence is useful evidence of NVIDIA’s outlook; it cannot by itself settle that wider question. Read future demand claims alongside customer economics and actual results, while keeping forecasts clearly separate from completed business.