From @garymarcus on BlueSky:
A computer scientist’s perspective on vibe coding:
From @garymarcus on BlueSky:
A computer scientist’s perspective on vibe coding:
@bsletten @garymarcus it lets people write programs using intuitive, plain English!
Like COBOL?
NO, it's not like COBOL at all!
@bdf2121cc3334b35b6ecda66e471 @bsletten @garymarcus Drats. You beat me to it with COBOL. Managers and lawyers were supposed to be able to read the code and understand it.
SQL. That was another one.
@bsletten @garymarcus I've been hired to fix 'vibe programs' that were anything but deteministic. Excel is notorious for producing code that can't be verified and only works if the original author runs it on the machine it was created on.
But the same problem happens whenever someone plays "keep twiddling until I can't see any problems".
At least they never did anything important with those programs, it was all just financial stuff.
@bsletten @garymarcus Very accurate.
But what we didn’t have at that time, was the interweaving of highest-level political power with software development industries.
Political corruption ticks stuff up several levels at once!
@gimulnautti @bsletten @garymarcus
The emperor's got no clothes, brain, or heart.
Agreed, I am not an expert at coding by any means but what I do know is that good programming requires skill, effort and pateience so your final software works properly.
Quite a few people have jumped on this bandwagon so yeah it is about promoting those platforms or making a quick buck somewhere.
Maybe it is another example of dumbing down.
@bsletten @garymarcus declaring that Visual Basic, Delphi (aka Object Pascal with a nice IDE) or Flash would break when you try anything complex, it's a wild ride that I don't agree.
VB and Delphi are full-fledged languages with nice IDEs. And we know how many things were made with Flash, including a lot of games. Other thing would be the quality of the Flash plugin...
@utterlymundane @FantasmitaAsex @bsletten @garymarcus having only touched VB as a "plain" language and never anything visual: the core language part is certainly "no, that's just a normal programming language", but what was the visual part like?
I assume it's not hypercard-like or UML-to-code-like and is more capable, but I'm wondering if there's literally any nugget of truth in there or if it's just totally wrong (like Flash - it definitely doesn't belong in that list).
@groxx @utterlymundane @bsletten @garymarcus the GUI building thing using drag&drop was really amazing. And, contrary to Delphi, QT creator, Swing GUI builders, etc... Don't generate user editable code. It's stored on an opaque .FRM file
PD : I'm not talking about the code that you could put on the events handlers of the GUI. I'm talking about "Have a windows of X size with the title Y and a button". Delphi, and other languages use code for this. Code that you can edit.
@FantasmitaAsex @groxx @utterlymundane @bsletten @garymarcus
These builders originated in Smalltalk, where everything was live coding. To create a window, you created a new instance of a window. But everything was also introspectable, so you could use the GUI to inspect and modify any properties of your new window, and put things in it.
NeXT’s Interface Builder did the same thing. It let you create a serialised object graph. This was richer than a lot of later tools because you could embed (and configure) controllers and default versions of models directly in the NIB file. When you coupled it with something like the EOF ORM, you could bind these things to data sources and create entire CRUD tools with little or no code.
AI is heavily funded by the fossil fuel industry.
https://www.bloomberg.com/news/articles/2024-11-06/saudis-plan-100-billion-ai-powerhouse-to-rival-uae-s-tech-hub
More people need to ask why the Saudis suddenly decided to invest $100 billion dollars in AI initiatives.
https://www.bloomberg.com/news/articles/2018-04-06/google-thiel-stand-out-in-saudi-prince-s-silicon-valley-tour
https://www.vox.com/2018/4/6/17206358/saudi-crown-prince-visit-tech-silicon-valley
https://pagesix.com/2018/03/24/saudi-prince-touring-us-to-meet-with-americas-moguls/
More people need to ask why Charles Koch suddenly decided to launch a Silicon Valley venture capital outfit.
https://www.axios.com/2017/12/14/koch-tries-its-hand-at-venture-capital-1513277500
1/2
Saudi Arabia is planning a new artificial intelligence project with backing of as much as $100 billion as it seeks to develop a technological hub to rival the neighboring United Arab Emirates, people familiar with the matter said.
2/2
AI is getting this massive push for a reason.
https://financialpost.com/pmn/business-pmn/climate-tech-funding-fell-40-in-2024-as-investors-rushed-to-ai
Redirecting investment away from climate action & putting it towards tech that maximizes fossil fuel usage.
https://www.theguardian.com/environment/2025/apr/03/trump-fossil-fuel-donors-data-centers
https://energydigital.com/articles/energy-water-data-could-google-make-ai-sustainable
https://theconversation.com/data-centre-emissions-are-soaring-its-ai-or-the-climate-240137
Vibe coding enables people who aren't well trained computer scientists to create complete, working applications. Is this a breakthrough? Not even close - there have been such tools since the late 1980s. See, for example: Apple HyperCard, Sybase PowerBuilder, Borland Delphi, FileMaker, Crystal Reports, Macromedia (and then Adobe) Flash, Microsoft VisualBasic, Rational Rose and other "Model Driven Development" tools, IBM VisualAge, etc. etc. And, of course, they all broke down when anything sightly complicated or unusual needs to be done (as required by every real, financially viable software product or service), just as "vibe coding" does (see https://lnkd.in/enhAE3Ri). The only difference is that the outputs of those older tools were actually deterministic and well documented and understood, while your AI prompts and models are not! To claim that "vibe coding" will replace software engineers, one must: 1) be ignorant of the 40 year history of such tools or 2) have no understanding of how AI works or 3) have no real computer science education and experience or 4) all of the above, OR, most importantly, be someone trying to sell something and make money off of the "vibe coding" fad. | 11 comments on LinkedIn
@bsletten @garymarcus Once again demonstrating the lack of historical perspective in academic computer science. It’s funny to see any form of Basic listed as originating in the late 80s, when Basic had already been in use for 20 years as a low barrier to entry tool for non professionals.
Anyway, vibe coding is qualitatively different than those things. They are closer to things like Matlab or the GUI based tools used by Ardino hobbyists.
@bsletten @garymarcus @ericsedge strong disagree that many of those examples qualify as equivalent to vibe coding.
To me, this is an elitist view of what qualifies as “real” coding being equated with getting a computer to make shit up.
Of the tools listed which I have used, *none* are the same as LLM or vibe coding.
We should do better, and stop gatekeeping on real people solving real issues with the tools that allow them to code.
ChatGPT, Copilot, etc. do not teach *any* coding principles in the way those tools do, and it’s beneath any sensible person to suggest there is an equivalence here.
0/10 (rating, not thread post count)
@europlus @bsletten @garymarcus
A better, and still weak example, might be paired programming with a very junior developer with great memory and little experience. The current level of LLM competency is more akin to “maybe try this?” rather than the performance I expect from human.
I’ve been experimenting a bit with vibe coding and have found it somewhat useful if my expectations are set properly.
2/2
@europlus @bsletten @garymarcus @ericsedge For me, the elitist part I noticed was the “slightly complicated or unusual,” about the tools he listed. I would agree if he had said “out of the scope of the tool,” instead.
People made full-featured multi-window applications in FileMaker, manage complex workflows in Excel, and that makes them computer programmers, and it’s orders of magnitude more valid than vibe coding.
@oscherler @bsletten @garymarcus @ericsedge absolutely!
Thanks for pointing that out, it’s certainly a choice to imply those systems can’t do complex (or unusual) things.
Most of the things I “code” are explicitly unusual, hence my need to automate them through custom code.
And yes, I forgot to add Excel (and SQLs) to my list, thanks for the reminder!
Vibe coding enables people who aren't well trained computer scientists to create complete, working applications. Is this a breakthrough? Not even close - there have been such tools since the late 1980s. See, for example: Apple HyperCard, Sybase PowerBuilder, Borland Delphi, FileMaker, Crystal Reports, Macromedia (and then Adobe) Flash, Microsoft VisualBasic, Rational Rose and other "Model Driven Development" tools, IBM VisualAge, etc. etc. And, of course, they all broke down when anything sightly complicated or unusual needs to be done (as required by every real, financially viable software product or service), just as "vibe coding" does (see https://lnkd.in/enhAE3Ri). The only difference is that the outputs of those older tools were actually deterministic and well documented and understood, while your AI prompts and models are not! To claim that "vibe coding" will replace software engineers, one must: 1) be ignorant of the 40 year history of such tools or 2) have no understanding of how AI works or 3) have no real computer science education and experience or 4) all of the above, OR, most importantly, be someone trying to sell something and make money off of the "vibe coding" fad. | 11 comments on LinkedIn
@bsletten @garymarcus gatekeeping software development to "well trained computer scientists" is just the other extreme, opposite to vibe coding. We should be happy more people embrace coding as a fun thing to do.
Besides, making good Flash applications where you don't just have to deal with code but also graphics, animation and sound is something even "well trained computer scientists" may have struggled with.
As I’m often reminded as a past full stack developer, just because I can doesn’t mean I should nor is it optimal. My graphics and UI skills are “meh”.
Vibe coding in a nutshell:
"So you know this coding thing which has plenty of problems with human errors? We have created this machine that takes 10^24 gigaterafuckwatts and creates code with random errors."
And the AI crowd want to use this in production software. Dystopic fiction didn't prepare us for cyberpunk dictatorships being cringe.
HyperCard was more user/novice-empowering than the current state-of-the-art LLM’s will ever be.
Access is the exemplar here. I worked with a few companies that, 20 years ago, produced in-house CRUD ‘applications’ entirely in Access using almost entirely the GUI, with no in-house programmers. You drew the GUI, connected it up to fields in the database, created queries with drag-and-drop, and drew forms to display the results.
Access had the world’s worst branding. It was a surprisingly good RAD tool for business apps, bundled with the world’s worst database. Microsoft sold it as a database. But Access could also work as a front end for SQL Server (which was a fairly good database, even 20 years ago). And that let you scale these business apps up to the requirements of a moderate-sized enterprise.
I think Power Apps are the newest thing in this space from MS. They use the Calc language from Excel and a visual GUI designer. Excel had a reactive programming model before anyone knew that was a thing with a name and now can even use remote SQL databases as a data source. There are a billion people writing programs in Excel.
I am 100% in favour of empowering users and reducing the user/developer distinction. I work on hardware that is designed to enable exposing lower-level things to end user programming environments without making it too easy for users to shout themselves in the foot. I started working on this as a result of building end-user programming environments and seeing the ways that making them more powerful also made them less reliable and concluding this needed better hardware to enable the things I wanted in software.
I just don’t think giving users a magical black box that produces something statistically related to the thing that they asked for, that is hard to debug, and that they don’t understand is empowering.