It might sound really crazy and stupid today, but when SQL came out, it's advertised as one of "program-generating" languages and was supposed to reduce the need to code.
https://en.wikipedia.org/wiki/Fourth-generation_programming_...
(I mean, in some sense it's true because it's much less code than writing our own database...)
That never happened. SQL is hard to master, but it’s the easy part of understanding the relational model and any particular schema. Instead Oracle and the other RDBMSs that followed created more jobs for programmers and database analysts and admins.
From my point of view, it totally did happen? Can you imagine how many programmers the company would've needed to get all the data a business analyst casually queries per day?
What you're looking at is the quantity of people actually employed in the industry, not how many SQL made obsolete. The industry just grew so much that it didn't become an issue.
With more flexibility in the database companies could collect, store, and use more data. And that data had to get into the DBMS somehow: more code.
Setting up and managing databases required DBAs, a role often filled by programmers in smaller companies, and still filled by programmers today. And only larger companies had business analysts. In smaller companies programmers or maybe technically-proficient managers did that work.
Anyone who had devoted their career solely to building bespoke database systems had to pivot or walk in the late ‘80s, but very few programmers only did that — it was part of the larger application development. If you were good enough to write robust database code pre-Oracle you had plenty of options.
In the ‘80s when RDBMSs started to take over I worked in enterprise logistics. Oracle created jobs at the places I worked, and no programmers had to leave because we got a better tool.
I’ve worked in large and small organizations for 40+ years. I’ve never worked with a manager or analyst who could write SQL more complex than “SELECT * FROM orders WHERE total > 10000.” In sure they exist, but not in threatening numbers.
Maybe ChatGPT will get good at designing relational database schemas, who knows?
Give what even pessimistic AI alignment people are saying, I think you're correct that there are a few more repeats possible before AGI.
(Whether the pessimists are correct that all of them will happen before the 30s, I cannot say).
I think a lot of smaller companies would struggle if that was still a requirement, but if relational/SQL had not come along we’d have something else like it.
Indeed SQL was not the only local maxima we could have gone for, but the point is that having an easy to use database with a powerful query language did not reduce the number of jobs, but instead increased it. Instead of a few companies hiring a lot of developers, we have a lot of companies hiring a few. The latter will usually mean more jobs.
Not really. There were a ton of ISVs competing with Oracle and pretty much every mainframe and minicomputer maker also had their own database products, many of them non-SQL.
All commercial databases pre-Oracle were non-SQL. Watching that history get rediscovered and pushed as an improvement by people mostly too young to know better — so-called NoSQL databases — brings to mind Alan Kay’s quip about the software industry constantly reinventing the flat tire.
The hash-based Pick OS/database came out before Oracle but only ran on a few computers from Microdata and later Pr1me and Honeywell. Pick-based systems remained popular into the early 2000s in some sectors. A friend of mine still works on one, for a state government agency.
You could construct a database management system from the ISAM-based libraries for COBOL or Fortran, but I wouldn’t call those database management systems comparable to Oracle. Mostly they didn’t have a query language per se — you had to write code to get anything in or out. And they ran as part of the application process, not as a separate server dedicated to running the database engine.
* language syntax
* the relational model (relatively easy? It is just "there"?)
* database schema, where I have 3 000 tables, or 50 0000 tables (that's how ERPs are made)
* actualy knowing how to use the language
* building those queries that join columns from 15 different tables together to deliver the needed data -> it sounds easy, but Im struggling to do it better/faster - I never saw a book or website that focuses on that (multiple joins from multiple tables), when it feels my work is is mostly that
* understanding what is needed
* actual "programming" problems (say: you have 100k items on stock, those are parts of BOMs, make a list of 'where used'... yes I know you can google it and try to copy from stockoverflow)
Seriously, I am trying now how to learn how to setup a simple DB to consolidate data (20 csv files on a shareddrive) and the guides are often like in the 'how to draw an owl' meme...
Controlling/reporting/analyst jobs feel in some way as "programmer" but without tools/training/salary - just figuring stuff on your own. Im doing it, but apart from that SQLzoo website I didnt manage to find any decent guide for the 'usual' problems. Also since those are like 10% of my work I cant focus on those problems properly - and try to learn it after work.
Also SQLzoo is a lot like the "how to draw the owl" meme.. the easy tasks are easy, the hard ones are impossible and there are no medium tasks. There also dont seem to ne any real life examples like those tasks like: 'join 15 tables to make some report'.
* Rewriting queries and also parts of the schema when it turns out the things that made sense in the design phase cause massive bottlenecks once you get more data, or just get it faster, than you tested with.
Of course the good news is now you can run the best RDBMS's on your laptop, or on a $5/mo VPS; or have a small managed one starting at <= $15/mo. Plus ChatGPT can help you remember how to do that inside join. ;-)
You can reach me at nominallyanonymous-at-protonmail-dot-com. From there I’ll give you my “durable” contact info - Slack, Discord, SMS, Signal, Telegram… whatever you use regularly, basically.
I recommend Chris (C.J.) Date’s books.
On a side note, if you have to join 15 tables for a report it is a sign you may go in the wrong direction. In 99% of the cases I never needed more than 5-6 tables if the data is well structured, but that takes years to learn.
I don't know of any modern environment that functions like that, on any platform.
I'm posting this to invite others to either a) correct me and tell me I'm wrong that tooling doesn't matter (a legitimate view, but wrong), and/or b) recommend some modern, OSS Access-like tools that might help flatten the learning curve for you. (And if you're more comfortable with a CLI and a REPL already, then you don't even need this hypothetical tool, but I myself am curious about it.)
EDIT: looks like LibreOffice Base is pretty close to MS Access circa 1997.
But there are good reasons where you don’t want random people running sql queries on production database or having direct access to the data.
https://www.zdnet.com/article/excel-errors-microsofts-spread...
I have seen this multiple times. Got called into a corp. audit once because CFO's spreadsheet didn't match reports from database. His SUM() function was missing some rows, ZIP Codes got converted to scientific notation and he didn't know why, that kind of thing.
If you get experienced db guy to do excel I bet he will do it correctly.
If you get “data jockey”, whatever tool he gets it will be a mess.
Getting basic queries or basic calculations in spreadsheet a person can learn in couple hours.
Somewhere after that comes zone "proficient enough to shoot himself in the foot".
Then there is level of excel or SQL where without months of experience you simply don't know what is happening.
What it didn’t do was reduce the need for programmers, because the new SQL users always wanted more data to answer more complex questions.
IA is the new Excel.
The funny/sad part about computer science is that people don't want to understand the costs of customization and sales/management/marketing forever want it as a differentiator.
SQL could have eliminated us from the flow of many niches as easily as the spreadsheet did from business operations. I think why it didn't has more to do with market timing.
It’s just that, with more powerful tools we can create more advanced and intricate things, so it never ends…
I not exactly welcome but rather anticipate, a near future were instead of reinventing every kind of pattern and library in terms of cloud and infrastructure management, we will see the same meta-management of different AI engines and services…
Even funnier - we've already lived through the great 'software development adoption' by business and hardly noticed, except some of us got work out of it. A lot of small businesses (legal practices and accountancies were particular suspects) grabbed MS Access with both hands and went wild in the 90s/early 2000s. Timesheets and other HR functions were popular applications.
SQL was a typical failure in this regard. It was seen as a savior but at the same time was designed for at least math-versed people. The prediction was right, its scope was wrong. Since then we failed many times and as I see it, we tend to diverge from this idea more and more. And the reason is there’s no one controlling our stacks vertically, so it takes more jobs than it could to make an app.
With SQL you need none of those.
But that's me looking into my crystal ball, nobody can say what will or will not be possible in a given time frame. But I chose to not worry about it - new developments will probably be accompanied with new opportunities, which I can jump on even if I didn't predict them ten years before.
And even if AI is able to do that... Well, then anyone out there could become a UberCycle CEO. Which means: customer requirements will be much more complex by then and coming up with an entire product in 1h won't be enough anymore.
I also expect, as you say, for this to create a lot of competition and increase the minimum quality that customers demand such that… heh, just realised, "Make Uber for bicycles" is basically going to be the next generation's pre-interview coding challenge to screen people that can't use AI effectively.
Which isn't actually different from today. We've just shortened the time to get your MVP to market. Rather than having to raise some seed and hire devs, you can do it with an AI in a few days perhaps. Everything after that seems it would be the same.
Why not let the people on the floor complain to a AI manager, the people on the floor know exactly what is killing their productivity. Like "yo ai manager, manually filling in those data field is taking quiet some time to check someones credit score is it possible the IT department can automate that process it might save 2 min a form" then AI manager can prioritise stuff like this.
The reality is that they frequently don’t.
I always found that overwhelming majority of professional academic educators are particularly bad in anything that has to do with the practical side of their domain. That is especially true for software engineering.
University lecturers are very detached from our field and it shows in how badly fresh grads are prepared to do any dev work. For that reason I almost always ignore their projections.
I feel like there's currently a movement towards slightly lower abstraction, or at least simplified, consistent APIs, less magic. The rise of Go and Rust are examples of this. Typescript could be another, although the abstraction isn't really lower, it is once again an attempt to coerce JS into something workable. I get really frustrated writing Python or Rails these days due to the sheer magnitude of hidden magic that sometimes works and sometimes doesn't.
To tie this thought in with generative AIs: Currently they seem to be much better at programming with relatively simple syntax. By far the largest success so far I had with shell. Basically I tested assisted writing the same tool in Python, Perl and shell, and the results in shell where close to perfect. ChatGPT was even able to accurately limit commands to specific OSs, and was able to accurately summarize the functions of other shell scripts.
So my prediction is that we will see a movement towards simpler, lower abstraction languages while Coding Assistants rise to take away the boring stuff from programming and programmers will have the time to think about delivering value to product, maintainability, and efficiency and correctness.
One last addendum: ChatGPT is really incredible at assisting with sys admin stuff, my guess for why is that there are a gazillion obscure forum entries going back to the 80s explaining basically everything there is to know, but these are hard to find or comprehend for humans. With and AI assistant, self-hosting becomes much easier, and another development could be for startups and smaller companies to move away from AWS et al.
Fwiw, GPT doesn't seem to write Clojure as well as Python or JS or even C++, which is obviously much more syntactically complex. That said there is probably an order of magnitude more training data for all of those.
The model I have is closer to “how many people can program using Excel vs how many can program in a ‘traditional ‘ language?” The difficulty in specifying exactly what you want is still there in Excel, but there’s far more people who can get started and they can make changes when they see it’s not doing what they want.
It doesn’t have to one shot “hey, implement these 59 Jira tickets for me” to be disruptive (in a good way). It’s extremely rare for something to get much cheaper and for there to be less demand for it. If you’re the top half of a field that’s in a lot of (and growing) demand, I think you’ll do fine.
Are we sure that an AI could not engage in enough back and forth conversation to firm up the spec? You’re kind of assuming that systems will be generated from a one-shot prompt, but it seems more likely that an interactive AI will identify the gaps in requirements and ask for clarification.
Alternatively, if the prompt-generate-evaluate loop is short enough the user can simply play with the running system and provide feedback to alter it.
This is essentially what developers do when they present a “weekly build” or whatever in an agile environment.
The process of solidifying requirements, stating them clearly and translating them into machine-executable formats are all language tasks and these models are really fucking good at those.
I’ve noticed in discussions like this that many software folks are assuming that AI capabilities will plateau soon, or will merely be extensions of what we already have (a better autocomplete, etc). I submit that we may reach a point where the AI is so compelling that we’ll reorganize teams/systems/businesses around it.
Until that day happens, i highly doubt that a business owner would just blindly trust an AI to generate their business code/software, without hiring someone to at least look after it. Therefore, software jobs could evolve, but not disappear.
This is the doomsday argument. What would I do if there's a nuclear apocalypse before lunch? I guess I'll die like everyone else.
An AI sufficiently advanced to do that is also sufficiently advanced to run the entire business in the first place, and also argue cases in court, do my taxes, run for president and so on.
You either believe that transformers models are "it", or you haven't actually removed the problem of specifying requirements formally. Which, you know, is actually much harder to do in English than it is to do in C++.
This is actually something that makes me happy about the new AI revolution. When my professor said that I thought he was an idiot, because no-code tools always make it harder to specify what you want when you have specific wants the developer didn't think about.
We give kids books with pictures because pictures are easier, but when we want to teach about more complex topics we usually use language, formulas, and maybe a few illustrations.
I still think no-code was always doomed due to the fact that any attempt at it lacked the interface to describe anything you want, like language does.
AI is finally putting an end to this notion that no-code should be clicky high-maintenance GUIs. Instead it's doing what Google did for search. Instead of searching by rigid categories we can use language to interact with the internet.
Now the language interaction is getting better. We haven't regressed to McDonald's menus for coding.
Or a single language that has both visual and textual components
Or a single language where each component can be viewed in textual or visual form (and edited in the form that makes most sense)
For example, this guy was able to do some amazing stuff with ChatGPT. He even managed to get a (mostly working) GPU-accelerated version of his little sample "race" problem.
No, the amount of handholding you have to do to get it to work effectively presumes you already know how to solve the problem in the first place.
The best way to use it is the opposite everyone is busy selling: as a linter of sorts that puts blue squiggles below my code saying stuff like "hey stupid human, you're leaking memory here", or even "you're using snake case, the project uses camel case, fix that".
That would actually lower my cognitive load and be an effective copilot.
I asked GPT-4 to give me an SVG map of my town. I then asked it to put dots on some local landmarks. The map was toddler level, but the landmarks were relatively accurate in terms of their relationship to each other and the blob that it drew.
So this is a language model that has some emergent notion of space in its code generation abilities.
The point I was alluding to above was that the prompts themselves will be recursively mined over time. Eventually, except for truly novel problems, the AI interpretation of the prompts will become more along the lines of "that's what I wanted".
Some things to think about: What happens when an entire company's slack history is mined in this fashion? Or email history? Or GIT commit history, with corresponding links to Jira tickets? Or the corporate wiki? There are, I'd guess, hundreds of thousands to millions of project charter documents to be mined; all locked behind an "intranet" - but at some point, businesses will be motivated to, at the least, explore the "what if" implications.
Given enough data to feed upon, and some additional code/logic/extensions to the current state of the art, I think every knowledge worker should consider the impact of this technology.
I'm not advocating for it (to be honest, it scares the hell out of me) - but this is where I see the overall trend heading.
In a world where we have the technology to go from two lines of prompt in a textbox to a complete app, no questions asked, then the same technology can run the entire company. It's kind of hard to believe transformers models are capable of this, given we are already starting to see diminishing returns, but if that's what you believe they are, then you believe they can effectively do anything. It's the old concept of AI-complete.
If you need to formally specify behavior, at any point in the pipeline, then we're back to square one: you just invented a programming language, and a very bad one at that.
This remains true for any version of a language model, even an hypothetical future LLM that has "solved" natural language. I would not rather write natural language than formal language given the chance.
But what if the "programming language" is not a general-purpose language, but a context/business domain specific language? One that is trained on the core business at hand? What if that "language" had access to all the same vocabulary, project history (both successful and unsuccessful), industry regulations, code bases from previous (perhaps similar) solutions, QC reports, etc.? What if the "business savvy" consumer of this AI can phrase things succinctly in a fashion that the AI can translate into working code?
I don't see it as a stretch "down the road." Is it possible today? Probably not. Is it possible in 5-10 years time, I definitely think so.
For starters I'd like Codex to be more than next word predictor, it should also "feel" the error messages, data types and shapes, file formats, so I don't have to explain the context. It should be part of the system, not just part of the text editor.
Especially the "debuging" part will be much harder. Noone can look under the hood to understand what is wrong and all you can do is shoot random prompts in the dark hoping it will create the right result.
It is scary right now how confidently and spectacularly wrong the chatGPT is and it will create disasters.
Not that we would have missed much from missing the particular implementation of this idea that Facebook gave us but the idea and what it unleashed is much more than that particular implementation.
Whatever the AI prompter brings to the table will quickly be provided by the AI itself. If a user doesn't really know what they want, there isn't a scenario where the AI prompter will suss it out but the AI itself won't
It seems prohibitely expensive to build and run transformer models with that much capacity.
I am in the process of adding VMs which the AI will be able to write software and fix compilation and other problems automatically.
Sounds like me get reorganized out of a job though...what does it mean to reorganize everyone around the A.I if it does everything better than us?
Off to feed my entire task list into GPT to see if it can prioritise me without saying I need to learn to juggle tasks
If you pinch this idea to do it better pls remember to only send the task metadata to GPT rather than the entire task so that I can also use it for business stuff, ta!
Here's my concern:
1. AI assistance thrive on structured data
2. Computer programs are some of the most structured data. And it's available abundantly out in the open.
3. Yes, you can't generate an Uber for bycycles with a single prompt, but you can fire half your development team and increase the productivity of the rest of your dev team with an OpenAI subscription.
Another perspective: Given that…
1. …OpenAI accelerates development productivity (especially for lower skill devs, but also for product and marketing);
2. …there is a notable arms race at the moment that is accelerating tech and business innovation at a blistering speed;
3. …there is still a lot of money looking for growth;
…then it seems that…
4. …higher standards of productivity will be expected across industries just to keep up;
5. …teams can complete existing work at a much faster pace;
6. …new product concepts can be prototyped in a period of hours/days;
…then, this could arguable result in…
7. …weak new hires pulling their weight like never before (ie, interns writing marketing copy is a good example, but so too is intern-level development or prototyping);
8. …existing employees delivering far more value;
9. …an overall increase in demand for human labor.
Looking around society, there is clearly a LOT of work to be done. Maybe the issue isn’t the lack of need for labor but the lack of management ambition towards a future vision. Being able to “lean in” and adjust to a higher rate of innovation and productivity brings a much lower set of risks than ever before. Of course, this puts more demand on management, product, marketing and sales.
Bottom-line: if OpenAI amplifies the quality of all employees, then maybe job-loss isn’t inevitable. Perhaps, just the opposite.
This is the same Fallacy that we hear since 50 years. All Program requirements are almost the same, just reuse and adapt an existing one. Guess why it has never worked? Because the premise is false. Structured data for X is not optimal for Y (and can be even very wrong).
Apart from the "personal blog software", everything else has various needs of accountability. AI Black box approach is not suitable for any of these so you have to manually verify the code. Veryfing code that you are not familiar with especially in complex interactions is much more difficult that writing it (from this comes the often "rewrite from scratch request", because institutional knowledge has been lost, imagine how much worse it is if this knowledge has never been there in the first place).
Finally and the most important one, all AI models rely on learning, if there is noone to learn from all you get is stagnation. Most of the breakthroughs come from a complete reimagining of the solution space. If the solution space is fixed because "AI has substituted all Engineers" there is no going forward.
For the tinkerer, no-code tools are (probably) great, but if successful, even tinkerers will hire managers who will hire translators.
There is definitely a tendency of some people to avoid tech at all costs if they can and trust in "tech people." Fine with me, I ask them about my people questions often enough.
The ergonomics of mainstream modern programming have been stable since at least the 90s. Any fundamental improvements that were going to happen have already happened.
https://en.wikipedia.org/wiki/The_Last_One_%28software%29
> The Last One is a computer program released in 1981 by the British company D.J. "AI" Systems.[1][2][3][4][5][6][7][8][9] Now obsolete, it took input from a user and generated an executable program in the BASIC computer language.
It was THE LAST ONE because it was the last program you'd need to buy. FOR-EV-ER. While its baseball-eating ability was unknown, it was "all the programs you'd ever need for £260" and it apparently focused on ETL jobs; that is, extracting data from files, transforming it, and loading it into some other file. Talend for the Doctor Who set, in other words, and it was set to EX-TER-MI-NATE programmers. Maybe it did; after all, we don't very well write programmes, now, do we?
https://modeling-languages.com/last-one-code-generator-basic...
https://archive.org/details/PersonalComputerWorld1981-02
It certainly got a lot of hype at the time but, like all Coming Things, it's hard to tell where it Went.
By all means, that is part of everyone's job. Sometimes I really can't remember how to do some specific thing. I know precisely what I want, but I don't remember the proper way to do it. I would have resorted to a web search until last year, now I ask chatGPT. It is faster, more concise, and surprisingly accurate. And when it's inaccurate it's easy to either refine my question or cross reference what it generated with online sources.
But I think it's a bit silly how people are treating it as if it was some kind of General AI. It is not, it can only give out known answers to known problems based on language statistics. As impressive as it is, it can't reason logically about problems without known solutions, it can't identify faulty, incomplete or inaccurate information, it can't evaluate on drawbacks and tradeoffs of different approaches.
ChatGPT made like 20% of my work a lot faster and less cumbersome. I like it enough that I pay for premium access. But even the notion that prompt engineers might replace software engineers is silly. I imagine people repeating this nonsense are either not engineers, or extremely junior in the profession to the point where their work is only writing code for clear specifications, no questions asked.
Are you sure you actually tried, or is this an a priori argument?
(btw, apparently GPT-4 does much better on mathematical reasoning, I've yet to try myself though)
It fails sometimes on known problems, not even particularly hard ones - i.e.: simple questions about the capabilities of programming libraries for which there is ample documentation available.
The thing is, by talking about its shortcomings I may be sounding dismissive. I'm not. I think ChatGPT is an amazing tool. I just recognize it's shortcomings to use it to its best capacity.
For things that don't have a known answer (i.e.: I couldn't easily find an answer online) it does generate plausible bullshit. On those cases it shines for things that don't really have any strictness requirements (for example, ask it to generate some fiction, ask it to generate a polite response to an email, etc). I used to suggest character names for a game I was going to play based on some parameters, and I loved the responses it gave me.
I think software engineering is going to look very different in a few years, and likely be a smaller field with lower paying jobs. But it's not going away in the near (5-10 years) future.
https://abcnews.go.com/Technology/openai-ceo-sam-altman-ai-r...
I neither know how LLMs work nor how our brains work. And I don't know what could be parallel between these two. For my very very limited knowledge of how properties can emerge from unique arrangements of constituent components (the S-R latch giving rise to state - i.e. memory - comes to mind), I would not at this point write off the possibility that a very large / very deep / very intricate neural network trained on language prevalence in very very large datasets could manifest properties that we would interpret as reasoning.
And I further wouldn't write off the we humans may owe no small part of our reasoning ability to language comprehension that we begin to ascertain from infancy.
ChatGPT is impressive, but gets many things wrong. If you know what you are doing it's an amazing programming assistant. It makes me noticeably more productive. It may lead someone who doesn't know what they are doing in weird rabbit holes that will lead nowhere however.
One silly example. I was using a library I hadn't use before, and I asked how I could get certain attributes. It gave me an answer that would't compile at all, the imports didn't exist.
Then when I mentioned that it didn't work, it game me a slightly different answer, that also didn't work, and explained that the previous answer was valid for 3.x. in 1.x or 2.x the new answer was the correct one.
But there's the catch. There's no version 3.x. there's not even a 2.x. It's language model just statically got to that conclusion.
Doesn't make it any less impressive to me. It gets things right often enough, or at least points me in a good direction. I effectively learned new things using it. But it can't replace a developer.
Using ChatGPT as if it was General AI is similar to eat a meal using a hammer and a screwdriver as utensils. You can probably do it, but nobody will have a good time.
If anyone thinks genuinely that ChatGPT can meaningfully replace a software developer should be starting a tech business right now. You can replace basically your largest cost (staff wages, especially for expensive pesky software developers) with a $20 account that will work overtime, will never get tired, and whose performance will never fluctuate.
No excuses for not getting insanely rich. Of course it's not happening, because it's bullshit.
I am genuinely impressed and even excited with ChatGPT. It's an amazing tool, that would have been massively helpful when I needed a good NLP in the past. It will certainly be massively useful in the future. Hell, it's being a great assistant right now.
But it's not General AI, and I do facepalm at people LARPing as if it were.
General AI will in day exist, and knowledge jobs will most likely be a thing of the past at that point. How far in the future it is is anyone's guess (my prediction is in between 12 and 235 years in the future). We're just not there yet.
I think as a society we don't really have precise words for describing the different levels of intelligence, except in a "I know it when I see it" way. I don't think I'm LARPing in any way, I'm probably even less excited with it than you are given that you seem to be using it more often than I am. I'm just saying I think GPT does exhibit some logical reasoning abilities and not merely remembering statistical patterns.
I agree with most of your reply, except this bit.
I mean, generating a response through statistical language patterns is a sort of reasoning, and ChatGPT has been accurate enough to replace internet search for me quite often. But it also generates bullshit that an untrained eye would miss (because the bullshit it generated was statistically plausible).
When it get things wrong it generates some comically wrong behavior. I had one case where it looped through variations of the same wrong response - precisely because it is unable to do any kind of logical reasoning upon faulty or inaccurate data.
- gpt doesn't need you to be exact
- one day gpt will be able to ASK YOU question to refine what you need
It may take years, but at some point, it will be able to do 80% of my job.
Not a big deal though, cause people will still hire me to do the remaining 20%, and pay me even more money because not a lot of devs will be able to do it.
ChatGPT can do more than we are lead to believe. Don't believe the canned responses OpenAI triggers. Yeah yeah I know, convincing parrot chinese room, overpowered autocorrect...but what is the difference between convincingly faking it to within epsilon and actually having it as an emergent property?
https://twitter.com/bitecode_dev/status/1637420579278954501
Now that's going to be interesting.
Sure one class of job has taken a huge hit (sys admins, supposedly ops), but it's been replaced by another that needs intimate knowledge of the tooling required to do what was being done previously (just in a more "scalable" way). DevOps have been demanding a premium for years now.
And there are still sys admins out there doing what they've always done..
The equivalent assumption would be that ChatGPT and MidJourney can't work because you have to specify every paragraph and every shape.
Clearly that's not true. Just because every object and every texture had to be defined rigidly by hand in 3ds Max or Blender by highly skilled designers doesn't mean that's how it works today.
In fact AI is the stage of software where abstraction itself is automated. The current Gen 1 AI tools do a little of this, but as Gen 1 they're just getting started.
They're going to be in a very different place ten years from now. And so are we.
> The current Gen 1 AI tools do a little of this, but as Gen 1 they're just getting started
The current tools are not Gen 1. You can already see diminishing returns.
But aren't managers there to coordinate software engineers. So if software engineers are out of a job, so will the managers. Assuming AI replaces software engineering, the ones left may be the product people and the founders, perhaps rolled into one person in several cases?
Which aligns somewhat with what Carmack is suggesting.
> Instead he made this prediction because he saw the rise of no-code tools replacing software developers
I think that people who make these predictions are not very good at actual programming (taken in the more wider meaning)
The latest fad I remember (before ChatGPT) was that with BDD testing non-technical people would be able to write tests and we all see where that went
But most fundamentally, the non-technical people don't have the time nor the expertise to learn all the details needed for shipping software (and why would they? their job is to look at other aspects of the business)
That might actually be what development is, teasing out the requirements by iterative entering them into a machine that does exactly what you told it to, and comparing the outputs to expectations.
This. I started programming in Machine Code, where the "editor" was a pad of graph paper.
I've watched management- and business-focused people sneer at us geeks for my entire career, and watched them drool over the thought of finally getting rid of us.
Hasn't happened yet.
> I'd be surprised if the next step is "Hi, I'm an ideas guy, please give me an app that does Uber, for bicycles, but better."
I get that, from "idea people," on a regular basis. They have nothing but contempt for us "Make It Happen" folks. They treat us, as if we grow on trees, and are suckers, to boot.
Inevitably, the above, is followed by something that translates to "Of course, you will do all the work, but I'll get all the money, because the idea is what's really valuable."
If I follow up on it at all, I'll be asked to sign an NDA, and meet in some clandestine place, so they can whisper to me about their AI-powered cheese straightener.
True. And once their product hits the market, if ever gets there, they don't thrive.
Prior to launch they are sooooo in love with their idea that they are meticulous about features all the while thinking they're smarter than the market.
They don't understand and appreciate the value of execution. Ideas are easy. Execution - because it involves people as well as adapting to change - is 10x harder.
Yes, those people exist. Unfortunately, that bias will ultimately undermine them, but they'll never admit it.
It tends to be a very long process, and often involves a lot of "trial balloons." I just went through that, in the last couple of years. The project we're realizing, looks absolutely nothing at all about what the CEO originally dreamed up, but everyone that has seen it, loves it.
The trick is to not start off by saying "It'll never work." That slams doors shut, right away.
It's more like, "OK, so let's walk through what we'll need to do, to make it work."
That will often result in changes being made, by the "idea person," as the plans are laid. We will also try to create test harnesses and prototypes. These often end up, with the idea person going "Hmm...it seemed like a good idea, but it doesn't work the way I wanted."
It's slow and painful, but works.
I could be a lot more of a pain in the butt than I am, as I have a lot of leverage, but one of my goals is to help others on the team (including a young CEO) to learn how to make ship software.
If I were working on my own, I probably would have shipped, eighteen months ago.
No one is making a dime. It’s a nonprofit effort.
When someone says "can we..." or spills their idea they've often identified a real need but they often don't seem to do any sort of analysis further than identifying a demand signal. Chances are, you're not the first person to observe the problem and an opportunity to capitalize on it and there may be good reason no one has yet (its economically infeasible given current understanding, resources, and value people see in it).
The trick is to hold their hand and walk them through a feasibility analysis. Given infinite time and money, I'm confident I can arrive at pretty much any solution or at least a useful approximation that's an improvement on where we currently stand, so we start there and walk backwards. How much time do we really have, how much money do we really have. What's your appetite for risk within those bounds. Let's start talking about some tradeoffs of what you want to see and what I think we might be able to actually achieve.
In my experience, they've identified a want. Either their want or a want they think the market has.
It's not until further analysis do they / you / we get to root need(s). This is why (allegedly) so many "IT projects" fail. The client got what they wanted...not what they needed.
Socratic Software development is the way to go.
The same reason why menial workers earn so little relative to how hard they work.
It would appear that both credit and money flow to those who enable things to happen rather than those who execute.
In top places though, it’s often the case that trainees have an idea, get shut down by PI, trainee demonstrates project can work, then PI changes mind.
Also in science, the usual progress of a project is often PI has an idea, trainees explores the idea, finds an even better idea, PI says great now write the paper.
But I agree with you in the sense that PIs act as the “quality control” or “selection process” whereby ideas get culled and refined. And their scientific taste is non fungible.
I'm reminded of reading about a Lisp machine company that ran into the ground because it was managed by techies. Their tech may have been great but their marketing and business skills were very arguably what killed the company.
Some MBAs are dumbasses. Some are brilliant business people. Some of them actually have tech skills and have been writing software longer than you’ve been alive. This is true for many degrees.
I don’t know how some of the anti-MBA commenters think it works: A smart software engineer goes off to business school, gets an MBA in addition to his undergrad technical degree/experience, and now… what? He’s somehow dumber? How does that work? Only on HN.
Now, if you already had relevant experience, then went and took the MBA course, and were able to identify when to lean on which, great! But, of course, part of that education is telling you that you don't NEED that experience in the first place, and if it and MBA principles disagreed, you should rely on the MBA principles.
So, if you took the course truly to heart, then, well, yes, you would be somewhat "dumber", so long as you define that as "unable to make correct decisions."
There was no expectation of being "air dropped" into an industry, and in fact, industry recruiters looked for candidates who already had prior experience when picking students for interviews and hiring. The students who had no relevant experience for the industry they were recruiting for (e.g. ex-Peace Corps or mgmt consultant looking to transition to tech or industry in general) were the ones who struggled to find roles. Your assumptions about what an MBA education entails are incredibly far off the mark.
But there’s one type of ideas person I’m thrilled to work with. Someone with deep and successful experience in sales. The head of sales at my company is also relatively product oriented and boy, he just has a knack for what customers want.
In general, I think many engineers can do some product management. We can figure out basic accounting and finance. But the idea of making a cold call to sell dental software, or chasing down school principals at an education conference is almost as foreign to most software engineers as CS is to most sales folks.
Salesperson + engineer = superpowers
(When I worked for a long-ago hardware company, system engineers (SEs) were pretty much the equivalent of SAs where I am currently.
Being the engineering contractor to SA organization is daunting to a programmer, but rather rewarding.
Right but how much of that is because humans don't scale? With the AI, if it can do a thing well enough it can do it 100 000 times per day.
In my mind, having individuals and small self managed teams go off and do whatever they feel drawn to do is exactly how a good company is run. - So long as those teams and individuals talk to (& seek advice) from the rest of the company when their work has impacts outside the team. The book Reinventing Organizations by Laloux talks about this a lot, and how it works in some companies today.
The reason it doesn't happen more is that upper management doesn't trust their employees and they don't feel in control when people just do things.
All those jobs do is push emails rewording other emails to other people pushing email.
I feel like eventually every company will have a single a figurehead that’s fed instructions but believes it’s original thought and is told by the AI what a great idea they just had all day long.
A bit like the Captain from Wall-E.
Not the full job, but it can probably be built right now.
Anyway, the good version of those three are completely different, and add a ton of value on very different places. I think they are rare enough that many people don't ever meet one of them.
Maybe it's going to be AI enables quality more than quantity?
He's still expected to put in the same amount of time, but now he has to expend brain power the entire time.
Obviously that didn't happen. And people actually made predictions like that once upon a time.
As software becomes cheaper, and the production of software can be done with fewer people, demand has always increased and I don't see any reason that's going to stop.
Reminds me of this: https://en.m.wikipedia.org/wiki/Jevons_paradox but the situation with software is a bit different.
Had Lisp machines (as an idea) remained viable, we might've come close! But managers just love having large teams to command, and we've invented process and ceremony to fill up the spare time necessary to do the work of ten programmers with a hundred, as well as rejecting technologies like Lisp and Smalltalk that give individual programmers tremendous leverage over the problem space.
Should AI actually make people more productive, instead of being an endless generator of messes for fleshbags to clean up, I imagine some successor of Scrum to come along to tie up most of staff programmers' work days with meetings. Hell, that may happen anyway.
Asking a human manager for anything is many orders of magnitude more frightening. It's not that I don't have questions! I have many, of which a lot are dumb questions, questions that I should be able to figure out without help.
Say i'm pondering taking a day off, with the human I would make up my mind, pick the day and ask for the day off. With a bot I would ask which day would be ideal before making up my mind. A human manager would think you've gone insane if you asked 20 such questions. You either want a day off, you don't or you say nothing.
I might like a different brand of coffee while we are trying to meet the deadline. I'm sure that question is going to go down well with an overworked human manager.
With people asking a question is never just asking a question.
Like with any job there is a sub set of tasks better done by a machine.
I never realised Bullshit jobs was written by an actual academic.
It's not a new idea.
I think it's an underestimated aspect of why CEO's get paid the big bucks.
They're paid money to not let this stupidity happen.
On their side, the engineers might even ask it “summarize my tickets and come up with an optimistic, business-friendly high level storyline.”
The only time I see synergies is when we talk about what task to take on next and why.
(which has been major time sink of engineering managers that I know).
Like when the brick of cheese comes out of the freezer with an obvious bend?
Or like when I don’t make a nice square cut and my sandwich is cheesier in some spots than others?
This is also one of the main reasons why all programming jobs were not outsourced to India.
There's a whole industry here in America that re-shores programming contracts. They know they can't underbid Indian/foreign body shops so they just wait a few months and call back the companies who went with cheaper programmers. If the company is still around it's generally a complete re-write.
Yes there are these "idea people" who don't have any clue about business side and don't have any clue about technical side and in the end don't even have the right connections or business network. But they think they can make it because they have an idea(TM). These could basically play lottery and outcome would be the same they might make it but chances are 1:1000000 at best.
Then there are these business people who have the right connections and have understanding of niche/business they are in .. that need technical help to execute their idea and these are worth their weight in gold.
Ideas are like assholes ... everybody has one and they're usually full of shit.
I've met my fair share as well. Jobs copycats too!
The trick is always to ask them how much they can raise, and do so assuming they raised successfully in the past and that they know VCs.
Either the conversation gets really interesting, or they clam up real quick!
They don't need to have the ability to raise anything in a B2B setting, if they can answer the alternative question:
"How many customers have you personally signed up for previous software sales".
I'll happily partner with someone who has profitably sold B2B software - having a list of existing contacts in existing companies as well as solid experience going through the entire sales channel is, to me, just as valuable as extracting money from VCs, because I'd rather have the money from customers than money from VCs.
I met a pair of such idea guys once, they ran their own idea company. Their specialty was brand consulting and online marketing, but if that involved something as menial as web development they'd help with that too. The thing that made me bristle was whenever they talked about the feasibility of software solutions for their ideas, they said "Oh, we'll run that through the machine." You know, as if their software team was just a big machine that ingested ideas and spit out code.
Their idea company is no longer around. One of them decided that he'd rather design hiking boots than anything computer related; the other is a head of UI design at Apple.
The best thing I find as a developer that it unblocks the resistance we have in starting up with something new. For instance, I just built my first Chrome extension [0] to make video players more accessible on webpages with keyboard shortcuts.
I had built a desktop app on similar lines long ago, but could never push myself to built an extension. Last night I gave my requirements to GPT and it guided me through the whole process – from creating manifests and js files needed, to the JS code (granted that I'd to fix/extend the code) and even how to install/enable the extension.
Within few hours I have my working extension without needing to go and read any extension docs.
0: https://github.com/varunkho/RamaPlayer/tree/master/Chrome
Doomsday prophecies are not limited to religious cults, you see them all around. It’s safe to say, from experience, AI is not going to outdated programmers but it’s going to make the mythical 10x engineer a common reality.
Some people can take a business goal and figure out how to turn it into a clear spec.
Sometimes these people are programmers.
However I know programmers who suck at this and I know non-programmers who don't.
As an aside, I'm getting a bit tired from the "programmers vs product/business people" trope. The average software engineer would be nothing without a good product manager in my experience. And going by this thread, there seems to be at least as much contempt from programmers/software engineers for product people than allegedly vice versa.
The average software engineer doesn't have a good product manager, he has an average product manager. They still get things done.
I think dividing work into programmer/UX/product manager is an actually huge regression.
Software is of lower quality today and requires far more do-overs than it used to. As someone who liked working closely with customers and making decisions in all these areas (for their benefit), it has made the entire career far more of a slog than it used to be.
Also, I find these endless internal cyclical conversations to be much more draining than actually writing code. Looping over and over with planners and ideas folks isn't energizing like actually writing the product is.
The same thing applies to trying to get business people to write "executable specifications" that can be used, if not to generate the code, to at least validate that the final system does what it's supposed to.
They always complain that the "tools are too hard" but the real problem isn't the tools but that they are unable / unwilling to precisely specify what the system should do. They just want to hand wave vague phrases and have their system magically perfectly defined. It's just not going to happen.
On an aside, this is the kind of bussiness idea a chatbot could generate now.
But I don't think it could have generated the idea for Uber when Uber* was the first...
* if you know the history of ride apps better, replace Uber with whoever was indeed first.
"Please give me a business model like Uber, for bicycles, but better"
Let us know how it goes :)
How does this manager-engineer look like now? Managers specify what they need in plain language, and both parties go back and forth with increasing complexity to create based on the shared vision. Managers can already pull off a similar dance with ChatGPT: give it a plain English prompt, it’s responds with what it thinks you want, and you refine until it’s got it. GPT4 can do this with website sketches, and who knows how much finer you could get from there by specifying your prompts and feedback more tightly over less than an hour. Remember: copywriters and brand marketers and marketing creative makers have a similar role of turning complex requirements into designs that sell (and their roles are dying fast).
A software engineers job is to realize products into software. ChatGPT is pretty much as capable as a high school programmer that moves really quickly and takes feedback to heart - and that can handle a good chunk of the software engineering job for a huge fraction of the price. Your job isn’t as bulletproof as you think, and especially not your amazing salary. I’m speaking as a software engineer turned vc backed founder, so I’ve seen both sides of this relationship.
Founders are weird because their job is just to “do everything that needs to be done” so the role flows like water into any spot that isn’t already filled in a company. Replacing that seems like by definition you have to replace almost everyone else first.
I’ve heard that exact same argument for software engineers.
As an outside observer it sounds like you’ve switched from SE to founder, so you’re failing to apply the same standards because you (consciously or not) want to defend your current position.
Do you think it failed to hold?
Founders are the jobs trying their hardest to replace themselves (with hires and go can do the work better) and yet still always find more to do. That’s very unique
That's not even close to unique. Software engineers have been doing that since they existed.
>but the founder role is kinda unique in that it’s just “do everything and if you’re successful you should be constantly replacing your job with an employee”.
That's maybe the ideal of a founder, it's definitely part of the mythology, but it wasn't true for the majority of the many founders I worked with back when I was consulting for early stage startups. Based on my experience with a large sample size, the job of a founder is to convince VCs to give them money.
>How do they apply to software engineers exactly?
It's one of the definitions of the singularity. When AI can write itself, it will improve so rapidly that humans won't be able to keep up.
Not really. They are not "replacing self", the coordination is always needed, just hire the people to the point coordinating the company is their only job. I.e. finding peasants to do their dirty work and they can just rule over them.
Uber was founded in 2009. Before that you could maybe hail a cab if you were in NYC, otherwise you had to call ahead. Someone would drive the cab to you, you'd get in it, and they'd drive you to where you wanted to go.
Now we have AI. Now instead of driving to a financial advisor's office or a bank loan officer's desk or my kid's school:
- I have an AI app that automatically invests my money and runs on my phone's neural processor
- I have an AI app that automatically tells me if I qualify for a mortgage based on a truckload data about me and runs on my phone's neural processor
- There's a booth in my kid's classroom that her teacher puts her into; he says, "hey Siri, teach this kid to stop throwing carrots", the AI teaches my kid in a kind, convincing way that carrots are food, not sports. This runs on the iPad inside the booth that is also the screen
- I have an AI app that tracks how long I'll live, and it ticks down 1.38 minutes whenever I set foot in an automobile, so I try and avoid it now. I can't believe people were ever so reckless.
The race now is who controls this. Probably what will happen is governments will deem it dangerous (disinformation, national security, export controls), and allow only megacorps like Google/Facebook/Microsoft to run them under strict regulations. As a result, these models won't run on our phones but on server farms, and we'll become (even more) dependent on these corporations. This will feel good to the US because it shores up our tech sector--a big economic advantage compared to every other country--while also looking sensible.
I'm bullish on recent ML advancements but it will be a long time before I give one my Schwab login.
I'm skeptical too, but I assume this will just be Darwinian: curmudgeons like us will shamble along with our 5% gains and ill-behaved children, and the bulk of society will reap greater rewards. The worst part is our judiciousness won't even save us if we turn out to be right: society will still collapse whether or not you turned your life over to AI.
Yes, the AI is good at taking low-level requests and turning them into reasonable code, and then refining them, but unless the CEO is going to sit down and spend days telling the AI "OK now make this button disabled when it's clicked until the table loads", you need someone to be doing that.
But still, the real product needs to be fleshed out, back and forth, interactively.
Regardless of whether it is the manager or engineer who commands the tool, you can't entrust the AI to get everything 100% right. Some one has to proofread it, and that is the bottleneck, or where the value of human kicks in.
A human brain can only hold that many details by itself. It can't maintain all the details live all at once, and some business do have overwhelming number of details. A one man team with help of infinite AIs to replace an actual team isn't useful.
At some point, that human will start to fail to verify the system generated by AI is actually what he/she wants.
Educators, generals, dieticians,
psychologists, and parents program. Armies,
students, and some societies are programmed.
by Alan J Perlis, the first Turing Award recipient.Managers, CEOs, Department Heads...are already 'programming'. As the abstraction moves higher up, some people that have excessively narrow expertise will be made redundant.
You can't do that today because fixing the specification after the fact means waiting 6 months and paying 60+ man-months. Once fixing the specification means waiting 10 minutes and paying $10 for inference compute, the idea guys can afford to learn by doing.
That said "software is a tool" is a good frame of mind. You shouldn't position yourself as a 'coder' (those are hired en masse from poor countries) but a problem solver who uses tech.
"Ideas guys" and "coders" being seperate only exist, imo, when development is viewed through the junior lens - once you progress you'll end up wearing all the hats from ideation to QA at different points.
Do I program when I retouch photos in Photoshop? It's a no-code environment that an artist can easily learn and use with no programmers needed.
What's new here is that you DO NOT NEED TO BE EXACT with AI. AI knows human nature and human speech and it can infer what you say and what you need, if you're approximately close, and then make the exact code to do what you need.
Even if you need programmers, you now need 1/100 of the programmers you needed before. What happens to the other 99?
Does it matter whether programmers are 99% unemployed or 100%? Same deal.
It takes drastically less programming effort to create programs than it did 50 years ago. Did that decrease demand for programmers? No, it drastically increased it.
I got a feeling programmer efficiency peaked in the 90s with VB or Pascal desktop GUI apps.
There is so much bloat demanded to create a minimal product now.
> Do I program when I retouch photos in Photoshop? It's a no-code environment that an artist can easily learn and use with no programmers needed.
They're obviously not talking about all GUIs being impossible. The difference between Photoshop and a no code tool like what OP was referring to is that Photoshop doesn't attempt to be Turing complete.
> Does it matter whether programmers are 99% unemployed or 100%? Same deal.
AI completely replacing programmers is pretty much the definition of AGI. There's no point in worrying about your own joblessness in that scenario because the entire economy will either collapse or be transformed so as to be unrecognizable.
Until that point, I'm not personally worried about any efficiency gains putting me out of a job.
You may have seen that there's now a load of drones flying around, doing things that were never possible until now. Dropping bombs on soldiers in trenches, giving support for assaults, targeting artillery.
The fact is war changes with technology. Spearmen and cavalry are obsolete now. Maybe pilots will be soon as well.
But the generals are not. There's always a need for someone to say "given the position we're in, and the logistics we have, and the resources we have, and what we know about the enemy, we should do this..." and this role is still recognizable through history.
Whether computer jobs become obsolete depends on whether you're closer to the general end or the footsoldier end.
I think it’s more probable that managers get replaced by AI first.
While absolute rigor in programming is a hard thing to find in reality, I don't think people understand the difference at all.
Won't that create stronger competition for fewer roles?
This means that an AI capable of completely replacing programmers is by definition an artificial general intelligence. I don't think we're at that point and I don't think we will be for a long time, but if we were there would be no point in worrying about our own joblessness. AGI would spell the end of the economy as we know it today, and it would be very hard to predict and prepare for what comes next.
However, as Carmack says, many developers lack product skills, it’s not just the managerial class that lacks them.
So I think the correct advice is not “engineers don’t need to worry at all”, it’s “make sure you are learning how to build a product not just write code”. For many that’s trivial advice, but not for all.
We have systems, today, that take a specification and generate code from it (and will even do so in a provably correct manner). There are scalability issues with such systems, but they exist and have in some form for decades.
None of these systems have taken the world by storm.
If what ChatGPT and their ilk do is make it easier to spew out bulk spaghetti code that is essentially unverifiable, I don't think that's going to transform the industry nearly as much as anyone thinks it will. For boilerplate code, sure—but that's exactly where we should be using better abstractions anyway. For small code which you can verify by hand, sure. But for anything larger than that, we're just watching a trainwreck in slow motion.
Large code bases written by humans are already hard enough to understand. How much more difficult will it be when your AI can spit out a million lines of code in a second?
Without specifications, it's all worthless. But the specifications are the hard part.
Think about the difference between Assembly and Python. Programming will continue to get higher level. Perhaps it won’t be so foolish to ask a truck driver to “learn to code.”
But here is an even more interesting comparison: the salary of someone writing assembly 30 years ago and someone writing Python today. Higher level != paid worse.
One day a corporate genie appeared to a middle manager and granted him one wish. The manager wished that their business logic could be edited by a graphical tool and that programmers were no longer needed. "Granted", poof!
The next day the manager came to work, all the programmers were gone, fired by upper management. The manager sat down and opened the new graphical tool that had magically appeared on his computer. The interface was beautiful, except for what appeared to be a hairball drawn in the middle of the screen. The manager asked his manager about the hairball, "oh, that's our business logic, you have to zoom in to see the individual rules". The manager zoomed in and surveyed 180,000 business rules, intertwined in a complete graph.
"How am I suppose to work with this?" the manager asked his manager. "You just click and move them around, it's easy", replied the upper manager. "Also, I've been meaning to ask, when do you think those new business rules we talked about last week will be implemented? We need them by Friday."
To reach the point where no-code is a no-brainer further implies inflexibility in the capabilities of applications and moreover, fungibility in their fundamental uses. The question to ask is are the upper layers of abstraction reducible to what's analogous to the outputs of a parse tree.
If we reach this point where creativity and actual innovation is lost and we're all simply attempting to make the next 'facebook for dogs', I anticipate my future personal career switch to starring in geriatric pornographic films.
Given the way we had tremendously productive tools like Visual Basic 6 and Delphi, that seemed to be a reasonable interpretation.
But those tools gave way to C++ and all the .NET insanity. I don't know why someone would give up Delphi, which worked flawlessly and didn't generate any (as far as I can recall) re-only generated boilerplate, to C++ which was a mess, but it happened en masse.
Then most people abandoned personal computing, and the windows desktop for always internet connected pads and phones. Tools have gone severely downward in the past 2 decades as a result.
I suspect we'll get some really powerful AI tools, and go back to worrying about null pointers in a few years anyway.
What’s great is that you could also have AIs write tests - for performance or correctness. Then future prompts could rely on these tests to stay in the correctness bounds for existing projects.
Emulating cognition is almost limitless in what it can do once it gets sufficiently good. So good that it can operate itself. You could hire an AI founder and give them cognition and compute resources to solve any problem.
Here’s an example:
You: Build me a profitable company that uses AI generated content to entertain humans.
AI: How much can I spend on cognition and compute? How much time do I have?
You: I want you to spend at most $X before coming back to me. You have until Monday. Along the way I want to see your plan before we decide to execute. Spend at most 5% of that making the plan and 1% calculating its probability of success with confidence interval.
… within some small timescale
AI: Here’s my plan as a video with rough examples demonstrating the product and strategy. Here’s a table with our probability distribution of predicted success metrics and confidence intervals.
You: Plan approved you can spend the rest of the $X
… on Monday
AI: Done. We’ve made a profit of $Y, with total revenue so far of $Z. We have a plan for reinvesting these proceeds to make even greater future profit. Here’s a table of our expected success metrics based on reinvested amount. How much should we reinvest?
You: Reinvest 50% into perpetuity and deposit the remainder in my account. Book me a trip around the world using 25% of the current deposits with accommodations for my family.
Now go enjoy your life.
I gave GPT-4 a budget of $100 and told it to make as much money as possible. I'm acting as its human liaison, buying anything it says to. Do you think it'll be able to make smart investments and build an online business?
The fallacy here is that programming is not a process of translation. It is a process of invention. Going from rough or high-level functional requirements to a functioning computer program requires inventing novel logic that has not been invented at any other step. It involves making business decisions that have not been made at any prior step.
Whatever can be categorized as menial "translation", i.e. a step in the business process where no new logic is added or business decision made, has already been automated away - compilers.
Nevertheless, AI is different. Because now you don't have to specify things at a programming-language level of granularity. It will select the programming language constructs most statistically likely to match a solution to the specified problem.
So now we're in a whole 'nother world where the programming, as such, will be done by machine while the role of designer or architect must still be done by humans. But the legions of programmers employed by companies are about to go the way of farriers.
> In essence what has happened in software development is that the level of abstraction has gone up while the machine has taken over more and more of the nitty gritty details. From punchcards, to assembly, to COBOL, C, Perl, Java, Python, Erlang, Rust.
I've had friends and peers contact me in regards to the AI, GPT, et al headlines and they're always surprised at my calmness. You've written out exactly what is happening and why there shouldn't be any worry regarding jobs. I think the worry is generally under the premise that everyone has to upskill themselves which is against human nature. Folks are feeling uncomfortable but this is another cycle.
> But the rigidity of what is needed to specify a program that really does what you want hasn't. Especially evidenced by the fact that recent programming language developments often have a specific area where they shine, but not raising the abstraction level that much.
Another great point. Like you wrote, there really is a need for "translators" for programming languages and it's not going away anytime soon. If anyone has played around or even just read/watch through folks using AI/GPT/Etc. you'll come to notice that the tech is not "stable". It's only good in limited instances.
Finally, the headlines and copywriting of most of the pieces and journalism coming out is definitely hit and miss. A mix of pessimism and optimism. It's important to remember what we are witnessing is another technological cycle and another cycle in general if you want to looking at a bigger scale but that's another discussion for another time.
His prediction was based on the trends he was seeing at the time. But it wasn't even AI. Instead he made this prediction because he saw the rise of no-code tools replacing software developers because managers could finally cut out the pesky "translators", i.e. software developers.
I said it then and I will say it now. If your managers could specify what they need in a manner that no-code tools, or now AI, can generate the code they want, they will have to be extremely exact in their language. So exact in fact that they will need to specify a program in a click and drag interface, or in human language.
Since they hire software developers to make the specification more rigid, and the managers don't seem to be getting better at this over time, why would you believe this skill set is going to go away?
In essence what has happened in software development is that the level of abstraction has gone up while the machine has taken over more and more of the nitty gritty details. From punchcards, to assembly, to COBOL, C, Perl, Java, Python, Erlang, Rust.
Of course I'm leaving out some languages here, but the level of abstraction has been rising.
But the rigidity of what is needed to specify a program that really does what you want hasn't. Especially evidenced by the fact that recent programming language developments often have a specific area where they shine, but not raising the abstraction level that much.
I'd be surprised if the next step is "Hi, I'm an ideas guy, please give me an app that does Uber, for bicycles, but better."