10 Comments
User's avatar
Zac Hill's avatar

It is wild to me that data harvesting isn’t, like, a giant national security priority. Like we are getting it but are discarding it because it’s costly to store!

Jordan Schneider's avatar

because it' sannoying...

Ruxandra Stefan's avatar

This was fantastic.

Mansi Kulkarni's avatar

What I find most interesting is how the bottleneck shifts once the technology becomes more capable. We tend to focus on the model itself, but getting useful AI into the real world seems to depend much more on the less glamorous systems around it — data collection, testing, governance and continuous feedback.

The “boring” part may actually be where much of the strategic advantage is built.

Vladimir Stojanovski's avatar

Keeping the data in the buyer’s hands is only part of avoiding lock-in. The labels, corrections and reasons for rejecting a model need to survive a supplier change too. Otherwise the next contractor gets the files while the organization loses the learning that made them valuable.

Stefan Mihaylov's avatar

The Ukraine thread stuck with me - years of data collection that people on the ground didn't even know was happening, and that's what made autonomy possible. Hard to imagine DoD stumbling into that the same way.

Abhishek Pan's avatar

Readiness belongs to a dataset plus a use case, not the dataset alone. The Maven story shows it: plenty of data, little of it containing the targets. A quality stamp with no use case attached certifies nothing.

Meghan Heintz's avatar

You really let Bharat say that making AI for using tanks to kill people autonomously was about saving lives and didn't push back on that?

Jordan Schneider's avatar

well one side of the war is saving their soldiers' lives...

Meghan Heintz's avatar

I'm not joking. Would you have let it slide if Agent Orange was framed as a tool for saving lives?