▲
3
0
▼

Rin has reversed his position on Machine Learning/AI


 invite response                
2020 Feb 17, 3:54pm   2,833 views  35 comments

by Rin   ➕follow (8)   💰tip   ignore  

Earlier on this thread ...

https://patrick.net/post/1327438/2019-09-23-the-rin-yang-2030-initiative-fire-80-of-white-collar-workers

I was upbeat about ML/AL.

I've reversed my position on Machine Learning/AI and its influence on our current business environment.

Originally, I was under the impression that intelligent ppl were in charge of implementing ML to solving real business problems and delivering results for their shareholders and C-level suites.

Over time, I've realized that these were not intellectuals but MBA-ologists, useless shitheads who use terms like synergy, value-added, etc, without understanding what those things really mean. What's happening is that a lot of corporations are throwing cash at ML/AI, because of the fear of missing out (FOMO) against their competitors.

In other words, they don't have real business cases to solve. Ones which could make 'em money.

So in the past 6 months, I've been to conferences around the country and I've seen countless MBA-ologists, talk about ML/AI and getting customer support centers, etc, to be automated and up to snuff. I asked myself (and a handful of them) ... 'Aren't those some $16-$23/hr jobs?' I mean most companies can afford to keep lowly paid staff. What about those so-called senior business analysts (BA), living in corporate silos, earning $35-$60/hr, doing dubious work like 'revenue recognition', 'forensic analytics', and other activities with hokey metrics?

And then, when you actually meet those BA retards, all they have are some Excel sheets with product codes, a few numbers associated with them, and then, they click macros for future expectations. In other words, many of these jobs are easily automatable but hidden because they sit on a person's PC and don't train others on how to use it. And then at meetings, they don't share how the metrics are derived.

With the above in mind, as a gestalt, I predict an "AI winter" will come 2022 or at latest 2024. When companies follow the herd and toss money at 'Bridge to Nowhere' projects, eventually, a crash ensues shortly afterwards.

The only person who's using ML/AI to produce real results (which are product enhancing), is Matt McMullen of RealDoll who's making his Harmony product a chatbot which can really interact with a horny guy. I predict great things for sex dolls but not for corporate America.

Comments 1 - 35 of 35        Search these comments

1   Tenpoundbass   2020 Feb 17, 4:44pm  

Rin says
And then, when you actually meet those BA retards, all they have are some Excel sheets with product codes, a few numbers associated with them, and then, they click macros for future expectations. In other words, many of these jobs are easily automatable but hidden because they sit on a person's PC and don't train others on how to use it. And then at meetings, they don't share how the metrics are derived.


PMs and Department managers gave them the spreadsheets, and they weren't ever in one single meeting. Because women and minorities needed to be empowered.
They hand the hard part off to someone smarter, though after they have shat on it and made a complete nonsensical mess out of it.

Update with my situation, our company got hit with a Ransomware attack. We've been down a week since this past Friday.

My code is all gone, but if they ever get the web server back up, I can try to rebuild the projects using reflector, or we're at a crossroads, where they can finally cut me lose and push harder to get on Netsuite.
2   Rin   2020 Feb 17, 5:26pm  

Tenpoundbass says
They hand the hard part off to someone smarter, though after they have shat on it and made a complete nonsensical mess out of it.


This is called the Non-Pareto Principle where 80% of the dolts make most of the work for the remaining 20%, who have to get it done.
3   CBOEtrader   2020 Feb 17, 6:15pm  

Every company peppers in AI and ML into their buzz word salad these days.

Use computers to run statistics? Engage in digital marketing? Scaling automated systems? Solving old business problems w new technology?

If a company can answer yes to 2 of the above, some MBA asshat will pitch the company as an AI/ML tech play.

^^somenof these companies are cash cows, regardless of how little AI or ML actually does for them.

Ex : http://www.investor.prudential.com/news/press-release-details/2019/Prudential-Financial-to-acquire-Assurance-IQ-Inc-a-leading-consumer-solutions-platform-for-health-and-financial-wellness-needs-for-235-billion/default.aspx

This company runs dinky as fuck digital ads at mass scale, then hand off the insurance leads they generate to insurance agents who are contract workers and therefore login to sell whenever they want. They are huge AF... but calling them an AI/ML insuretech play is just stupid. Assurance is a scaled modern business. They are not an AI insuretech gamechanger.

Big difference imho.
5   CBOEtrader   2020 Feb 17, 6:18pm  

Assurance ad flow sample pages... not exactly high tech shit.
6   Bd6r   2020 Feb 17, 6:19pm  

In my field of research "machine learning" is the new buzzword. I have not yet seen any significant discoveries with people applying "machine learning" to difficult research problems.
7   CBOEtrader   2020 Feb 17, 6:21pm  

rd6B says
In my field of research "machine learning" is the new buzzword. I have not yet seen any significant discoveries with people applying "machine learning" to difficult research problems.


We used it a lot in trading. It's just statistics w a lot of computer muscle. So what? It's just a damn technique. It still requires a genius to direct that technique towards a genius end. Otherwise it's just nonsense data.
8   CBOEtrader   2020 Feb 17, 6:26pm  

https://deeplawfirm.com/about.html

Heres one :) gimmick or $billion AI tech firm? You decide
9   Rin   2020 Feb 17, 6:29pm  

CBOEtrader says
We used it a lot in trading. It's just statistics w a lot of computer muscle. So what? It's just a damn technique. It still requires a genius to direct that technique towards a genius end. Otherwise it's just nonsense data.


Well, it requires 'know how' and in most cases, it applies more to risk management than profit taking as most profitable trades are exactly that, highly profitable, however, if one can manage risk, then the profits take care of themselves.
10   Tenpoundbass   2020 Feb 17, 6:44pm  

CBOEtrader says
It's just a damn technique.


That's what I've said since the beginning. It's just a new programming pattern and new algorithms.
At the end of the day it's still a bunch of loops with nested if statements inside.


I modeled a process using an outcomes rubric. Were decisions were made by a weighted rankings.
At the end of the day the data has to be populated with human input. There's just really no way to program wisdom.
Where the process would make new connections based on previous input, and derive a new conclusion for a problem.
So instead the process relies on impressions, input and stimuli to make new records for learned behavior. Not by analyzing data and intelligence that is already has to make new records for learned behavior.
Basically it can learn to sense it's Hot, and move away from the heat source. It just means it wont move near that spot any more. And marked it with gps coordination and noted the temperature. But those were behaviors that had to be in the initial data load. The AI subject without any instructions given Thermal and Temperature sensors, and ways to move around. If you started a fire next to it, it would not create a new set of records to compel it to move from its spot. Those thermometers and wheels are already programmed to work together.

Even bacteria will move away from a heat source. That's not intelligence.
11   Bd6r   2020 Feb 17, 6:45pm  

CBOEtrader says
We used it a lot in trading. It's just statistics w a lot of computer muscle. So what? It's just a damn technique. It still requires a genius to direct that technique towards a genius end. Otherwise it's just nonsense data.

in my field, people were trying to feed data to computer about a set of reactions and to predict the outcome of an unknown one, or to develop a new route to some compound based on a dataset. It works for close analogies, which a chemistry grad student can predict without a computer. So far nothing new has been discovered despite massive posturing.
12   Rin   2020 Feb 17, 6:51pm  

rd6B says
It works for close analogies, which a chemistry grad student can predict without a computer. So far nothing new has been discovered despite massive posturing.


For our hedge fund work, a lot of that was based upon control theory, a pre-existing discipline, before ML became popular.

The difference is that ML is attempting to develop a type of control algo, without the modelling aspects, which while it may seem innovative, doesn't change the fact that it's only trying to do what engineers have been doing for decades.
13   Tenpoundbass   2020 Feb 17, 7:03pm  

I think a big part of AI computing is missing from the research to even create an intuitively aware and astute AI. The focus is on the brain, retention and storage.
The Gut is missing. The Gut is life's second brain.
I modeled a Computer with a gut, bacteria, gases, acids and alkaloids all making reactions that sensors react to to stimulate AI responses.
A programmable gut that could be programmed to sense when something isn't right and create an upset stomach, or even butterflies when something exciting is about to happen.
These sensors could be triggered before other telemetry or optical sensors can even have to sense or react. Or react on stimuli those sensors may have missed.
14   Rin   2020 Feb 17, 7:11pm  

Tenpoundbass says
The Gut is missing. The Gut is life's second brain.
I modeled a Computer with a gut, bacteria, gases, acids and alkaloids all making reactions that sensors react to to stimulate AI responses.
A programmable gut that could be programmed to sense when something isn't right and create an upset stomach, or even butter flies when something exciting is about to happen.
These sensors could be triggered before other telemetry or optical sensors can even have to sense or react. Or react on stimuli those sensors may have missed.

Ok, but we're talking about loser MBA-ologists; these ppl think that 'the gut' is some so-called synergy between two depts who hate each other.

The current business environment is setting up all new ML applications for a severe bust. I predict that that bust is near.
15   Booger   2020 Feb 17, 7:11pm  

Tenpoundbass says
My code is all gone, but if they ever get the web server back up


No backup on a home computer???
16   Tenpoundbass   2020 Feb 17, 7:18pm  

Booger says
No backup on a home computer???


I have my projects at home, the new CIO/COO guy was supposed to take care of all of that. There must be at least 8 Corporate In Yo Bidnessware icons in the system systray. If they weren't backing up every file I saved on my computer somewhere on the network. That's on them. When he came we were using TFS, the new network guy killed that server.
I'm not going to second guess them, when all of the Watchware apps started showing up in the systray I fully expected every picture I had to save to my pictures folder even, was being written to a directory somewhere. Even if my computer wasn't being backed up. After he came everyone's USB's were disabled.
17   CBOEtrader   2020 Feb 17, 8:05pm  

Rin says
rd6B says
It works for close analogies, which a chemistry grad student can predict without a computer. So far nothing new has been discovered despite massive posturing.


For our hedge fund work, a lot of that was based upon control theory, a pre-existing discipline, before ML became popular.

The difference is that ML is attempting to develop a type of control algo, without the modelling aspects, which while it may seem innovative, doesn't change the fact that it's only trying to do what engineers have been doing for decades.


Totally. Its FAR more important to identify the relevant variables than to crunch huge numbers on random data.

I used a mini machine learning tool called cell analysis to do most of my risk management reports as a trader. It was limited to Csv file uploads so small data sets compared to ML but I was able to find far more relevant relationships.
18   Minime   2020 Feb 18, 3:19am  

AI is very good at finding patterns. Thats pretty much short description for you. Just like in dot com days there will be lots of shitty companies that scam investors on buzzword. They all will collapse. Btw you already starting to see cooling in VC markets toward it. I agree that winter is coming for the buzzword believers.

But i tell you that ones that will survive next recession will become next Amazons and Ebays of the post dotcom era.
19   HeadSet   2020 Feb 18, 6:48am  

For those old enough to remember, "Artificial Intelligence" was quite the buzzword in the late 1980s. Everyone was calling their code projects "AI," even down to the level of spreadsheet macros.
20   B.A.C.A.H.   2020 Feb 18, 7:32am  

Oh gosh remember when Nano-Technology was the buzzword?
21   Bd6r   2020 Feb 18, 8:44am  

B.A.C.A.H. says
Oh gosh remember when Nano-Technology was the buzzword?

yes, and nothing came out of it other than funding for people who claimed that they will cure cancer, SARS, HIV, and indigestion.
22   Tenpoundbass   2020 Feb 18, 9:48am  

Headset it was called Fuzzy Logic at first
Which makes sense auto correct, auto fill lists and predictive writing came out of that.
23   Rin   2020 Feb 18, 10:49am  

Tenpoundbass says
Fuzzy Logic


Which is little more than an array, if you think about it, a list of values ranging from low adherence to a high one. I never understood how that was different from let's say a supersized case statement.

B.A.C.A.H. says
Oh gosh remember when Nano-Technology was the buzzword?


You can blame that one on the Cult of Eric Drexler. This was a guy from the MIT Media Lab, with no background or even practical know-how in applied chemistry, talking about assembling "anything" on an atom by atom basis. Seriously, if a device could do that, outside of some Sci-Fi version of the year 3000 AD, then we're in effect, living in some world of magic and sorcery.

So what happened was that a lot of labs and startups, absconded money to make things which can keep your car's paint finish last longer. So yes, it was an fix for a paint job. And as far as the work on fibers go, that's a part of an existing discipline called material science, nothing 'nano' about that.
24   Heraclitusstudent   2020 Feb 18, 12:19pm  

Rin says
Tenpoundbass says
The Gut is missing. The Gut is life's second brain.

Talk for yourself.
Are you hungry right now?
25   Rin   2020 Feb 18, 8:00pm  

CBOEtrader says
I used a mini machine learning tool called cell analysis to do most of my risk management reports as a trader. It was limited to Csv file uploads so small data sets compared to ML but I was able to find far more relevant relationships.


That's cause your due diligence on smaller sets prevents ML's biggest problem and that's that of over-fitting a curve.

And as for price ticks, even with let's say 4-5 price ticks per second (highly unlikely unless you're right on the switch itself), for a year's worth of data, that's 157M data points. Anyone can develop a function looking like ax + bx^2 + cx^3 + d^4 + e^5 + f*sin(t) + g*cos(t) (+ some log series ) over-fitting that data set, making it totally useless for any future additions as it would over fit the past and fail to pattern ahead for price action.

And sure, for let's say telecom systems, that plan may in fact work because one will be able to deduce when and where the highest call volumes were placed and thus, place additional bandwidth support for those periods of time. Though in my experience, a simple SQL script does that job nicely so again, that was invented during Codd's era and doesn't need to be reinvented.

I'm surprised why so many ppl haven't figure that out by now.
26   CBOEtrader   2020 Feb 18, 8:38pm  

I used to do a lot of stuff like this, wherein i tried to identify the ideal medium term risk exposures to our options portfolio. This is half of 1 page in a 15 page "santa rally" risk management report.
27   CBOEtrader   2020 Feb 18, 8:40pm  



I used my mini ML cell analysis tool daily. Loved it for empirical options modeling
28   just_passing_through   2020 Feb 18, 10:01pm  

rd6B says
I have not yet seen any significant discoveries with people applying "machine learning" to difficult research problems.


Checkout Phred:

https://cs.stanford.edu/people/eroberts/courses/soco/projects/2000-01/computers-and-the-hgp/phred.html

It's behind nearly every genetic discovery over the past 25 years. It's still being used in illumina's NGS machines which don't use gels.

Now that we have high quality genetics data I think the next decade will reach new heights with polygenic risk scores:

https://www.genome.gov/Health/Genomics-and-Medicine/Polygenic-risk-scores

Not all are ML or AI but many are. Also, these methods don't work out the underlying fundamental molecular biology but they are getting better and better with predictions all the time.

Being a chemist maybe you have an opinion of protein folding prediction tools that are out there these days?

I think that ML/AI will transform the biology space. It's too hard for humans to do otherwise in any rapid fashion.
29   just_passing_through   2020 Feb 18, 10:08pm  

I too think the buzz will die down. However, modern AI methods are a revolutionary break through vs. what we had a decade ago.

Free material:

https://www.fast.ai/

-If you're into python.

The reason it's all over-blown is AI doesn't have the concept of intuition/imagination. It's still going to be a great tool but there will never be some 'generalized' intelligence that will cause most humans to screem: "dey tuk are yobs!". Not without another level of revolution that brings in the missing parts.

Good luck with that. Some of the recent neuroscience on that and just consciousness in general is becoming pretty complex. Some recent theories believe all of that takes place at the quantum level in our biology. Wouldn't be the first time that happened if proved true. Plants already make use of quantum laws. I'm probably not describing this entirely correctly but in essence they make use of the fact that a photon can be in >1 place at the same time and extract extra energy from that feature of physics.
30   CBOEtrader   2020 Feb 19, 5:49am  

just_dregalicious says
The reason it's all over-blown is AI doesn't have the concept of intuition/imagination.


A computer can only deal w variables fed to it by a human. What is relevant or not relevant to modeling a prediction? That's where you need an expert AND a problem that isnt overcomplex such as predicting words in a google search given previous words (solvable by machine learning) vs global warming.
31   Bd6r   2020 Feb 19, 8:15am  

just_dregalicious says
Checkout Phred:

https://cs.stanford.edu/people/eroberts/courses/soco/projects/2000-01/computers-and-the-hgp/phred.html

It's behind nearly every genetic discovery over the past 25 years. It's still being used in illumina's NGS machines which don't use gels.

Now that we have high quality genetics data I think the next decade will reach new heights with polygenic risk scores:

https://www.genome.gov/Health/Genomics-and-Medicine/Polygenic-risk-scores

Not all are ML or AI but many are. Also, these methods don't work out the underlying fundamental molecular biology but they are getting better and better with predictions all the time.

Being a chemist maybe you have an opinion of protein folding prediction tools that are out there these days?



Phreds looks like a statistical data analysis program, but I may be wrong about it.

Protein folding is way out of my field. It seems that there are way too many variables for this to be easily solvable.
32   HeadSet   2020 Feb 19, 12:25pm  

HEYYOU says
I've reversed my stand on being able to be a'learnin' RepCons anything.



I've reversed my stand on being able to be a'learnin' anything.

Fixed it for you.
33   just_passing_through   2020 Feb 21, 9:43pm  

rd6B says
Phreds looks like a statistical data analysis program, but I may be wrong about it.


Yeah, sort of. It's definitely a machine learning tool. You have to train it. It learns. For modern DNA sequencing (no peaks no gel) we had to come up with our own predictors. Things like signal to noise, among other things. Then you train it on billions of data points and it pops out a lookup table. After that when a machine is writing out the DNA bases it assigns quality scores (probability this is an error is 1:100, 1:1000, 1:10000 etc.) to each "base-call". You want those to be accurate. In the early days of NGS the data sucked. It's pretty damn good now but there are still crap reads in there to throw out.

Newer types of machine learning/AI actually learn much more like a human brain. The hierarchical levels of neurons are modeled. Each layer has... Man I'm tired, I forgot the name. Essentially 3D vectors (where vectors in math are only x/y) and so on. You have to train those as well. That training session like I show above has an example of .. well I can't remember exactly that either. If I recall it called a dog a wolf. When queried 'why' it drew an outline around the snow in the picture. So one must be careful training these things.

rd6B says
Protein folding is way out of my field. It seems that there are way too many variables for this to be easily solvable.


Yeah, I stay away from proteomics. Too much for my wee genomics brain to handle. It's pretty amazing to me what chemists are able to do. There were images of the folded coronavirus out nearly immediately after the DNA sequence became public outside of China. Later the spike was modeled by non-computational means. Some super special microscope thingy.

There have been apps out there for nearly a decade for protein folding. At least one is a game. Some little old lady in the UK liked to play it. She's the best protein folder on the planet despite not having any STEM background at all. AI has a long long way to go haha..
34   Booger   2020 Feb 22, 4:17am  

More importantly, how can AI improve sex robots?
35   Tenpoundbass   2020 Feb 22, 5:18pm  

Booger says
More importantly, how can AI improve sex robots?


By recognizing patterns in your "Oh Face!" Than say stuff like "Tear that Latex up!"
Then have a compressed air when you pull out for simulated quiff

Please register to comment:

api   best comments   contact   latest images   memes   one year ago   random   suggestions