The Slider of Autonomy

Andrej Karpathy has a concept he calls autonomy sliders. The idea is simple: AI shouldn't be binary. Not "AI on" or "AI off." Not "manual mode" or "autopilot." A continuous dial. The user decides where it sits, and they can move it anytime.

The research behind this goes back to 1978 -- Sheridan and Verplanck outlined ten levels of automation, from "the human does everything" to "the computer acts and tells you after." Most software today treats those as two options. You either have full control or you hand it off. Karpathy's insight is that the dial matters more than the endpoints.

Two things make this pattern work. First: trust calibration. You don't fully trust a new system on day one. You give it a small task, see how it does, and extend the leash gradually. The slider lets you do that. Second: risk matching. A low-stakes task deserves more autonomy. A high-stakes decision warrants more human judgment. The right setting depends on what you're doing, not just who you are.

Cursor figured this out for code. You start with tab completion. When you trust it more, you move toward agent mode -- multi-file edits, running tests, committing changes. Same tool, same AI, different position on the slider. That's why it works for a junior developer and a senior engineer simultaneously.

Analytics has the same problem. And nobody had solved it.

Every other tool picked a side

Legacy BI tools (Tableau, Power BI, Looker) were built for analysts. Shelf-based builders, drag-and-drop fields, total control over every calculation. Powerful. Also impenetrable to most business users.

Then came the "AI analytics" wave. Tools that promised you could just ask a question and get a chart. Great for executives, useless for analysts who need to trust the output.

Both camps assumed the answer was picking a user type and designing for them. The slider says that's the wrong frame. The answer is designing for a position on a dial -- and letting users move it.

What the slider looks like in Golden

In Golden, the slider runs across every stage of working with data. At each stage, you choose how much the AI does.

Discover

You Drive: Browse your data table, apply filters, sort columns, and slice by dimensions. Full SQL access for power users.

AI Assist: As you explore, AI surfaces relevant insights -- anomalies, trends, and correlations related to what you're looking at.

AI Creates: AI generates a comprehensive analysis report with key findings, visualizations, and recommended next steps -- all in seconds.

Connect

You Drive: Upload files, configure connections, and map columns yourself. Every setting is exposed -- nothing is hidden behind AI.

AI Assist: AI auto-detects column types, suggests table joins, and flags missing data -- you review and approve each suggestion.

AI Creates: Point Golden at a data source and it configures everything: types, relationships, and initial cleaning -- ready to explore in seconds.

Prepare

You Drive: Filter, pivot, join, and reshape with a point-and-click interface. See every transformation step and reorder them freely.

AI Assist: AI highlights outliers, suggests type conversions, and recommends cleaning steps based on data quality analysis.

AI Creates: Describe your goal -- "combine these two tables and remove duplicates" -- and AI builds the full transform pipeline.

Analyze

You Drive: Drag fields onto shelves, pick chart types, configure axes, colors, and formatting. Pixel-level control over every element.

AI Assist: AI recommends chart types based on your field selection, auto-formats axes, and suggests color palettes that match your data.

AI Creates: Tell Golden what you want to see in plain English. AI picks the chart type, maps the fields, and styles it -- then you fine-tune with point-and-click.

Communicate

You Drive: Drag, resize, and layer components manually. Snap-to-grid, alignment, and rich formatting controls.

AI Assist: AI suggests component arrangement, auto-aligns, and recommends which insights to highlight on your dashboard.

AI Creates: AI analyzes your session and generates a complete dashboard -- KPIs, charts, and narrative -- ready to present or customize.

The columns are the slider. Most users don't live in one column all day. They slide around depending on the task.

The AI never hides what it's doing

This is the part that makes the slider actually work -- and what most AI tools get wrong.

When the AI suggests a calculation, you see the formula. You can edit it. It's not a number that appeared from somewhere. It's a starting point you own.

When it surfaces an insight while you're exploring, you click it and it drops into your analysis. You can manipulate it, drill into it, build on it. It's a suggestion, not a conclusion.

When you import a data dictionary and the AI reads your context, it shows you exactly what it's planning to change before it changes anything. Every field it creates, every calculation it sets up -- you can open and modify it manually.

When it recommends data cleaning steps, it tells you what it recommends, why, and what the impact is. You accept what makes sense. Skip what doesn't.

When it proposes a dashboard, it shows you the prompt it's going to use first. You can rewrite that prompt. And the dashboard it generates is fully editable -- every chart, every layout, every word of the narrative.

AI Creates doesn't mean AI decides. It means AI does the first draft, fast. You drive from there.

The slider is for everyone

The point isn't that analysts live on the left and business users live on the right. The point is that nobody stays in one place.

An analyst who usually wants full control might ask the AI to generate a first-draft dashboard when they're in a hurry. A business user who normally asks questions in plain English might want to dig into the underlying calculation when a number doesn't look right. A founder prepping for a board meeting might want AI to create everything -- and then spend an hour fine-tuning every detail.

The slider adapts to your skill level, your time, and what's at stake in the moment. Low-stakes exploration with a familiar dataset? Slide right. First time touching this data? Slide left, understand what you're working with. Tight deadline? Let the AI create a draft and review it after.

And critically -- no black boxes, ever. Whatever position the slider is at, you can always look inside. Every AI suggestion is inspectable. Every generated output is editable. Every calculation the AI wrote is a calculation you can open and change. You're not trusting the AI blindly. You're reviewing its work, accepting what's useful, pushing back on what isn't.

That's what makes it different from every other AI analytics tool. It's not about how much you know. It's about how much control you want right now. And that changes. The slider moves with you.