Comment by djous | Hacker News Reader

Comment by djous | Hacker News Reader

During my master's degree in data science, we had several companies visit our faculty to recruit students. Not a single one was a specialized NLP company, but many of them had NLP projects going on.

Most of those projects were the usual "solution looking for a problem to solve". Even those projects that might have had _some_ utility, would have been way more effective to buy/license a product than to develop an in-house solution. Because really, what's the use of throwing a dozen 25-30 years old with non-specialized knowledge, when there are companies full of guys with PhDs in NLP that devote all their resources to NLP? Yeah, you can pipe together some python, but these kind of products will always be subpar and more expensive long-term than just buying a proper solution from a specialized company.

To me it was pretty clear that those projects were just PR so that c-levels could sell how they were preparing their company for a digital world. Can't say I'm sorry for all the people working on those non-issues though. From the attitude of recruiters and employees, you'd think they were about to find a cure for cancer. Honestly, I can't wait for GPT and other productivity tools to wrech havock upon the tech labour market. Some people in tech really need to be taken down a notch or two.

  those projects were just PR so that c-levels could sell how they were preparing their company for a digital world

This is exactly it. The 2017-2019 corporate version of "invest in AI" meant to build an in-house team to do ML experiments on internal data, and then usually evolved a bit to get some "ml-ops" thrown in so they could "deploy" the models they built. I spent some time with a few companies doing this and it always reminded my of "the cat in the hat comes back" when the cat let all the little cats out of his hat and they went to work on the snow spots... just doing busy work...

Anyway it's a symptom of the hype cycle - AI was the next electricity, but there were no actual products and nothing clear to do with it, just hire a bunch of kids to act like they were in a kaggle competition, or worse a bunch of PhDs to be under-utilized building scikit-learn models.

Now that there are (potentially) products coming along that at least bypass the low-level layer of ML, having an internal team makes no sense. Maybe the most logical thing that will happen is the pendulum will swing too far, and this bubble will consist more of businessy types using chatGPT without remotely understanding it or realizing it's just a computer program.

You tend to oversimplify the GPT's - they don't just work all the time, you got to test how well they work, then you got to select the best prompt and demonstrations, then you got to update your prompt it as new data comes along. There is plenty of work parsing various inputs into a format it could understand and then parsing its outputs, especially for information extraction.

> having an internal team makes no sense.

Disagree. I was on one of these R&D/prototyping teams running ML experiments and you're right, it was the company wanting to present itself as future-leaning, ready to adapt, and I would say that at this point it was a good move to have employees who understand where the tech is going.

Companies with internal teams that are able to implement open source models are in a much better negotiating position for the B2B contracts they're looking at for integrating GPT into their workflow, they won't need GPT as much, if they can fallback on their own models, and they will be better able to sit down with the sales engineers and call bullshit when they're being sold snake oil.

They’ve been being sold snake oil for years internally, is I think the point being made.

>The 2017-2019 corporate version of "invest in AI" meant to build an in-house team to do ML experiments on internal data, and then usually evolved a bit to get some "ml-ops" thrown in so they could "deploy" the models they built.

You nailed it, although very few models actually ever got deployed to Prod at Fortune 500 non-tech companies and the few that did delivered little value. I'm a consultant and most internal AI/ML/DS teams that I interacted with were just running experiments on internal data as you said, and the results would get pasted into Powerpoint, a narrative created, and then presented to executives, who did little or nothing with the "insights". Reminded me of the "Big Data" boom a few years earlier where every company created a Big Data Team who then promptly stood up a Hadoop cluster on prem, ingested every log file they could find, and then..................did nothing with it.

> Honestly, I can't wait for GPT and other productivity tools to wrech havock upon the tech labour market. Some people in tech really need to be taken down a notch or two.

You have to remember that when these sorts of things happen, the ones who get "taken down" in ways that actually affect their lives are invariably the ones who already have the least. The ones who "need" that takedown will be just fine, unless they've made incredibly stupid investment decisions.

> the ones who get "taken down" in ways that actually affect their lives are invariably the ones who already have the least

I'm not sure that was the case with personal computing in 1980-s. What was the significant part of society which had the least and got "taken down"?

Personal computing didn't automate too many things that only humans could previously do. Personal computer enabled you to move the data haystack from paper medium to digital but you still had to know the right SW incantations and meticulously dig through it to find the needle.

ChatGPT and other ML apps can find you the needle in the data haystack. To look up stuff on the PC you still needed to know the location of your stuff, filesystem info and how to formulate queries. You no longer need to learn to "speak machine language" but finally the machines can now understand human language to do what you tell them to do.

Of course, ChatGPT & friends can also say dumb shit or just hallucinate stuff up so you still need a human in the loop to double-check everything.

The question was about the case "when these sorts of things happen". This is a generalization which assumes something analogous happened in the past. We're not talking narrowly about something which could "automate too many things that only humans could previously do", which we didn't have before ChatGPT, at least in such a short time. We're talking about similar, but not the same, things and PC revolution seems like a good example.

So it would be good to see the parallels with 1980-s here, if this generalization holds.

I was talking specifically about "when this happens in such a way as to 'take down' people", not saying "any time automation happens, workers suffer".

>> "solution looking for a problem to solve"

I wonder if this is a bad as everyone thinks. When a new technology arrives which is not completely understood, isn't the right approach to try to find some applications for it? Sure, most will fail, but some valid use cases will likely emerge.

I'm pretty sure almost all technologies at some point were solutions looking for a problem to solve. Examples include the internet, the computer and math.

R&D is fraught with risk, but some risks are more rewarding than others. These companies don't just sit on useless knowledge. Take Google who now sits as a "loser" in the current AI "competition"; their projects are far from worthless. Because they've built up expertise, they're now in a very good position to overtake Microsoft on AI, even though they currently seem a bit behind. (And frankly on many fields they're already far ahead.) So OK, perhaps the behemoth that is Google is a bad example, but I still think the same thing is true for smaller companies. If you just read the news, you would think that a technological race like this only has one winner, but that just isn't true. Even quote unquote "worthless projects" can help increase the understanding and expertise in quite important areas, that while not "worth" anything currently, may still have huge value in the future. The only way to know, is to stay in the race.

> I wonder if this is a bad as everyone thinks.

I think it is. If they actually do end up finding a problem to solve, that would be serendipitous but I imagine the vast majority of the time they find themselves in the business of trying to convince the rest of us to buy a thing that we don’t need. And while the latter may drive the economy to some degree as I get older I detest it more and more.

This appears to be the computing model of the past 20 years, from what I can tell?

There have been no real advancements since the desktop model of the late 1990s. We might have more animations and applications running in virtual machines for security purposes, but literally nothing new has come out.

Even all the web apps are reimplementation of basic desktop capabilities from the decades before, but slower and with more RAM usage. They might be easier to write (I personally don't think so - RAD apps from the 90s were quicker to write and use) but the actual utility hasn't changed; if anything it's just shoving all of your data from your microcomputer to someone else's microcomputer, and being tracked and losing control of said data whilst you're at it!

And we have easier access to videos on the Internet, I guess??

It all seems to be missing the point of actually having a computational device locally. There is no computation going on. It's all digital paper pushing.

No one actually needs anything - perhaps food and water but even survival is not strictly necessary.

The problem with “stuff we don’t need” arguments is they are fundamentally nihilistic.

Everyone needs a flying car so let’s get on with it.

The computer was always designed to be a computational machine. It didn't just appear and then someone thought "what could I actually use this for?"

Also the Internet came out of DARPA which was a method of sharing data between geographically remote military facilities. It wasn't like they wired up devices and thought "what could we use this for?".

Do you assert that we had a good understanding of all of the problems that a computer could solve before making it? This seems absurd to me.

GPs point is that the technologies you've mentioned solved real problems before they were adapted for different use cases. They didn't make Darpanet and then think "man, if only there was some use for this" until the Internet came along. They designed it to send signals between distant nodes while being resilient to individual nodes being nuked.

Only after DARPAnet solved that problem did it get adapted to some other problems (ex: how do I send cat pictures to people)?

Fair. Maybe let’s look at it a different way. The computer is a solution to N known problems and M unknown problems at the time of creation. I’d say the N/M ratio is vanishing small.

AI was developed with solid used cases - mostly image recognition or other classification problems. Remember when you couldn’t select text in an image file and paste it into a document? Well that’s a solved problem. The fact that AI has been overhyped and business struggle to find uses for it doesn’t diminish the cases where it does work well.

It might not be optimal if we knew the future but to me its just a natural organic process, organizations and factions inside of organizations are slime molds. A new value gradient appears in the environment and we all spread out and crawl in a million different out growths feeling blindly in the general direction of something that feels like a good idea until one of the tendrils hits actual value and becomes a path of least resistance and the other ones dry out and die.

> I'm pretty sure almost all technologies at some point were solutions looking for a problem to solve. Examples include the internet, the computer and math.

I think the opposite -- nearly all technologies came about as a result of people trying to solve existing real problems. Examples include the internet, the computer and math. (Although I don't think "math" counts as a technology.)

The internet came about from darpanet, which was solving the problem of network resiliency. Computers automated what used to be a human job ("computer") of doing very large amounts of computations. That automation was solving the problem of needing to do more computations than could be done with armies of people.

Why start startup of any kind if there's a bigger company full of people already competing in the same space?

Because product execution at SO many places sucks. LLMs won’t help with that, either. They’ll just help people market their crappy products more cheaply. Woe to the marketers, however.

> Honestly, I can't wait for GPT and other productivity tools to wrech havock upon the tech labour market. Some people in tech really need to be taken down a notch or two.

That's an odd reason to want this.

Less bullshit jobs. Society needs doctors, nurses, plumbers and teachers not tech bros.

Society definitely needs those, but the incentives of the system most societies live under do not align to those needs. We are 100% into a society of wants, not needs, and the rewards are for those who sell stuff for these wants. Our needs went into the "cost center" of society's calculation, not an investment, and so it's been a race to the bottom for those professions.

While adtech, crypto and other bullshit gets massive funding because it can turn a profit.

The incentives to have a good society don't align with the incentives of financial capitalism.

I agree with your sentiment but disagree that AI research is in any way the domain of tech bros.

I’m starting to see the term “tech bros” appear more and more in HN - before hand I more frequently saw it outside of this site.

Some people on HN I have seen really come down on those that use the term. I don’t.

Perhaps those of us in the industry ought to recognize that the term exists because of a growing resentment among people outside of the tech industry.

Your comment hints too as to why that is.

It's evidence of resentment, but not of well reasoned discourse against something the tech industry is doing. Characterizations like this anthropomorphize a group into a single entity that is easier to hate and assign intentions, too. It's not constructive to any conversation that moves a discussion forward. A person who is mad at "tech bros" is likely more upset about systemic forces that they want to blame on a target. It's logically equivalent to making sweeping statements blaming immigrants for suppressed wages.

Comparing affluent ivory-tower digital landlords to vulnerable people being blamed for things outside their control is definitely one of the decisions of all time. It also seems like a lot of exercise just to feel justified in discarding a large group of opinions.

People start generalizing about groups like this when they've stopped caring about negative policy consequences which affect those groups. Politicians who blame wage stagnation on immigrants do not expect to have those immigrants who gain citizenship vote for them. Why do you think people might have stopped caring what happens to the group designated "tech bros"?

take it from me; "doc bros" are far, far worse.

AI or technology won't reduce bullshit jobs. To the contrary, they might increase bullshit jobs, because there would be more resources to allocate for those jobs.

Yes, yes, yes. My wife (who teaches humanities) keeps asking me if AI is going to take all the jobs. I keep saying society has already had to deal with the fact that not that many people are required to build the things and grow the food, and we adapted by inventing all kinds of vaguely bullshit jobs, and jobs to provide services to the people who get their money from bullshit jobs, and all of those jobs spread the wealth around. There will be more and more of those (and, to be clear, many more people who aren't in a position to get one.)

Still better to train lots of electricians and refrigerant techs and solar installers and all the other workers that the energy transition will need, of course.

Also have yet to Graeber's book ("Bullshit Jobs: A Theory").

Two problems: Who is going to pay to retrain people for those jobs?

Except for perhaps doctors (and even then residency is BS) all of those jobs are treated or paid like crap.

You do realise those tech "bros" are what enables doctors, nurses, plumbers and teachers to have a better work-life?

I am not sure I can agree.

Doctors and nurses now spend more time entering data than talking to patients.

Teachers now spend more time entering grades into online systems and fielding messages from parents.

Not sure how tech is helping or hurting plumbers except for the standard GPS tracking that bosses use to follow them around.

With this kind of mindset, we'd still be using lead pipes and letting blood.

Doctors and plumbers might make society work, but technology drives society forward.

Sure. But recruiting scheduling coordinators do not. Those people would better serve society stringing up new HVDC lines, which the current model does not incentivize.

I'm not a 'tech bro'.

My job is to optimize workflows which reduce cost, CO2 and make the world better and easier.

It's not my fault that society is unable to optimize everything for less work for everyone while everyone can have a good live.

Counterpoint, if one doesn't have their own baseline model how does one know the vendor is providing value.

Yeah having a whole big team create the internal baseline is not cost effective, but having at least one or two people work on something to actually know the vendor is worth their cost is important.

Reply on news.ycombinator.com