AI might be the end of our careers

(baldurbjarnason.com)

3 points | by meetpateltech 3 hours ago ago

2 comments

  • spacedcowboy 2 hours ago ago

    I think you could title that article "The Luddite Manifesto".

    The Luddites weren't wrong, of course; technology ate their lunch, and I do expect AI to eat software-development's lunch as well, because AI de-skills the job as it stands. I don't think that that's necessarily a terrible thing though...

    There are two ways this could develop - I'm discounting the non-use of AI, there's too much money and momentum behind it now to matter what people think, unless governments get involved and literally make AI illegal.

    1) The demand for software engineering takes a nose-dive, because you only need a few "architect" (or "project manager" as things develop) types to oversee the AI results. Specs are currently important in swdev, but they become ever-more critical and software engineering as a subject in its own right bifurcates into academic pursuit, and practical AI management.

    2) The business need grows along with the capability. My money is on this one. I don't think it'll grow as rapidly as AI destroys the sphere, at first, but once the requirement to conform to whatever formats and capabilities are provided by software companies is removed, bespoke software will rise and people and companies will have applications customised directly to their needs, not running through some generic system that gives an 80-90% coverage of everyone. There will still be a lot of demand, but the skillset will be different, note though that knowledge will still be important, even one level removed from the engineering cliff-face. A lot of what makes GenAI useful is built into the model, but limitations are too - like the inability to use and act upon a null hypothesis - once set on a path, the GenAI will continue down that path, akin to ancient astronomers constantly adding epicycles to their reasoning ... and recognising when to abandon the path, which takes exactly that knowledge, is exactly the skill that remains human.

    Note that (2) still implies a reduction in swdev as both a desirable (and hence well-paid) career and in the quantity, it's just nowhere near as devastating as (1). It also has a possible path to grow into something new that we haven't really seen yet, in the sparse-solution space that swdev has existed in to date.

    I agree with him that AI doesn't need to reach AGI to do this damage. My optimism, though, rests on my view that current GenAI and AGI are not on converging paths. If I'm wrong about that and true AGI is on the cards, then it's simply over, swdev had a good run and it's no longer necessary. GenAI, unlike AGI, does not provide the introspection and information-extraction from sparse samples that humans do. It runs, fundamentally, through an ever-feed-forward network, without the inherent and internal ability to continually learn, and react to that learning. Stuart Kauffman would say its weights were stuck in the frozen realm of order, and that it takes a state-change to get to where the real action is - at the boundary of chaos and order.

    GenAI has made some attempts to help this - the context is read/write and can store, well, context, but even J-space (a shared "workspace" for deliberate, multi-step reasoning) is ephemeral; it doesn't inform the weights in the base model, and when we try to do weight-space programming, the model needed is far and away larger than the one being operated on, and once a change has been made the operator model is now out of sync with its original representation. We have not solved "learning", we just have a high-probability path between states in weight-space which we follow and feed back.

    It's been 30-odd years since I was playing with neural networks on a Meiko computing surface at college. Back then I used to think that intelligence was an emergent property of scale, but it turns out that recurrence and complexity play their parts too, and we've only really solved one of those three.

    Bubbles pop, but the technology doesn't disappear. The old .com bubble and subsequent crash were net positives in the long-run, creating and funding the internet we use (and "love") today. I was around when the web was born, when you had to send an email to CERN to be put on the list of websites available in this world. Things have changed. Dramatically. Sometimes even for the better :)

    So, yes, the Luddites were correct, and so is Baldur, but all is not lost, and in fact we may be on the cusp of something far greater than anything we've seen before. The key human trait that has let us get to where we are is that we're adaptable. Well, now is the time to adapt - standing in the way of a juggernaut is a good way to get crushed, as those weavers found to their cost.

  • DivingForGold 3 hours ago ago

    You can always do porn, or security ...