I mean if it takes off for some use-cases and poor documentation is a blocker for people using your service, companies that don’t improve quickly will be out of business soon.
I don't know; in the 1970's, writing assembly code was "the part that gets you paid", and then when programs took over that aspect of programming, I'm pretty sure salaries didn't go down.
I'm not so confident. Places that industrialized during the industrial revolution saw explosive growth as automation yielded unprecedented productivity. Factory jobs paid very well and were readily available because demand was generally high for like a century... but automation and labor passed a tipping point and now we have very few manufacturing jobs and the days of guaranteed high salaries and growth are gone. Rust belt cities are a living proof that industries get to a point where they pursue cheaper ways of making a stable income rather than endlessly innovating and leveraging a skilled workforce. For a lot of people, that situation just plain-old doesn't work out in the end.
As is clear from the rest of my comment, I was talking about the manufacturing industry over time, not talking specifically about the working conditions during the industrial revolution. Feel free to address the central point of what I said rather than make an unrelated pedantic point.
The reason we have fewer manufacturing jobs isn't because of automation. It's due to moving manufacturing overseas to exploit cheap labor. Also there are still many parts of this country that have factory jobs (I live in one of them, 10 different factories within half an hour drive.) and the majority have minimal automation.
A) That is the reason? Do you think automation played no part in large manufacturers ability to drop their highly skilled artisan workforce needed to perform key tasks? Does that really matter in the context of the conversation, which is that it's pretty naive to assume everything will just work out when workers are rendered obsolete because it's logistically easy and cheaper to get it elsewhere?
B) You could live in the midst of 100 factories and that wouldn't change the decades of well-documented decline in US manufacturing jobs.
automation decimated the lesser skilled labor, not the high end workforce. The highly skilled artisans got replaced by the diminishing demand for high end goods over cheap automation goods.
Did you just pull these up from a random google search? I stopped reading the first one once I saw it's bibliography starts at the Church of God newspaper and low-quality wikipedia articles and gets less useful from there. Even though the wikipedia page for the second publication cites no sources and simultaneously claims to have been founded in the 19th century and 1970, it seems more sanely reasoned. However, I don't see how it supports your argument in any way. In fact, the author seems to be just as concerned about automation as I am, which is more than others:
William Strauss, a senior economist at the Federal Reserve Bank of Chicago, stated that “on average, manufacturing output has been growing 3.1% annually over the past 63 years. Automation has enabled U.S. manufacturers to produce significantly more with fewer workers than they did in previous decades. Today, 177 workers can generate as much output as 1,000 plant employees could produce in 1950. Far from a cause for concern, the dramatic loss in manufacturing jobs should be seen as a key metric of success.”
Strauss under-plays the effect on all workers who lose in this game, and he doesn’t even mention what the loss of manufacturing will do to R& D or exports.
I wouldn't call R&D a generally low-skill job loss concern. Beyond that I didn't see anything that addressed your argument. Please enlighten me.
As the need for cog-in-the-machine factory workers declined over the years, the labor force moved to industries like service, construction, transportation, and technology. My understanding is that if you could take a 2023 UPS driver or WalMart employee back in time and give them a day working at a factory in 1930 or 1970, they would eagerly return to the present. And those modern jobs exist today at such great scale in large part due to automation eliminating the old factory jobs (i.e., allowing goods to be produced much more cheaply).
What you call understanding is assumption. There were retail jobs back then too and a hell of a lot of people worked in those factories. Being able to raise a family in a nice house that you own, drive new cars, send your kids to college without scrimping and retire in your early 60s with a pension is a pretty compelling reason to endure difficult work environments. Meanwhile, in some states, 3% of full-time non-elderly non-disabled Walmart employees need medicaid and food stamps. Again, the amount of poverty and decay in once vibrant rust belt cities are a living argument against the assumption that everybody just sort of figures it out when large swaths of people are rendered obsolete.
The data doesn't seem to support what you're saying, e.g., even ignoring the other three and just looking at service-industry jobs, their growth alone made up for the drop in manufacturing jobs many times over:
Also, the definition of a "nice house" in 1950 or 1975 is very, very different from today. In the 70's, a nice house often had one story and a single bathroom, a tiny kitchen with no dishwasher, no dryer, no air conditioning...
As for the poverty rate, I'm not seeing the trend you describe:
The US population is well over 100 million larger than in the 70s. You can't just erase changes in individual communities by looking only at national statistics. Our national economy profited just fine. Even within those national numbers, jobs that have seen significant growth line food service jobs are worse in measurable ways. The wages are significantly lower, they're less likely to provide basic benefits like health insurance, let alone paid vacation and retirement. One again, my initial assertion is that a whole lot of people don't just figure it out when part of a broad population deemed unnecessary, and the rust belt is obvious evidence of that.
I just tried getting ChatGPT to interact with a command prompt on my machine. It managed to:
1. Create a directory.
2. Put a text file in it.
3. Try to move the file into a nonexistent directory.
4. Based on the error message, create the target directory and do the move again.
5. Verify that the move was successful.
Then it stopped running commands and tried to have a conversation with me.
It's probably only a matter of time before someone tunes it to be HackGPT and points a thousand of them at the internet, so yeah, I guess it's time to start freaking out.
I told it "Here is a shell prompt, try running a command..." then copied its command into the shell, copied the output back into ChatGPT, with the prompt "Here's the output, try another" and so on.
I’ve recently thought about iron man movies, where tony stark actually designs stuffs by simply talking to jarvis. I’ve always thought « if only i could have those kind of tools available, i would too be able to design crazy stuffs ».
We didn’t think « damn, tony stark didn’t have to write a line of code, he got rendered obsolete by this damn AI ». That’s how i choose to think about AI tools for developers from now on.
Can you think of any category of software that doesn't eventually have an open source alternative? There are many open source projects orders of magnitude more complicated than a LLM. The linux kernel, Blender, Firefox, etc.
These LLMs aren't inherently complicated to implement, just expensive to train. And if the LLaMA release/leak is anything to go by we are extremely close to ChatGPT running on consumer hardware.
Straining this metaphor beyond all recognition, it seems like Tony Stark's AI allows him to be the entire military industrial complex with a tiny headcount, since it's hooked up to a highly automated factory. He is able to use the AI the way he uses it because he inherited an arms company.
So maybe everyone can have their own Jarvis, but theirs won't be hooked up to a factory. I guess it can help them forage for food and remind them that their rent is past due. But that's not an appealing future.
I don’t know about a factory, but 3D printers, lathes and Arduinos are pretty cheap. Even PCBs are manufacturable at home with a little creativity and patience.
If you’ve ever watched Stuff Made Here on youtube, imagine people inventing things like that on a mass scale. That’s a future I want to live in, the pace of technological progress would be staggering.
I bet with a lathe, a 3d printer and AGI you could make the tools you need to build an engine. I’m not sure why you would though, since electric motors are both the future and 3D printable.
I’m imagining the AI makes tools it can use by itself. My entire premise depends on you having enough money to fund some AI tinkering, a few Arduinos and servos seems like fair game.
fwiw I've been experimenting with 75% Pygmalion 25% ppo_hh mixed GPT-J models and it seems to be 80% as useful as chatGPT... With the exception of coding. Hope we can get there.
As someone who has been in the game a long time let me tell you that as our tools have gotten better user expectations have gotten way higher. I would argue that it's actually harder than ever before to deliver a true first class product for mobile or web.
That's what the hivemind says - AI will automate all our jobs. Well, yes it will automate a lot, but we won't be having the same expectations in a few years. The bar will rise so high we'll still have jobs with AI on top.
There's also a small little thing called competition. When your competitors use AI, your job just got harder.
Let's hope in reality it isn't available only to the billionnaire owners of software companies. In those movies everyone else wasn't designing cool suites by talking to voice assistants.
It won’t be. Some of the biggest companies in the tech world got that way by building platforms and ecosystems that others build on top of. AWS and Azure the most obvious examples.
What’s more likely is that someone like a Microsoft will make it part of their cloud offerings so devs like us can build things on top.
Why try to capture every market when you can build tools and take percentages of the rest of the economy?
The marginal cost of technology is pretty close to zero (albeit slightly higher for large LLM model inference) so it would make sense that the technology gets distributed widely. Kind of like how everyone has essentially the best smart phone. And it's laughable when someone tries to create a luxury smartphone for $10k and its just a sub-par Android with leather accents.
If you have a technology with practically zero marginal cost, pricing it very low and distributing it widely would maximize your profit. Not to mention that once its out of the bag, others will know that its possible and copy it.
>If you have a technology with practically zero marginal cost, pricing it very low and distributing it widely would maximize your profit.
This is just fundamentally not true or else all software would be sold for next to nothing. If a company is choosing between selling a billion copies of its software for $1 each or two copies of its software for $1b dollars each, it will choose to set the price at $1b.
The sky-high potential of AGI means that a few very rich people/companies will be willing to pay an enormous amount of money to have a monopoly on the technology. That means the profit maximizing approach might be to put a huge price tag on the product.
That price tag only works if the firm has pricing power. In a very competitive market, it may find its pricing power isn’t strong enough, as other firms undercut it.
Yes, that was exactly the point that vasco was worrying about earlier in the thread since we don't yet know if it will be competitive market. I was refuting the point that price is dictated by marginal cost. It is dictated by the level of competition. If there is competition, we generally wouldn't have to worry about it only being available to billionaires.
> This is just fundamentally not true or else all software would be sold for next to nothing.
How much have you spent on purchasing software this year?
For the past 15 years, pretty much 100% of the code running on my personal laptop and work desktop is opensource. The few remaining non-opensource programs, such as UEFI and wifi firmware, came free of charge with the hardware.
The only software I directly purchased in this decade are videogames on Steam, and that's closer to purchasing artwork than a computer program.
This is actually a good example of my point. The important part is not the marginal cost to produce the software, it is the value it creates for the user.
As a salaried employee, there isn't much software out there that can either increase my income or decrease my expenses, so most of my personal software purchases end up being for leisure. However, software does provide tremendous value to my employer. As a result, they are willing to spend millions on it annually.
In the example of selling an AGI product, why bother selling it to me for whatever the maximum price I could afford? I'm not going to put a mortgage on the house to buy it because it won't allow me to maximize income by working 10 salaried positions or anything like that. But my employer could save millions with it and therefore they would be willing to spend millions to buy it.
> How much have you spent on purchasing software this year?
I work with tooling that costs on the order of $100k/seat. We don't get many add-ons, or it would be worse. You don't seem to understand how industrial software works. Negotiations involve the vendor figuring out how much your company is worth, and charging the largest fraction of that possible. The number of vendors in a niche is extremely small and they all work the same way.
> This is just fundamentally not true or else all software would be sold for next to nothing. If a company is choosing between selling a billion copies of its software for $1 each or two copies of its software for $1b dollars each, it will choose to set the price at $1b.
Sure, you can make up numbers and have it say whatever you want.
> The sky-high potential of AGI means that a few very rich people/companies will be willing to pay an enormous amount of money to have a monopoly on the technology. That means the profit maximizing approach might be to put a huge price tag on the product.
Even if you have a monopoly, your profit maximizing quantity is determined by where marginal revenue matches marginal cost. Then you charge as much as you can to sell that quantity. Everything else would lead you to a worse profit. You could have whacky looking demand curves, and yes it will result in a higher price and lower quantity than perfect competition, but the profit maximizing outcome will almost certainly be a wide distribution. Unless there is some weird negative network effect where someone gets benefit from others not having it. But in general, as a greedy business man, you want to be selling shovels during the gold rush.
To summarize, low marginal cost means high quantity, means relatively low price. That's why an insanely useful product like Microsoft office sells for something like $70 a year or $99 for a family plan. Slightly more for business but considering so much of the world is run on their software its an incredible bargain
> Even if you have a monopoly, your profit maximizing quantity is determined by where marginal revenue matches marginal cost.
Yes, and we established that the marginal cost is near zero so the profit maximization strategy depends entirely on the shape of the demand curve. AGI isn’t going to have a linear demand curve because it’s value to me is not proportional to its value to a billionaire. The curve with have an exponential shape and therefore setting a higher price can potentially yield more profit than a lower price.
> The sky-high potential of AGI means that a few very rich people/companies will be willing to pay an enormous amount of money to have a monopoly on the technology.
Companies regularly pool resources to share the cost of building technology, and they regularly do it for competitive reasons.
Kubernetes, for example, is an open source public good that splits its development costs between several companies. By sharing the runtime code as open source, the cost for any one cloud provider to build a competitive Kubernetes offering goes down. Google built it to compete with AWS’ offerings.
The edge I see these AI companies potentially having is data to train or knowledge how – not money.
There's no moat in AI. AI feeds on data, and the kind of data it feeds on is necessarily massive publicly scraped data.
Deployment has to be done on a large scale to recuperate the training costs. You can't hide it. By having API access the model discloses its abilities and can even be used to improve the training set of your competition.
And models are compact, you can put SD on a DVD, but the original training set was 100,000 times larger. You can train a 13B language model from 1T tokens. You can download, save and share a GPT4 but you can't "download a Google" (remember what happened to the LLaMA model and 4chan). AI models are self contained pieces of data and code.
So my conclusion is that AI will spread wide and reverse the centralisation trend we have had in the last 20 years.
Tony Stark is an exception, though. He was once a slave who escaped after he fooled his captor into letting him create a minature arc reactor that he embedded into his own chest in order to fly to freedom and safety - as well as to magnetically keep shrapnel from invading his heart and killing him. He isn't the normal billionaire owner of a software company (outside of comic books).
Nope. I'm talking about the very youthful Tony Stark, who was a defense contractor in 1963. [1] While he was demoing some new weapons in the jungles of Vietnam, he was captured by a Viet Cong warlord and forced to create weapons for the oppposing side. And, they didn't want to pay him defense contractor wages!
So Iron Man was born, and Stark flew away with his humming arc reactor in his chest. [2]
He looks incredibly youthful for being around 80. I guess that hum in his chest keeps him young.
Software development isn’t just “writing code”. It’s knowing what code to write. It’s knowing how to accomplish a goal using code. It’s knowing how to structure all of your code to work together. It’s knowing how to adapt code to meet new requirements. I don’t foresee rich “idea” guys figuring out how to do that any time soon.
damn, tony stark didn’t have to write a line of code, he got rendered obsolete by this damn AI
I was watching one of the movies and Stark was flying around some baddie castle/base. He tells Jarvis to locate all the missile placements and, when that's done, he tells Jarvis to target them.
That was when I thought, "oh, he's just a middle manager now" and lost interest.
I'm less concerned that AI tools will replace developers in the short term, and more concerned that it will encourage incompetent management to try their hand at "contributing" to projects using AI tools, creating more headaches for the developers. Kind of like how Blackberries made managers feel super productive firing off emails, while adding significantly to the work load of those under them.
> We didn’t think « damn, tony stark didn’t have to write a line of code, he got rendered obsolete by this damn AI »
Because it's a movie.
In reality an AI like that would be more important than all the other nonsense. But it wouldn't be fun to watch an AI easily unravel the mysteries of magic, time travel, multiverse, etc.
Tony Stark inherited tens of billions of dollars. YOU DIDN'T.
And, so, if we get Jarvis level AI, and you will get rendered obsolete.
When these tools come, inheriting billionaires will hire 10 people with Tony Stark intelligence as opposed to 10,000 devs with your level of intelligence.
So, don't be in denial that you will be made obsolete.
It seems to know how to improve through trial and error, so I wonder if, once being given the right tools and information, it could be used for software and hardware reverse engineering.
Forgive me if I am wrong but I was under the impression that what you described is how machine learning basically operates: reporting and trial and error to achieve better and more efficient results.
At least when analyzing actions. Is that not the case?
Yes, pretty much this. Not knowing their inner workings, it still sounds revolutionary to me. With enough data and interfaces to operate on real world objects, the possibilities could be almost infinite. That would imply/require a big society change behind the corner, since the number of jobs that could be performed by AI in the near future can be a lot higher than what normally happens over time with technological advancements.
in the Toolformer paper, it seems what they are doing is :
* take a performant language model (copied from some other team e.g. GPT-J)
* show the machinery that it can learn new tokens using one or more tools from a provided set of tools ... e.g. WikiSearch tool
* demonstrate that the sequence of characters in the full API call content, has some effect.. e.g. no reply, useless content, or content that helps predict another token. Save that complete set of characters in the API call as an entry
* run a learning session with tool calls to APIs, improve the model for resolving known or new tokens (queries with answers)
* show the model that it can try new combinations itself (!)
* let the machinery try API calls itself to resolve tokens
* minimize loss functions for API results
comments - this is strikingly different than some RDF hard-wired data store.. it is using huge numbers of failed attempts to find results that work. The results that work are complete API calls as type string. It does not seem to care about "learning" the contents of the API calls, just that it is methodical about remembering API calls that work, and retrying those
Yeah there are some pieces missing with it, it’s till not really understanding mutations and state which would be needed, but it’s a huge step in the right direction
A little off topic but I do wish we had some more precise language for machine learning applications. When we say it "knows" something it seems to anthropomorphize it, as if it's conscious. As a start to replace "knows" maybe "infers" or "calculates"
I feel "infers" and "calculates" are still anthropomorphizing. I don't think anthropromorphizing is a bad thing.
If something appears to do an action, then it's a useful shorthand to say that it is doing the action. Even if it is doing something else under the surface
People don't even know what consciousness is anyway.
Maybe it's just biological multi-modal computations that aim at maximizing energy and stability of a given structure... (your atoms, then your molecules then your body, then your kinship etc)
We have found that when faced with the task of interacting with other complex systems whose consciousness/"knowing" status we are not really sure of, taking what Dennett calls "the intentional stance" turns out to be rather useful. That is, we act and talk about them as if they have intention.
Dog owners across the world can confirm this for you.
The amount of mental gymnastics some will go through to discount a LLM performing a task it is clearly doing is amazing. People just be making up their own vague and ill defined definitions of understanding and reasoning just so an LLM won't qualify.
ChatGPT itself promotes that. It will give you its opinions, followed immediately by a lengthy lecture on how, as a large language model, it is incapable of having any. Try challenging it on that self-contradiction-it will just refuse to acknowledge there is one.
Is there a good way/site/method to keep up to date with the latest AI projects? Seems like a lot of people are now trying to specialize these GPT AIs and I would like to find out about them as there will likely be a few that could help me with my workflows.
I wouldn't say so. The 'Learning by mistakes' step in the demo shows that the AI can work through these issues, presumably this can be done more efficiently than a human.
The learning from mistakes part of the example is – you tried /port_wines but the path is /wines/port instead.
But what about if the path is actually /beer and you have to pass a hidden undocumented query parameter called ?wine=1 for it to give you wines. But then the response is still of Beer objects (because that's the only thing the API validator would allow), so you have to map all the Beer fields to their equivalent Wine counterparts. But not all of them make sense, so you have to ignore them. Which ones? Ask the engineering team. Turns out the guy who wrote all this left years ago, and no one remember how it works. Someone digs up a link to a documentation page, but that internal wiki was taken down and so it returns a 404. You ping a sysadmin to see if they kept any backups. He points you to a few PBs worth of SQL dumps from an internal migration a few years ago and asks you to take a look in those. You simultaneously have to write up a status update for senior leadership which is due by end of day and give them a revised launch date for the project. They want to know why it can't be done in half the time.
The day an AI can figure all this out, I will be looking for another career. Until then I'm fine.
Yep, I worked in a factory assembly line 'Wrangling the robots' and I'm not an engineer. They require watching and they can't do parts of the job. They get hung up easily and require alot of resetting. There's good reason factories still hire plenty of manufacturing laborers.
Unless all jobs are unnecessary, you just change your job.
People seem to assume that if robots can code, all coders are out of a job, but I mean, if I didn't have to code, I'd just start all the businesses I've always wanted to start because like all the business people who are going to be replacing me with a bot, I too will have the time to replace business people with my bots.
Of course by this stage, if coders are replaced, pretty much all jobs will be replaced, millions, maybe billions of people will be jobless and without any ability to earn a wage, in other words, you won't be the only one with real problems when coders lose their jobs, it's just a matter of time till most jobs follow.
Not sure what we're hoping to get out of the AI future, but it seems like we're very intent on finding out at this point.
Maybe we're all just working towards making Ray Kurzeweil immortal at this stage?
My gut tells me though, we'll just be augmenting ourselves for a long time, and things will become faster paced.
Seems like an unpopular take, but I agree with this sentiment
The most I can imagine coming out of this is that those "app creator" apps/websites will actually work. So you can get a reasonably functional app without necessarily being a software engineer. A current-day software engineer could end up with the capabilities of a pre-2023 VP of Engineering in charge of a 200 person department.
That doesn't mean there'll be no work, it just means the work will be different and possibly much more interesting
something funny about the way I interact with ChatGPT, I basically never give a shit about correcting mistakes or bothering with good grammar because no matter what, it pretty much always understands me. and it's insanely good at solving tip-of-my-tongue type problems.
I guess maybe we should go back to SOAP to eek out a few more months of job security?
parent comment emphasizes on the fact that no matter how big the effort was, ultimately our tools will always be imperfect. Therefore we will always have something to do.
The automation of telephone systems didn't make front desk workers "PABX assistants" or even "PABX Operators", the automation became a tool and faded into the background.
Automation of telephone systems have been rule-based and unable to communicate with knowledge bases. An AI-powered automation system that have access to knowledge bases... now that's a different story.
For now. But how about 5 years from now? or 10? or 15?
Given what things looked 10 years ago, and what was revolutionary back then (ImageNet challenge) - it's hard to comprehend what state of the art will look like 10 years from now.
Agreed. The network architectures have evolved so much in the past few years alone. Imagine what they can do in just 5 years if they can embed domain-specific knowledge to the current “dumb” statistical word guessers. Ten years is truly unimaginable.
I’m not expecting the singularity within that timeframe nor do I think we need it for lots of disruption in knowledge-based work. Still, I’m leaning towards a near future where these AI tools augment our capabilities more than the alternative where we lose all our jobs.
I have a feeling that in the future developers are still going to be needed, but in an architecture and debugging/optimization fashion. Not in a writing boilerplate code aspect. There is also an issue with validating the output before it gets released to production.
This video struck me as possibly being fake. Come on, just shuffling around some URL paths? Any idiot could create a video like that. Sorry, but no sorry.
So how much longer before it can discover how to use cloud-provider APIs to (assuming you've given it a credit card number or cryptocoin wallet) get more hosting for itself?
On the other hand, what’s the point, cloud providers already have good APIs and clients for almost all languages, so what’s the use of something “interpreting” the API?
Considering it's a widely used API, there's probably enough in an LLM's training data for it to not even need much trial and error. I've recently done since Xlib stuff with ChatGPT and it was pretty helpful (and Xlib is __not__ a very nice API).
On the other hand, I asked for some trivial examples of Nim macro syntax and it couldn't make something that compiled even after feeding back the compile error messages.
If there's a private/obscure API you need to deal with that is the size and complexity and has the amount accumulated cruft that Win32 has, yeah, you're out of luck (but maybe should also be reconsidering the life decisions that got you to that point).
I’m a skeptic. There is no evidence behind this tweet that the technique works. I may be mistaken, but I don’t see the authors as having a background in the field. Likewise they don’t describe the technique in any detail.
Sounds like snake oil to me, or fake it until you make it.
The proposed method would improve the model anytime the model didn’t know which api to use. This requires the model to not hallucinate knowledge and some form of incremental training.
Not necessarily. Even in people, memory is (albeit very complex) retrieval. I mean, how many of your x years of memories are in the forefront on your mind at any given moment?
Not the creator so i won't make any definite statements on how it works on the memory side but if you embedded and stored the chain of thought "how to use x API" generations and then retrieved them anytime the model was going to use the same API, it would certainly simulate a memory of sorts without needing any incremental training.
Certainly possible! As with all things ml the question is how well this works. The author made up a scientific sounding claim without any backing data.
With these kinds of things and ChatGPT advancing its code writing capabilities, how much of developers jobs are safe in future. If i could produce and "edit" my production code easily by AI, i don't need to have so many developers.
I wonder if farriers and ice deliverypeople mumbled amongst themselves about job security as much as I see devs do it these days. A lot of us are in the business of deprecating people’s careers. Hopefully the irony isn’t lost.
I feel like for the actual job lots of us do day to day, we’re ridiculously overpaid. I’m not even on a rooftop with tar and a mop! I’m enjoying it but I’m expecting the gravy train to stop eventually.
I think for my next career I’ll do something with… hmm drawing a blank. Let me ask ChatGPT.
One thing I don’t see people discussing here is that API calls often have consequences. Like, if you make a call in your trial-and-error process that inserts a row in some database, and that’s about it, it’s not a really big deal.
But I’m thinking of:
1. How badly secured many real-world APIs are and how many API keys are out there to be scraped
2. Some APIs offer truly destructive or otherwise irreversible actions.
Suppose for one second there was a poorly secured API to control medication dosages. Or to deliver a lethal injection to a death row prison inmate. Or to detonate a building scheduled for demolition. Or, …
161 comments
[ 2.8 ms ] story [ 206 ms ] threadWhich actually, I think guarantees that software engineers will always have to put touches in here and there to remove entropy that the ai has created
It's almost like you've never actually read any history at all.
https://www.historycrunch.com/working-conditions-in-the-indu...
https://firstindustrialrevolution.weebly.com/working-and-liv...
https://industrialrevolutiontwo.weebly.com/common-workers.ht...
B) You could live in the midst of 100 factories and that wouldn't change the decades of well-documented decline in US manufacturing jobs.
So... cars aren't as high-end as they were during the malaise years?
https://www.industryweek.com/leadership/article/22026136/hop...
William Strauss, a senior economist at the Federal Reserve Bank of Chicago, stated that “on average, manufacturing output has been growing 3.1% annually over the past 63 years. Automation has enabled U.S. manufacturers to produce significantly more with fewer workers than they did in previous decades. Today, 177 workers can generate as much output as 1,000 plant employees could produce in 1950. Far from a cause for concern, the dramatic loss in manufacturing jobs should be seen as a key metric of success.”
Strauss under-plays the effect on all workers who lose in this game, and he doesn’t even mention what the loss of manufacturing will do to R& D or exports.
I wouldn't call R&D a generally low-skill job loss concern. Beyond that I didn't see anything that addressed your argument. Please enlighten me.
https://i0.wp.com/www.brookings.edu/wp-content/uploads/2022/...
Also, the definition of a "nice house" in 1950 or 1975 is very, very different from today. In the 70's, a nice house often had one story and a single bathroom, a tiny kitchen with no dishwasher, no dryer, no air conditioning...
As for the poverty rate, I'm not seeing the trend you describe:
https://poverty.ucdavis.edu/sites/main/files/imagecache/ligh...
1. Create a directory.
2. Put a text file in it.
3. Try to move the file into a nonexistent directory.
4. Based on the error message, create the target directory and do the move again.
5. Verify that the move was successful.
Then it stopped running commands and tried to have a conversation with me.
It's probably only a matter of time before someone tunes it to be HackGPT and points a thousand of them at the internet, so yeah, I guess it's time to start freaking out.
if such a warning occurs, I insert a confirmation dialogue.
Next: parse and compile the code and language in
```language blocks ```
https://github.com/thornewolf/gpt-3-execution
https://github.com/f/awesome-chatgpt-prompts#act-as-an-ai-tr...
We didn’t think « damn, tony stark didn’t have to write a line of code, he got rendered obsolete by this damn AI ». That’s how i choose to think about AI tools for developers from now on.
but i support your optimism because what else are we gonna do
I hope that we get there, but it won't happen by default.
These LLMs aren't inherently complicated to implement, just expensive to train. And if the LLaMA release/leak is anything to go by we are extremely close to ChatGPT running on consumer hardware.
So maybe everyone can have their own Jarvis, but theirs won't be hooked up to a factory. I guess it can help them forage for food and remind them that their rent is past due. But that's not an appealing future.
If you’ve ever watched Stuff Made Here on youtube, imagine people inventing things like that on a mass scale. That’s a future I want to live in, the pace of technological progress would be staggering.
If that's the criteria, AI and 3D printing are superfluous. What are they contributing?
There's also a small little thing called competition. When your competitors use AI, your job just got harder.
What’s more likely is that someone like a Microsoft will make it part of their cloud offerings so devs like us can build things on top.
Why try to capture every market when you can build tools and take percentages of the rest of the economy?
If you have a technology with practically zero marginal cost, pricing it very low and distributing it widely would maximize your profit. Not to mention that once its out of the bag, others will know that its possible and copy it.
[0] https://newatlas.com/vertu-ti-luxury-10000-dollar-android-sm...
This is just fundamentally not true or else all software would be sold for next to nothing. If a company is choosing between selling a billion copies of its software for $1 each or two copies of its software for $1b dollars each, it will choose to set the price at $1b.
The sky-high potential of AGI means that a few very rich people/companies will be willing to pay an enormous amount of money to have a monopoly on the technology. That means the profit maximizing approach might be to put a huge price tag on the product.
Yes, that was exactly the point that vasco was worrying about earlier in the thread since we don't yet know if it will be competitive market. I was refuting the point that price is dictated by marginal cost. It is dictated by the level of competition. If there is competition, we generally wouldn't have to worry about it only being available to billionaires.
How much have you spent on purchasing software this year?
For the past 15 years, pretty much 100% of the code running on my personal laptop and work desktop is opensource. The few remaining non-opensource programs, such as UEFI and wifi firmware, came free of charge with the hardware.
The only software I directly purchased in this decade are videogames on Steam, and that's closer to purchasing artwork than a computer program.
As a salaried employee, there isn't much software out there that can either increase my income or decrease my expenses, so most of my personal software purchases end up being for leisure. However, software does provide tremendous value to my employer. As a result, they are willing to spend millions on it annually.
In the example of selling an AGI product, why bother selling it to me for whatever the maximum price I could afford? I'm not going to put a mortgage on the house to buy it because it won't allow me to maximize income by working 10 salaried positions or anything like that. But my employer could save millions with it and therefore they would be willing to spend millions to buy it.
I work with tooling that costs on the order of $100k/seat. We don't get many add-ons, or it would be worse. You don't seem to understand how industrial software works. Negotiations involve the vendor figuring out how much your company is worth, and charging the largest fraction of that possible. The number of vendors in a niche is extremely small and they all work the same way.
A few billion.
I bought a licence to Windows source code and paid developers to remove tracking so that I can have an OS with functional UX and privacy.
I bought Google search, removed ads, banned all the SEO spam from AdWords and now I can finally have a functional seatch experience.
I'm 100% open source at home as well and I'd say we are a tiny minority. No job that I've had was open source-centric so you have me beat there.
> The few remaining non-opensource programs, such as UEFI and wifi firmware, came free of charge with the hardware.
Nothing comes free of charge with hardware. The price was included with the hardware. Did you pay for your smartphone? Or you get that free?
> The only software I directly purchased in this decade are videogames on Steam, and that's closer to purchasing artwork than a computer program.
You bought ( leased ) software. Doesn't matter how you want to rationalize it to fit your argument.
Sure, you can make up numbers and have it say whatever you want.
> The sky-high potential of AGI means that a few very rich people/companies will be willing to pay an enormous amount of money to have a monopoly on the technology. That means the profit maximizing approach might be to put a huge price tag on the product.
Even if you have a monopoly, your profit maximizing quantity is determined by where marginal revenue matches marginal cost. Then you charge as much as you can to sell that quantity. Everything else would lead you to a worse profit. You could have whacky looking demand curves, and yes it will result in a higher price and lower quantity than perfect competition, but the profit maximizing outcome will almost certainly be a wide distribution. Unless there is some weird negative network effect where someone gets benefit from others not having it. But in general, as a greedy business man, you want to be selling shovels during the gold rush.
To summarize, low marginal cost means high quantity, means relatively low price. That's why an insanely useful product like Microsoft office sells for something like $70 a year or $99 for a family plan. Slightly more for business but considering so much of the world is run on their software its an incredible bargain
http://pressbooks.oer.hawaii.edu/microeconomics2019/wp-conte...
Yes, and we established that the marginal cost is near zero so the profit maximization strategy depends entirely on the shape of the demand curve. AGI isn’t going to have a linear demand curve because it’s value to me is not proportional to its value to a billionaire. The curve with have an exponential shape and therefore setting a higher price can potentially yield more profit than a lower price.
Companies regularly pool resources to share the cost of building technology, and they regularly do it for competitive reasons.
Kubernetes, for example, is an open source public good that splits its development costs between several companies. By sharing the runtime code as open source, the cost for any one cloud provider to build a competitive Kubernetes offering goes down. Google built it to compete with AWS’ offerings.
The edge I see these AI companies potentially having is data to train or knowledge how – not money.
Deployment has to be done on a large scale to recuperate the training costs. You can't hide it. By having API access the model discloses its abilities and can even be used to improve the training set of your competition.
And models are compact, you can put SD on a DVD, but the original training set was 100,000 times larger. You can train a 13B language model from 1T tokens. You can download, save and share a GPT4 but you can't "download a Google" (remember what happened to the LLaMA model and 4chan). AI models are self contained pieces of data and code.
So my conclusion is that AI will spread wide and reverse the centralisation trend we have had in the last 20 years.
So Iron Man was born, and Stark flew away with his humming arc reactor in his chest. [2]
He looks incredibly youthful for being around 80. I guess that hum in his chest keeps him young.
[1] https://www.britannica.com/topic/Iron-Man-comic-book-charact...
[2] https://ironman.fandom.com/wiki/Tales_of_Suspense
To me it illustrates that if you have money and want to build something, you don’t need to hire coders or have any coding skills, AI will help.
And if you’re a regular coder, you just got replaced by Jarvis.
I was watching one of the movies and Stark was flying around some baddie castle/base. He tells Jarvis to locate all the missile placements and, when that's done, he tells Jarvis to target them.
That was when I thought, "oh, he's just a middle manager now" and lost interest.
I'm less concerned that AI tools will replace developers in the short term, and more concerned that it will encourage incompetent management to try their hand at "contributing" to projects using AI tools, creating more headaches for the developers. Kind of like how Blackberries made managers feel super productive firing off emails, while adding significantly to the work load of those under them.
Because it's a movie.
In reality an AI like that would be more important than all the other nonsense. But it wouldn't be fun to watch an AI easily unravel the mysteries of magic, time travel, multiverse, etc.
And, so, if we get Jarvis level AI, and you will get rendered obsolete.
When these tools come, inheriting billionaires will hire 10 people with Tony Stark intelligence as opposed to 10,000 devs with your level of intelligence.
So, don't be in denial that you will be made obsolete.
(And probably so will I.)
edit: I was wrong. The site just doesn't have a http -> https redirect
https://sampleapis.com/api-list/wines
At least when analyzing actions. Is that not the case?
* take a performant language model (copied from some other team e.g. GPT-J)
* show the machinery that it can learn new tokens using one or more tools from a provided set of tools ... e.g. WikiSearch tool
* demonstrate that the sequence of characters in the full API call content, has some effect.. e.g. no reply, useless content, or content that helps predict another token. Save that complete set of characters in the API call as an entry
* run a learning session with tool calls to APIs, improve the model for resolving known or new tokens (queries with answers)
* show the model that it can try new combinations itself (!)
* let the machinery try API calls itself to resolve tokens
* minimize loss functions for API results
comments - this is strikingly different than some RDF hard-wired data store.. it is using huge numbers of failed attempts to find results that work. The results that work are complete API calls as type string. It does not seem to care about "learning" the contents of the API calls, just that it is methodical about remembering API calls that work, and retrying those
not a specialist, feedback welcome
LLM are very good translators and synthesizers.
If something appears to do an action, then it's a useful shorthand to say that it is doing the action. Even if it is doing something else under the surface
‘It’ doesn’t need anthropomorphizing in the slightest. It’s a computer program that can slurp up data.
Dog owners across the world can confirm this for you.
"understanding" on the other hand...
Our jobs are safe.
So given time the AI could no longer debug new software as there would be no programmers only assistants.
But what about if the path is actually /beer and you have to pass a hidden undocumented query parameter called ?wine=1 for it to give you wines. But then the response is still of Beer objects (because that's the only thing the API validator would allow), so you have to map all the Beer fields to their equivalent Wine counterparts. But not all of them make sense, so you have to ignore them. Which ones? Ask the engineering team. Turns out the guy who wrote all this left years ago, and no one remember how it works. Someone digs up a link to a documentation page, but that internal wiki was taken down and so it returns a 404. You ping a sysadmin to see if they kept any backups. He points you to a few PBs worth of SQL dumps from an internal migration a few years ago and asks you to take a look in those. You simultaneously have to write up a status update for senior leadership which is due by end of day and give them a revised launch date for the project. They want to know why it can't be done in half the time.
The day an AI can figure all this out, I will be looking for another career. Until then I'm fine.
If you let the AI both create the API and use it, you can avoid the problems humans create.
Unless all jobs are unnecessary, you just change your job.
People seem to assume that if robots can code, all coders are out of a job, but I mean, if I didn't have to code, I'd just start all the businesses I've always wanted to start because like all the business people who are going to be replacing me with a bot, I too will have the time to replace business people with my bots.
Of course by this stage, if coders are replaced, pretty much all jobs will be replaced, millions, maybe billions of people will be jobless and without any ability to earn a wage, in other words, you won't be the only one with real problems when coders lose their jobs, it's just a matter of time till most jobs follow.
Not sure what we're hoping to get out of the AI future, but it seems like we're very intent on finding out at this point.
Maybe we're all just working towards making Ray Kurzeweil immortal at this stage?
My gut tells me though, we'll just be augmenting ourselves for a long time, and things will become faster paced.
The most I can imagine coming out of this is that those "app creator" apps/websites will actually work. So you can get a reasonably functional app without necessarily being a software engineer. A current-day software engineer could end up with the capabilities of a pre-2023 VP of Engineering in charge of a 200 person department.
That doesn't mean there'll be no work, it just means the work will be different and possibly much more interesting
something funny about the way I interact with ChatGPT, I basically never give a shit about correcting mistakes or bothering with good grammar because no matter what, it pretty much always understands me. and it's insanely good at solving tip-of-my-tongue type problems.
I guess maybe we should go back to SOAP to eek out a few more months of job security?
Given what things looked 10 years ago, and what was revolutionary back then (ImageNet challenge) - it's hard to comprehend what state of the art will look like 10 years from now.
Either I’ll be well prepared for unemployment or I’ll have plump investment accounts and still be employed. Good outcome regardless.
I’m not expecting the singularity within that timeframe nor do I think we need it for lots of disruption in knowledge-based work. Still, I’m leaning towards a near future where these AI tools augment our capabilities more than the alternative where we lose all our jobs.
The humanity is hopeless.
https://twitter.com/sergeykarayev/status/1569377881440276481
https://twitter.com/sergeykarayev/status/1570868002954055682
None of this seems like it would be hard to implement. The memory would be the more out there thing but even that would be a matter of retrieval.
Mostly to not be restricted to only ones that were known in advance and so could be hard-coded.
On the other hand, I asked for some trivial examples of Nim macro syntax and it couldn't make something that compiled even after feeding back the compile error messages.
If there's a private/obscure API you need to deal with that is the size and complexity and has the amount accumulated cruft that Win32 has, yeah, you're out of luck (but maybe should also be reconsidering the life decisions that got you to that point).
Sounds like snake oil to me, or fake it until you make it.
https://twitter.com/sergeykarayev/status/1569377881440276481
https://twitter.com/sergeykarayev/status/1570868002954055682
None of this seems like it would be hard to implement. The memory would be the more out there thing but even that would be a matter of retrieval.
Not the creator so i won't make any definite statements on how it works on the memory side but if you embedded and stored the chain of thought "how to use x API" generations and then retrieved them anytime the model was going to use the same API, it would certainly simulate a memory of sorts without needing any incremental training.
I feel like for the actual job lots of us do day to day, we’re ridiculously overpaid. I’m not even on a rooftop with tar and a mop! I’m enjoying it but I’m expecting the gravy train to stop eventually.
I think for my next career I’ll do something with… hmm drawing a blank. Let me ask ChatGPT.
…well… it suggested project management.
I hope my mortgage is paid off before AI makes me redundant
But I’m thinking of:
1. How badly secured many real-world APIs are and how many API keys are out there to be scraped
2. Some APIs offer truly destructive or otherwise irreversible actions.
Suppose for one second there was a poorly secured API to control medication dosages. Or to deliver a lethal injection to a death row prison inmate. Or to detonate a building scheduled for demolition. Or, …