Oct 27, 2023 · 43m · latent-space
Powering your Copilot for Data - with Artem Keydunov from Cube.dev
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Cube co-founder Artem Keydunov joins hosts Swix and Alessio Fanelli to discuss how semantic layers bridge the gap between large language models and tabular databases. He details the technical architecture, historical lessons from StatsBot, and software engineering best practices necessary to build production-grade AI data copilots.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 21.1% of the talking time here. How this is scored →
The hosts as informed peer 6.1 Guest teaching 4.0 Guest disagreement 1.3 The hosts pushing back 2.3
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 16:27 Org chart conflict challenge
Swix directly confronts the clean narrative of semantic layers, arguing that conflicting cross-team metric definitions lead to shipping an org chart into the code.
Hardest push from the hosts ▶ 16:27 Swix presses on conflicting departmental metricsSwix refuses to treat semantic modeling as purely technical, forcing Artem to address how real-world political friction between finance and sales impacts metric definitions.
Biggest teaching moment ▶ 13:30 Artem details SQL fan traps and semantic constraintsArtem breaks down the exact database pitfalls like fan traps and chasm traps that cause direct LLM-to-SQL generation to produce silent analytical errors.
The host holds their own ▶ 16:00 Alessio articulates metric abstraction mechanicsAlessio demonstrates his technical grasp of data modeling by succinctly illustrating how the semantic layer abstracts multi-table joins into clean single-metric queries.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| The Origins of StatsBot and Early Text-to-SQL Limitations | 6 | 3 | 1 | 1 | Alessio and Swix demonstrate solid contextual knowledge of early text-to-SQL products and the 2016 chatbot era. Swix bonds over using regular expressions for parsing financial data while Artem explains StatsBot's origins. | |
| The Evolution of Cube and the Fundamentals of OLAP Cubes | 5 | 5 | 1 | 1 | Swix prompts Artem to define OLAP and multidimensional cubes for the audience, establishing baseline definitions. Artem explains the distinction between 2D relational SQL tables and multidimensional metric cubes. | |
| Bridging LLMs and Tabular Data via the Semantic Layer | 7 | 6 | 1 | 2 | Artem gives an in-depth breakdown of how LLMs fail at complex SQL joins (fan traps, chasm traps) and how semantic layers act as structured context. Alessio contributes concrete architectural summaries to ground the explanation. | |
| Treating Data Metrics as Code to Resolve Organizational Discrepancies | 6 | 4 | 2 | 6 | Swix pushes back on the idealistic view of semantic layers by raising organizational dysfunction and conflicting departmental definitions of core metrics. Artem addresses this by arguing metrics should be governed as code with pull requests. | |
| Data Copilots versus Natural Language BI Interfaces | 6 | 3 | 2 | 3 | Alessio queries how BI vendors can differentiate if natural language query interfaces become commoditized. Artem explains that BI fragmentation will persist because natural language query is just another commodity visual interface like bar charts. | |
| AI Agents, Smart Summaries, and the Modern Data Stack Evolution | 7 | 4 | 1 | 2 | Alessio draws parallels to Linus's UI ideas at Notion and discusses modern data stack automation across dbt and ETL pipelines. Artem provides a measured assessment of where LLMs genuinely accelerate data engineering versus standard consolidation cycles. | |
| Technical Nuances of Data Chatbots and Production Stacks | 6 | 4 | 1 | 1 | Artem elaborates on practical production implementations, emphasizing that effective data bots ask clarifying follow-up questions and rely on external Python logic for accurate arithmetic rather than raw LLM generation. | |
| Embedded Analytics Monetization and the Lightning Round | 6 | 3 | 1 | 2 | Swix raises the monetization struggle in embedded analytics, prompting Artem to analyze market saturation and custom build dynamics. The segment transitions into rapid-fire lightning round questions on foundation models and AI engineering practices. |
Statements from this episode (10)
Keydunov: The semantic layer is the ideal repository for AI data context
“So our take of that and my take is that semantic layer is just really good place for this context to live because you need to give this context to the humans. You need to give that context to the AI system anyway, right? So that's why you define metric once an…”
Keydunov: Querying a semantic layer over raw SQL reduces LLM errors
“Now, the query, I believe, should be created against semantic layer, because it reduces the room for the error, because what usually happens is that your query to semantic layer would be very simple. It would be like, give me that metric grouped by that dimens…”
Keydunov: Semantic layer metrics should be managed as code via pull requests
“So I think we should treat our metrics and we are in semantic layer as a code, right? And then collaboration is a big part of it. You know, like if there are like a multiple teams that sort of have a different opinions, let them collaborate on the pull request…”
Keydunov: Natural language AI will not consolidate the fragmented BI market
“I don't think it's going to be something that will change the BI market, you know, like something that will can take The BI market and make it more consolidated rather than, you know, like what we have right now. I think it's still re will remain fragmented.”
Keydunov: Modern data stack will not be dramatically disrupted by AI
“I don't see a lot of being compacted by AI specifically. I think, you know, that space is being compacted as much as any other space in terms of, yes, we'll have all this copilot capabilities, some of AI capabilities here and there, but I don't see anything so…”
Keydunov: The modern data stack will enter a consolidation cycle
“I feel like right now we should go through the cycle of consolidation. And you know, like I mean, if Fivetrend and dbt merge, they can be alter-ex of a new generation or something like that. And you know, probably some ETL tool to there. But I feel it might ha…”
Keydunov: Production AI data apps require extensive Python scaffolding around models
“When we're talking about production level use cases, it's quite a lot of Python code around, you know, like, your model to make it work, to be honest. It's, like, It's not that magic that you just throw the model in it, like it can give you all this answers. F…”
Keydunov: Data-driven AI apps should start on a proper warehouse from day zero
“I would just recommend going through to warehouse as soon as possible. I think a lot of people feel that MySQL can be a warehouse, which can be maybe on like a lower scale, but you know, like, definitely not from a performance perspective. So just kind of havi…”
Keydunov: Large enterprises abandon embedded analytics vendors to build in-house
“If you look at the embedded analytics market, the bigger organization, the big gets, They're really more custom, you know, like it becomes, and at some point I see many organizations, they just stop using any vendor, and they just kind of build most of the stu…”
Keydunov: AI will not make embedded analytics easier to monetize
“I don't think AI really going to change that just because it's using model, you just pay to open AI and that's it. Like everyone can do that, right? It's not much of a Competitive advantage. So it's going to be more like a commodity features that a lot of like…”