Hash – Complex Systems Simulation | Hacker News Reader

Hash – Complex Systems Simulation

hash.ai

108 pointscocoflunchy34 comments

This looks like a bunch of "Sales-Foo". What is the underlying algorithm, and how does it work to help business solve specific business problems? Machine learning techniques are rarely revolutionary, so what is this based off of, and what are the pros/cons of that approach?

I don't know, but this "coming soon" code base always makes me think it's vaporware.

> As much as 10-50x faster than existing best-in-class tools. We’ve flipped simulation software on its head through modern functional programming.

I'd guess they accelerate simulations on the GPU and/or over multiple cores.

Rarely revolutionary since this is certainly best practice in any serious application domain. Might make it much more accessible to implement though.

We'll see, I also remain rather sceptical.

It looks like they are using Rust, which is at least a positive.[1] Further, there are a few "wins" for agent-based modeling, like traffic congestion and supply chain optimization (wikipedia). However, in agent based modeling, you are looking for emergent behavior (i.e. simulating sheep farmers and observing a "tragedy of the commons") and require you to correctly assume the proper agent actions for given states. Further, explainability (why the model says X given conditions W,Y,Z) is pretty poor for ABM, but the solutions could still be useful in a black box way.

There's no doubt in my mind that you can run 10-50x faster ABMs than whatever tool exists using by leveraging advances like cloud computing GPUs, I'm just not sure what that actually get you in terms of problem solving capability. It's not like 50x speed like bring these models onto my phone, they will still be the domain of data scientists and subject matter experts with the patience to wait.

[1] https://www.meetup.com/Rust-NYC/events/266627924/?aff=garysg...

I imagine you can do 50x faster ABM just by tuning, writing in a compiled language, owning the low level library routines that do the work etc.

I mean, most ABM is probably sub-optimally implemented. There must be lots of “leading” software that squanders the resources it is given to run on.

So, just implement with some mechanical sympathy and you get order of magnitudes improvement.

>through modern functional programming

I find it odd that the only technology they use and mention is "functional programming", which realistically only tells me something about how their code is structured.

Your comment looks like someone who wouldn't use the service even if you understood what simulations as a service actually were. You still get your moment to express a distrust of marketing copy and a chance to vent about bad decisions you've made in the past. Enjoy.

Simulation... of what?

"General purpose simulation" doesn't really make sense to me as a concept, unless maybe there is some other meaning of the word simulation that is meant here?

I saw a talk from one of the founders just earlier this week at a meetup; it's essentially an engine that lets you model agent-based simulations and then multiple frontends that let you design them. He did several demos of their web-based UI, which could model simple things like Conway's Game of Life as well as much more complex things, like a 3D-representation of a thunderstorm in a rain forest. It was really cool!

Like Netlogo and GAMA?

From reading the link, it looks more like https://gym.openai.com/

There is a Machine Learning component that is not in NetLogo

That doesn’t actually answer the question, although the examples inch forward towards an unknown answer. Agent-based simulations of... what, exactly?

Assuming you can code an agent representing the actor or component of what you are simulating, it sounds like “what” gets defined by you, the user.

This ^^ Our starter models are pretty diverse and span real-world problems like optimizing controlled burns w/r/t wildfire containment, and improving Stack Overflow community dynamics (sorry... not a problem!)

In addition we have several dozen classic ABM toy models such as Boids, Schelling segregation, and Conway's GoL. These individually showcase different components of the system.

I'm guessing it's similar to this, at least in purpose: https://github.com/NotSoOld/OpenGPSS/blob/master/README.md

"Examples of such systems: shops (one can move among shelves with food, then stay in queue, then process at the cashbox), computer systems (transacts are electric signals or information messages), public transport (people move from stop to stop)"

Sounds like it is mostly business operations and organizational planning concerns. From the solutions page...

Here are a few problems the community is working on:

Cloud infrastructure orchestration

Sales team remuneration

Mergers and acquisitions

Price sensitivity and revenue

Warehouse operations

Fleet efficiency

Supply chain risk

Macroeconomic policy

Call center operations

Store design/layout

Critical national infrastructure

Evacuation planning

Political risk

Marketing campaign efficiency

Public health

All the atoms in the universe!

Agreed. This is one of the most bizarrely impenetrable landing pages I've ever seen. The "solutions" page is a wide-ranging list of vague topics, not solutions. Even the "demo simulations" link at the bottom provides no information without signing up.

This is the new company by Joel Spolsky, the founder of Stack Overflow (and notorious blogger), which is why it's relevant to HN.

He’s chairman, I’m not sure if it’s “by” him exactly. He seems to be describing Hash as “they” rather than “we”.[1]

1: https://www.joelonsoftware.com/2019/12/05/so-hows-that-retir...

I wonder the extent to which it builds on or learns from Joel’s Monte Carlo project estimates?

(https://www.joelonsoftware.com/2007/10/26/evidence-based-sch... 2007)

Or perhaps it’s a completely different thing, made by different people, who had no knowledge of all that?

Perhaps Joel is just the accidental sponsor like he was with Trello?

A bunch of interns working for him invent something cool and present it to him and then let him take over the world? (Not that I think that a bad thing!)

Hey William! Monte Carlo is one experiment type we support. :)

You're half right -- Jude started as an intern at Fog Creek eleven years ago and rose over a decade to become CTO. So technically he once was an intern...

But the startup originated outside of Fog Creek, and Joel's involvement is very much intentional.

Simulation of what, ffs??

I guess it's a generalized environment to simulate any (or a subset of) complex system https://en.m.wikipedia.org/wiki/Complex_system

Hey Åsmund, all. Tom's got it ^^ In theory you can model all of the above, but simulating complex environments in an agent-based fashion is the initial thing we've focused on creating a really streamlined experience for.

That was my question as well. Can this thing solve toy simulation problems like Stokes flow around a sphere, stress in a cantilever beam, or even a van der Pol oscillator?

traffic, packages moving through sorting centers, shoppers in a store, people choosing where to live, in general, the what, in agent-based modeling is just an agent, which will interact with the environment through a set of defined rules.

Agent Based Modelling, an approach for modelling complex systems - The most popular open source tech is Java based NetLogo. Here is the web version, hit setup and then run to see an example simulation of ants collecting food - http://netlogoweb.org/launch#http://netlogoweb.org/assets/mo...

I've been using cadCAD lately, might be interesting for others here: https://community.cadcad.org/

So does it replicate what Mujoco and Pybullet do? What exactly do you mean by simulation?

This is a cool idea. It seems like currently only the major players have troves of real data to feed their ML algorithms. Google has clocked how many real hours of driving for their cars? But if an upstart wants to bring a better driving AI/algorithm to the market they won't be able to because they don't have a Google-scale fleet with unlimited cash to burn on driving hours. In those cases simulation becomes a really good substitute. Driving is one example but I can imagine other cases where simulating the real word as data for an ML model is just as good as clocking the real world hours.

This is a great example. Relatedly Josh Tobin and team (OpenAI) have done some awesome recent work on how domain randomization can be applied to simulated worlds to improve the generalizability of algorithms as well: https://arxiv.org/abs/1710.06425

This is one of the game-changer use-cases of wasm. Not sure how such a product would work without that technology.

Probably just the same but a bit slower. People have been compiling to Javascript (and asm.js) for many years.

I think it's fair to say we would have chosen an entirely different architectural approach had it not been for Wasm. Whilst simulations would have run slower, our dev progress would also have taken a hit.

We're big fans and optimistic about its future :)

Reply on news.ycombinator.com