High Signal aims to provide deep, applied data science conversations for practitioners who lack a good source of practical advice.
2
Data science teams often fail to deliver expected returns because infrastructure, research, and business execution all have to work together.
3
Generative AI assistants can handle code and SQL today, but data scientists still need agents that can explore data, build models, and support the full experimental workflow.
Summary
Demetrios Brinkmann, Hugo Brown Anderson, and Duncan discuss why they created the High Signal podcast. Hugo says the show is intended to examine what data science learned during the 2010s and what changes in the age of generative AI and large language models. Duncan describes a gap between deep academic material and shallow, highly applied blog content. High Signal aims to provide material that is both deep and useful to practitioners. The guests also discuss why data science has produced strong returns at some companies and weak returns at others. Successful work requires usable data, infrastructure, exploratory research, and delivery into business operations. They argue that many data teams are still treated as service centers, which limits their ability to find valuable problems. Current copilots help with SQL, code, and notebooks, but they do not handle the experimental nature of data science. Delphina is betting on agents that can support exploratory analysis, data preparation, modeling, and iteration.
High Signal fills a gap between academic depth and practical data science
Hugo Brown Anderson says the podcast grew from a need to reassess data science after the 2010s and the arrival of generative AI. Duncan describes the content market as a two-by-two of depth and application. Academic work can be deep but hard to apply, while much of the data science blogosphere is applied but not deep. He says this leaves a gap for content that is both practical and technically substantial. The podcast is aimed at practitioners who need advice that can materially change their work.
Intelligent infrastructure needs statistics, computation, and economics
Hugo says Michael I. Jordan argues that current generative AI systems are good at capturing crowd knowledge but poor at reasoning under uncertainty and expressing uncertainty. Jordan's idea of intelligent infrastructure requires more than statistics and computation. It also needs economics and economic principles. Hugo mentions the phrase "recycled intelligence" as a way to describe systems trained on existing human-generated data. The podcast uses this discussion to question how AI systems should be designed and evaluated.
The early High Signal episodes connect technical work with organizations and business
The first episodes cover statistical concepts such as data-generating processes and simulation, organizational structures that help teams ship products, online experimentation, and data foundations. Hugo names conversations with Andrew Gelman, Kiara Faron, Ramesh Johari, Gabriel Weintraub, and Hillary Mason. Duncan says the discussions have reinforced how early data science still is across industries and countries. They also cover how academics should teach inference, experimentation, simulation, and noise rather than presenting data science as a finished discipline.
Data scientists still spend too much time on preparation
Duncan says most major companies now have data in cloud warehouses, which makes serious data science more possible than it was a decade earlier. He cites a recent survey in which 40% of data science time still goes to preparation, including SQL writing and cleaning data. He describes Delphina's work as an effort to make that routine work easier. His larger goal is for data science to move from a cost center to a business function that identifies valuable problems, manages initiatives, and eventually drives business units.
The data science workflow is more experimental than ordinary software development
Duncan says software copilots are widely used, while data science has a harder workflow. A data scientist may start by trying to predict churn without knowing whether a useful model can be built in a week or a month. Current assistants can write SQL and set up a notebook, but they do not carry out the open-ended investigation. He expects future agents to help with this work. Delphina's bet is that AI-powered agents can make data science move faster.
Data science has to connect infrastructure, research, and business delivery
Duncan explains why returns from data science vary so much between companies. High-value applications need clean data and data infrastructure, researchers who can do exploratory work, and a way to deliver the result into business operations. Research is slow, exploratory, and uneven in quality, so hiring a team does not automatically solve the problem. He calls these parts the four legs of the data science chair. Many companies have not built all of them, which makes it difficult to reproduce results such as the value attributed to Amazon's recommendation system.
Data teams need space to discover problems instead of only answering tickets
Hugo discusses Eric Colson's criticism of data functions that become service centers. A ticket-based model can help teams answer requests, but it does not give data scientists enough time to explore or participate in business discussions. Hugo gives the example of customer segmentation, where customer teams may ask for segments that do not create business value. A data scientist in the meeting might use clustering to identify more useful groups and recommend different actions for each one. The point is to let data scientists help find valuable work rather than waiting for assigned tasks.
Agents for data science remain harder than code assistants
Duncan says early agents work best on text-heavy tasks such as sales and customer support. Software engineering agents are more difficult because they must work through large codebases and existing systems. Data science agents face a similar problem because they must understand complicated data, models, and analyses. He says generative AI can perform this work in isolated examples, but the harder challenge is guiding it through the complexity of real business situations.
"A far better model is making sure data scientists have time and space to explore, to think, and have their ideas bring them into business meetings."Hugo Brown Anderson23:36
Who should watch
You are deciding whether to start or follow a data science podcast and want conversations that connect technical methods with real organizational work.
Your data team spends much of its time on SQL, cleaning, and service requests, and you want to understand what a broader operating model could look like.
You are evaluating AI agents for data science and need a clear account of why exploratory analysis and business context are harder than code completion.
