[HN Gopher] Scribble-based forecasting and AI 2027
       ___________________________________________________________________
        
       Scribble-based forecasting and AI 2027
        
       Author : venkii
       Score  : 50 points
       Date   : 2025-06-30 16:15 UTC (6 hours ago)
        
 (HTM) web link (dynomight.net)
 (TXT) w3m dump (dynomight.net)
        
       | keeganpoppen wrote:
       | this is actually quite brilliant. and articulates the value and
       | utility of subjective forecasting-- something i too find somewhat
       | underrated-- extremely clearly and convincingly. and same goes
       | for the biases we have toward reducing things to a mathematical
       | model and then treating that model as more "credible" despite
       | there being (1) an infinite universe of possible models, so you
       | can use them to "say" whatever you want anyway and (2) it
       | complects the thing being modeled with some mathematical
       | phenomenon, which is not always a profitable approach.
       | 
       | the scribble method is, of course, quite sensitive to the number
       | of hypotheses you choose to consider, as it effectively considers
       | them all to be of equal probability, but it also surfaces a lot
       | of interesting interactions between different hypotheses that
       | have nothing to do with each other, but still have effectively
       | the "same" prediction at various points in time. and i don't see
       | any reason that you can't just be thoughtful about what "shapes"
       | you choose to include and in what quantity-- basically like a
       | meta-subjective model of which models are most likely or
       | something haha. that said, there's also some value in the low-res
       | aspect of just drawing the line-- you can articulate exactly what
       | path you are thinking without having to pin that thinking to some
       | model that doesn't actually add anything to the prediction other
       | than fitting the same shape as what is in your mind.
        
       | groby_b wrote:
       | At least for me, the core criticism of AI 2027 was always that it
       | was an extremely simplistic "number go up, therefore AGI", with
       | some nice fiction-y words around it.
       | 
       | The scribble model kind-of hints at what a better forecast
       | would've done - you start from the scribbles and ask "what would
       | it take to get that line, and how'd we get there". And I love
       | that the initial set of scribbles will, amongst other things,
       | expose your biases. (Because you draw the set of scribbles that
       | seems plausible to you, a priori)
       | 
       | The fact that it can both guide you towards exploring
       | alternatives and exposing biases, while being extremely simple -
       | marvellous work.
       | 
       | Definitely going to incorporate this into my reasoning toolkit!
        
         | ben_w wrote:
         | To me, 2027 looks like a case of writing the conclusion first
         | and then trying to explain backwards how it happens.
         | 
         | If _everything_ goes  "perfectly", then the logic works (to an
         | extent, but the increasing rate of returns is a suspicious
         | assumption baked into it).
         | 
         | But everything _must_ go perfectly to do that, including all
         | the productivity multipliers being independent _and_ the USA
         | deciding to take this genuinely seriously (not fake seriously
         | in the form of politicians saying  "we're taking this
         | seriously" and not doing much), and therefore no-expenses-
         | spared rush the target like it's actually an existential
         | threat. I see no way this would be a baseline scenario.
        
       | MarkusQ wrote:
       | Another useful trick: plot the same data several ways (e.g. if
       | you were playing with Moore's law you might plot (log)
       | {transistors/cm2,"ops/sec","clock speed","ops/sec/$" etc.} their
       | inverses vs time, as well as things like "how many digits of p
       | can you compute for $1", "multiples of total world compute in
       | 1970") and do the same extrapolation trick on each.
       | 
       | You _should_ expect to see roughly comparable results, but often
       | you don't and when you don't it can reveal hidden
       | assumptions/flawed thinking.
        
       | crabl wrote:
       | Interesting! My first thought looking at the scribble chart was
       | "isn't this Monte Carlo simulation?" but reading further it seems
       | more aligned with the "third way" that William Briggs describes
       | in his book Uncertainty[1]. He argues we should focus on direct
       | probability statements about observables over getting lost in
       | parameter estimation or hypothesis testing.
       | 
       | ^[1]: https://link.springer.com/book/10.1007/978-3-319-39756-6
        
       | empiko wrote:
       | To be honest, I expected the punchline to be about how randomly
       | drawing lines is the same nonsense as using simplistic
       | mathematical modeling without considering the underlying
       | phenomenon. But the punchline never came.
       | 
       | Predicting AI is more or less impossible because we have no idea
       | about the its properties. With other technologies, we can reason
       | about how small or how how a component can get and this gives us
       | psychical limitations that we can observe. With AI we throw in
       | data and we are or we are not surprised by the behavior the model
       | exhibits. With a few datapoints we have, it seems that more
       | compute and more data usually lead to better performance, but
       | that is more or less everything we can say about it, there is no
       | theory behind it that would guarantee us the gains for the next
       | 10x.
        
       | Fraterkes wrote:
       | Im sorry, I think the line scribbling idea is neat but the most
       | salient part of this prediction (how longs this going to take)
       | depends utterly on the scale of the x-axis. If you made x go to
       | 2200 instead of 2050 you could overlay the exact same set of
       | "plausible" lines.
        
       ___________________________________________________________________
       (page generated 2025-06-30 23:01 UTC)