Gideon Mann

The through-line

Mann's consistent argument is that data science creates value only when it is close to the problem. He makes this case in two registers that turn out to be the same case: commercially, where the interesting shift is from selling data to people toward selling it to machines [1], and in the public and nonprofit sector, where he insists that the hard part is not the modelling but the long, embedded work of figuring out what the problem actually is [1]. Running alongside that is a conviction that no single actor can solve the problems worth solving, so the practical question is always about interfaces, incentives and partnerships between organisations that hold different pieces of the information [1]. His later work sits at the firm level, where he has described how a global investment firm approaches AI and how his team is organised around it [2].

On data science inside a data business

Mann describes Bloomberg as doing "a surprising amount of data science inside of the company," with "many more than a few handfuls of data scientists machine learning n people" working on analytics [1]. His worked example is news sentiment: as a story breaks, is it positive or negative with respect to the companies named in it [1]. The client value he identifies is not the event itself but the reaction to it, since customers "want to know not just what's happening but also how is the market reacting and what are people thinking about it" [1]. The same analytic output appears in two forms, inside the terminal and as a data feed, and Mann flags the second as the more consequential development: "an interesting kind of emerging model for our business is how do we provide data not just to people but also to machines" [1].

On applying data science to city problems

Mann's favourite illustrations of civic data work are cases where an ordinary administrative record becomes an inference. To find overcrowded buildings, compare stated occupancy against water flow rates, since consumption tells you how many people are probably there and the permit tells you how many are supposed to be, and a sharp mismatch is what sends an inspector [1]. To find restaurants illegally dumping kitchen grease into storm drains, combine GIS locations of restaurants with the list of premises licensed to have waste carted away, then look at which unlicensed establishments sit near a drain that clogged and overflowed [1]. He traces this lineage to the mayor's office of data analytics and to the difficulty of getting city agencies, which "typically have difficulty talking to each to one another," to share data at all [1]. What Works Cities, in his description, is the attempt to take that way of thinking and spread it [1].

On why embedding beats posting the data publicly

Asked directly whether the data should simply go up on Kaggle or a similar contest platform, Mann gave the sharpest statement of his method. He points to the data for social good efforts that have worked and observes that "the way that they're successful is long-term engagements with nonprofits or with public sector organizations," with one organisation pushing for "an even tighter relationship where someone goes in and is embedded inside of an organization for a long time period" [1]. He gives two reasons. One is that some data "simply cannot leave for privacy reasons" [1]. The other is more fundamental: "often is part of the problem is figuring out what the problem is and in order to really understand what's going to add value to the organization you need to be in the middle of it you need to be talking to people you need to have that conversation and that can't happen once you get to the point of putting it on kaggle" [1]. He does not treat competition platforms as worthless, only as downstream. His stated hope for the UNICEF researcher-in-residence role is precisely that an embedded person will "unlock all of these problems inside of Unicef and then make them more accessible to data kind and to the community generally to attach on to" [1]. The requirement is heaviest for organisations without in-house capability: most nonprofits do not have data science teams on the scale of a Bloomberg, so they need a continuing long engagement rather than a handoff [1].

On partnership between the private and public sectors

Mann's view of cross-sector work is that neither side has enough information to solve the problem alone, an argument he attributes to Aneesh Chopra and endorses in his own terms: "it would be arrogant to think that that either the government could solve all the problems or that the private sector could solve all the problems it really has to be a partnership" [1]. What makes such partnerships work, in his account, is "figuring out the right uh interface points and the right incentives to all of the players," which he treats as the key to effective public sector interventions [1]. He extends the same logic inside the corporation, arguing that employee volunteering should be structured as long-term relationships in which staff contribute to a partner's core mission rather than one-off events [1].

On convening the field

Mann treats gathering people as an intervention in its own right. The pre-conference held alongside KDD drew roughly 900 people and was organised into tracks on urban computing, data science for social good, and data frameworks [1]. Its successor, the Data for Good Exchange, was set deliberately in New York data week alongside Strata Hadoop, with the top papers given visibility there, a best paper award run with NYC Media Lab, a poster session, and no cost beyond registration [1]. The call for papers was targeted at four strategic areas, government innovation, education, public health, and environment, chosen to match what the philanthropy actually supports, with Bloomberg Philanthropies giving direction on priorities [1]. He also uses the platform as a labour market, telling the audience which speakers were hiring data scientists and urging submissions from anyone doing work reasonably related to the topics [1].

Takeaways

• The commercially significant shift in data products is from serving humans to serving machines: "how do we provide data not just to people but also to machines" [1].
• News analytics earn their keep by answering the second-order question, how the market is reacting to an event, not just what happened [1].
• Civic data wins come from joining two mundane records, occupancy versus water flow to find overcrowding, restaurant locations versus waste-carting licences to find illegal grease dumping [1].
• Publishing a dataset to a contest platform skips the hardest step: "part of the problem is figuring out what the problem is" [1].
• Embedded, long-term placements work because some data cannot leave the organisation for privacy reasons and because value definition requires being in the room [1].
• Effective public-private work is a design problem about interface points and incentives, on the premise that neither side alone has enough information [1].
• Structure corporate volunteering as sustained contribution to a partner's core mission rather than as one-off events [1].