[nickpsecurity]

They apparently say it right here about the ontology approach:

https://www.palantir.com/palantir-gotham/technologies/

That's especially hilarious given that approach's failures are what led to investment in machine learning in the first place. Such approaches tend to assume precise information, variables, and rules about the world. Most problems Palantir wants to address... the hard ones... are imprecise with hidden variables/relationships. The machine learning techniques did very well on those kind of mess problems. So, research shifted.

If Palantir is using ontologies for that stuff, then that would certainly be a sign for buyers to run. I still encourage academics to look into such approaches with probabilistic, simple methods in case any advances come up. Fuzzy logic was main one in my day. Just stumbled on a claim today a drone AI did human-level performance using that. Some corroboration for R&D in underdog solutions but not production apps. Haha.

[bazqux2]

Huh, so they are open about it now. They definitely were not when I talked to them. I guess enough time has passed that people have forgotten the hard won lessons of the past.

I worked on scaling and generalizing ontologies at university and had already switched to working with Big Data / ML at a big company when Palantir tried to recruit me. I talked to some of their senior engineers about their tech and made the point that their tech sounded just like ontologies. I tried to get them to admit what it was so I could be sure I was having an honest conversation with them. They flatly denied it and made it out like the whole thing was their great new idea. I was unimpressed.

I was still interested in working for them. Access to hard interesting problems can be hard to come by. In the end I couldn't take their legendary arrogance and insecurities - to me these are bright red flags of a toxic corporate culture. And they low balled me. I would have temporarily put up with the toxic culture for large piles of money.

[nickpsecurity]

"I was still interested in working for them. Access to hard interesting problems can be hard to come by. In the end I couldn't take their legendary arrogance and insecurities - to me these are bright red flags of a toxic corporate culture. And they low balled me. I would have temporarily put up with the toxic culture for large piles of money."

Smart decision. Far as ontologies, the Cyc project to create common sense in machines was my favorite at the time. Used ontological, knowledge base if my broken memory is accurate. I was and still am firmly convinced that finding an architecture suitable to solve that problem is a pre-requisite for the AI's we really want. Deep-learning is approximating it but closer to how brain does vision than common sense. Minsky noted at one point he could count number of researchers doing common sense on a single hand or so. That's a hard problem if you want one. Also unbelievably hard to get funded. (sighs)

[bazqux2]

Thanks. There is some interesting work being done with Deep Learning mixed with Bayesian Models and external databases of facts. The training apparently ungodly slow though. My work these days is visual so CNN works well enough. I'm not an authoritative expert in this field though so I'm taking 6 months off to study it intensely, after which I hope to be able to make some meaningful contributions.

[argonaut]

There is nothing wrong with the ontologies approach. Palantir is not aiming for a fully automated approach. Palantir's core product is a an entity graph mostly used and manipulated by humans for analysis. Basically their software is only meant to augment human analysis.

That being said, from my conversations with them, they also have a traditional machine learning team for whenever that approach is needed for a product. But their core product is meant to only help analysis that is mostly done by humans.

[bazqux2]

There is usually something wrong with the ontologies approach as it rarely works. There is roughly two decades of evidence for this for anyone who cares to look. Five decades if you loosen the definition to include the family of logic and constraint programming - see AI Winter. There is nothing new about these ideas. It always looks and feels like it's going to work which is why humanity has persisted with it for so long and will likely continue to persist for some time to come.

There is a whole generation of better techniques that have come out of machine learning that totally eclipses ontologies and I know Palantir isn't using them. Their corporate culture isn't set up for fostering that kind of applied research.

No-one is advocating for a fully automated approaches. I don't know where that notion came from.

In my view is that Palantir is a consulting company that is pretending to be a tooling company. And their consultants are not worth the money they charge. Just one of many Silicon Valley based frauds.

[dredmorbius]

Is this ontologies within the field of AI not working, or more generally?

Do you have references to any specific discussions on this?

Curious as I'm doing some work of my own (well outside AI) in which developing ontologies strikes me as useful, though I'd prefer not falling into any well-worn traps.

(My use is largely comping up with useful descriptive models of otherwise hairy concepts.)

[nickpsecurity]

What bazqux2 said is accurate. I'll go further to say that the kinds of work Palantir is involved in is mostly probabilistic. Especially intelligence work. So, use of models requiring certainty or straight logic in areas rife with uncertainty & degrees of truth seems set up to fail outside easy inferences. One can encode the logical stuff in probability models but harder to do reverse. Hence, their underlying tech should be probabilistic, fuzzy logic, or something similar for best results instead of just some results.

Far as ontologies in general, they have a mixed, track record. They take a lot of work to create. Then, they have to be mapped to real world inputs and outputs. One way they got applied is so called business rules engines or business process management. It's like a subset of ontology approaches of past. Here's a company that uses the real thing for enterprise software with Mercury language for execution part:

http://www.missioncriticalit.com/development.html

Also, Franz Inc, of Allegro Common LISP, covers many of the same use cases as Palantir with their ontological tooling.

http://allegrograph.com/solutions-by-use/

So, there's definitely companies using it for long periods of time for real-world, use cases. Palantir just seemed to be mixing it with hype and secrecy to maximize their sale price later. ;)

[bazqux2]

I meant generally they are generally not useful. Sometimes they are. It depends on the purpose and what you want to build and who it's for.

Given that you're building a descriptive model it would depend if you're working with facts or with probabilities. If it's facts then Ontologies should work fine, for probabilities I'd recommend Bayesian techniques.

The input for these are usually small. From the sounds of it you're generating the input yourself so you should be safe.

[dredmorbius]

An ontology of technlogical mechanisms (or dynamics):

https://ello.co/dredmorbius/post/klsjjjzzl9plqxz-ms8nww

Particularly in economic and policy discussion, technology is just "technology". A black box. In economics, Solow's Residual is described, by Solow, as "the measure of our ignorance" of factor productivity growth influences -- it's quite literally, statistically, what's left over after accounting for labour and capital.

I see a few quite evident classifications which strike me as useful:

1. Fuels. Apply more energy to something, it tends to happen faster. Wood, plant and animal oils, fossil fuels, nuclear fission, possibly fusion.

2. Material properties. Some things are highly dependent on specific material properties. Conductivity of gold, silver, copper, and aluminium. Ferromagnetism. Hardness of diamond. Softness of graphite. Semiconducting of silicon. Fertilising properties of nitrogen, phosphorus, and potassium. Many others. Point being, you're now locked into availablity and other properties of that material.

3. Specific process knowledge. What used to be called "arts". Most of what's now considered "technology", from agriculture to zymurgy (though zyumurgy's actually fairly close to agriculture...). These approach theoretical efficiency limits.

4. What seem to be dendritic or web structured aspects. Computer chips and Moore's law are today's classic example, but I'd count communications, transport, and trade networks, cities and urbanisations, knowledge itself, and other elements among these. What they have in common is an increasing rate of progress with greater accumulation, modulo retarding factors.

There are several other elements. Sensing and measurement increase various capabilities -- navigation and fine metal machining come to mind. Symbolic processing, from speech and writing to abstract maths and programming. Organisation -- of people, states, business, and finance.

The final element, and one which popped out at me whilst devising the ontology, was the concept of hygiene or pollution factors. They're a distinct class of phenomena which if not addressed tend to put a damper on further growth, everything from infectuous disease in cities to heavy metal pollution, salination of croplands, traffic congestion, spam and fraud in communications and business networks. It's a superset of common categories such as "pollution" or "disease" or "social breakdown".

Anyhow, that's what I'm working on. I find it a useful organising tool, still developing the idea.

[nickpsecurity]

1. In semiconductors, we get more out of stuff when we put in less energy due to shrinking the transistors. Even increasing transistors in same node doesn't always result in more work since bigger chips have slower clock rates. I think you need to look at inputs, which include time, more than fuel given it doesn't apply to a lot of things. Even human body which, as you increase fuel, will work slower due to being gorged and then die with exploded stomach.

2. This is true. It's worth noting such dependencies.

3. Elaborate on that.

4. That's true. There's a lot of work on that topic already that you can draw on. I remember some showing that how the cities grew was similar to how bacteria looked. Weird stuff.

Re waste. You can model it as a separate thing that goes up when certain actions happen, then starts bringing them down. Definitely should be considered.

[argonaut]

Again, Palantir is not an AI company. They are a data visualization and analytics company. So all your perfectly fine points about ontologies and AI winter are not relevant.

[bazqux2]

It is an ontology company - see their website. This is how they derive their analytics and visualizations. So my points are relevant.

[willalden]

Hey -- I wrote that BuzzFeed article in May. I'm always looking to learn new things. If anyone on here wants to chat and compare notes, off the record, please don't hesitate to reach out: [e-mail]