Multi-modality is pretty low hanging fruit so i'm glad we're finally getting started on that. Imagine if GPT-4 could manipulate sound and images even half as well as it could manipulate text. We still don't have a large scale multi-modal model trained from scratch so a lot of possible synergistic effects are still unknown.
I’ve used GPT4 (text) heavily as part of my business, including for front end stuff.
The technology is very impressive - but honestly Twitter examples are super cherry picked. Yeah, you can build some very ugly, basic front end web pages and functionality right out of the box. But if you want anything even slightly prettier or more complicated, I’ve found you need a human in the loop (even an outsourced dev is better). I’ve had GPT struggle with even basic back end stuff, or anything even a bit out of distribution. It also tends to give answers that are “correct” but functionally useless (hard to explain what I mean, but if you use it a lot you’ll run into this - basically it will give really generic advice when you want a specific answer. Like, sometimes if you provide it some code to find a bug, it will advise you to “write unit tests” and “log outputs” even if you specifically instruct it to find the bug).
Plus, in terms of capabilities, tools like Figma already have design to code functionalities you can use - so I don’t think this is really a change in usable functionality.
If AI continues to get better it won't just be you who's in trouble.
However, keep in mind that these are cherry-picked. If someone just took that output and stuck onto a website, it'd be a pretty horrible website. There's always going to be someone who manages the code and actually interacts with the AI, so there will still be some jobs.
And your boss isn't going to be doing any coding. I'm pretty sure that role is still loaded and they'll still be managing people rather than coding, and maybe sometimes engaging with an AI.
Another prediction: I'm pretty sure specialists are going to be significantly more important as your job will be to identify the AI's deficiencies and improve on it.
There's an idea with some truth to it. The first 90% is easy, finishing and getting it to 100% is hard, maybe almost impossible. So asking "can it do x" is relevant. Becasue it might not
> It will be able to do it even faster, better and more cheaply than a human can.
Take what you did in the past year. Write down every product decision taken, every interaction with other teams figuring out APIs you had, all the infra where your code is running and how it was setup and changed, all the design iterations and changes that had to be implemented (especially if you have external partners demanding it).
Yes. All that you'd have to input into the AI, and hope it outputs something decent given all that. And yes, you'll have to feed all that into AI all the time because it has no knowledge or memory of "on Monday the new company bet was announced in the all hands"
So ... in this example, your job is continually feeding information to the AI from various sources. Why would the AI not be automatically hooked up to all those sources? Building a system that can do that is essentially trivial with the OpenAI API.
We have 32k contexts now, how big do you want to bet the context will be in 10 years?
That’s before you add any summarization, fine tuning or other tricks.
The thing that computers have always done much better than humans is deal with much larger volumes of information. The thing that humans have always done much better than computers is reason on that information better. Now the computers are coming for that too.
Can any grunt then check that the AI actually produce what was actually required?
And yes, how can we forget that any audio of a meeting has just the correct and final specifications, and not meandering discussions about anything and everything. Can't wait to see a canhazcheeseburger in a financial app because people in the meeting had cats on camera, and people demanded to see them.
- six months of discussions involving almost 20 people
- 4 new BigTable tables
- Deployment of 4 new Dataflow jobs, and fixes to two other Dataflow jobs
- Several complex test runs across the entire system including a few recreations of last year's full data to test that nothing broke
Not a grunt job, definitely. And I'm 100% sure that people doing that would still have their jobs 10 years from now, even with AI.
It amazes me, really, that people who would otherwise boast about how rational they are, and how they follow logic etc. completely replace all their knowledge and expertise with child-like belief in magic when it comes to anything AI-related.
I'd be sacred too, but at least would be taking a rational approach to it. It's an adapt or die situation. Putting your head in the sand is just gonna get you mowed down.
Ad hominem is not as good an argument as you think it is.
The only ones scared in this conversation are you and others who literally say you're scared for your jobs and your careers because of a magical boogey man.
But if this development continues AI will surely be able to just parse the entire frontend, then look at the repo, then look at the specifications, then when you ask for a specific feature it will instantly implement it gracefully.
I don't see why it wouldn't understand piles of hotfixes on top of each other, or even refactor technical debt in tight coupling with existing or historical specification.
Or is there a reason this is not going to happen in a few years?
I very much doubt it. "Revert Norway tax code" or "add content restrictions for Turkey" cannot necessarily be deduced from the codebase. And there are thousands of product requirements like that.
It might become a better code assist tool some 10 years from now, but it won't be able to implement product decisions.
It depends. Did you ever work in Development Support? Understanding requests or bug reports from customers is quite challenging, for trained and experienced developers. In my eyes that would require AGI, which we do not know of, if that can be achieved with the current approach.
I actually did , and yes it's extremely challenging and can be rather rage inducing; "it doesn't work" --> what doesn't work --> "the thing" --> what thing --> ∞ ...
But the thing is conversations like the above ie. both external support and internal feature requests could theoretically be handled by a GPT-like system also ending up in a ai created custom specification that could both be implemented and documented by the ai system instead of humans?
If just being persistent and willing to iterate solves the issue, then yes, GPT-like systems could do that. If you have to employ creative thinking to even find out what the customer wants, then check it in the system, debug some processes and derive the desired feature or correction, then we are very far from having such a tool, IMHO.
Not yet, but give it time. The concept of self-driving vehicles even a decade ago seemed absurd (or even AI for that matter), but now it all seems like a reality.
And that's not even taking into account all the advances we'll have with AI within the next decade that we haven't even thought about.
> The concept of self-driving vehicles even a decade ago seemed absurd (or even AI for that matter), but now it all seems like a reality.
Nope. It's still not close to reality. It's as close to reality as it has been for the past 10 years while it was being hyped up to be close to reality.
> And that's not even taking into account all the advances we'll have with AI within the next decade that we haven't even thought about.
As with FSD, we may approach an 80% with the rest 20% being insurmountable.
Don't get me wrong, these advances are amazing. And I'd love to see an AI capable of what we already pretend it's capable of, but it's not even close to these dreams.
> Cruise and Waymo are in production in really difficult cities.
Cruise and Waymo are in production in very tightly fine-tuned and carefully monitored situations in two cities. We've yet to see if that can be easily (or at all) adapted to driving anywhere else.
The more people say that, the less convincing it is
There is no way I would have a UI developer onboarded when I can generate many iterations of layouts in midjourney, copy them into chatgpt4 and get code in NextJS with Typescript instantly
non devs will have trouble doing this or thinking of the prompts to ask, but the dev team asking for headcount simply wont ask for headcount, and the engineering manager is going to find the frontend only dev redundant
Yeah, I'm also skeptical about the actual value of specialists in the future. To me, AI brings a ton of power to generalists, who now have access to very powerful tools that would have taken them a long time to learn otherwise.
I would even go further and say the generalist gains a powerful tool belt that previously could not have existed. Not enough hours in the day or years in a lifetime.
Will you then use the AI to scale your platform ? Optimise your database ? Improve your test coverage, implement new features, write new backend services, integrate with old difficult but critical systems?
At some stage you must realise that you’re still working…
I’m going to say you edited or amended you comment because that second paragraph wasn’t there , if it was. I was so underwhelmed with the first I guess I didn’t bother with the second.
> when I can generate many iterations of layouts in midjourney, copy them into chatgpt4 and get code in NextJS with Typescript instantly
Have you actually tried this?
I did the first step and even that didn't work well. The "iterations of layout in MidJourney" step. If people can make it work, well bless them, but we're not getting rid of our graphic designer now.
The best counterargument to “GPT4 is going to replace us all” is actually using it for a couple of weeks.
It has a few neat tricks but it’s not reliable and at least half of what it generates is totally unusable, the other half requires heavy intervention and supervision.
In your twitter's comic book link the first image has a women with a huge ass with no pants on. The ass is colored grey and so the lack of clothes is not noticeable until you click into the image.
How many children here on hacker news are going to see this and get addicted to porn? Perhaps a few. You deserve to be banned.
Can do UI to frontend. Seems to understand the UI graphical elements and layout, not just text https://twitter.com/skirano/status/1706823089487491469
Can describe comic images accurately, panel by panel - https://twitter.com/ComicSociety/status/1698694653845848544?...
Lots of examples here also - https://www.reddit.com/r/ChatGPT/comments/16sdac1/i_just_got...
It's Computer Vision on Steroids basically.
Multi-modality is pretty low hanging fruit so i'm glad we're finally getting started on that. Imagine if GPT-4 could manipulate sound and images even half as well as it could manipulate text. We still don't have a large scale multi-modal model trained from scratch so a lot of possible synergistic effects are still unknown.