home.social

#openlineage — Public Fediverse posts

Live and recent posts from across the Fediverse tagged #openlineage, aggregated by home.social.

fetched live
  1. CW: On AI models driving layoffs in tech

    @davidculley that’s a conundrum I’ve been going through myself lately.

    Those of us who actually enjoy to solve problems, build solutions and automate things through software would also like to do it for a living, as engineering jobs have been high-demand and well-paid for a quite long time.

    However business-school Dunning-Kruger tax-cheater borderline-sociopath types have taken over most of the management positions in our industry in the past decade. So the most common option (build software for a salary) nowadays means having to interact mostly with that kind of people, rather than early Silicon Valley geeks who genuinely enjoy tinkering and advancing the world through technology. And IMHO it doesn’t make much of a difference to work on a permanent job or as a contractor (indeed, a contractor position usually gets all the short-termism issues even more amplified, while losing the relative financial stability).

    A thought that keeps buzzing through my mind is that of starting my own business - Platypush for instance is big enough and with enough added value to have a small company built around it. But, besides the downsides that that normally entails (like the loss, at least in the short term, of a stable income that can pay your mortgage and bills, and having to focus more on business/bureaucracy problems of kickstarting and running a business and marketing the product rather than the fun part of building stuff), what discourages me the most is that such a company would have to also build a sound business model that can compete with those leveraged by the unethical sharks out there, without compromising too much on our values and ethos.

    A sensible trade-off may be to work for a large company that is (at least for now) “less evil” than others, and try to carve your own space inside of it in such a way that you can give your little push of a tech giant in a more ethical direction. That’s what e.g. @timnitGebru tried to do at Google, by pushing the company’s AI plans towards a more ethical and sustainable direction. That’s also what I’m currently trying to do at Booking, by building solutions that create consistent data lineages in order to increase the accountability and “explainability” of data+decision flows that often involve dozens of different frameworks. And I’m trying to do that in the open whenever possible, by contributing to the #OpenLineage project. Giving back to the community is the least you can do when you work for a business with a multibillion revenue. Such a decision obviously comes with its risks, at Timnit knows quite well: the company you work for can get rid of you if enough business school jerks are called into a room and decide that your project / line of research affects too much the bottom line of “their” business.

    An even better bet may be to reach out to employers that built their business on strong values and are unlikely (at least on the short/medium term) to enshittify, lay off thousands of employees after the CEO got galvanized by a demo of a new AI prototype, or switch to a user data hoarder-and-reseller business model ( #Nextcloud and #Fairphone B.V., just to name two such business on top of my mind, or also Blender). But that usually means lower salaries than big-fish competitors, and probably an impact on a more constrained niche. I hope that this will change in the future though. If both enough engineers and enough users get sufficiently disgusted by the most common business models in the industry they work for/consume, then sufficient demand may push the rise of even more companies focused on alternative models and a stronger ethos. And institutions like the #EU can definitely do more - rather than advertising “European tech champions” along the lines of US and China, they should probably foster (and directly fund much more generously than they currently do) an alternative ecosystem based on sustainability, openness and fairness. By picking the will, not the winners. A combination of a steady demand/supply market for alternative solutions and more direct funding can also increase the salaries that these companies can offer to their engineers, thus contributing to a bigger “drain” of talent away from today’s Big Tech firms.

    For now, the more I think of all these trade-offs, the more powerless I feel, the more I contemplate leaving the industry altogether, move from the big city to the countryside, and start growing my own watermelons for a living.

  2. @Lobrien a good read that outlines some of the challenges towards transparency/explainability of ML models.

    IMHO these problems can be fit into three categories:

    1. What data the model was trained on.

    2. Why a certain model with a certain architecture trained on that data made a certain decision.

    3. What could go wrong when we deploy a model with these characteristics into the real world.

    In my current job I’m focusing on solving problem (1) (which is also the one this article focuses on), because I believe that you can’t answer the remaining two questions until you get a grasp of what data was actually used to train the model.

    As part of the #OpenLineage initiative, our goal is to automatically grasp lineage data from all the processes used in an ML pipeline. A particular picture, a db table or a data lake collection all need to have a unique identifier, and all processes that consume them or produce them should automatically report those dependencies. The only way to explain the decisions of a model or study its potential biases is to have a full picture of all the data consumed and generated throughout the whole pipeline - and that often includes any intermediate engineered features as well.

    I feel like such an architecture should be an industry standard enforced by external regulation - and I hope that sometimes we’ll get to that point.

    Unfortunately, such ideas have also come only when the genie is already outside of the bottle. Unfortunately, there are already huge training sets out there built on scraped data and put together absolutely without any supervision. Unfortunately, large models have already used them.

  3. OpenLineage soll einen offenen Standard zum Erfassen von Meta- und Stammdaten setzen, um die Steuerung von KI- und Datenprojekten zu vereinfachen.
    Meta- und Stammdaten: OpenLineage rückt in die Sandbox der Linux Foundation