I don't know about other people, but to me these constant salvos of AI hype are like watching an old silent comedy where people are throwing custard pies at each other. The first couple of pies are funny and then it just gets tedious.
Over 50 years I have written over a million lines of code in over 20 languages. AI gives me a step up in blocking out preliminary solutions especially in the earliest stages. However it sucks at debugging and I mean really sucks. The worst is an inhuman 'proximity bias' that often, when asked for a simple solution to a small issue will offer a complete refactoring, entailing starting from scratch and when challenged offers a transparently self congratulatory "applogy".The hazards barely exceed the opportunities.
Agree about the debugging. I wrote my first computer program in 1958 (in Bell Floating Point Interpreter for the IBM 650) and have been at it ever since. I use AI a bit differently: I ask it to produce code snippets after I have blocked out the logic of what I want, which saves me cudgeling my ancient brain for the right set of syntax details. After one or two total time-wasters trying to debug a wrong 'un, if the code isn't transparently correct I rephrase the prompt and ask again.
And this, unfortunately if not tragically, is what will happen when those DOGE 'experts' walk away from the chaos they created across all government platforms, networks, and architectures.
I work in a company with a lot of people over 50, and even though I'm in my 30s, I trust them with just about everything more than I trust our younger colleagues.
Yeah. I'm so worried about the future at the moment. Elon is clearly high on ketamine, and that is no joke. I never took it, but I imagine it helps with his anxiety, at the expense of making him (even more) delusional.
They are totally going to feed massive government databases into an LLM and form public policy based on the results. I think till now the government has made limited use of "AI" even though palantir have wanted to change that. Now these companies have their entry point and free reign to just do what they like.
Roose and company are further evidence not just of "the death of expertise", but a disdain for the effort required to master a skill (or any skill other than marketing hype). The impact of this attitude is visible across the spectrum with the notion that all one needs is a few "prompts" to be a programmer being one of the most public and egregious, but the sentiment is visible in the declining desire to master skills as disparate as welding and accounting. Shipyards can't find people who have the skills to build ships, government agency can't find competent forensic accountants, hospitals scramble to find nurses, particularly those with specialties. Because "mastery is hard". According to the snake oil set we don't need cardiovascular surgeons, there's a prompt for that.
Another very important point to keep in mind is that LLMs have a significant advantage building clones of classic arcade games because of how many programmer blogs and GitHub accounts have already done this exact task.
The fact that the LLMs still struggle with this task proves that they have very little abstract reasoning capabilities. These kinds of games are simple compared to modern games but still have more LOC than can fit in a SoTA LLM’s context window, so they can’t just memorize the answer in their fuzzy lookup table.
I’m curious whether there are any SWE benchmarks out there that test writing full apps/games from scratch, like 10k+ LOC that require a full test suite (1k+ tests) to pass. This is a task that every SR dev I know can do, in fact, IMO, it’s what separates the real SWEs from the devs that just memorize Leet code answers to get through the interview only to slow the team down post hiring. I bet even the most expensive LLMs would fail badly on this benchmark.
Yeah, given the vast amount of training data, Its reasonable to assume they should be better, the reason I feel 'games' in particular are hard is because they have to represent sprites and level design in code; which can require actual spacial reasoning to do correctly
much more complex than (web)apps which are cookie cutter and readily available for every niche need
Somewhat a counterexample to this is the 'good' performance in generating minecraft structures in that one benchmark.
I suppose its because these tasks are more openendedish than a replication of pacman and its easier to make a semantic stitching? of things, when you dont have clear requirements
I'm sure it's been said but, to get AI to write correct and sufficient code requires that complete, unambiguous requirements are provided. Good luck with that.
So true. Those of us who work in software (data engineering and analysis in my case) know that most of the "secret sauce" of programming comes from turning vague user demands into concrete and maintainable solutions - not from typing out thousands of lines of code.
This is a great example! Programming languages are obviously designed to strike some clever optimal balance between precision and expression, akin to math notation. For example, it would be hard to write a specific SQL query in plain English without losing precision and/or writing a much longer statement than the query itself. Just like how trying to do long division in plain English prose is feasible but foolish.
I used a pencil to write my Assembly Language code (or sometimes FORTRAN or COBOL or JCL or even channel programs) onto coding forms that were sent to (the all-woman) Keypunch Department to be, ummm, keypunched onto cards. The cards were rubber-banded together, then carried down the hall, where they waited patiently to be fed into our card reader, and run on our mainframe. The printed output (11 x 15) was rubber-banded together with my cards and returned to my desk via internal mail.
I wonder how long the whippersnappers would hang around
I must be a bit younger than you. By the time I was programming the Apple II, Commodore 64 and TRS-80 were things. Of course it was still easier to write out code on paper beforehand or print it out and check it over that way. Those black and green CRTs were an eyesore.
If you read Ethan Mollick’s post on this, it’s probably worse than Roose. He rhapsodizes about “vibecoding” (cringe) in which he was able to make something half assed that no one would want for only $13!
Let's take this back to 1975. Around then sci-fi writer Douglas Adams wrote "The hitch-hikers guide to the galaxy" a satire on Lonely Planet guidebooks.
As usual the book had androids .. built by the Sirius Cybernetics Corporation .. with Genuine People Personalities. The Sirius Cybernetics Corporation described them as "your plastic pals who are fun to be with".
Unfortunately, the SCC created personalities were all flawed (but in a comic way).
Marvin the Android was permanently pessimistic and depressed. He despised humans for giving him tasks that he thought beneath him. ("Brain the size of a planet and all you want me to do is compute a trajectory ....").
Self-opening doors and lifts were given AIs too .. all they could do, was open and shut the door and say "Glad to be of service" . All the doors chose to do this in a slow servile voice that soon got irritating.
A fully automated space craft, it's AI refused to authorise lift off as the passengers snacks hadn't been delivered ... It apologised to passengers for the delay. It was still apologising when the fossilised space ship was re discovered 300 years later, pax mummified.
You get the picture. PacMan wasn't even around in 1970s, , but already Silicon Valley was high on hype.
Shouldn't a good agentic reasoning model figure out for itself that it can't do a good job and just go on the web to grab the first plausible-seeming Pac-Man clone? Plagiarism is fully expected, not to mention the explicit forms of cheating reported recently. LLMs often say they're sorry, but they don't really mean it.
It makes you wonder why we (devs + PMs) spend hours understanding and formalizing a business feature request. In the end they usually are close to what the business actually wants but far from what goes into the code. I think it will require some pseudo code/LLM language to succinctly describe what AI should create.
o3-mini-high just solved it? I now have a perfectly fine pac man running. I don't think dario meant that literally anybody could code up anything (e.g., a networked air traffic control system). if you put in some effort into your prompt, one should not downplay what already can be achieved w/o any coding experience imho. I do not believe all coders will be replaced soon, but to be fair a lot of coders for the longest time could get away with pretending to be more skilled than they actually were. I'm convinced only people who are great at it will be coding in 2030, and it will be 80% less than are today.
The world is awash in frustrating code that often makes it necessary to call customer support. Banks, phone companies, others where actually talking to someone is hard to figure out. What possible use will the next gen of autocomplete software do to improve functionality, UX? What will AI produce but more frustrating software.
Trillion dollar company software sucks pre-AI and it will be like quicksand post-AI
I don't know about other people, but to me these constant salvos of AI hype are like watching an old silent comedy where people are throwing custard pies at each other. The first couple of pies are funny and then it just gets tedious.
Over 50 years I have written over a million lines of code in over 20 languages. AI gives me a step up in blocking out preliminary solutions especially in the earliest stages. However it sucks at debugging and I mean really sucks. The worst is an inhuman 'proximity bias' that often, when asked for a simple solution to a small issue will offer a complete refactoring, entailing starting from scratch and when challenged offers a transparently self congratulatory "applogy".The hazards barely exceed the opportunities.
Agree about the debugging. I wrote my first computer program in 1958 (in Bell Floating Point Interpreter for the IBM 650) and have been at it ever since. I use AI a bit differently: I ask it to produce code snippets after I have blocked out the logic of what I want, which saves me cudgeling my ancient brain for the right set of syntax details. After one or two total time-wasters trying to debug a wrong 'un, if the code isn't transparently correct I rephrase the prompt and ask again.
Earl, I was born in 1974, and you are my new hero.
And this, unfortunately if not tragically, is what will happen when those DOGE 'experts' walk away from the chaos they created across all government platforms, networks, and architectures.
I work in a company with a lot of people over 50, and even though I'm in my 30s, I trust them with just about everything more than I trust our younger colleagues.
The current neural network AI crowd are plug and pray practitioners.
They are merely tweaking LLMs without really understanding any of the stuff they are producing.
They are like the people who plug data into statistics packages without knowing anything about statistics.
This is what AI “science” has become
Yeah. I'm so worried about the future at the moment. Elon is clearly high on ketamine, and that is no joke. I never took it, but I imagine it helps with his anxiety, at the expense of making him (even more) delusional.
They are totally going to feed massive government databases into an LLM and form public policy based on the results. I think till now the government has made limited use of "AI" even though palantir have wanted to change that. Now these companies have their entry point and free reign to just do what they like.
Roose and company are further evidence not just of "the death of expertise", but a disdain for the effort required to master a skill (or any skill other than marketing hype). The impact of this attitude is visible across the spectrum with the notion that all one needs is a few "prompts" to be a programmer being one of the most public and egregious, but the sentiment is visible in the declining desire to master skills as disparate as welding and accounting. Shipyards can't find people who have the skills to build ships, government agency can't find competent forensic accountants, hospitals scramble to find nurses, particularly those with specialties. Because "mastery is hard". According to the snake oil set we don't need cardiovascular surgeons, there's a prompt for that.
Another very important point to keep in mind is that LLMs have a significant advantage building clones of classic arcade games because of how many programmer blogs and GitHub accounts have already done this exact task.
The fact that the LLMs still struggle with this task proves that they have very little abstract reasoning capabilities. These kinds of games are simple compared to modern games but still have more LOC than can fit in a SoTA LLM’s context window, so they can’t just memorize the answer in their fuzzy lookup table.
I’m curious whether there are any SWE benchmarks out there that test writing full apps/games from scratch, like 10k+ LOC that require a full test suite (1k+ tests) to pass. This is a task that every SR dev I know can do, in fact, IMO, it’s what separates the real SWEs from the devs that just memorize Leet code answers to get through the interview only to slow the team down post hiring. I bet even the most expensive LLMs would fail badly on this benchmark.
Yeah, given the vast amount of training data, Its reasonable to assume they should be better, the reason I feel 'games' in particular are hard is because they have to represent sprites and level design in code; which can require actual spacial reasoning to do correctly
much more complex than (web)apps which are cookie cutter and readily available for every niche need
Somewhat a counterexample to this is the 'good' performance in generating minecraft structures in that one benchmark.
I suppose its because these tasks are more openendedish than a replication of pacman and its easier to make a semantic stitching? of things, when you dont have clear requirements
Even better ! The AI not only can't code .. it can't even copy someone else's code ! Code sitting there, waiting to be found and copied !
I'm sure it's been said but, to get AI to write correct and sufficient code requires that complete, unambiguous requirements are provided. Good luck with that.
So true. Those of us who work in software (data engineering and analysis in my case) know that most of the "secret sauce" of programming comes from turning vague user demands into concrete and maintainable solutions - not from typing out thousands of lines of code.
This is a great example! Programming languages are obviously designed to strike some clever optimal balance between precision and expression, akin to math notation. For example, it would be hard to write a specific SQL query in plain English without losing precision and/or writing a much longer statement than the query itself. Just like how trying to do long division in plain English prose is feasible but foolish.
Just brilliant ! AIs can't even code a basic video game from 40 years ago !
I love these elegant debunkings of the AI hype
You wrote assembly for the 6502? I'm duly impressed.
yep, on Commodore 64.
A great machine as was the Apple ][
I worked in the IBM mainframe ecosystem, developing database products for software vendors. Glad I got out before GPT 4.5 was released ...
In our day we had to use a pencil to program computers!
Not to write, but to adjust the tension of the audio tapes!
What's an audio tape?!?! Get off my lawn you little whippersnapper!
I used a pencil to write my Assembly Language code (or sometimes FORTRAN or COBOL or JCL or even channel programs) onto coding forms that were sent to (the all-woman) Keypunch Department to be, ummm, keypunched onto cards. The cards were rubber-banded together, then carried down the hall, where they waited patiently to be fed into our card reader, and run on our mainframe. The printed output (11 x 15) was rubber-banded together with my cards and returned to my desk via internal mail.
I wonder how long the whippersnappers would hang around
I must be a bit younger than you. By the time I was programming the Apple II, Commodore 64 and TRS-80 were things. Of course it was still easier to write out code on paper beforehand or print it out and check it over that way. Those black and green CRTs were an eyesore.
Complete history of AI in tl;dr:
We don't have jam today but we will have jam in three years! And it will BE GREAT!!!
If you read Ethan Mollick’s post on this, it’s probably worse than Roose. He rhapsodizes about “vibecoding” (cringe) in which he was able to make something half assed that no one would want for only $13!
Let's take this back to 1975. Around then sci-fi writer Douglas Adams wrote "The hitch-hikers guide to the galaxy" a satire on Lonely Planet guidebooks.
As usual the book had androids .. built by the Sirius Cybernetics Corporation .. with Genuine People Personalities. The Sirius Cybernetics Corporation described them as "your plastic pals who are fun to be with".
Unfortunately, the SCC created personalities were all flawed (but in a comic way).
Marvin the Android was permanently pessimistic and depressed. He despised humans for giving him tasks that he thought beneath him. ("Brain the size of a planet and all you want me to do is compute a trajectory ....").
Self-opening doors and lifts were given AIs too .. all they could do, was open and shut the door and say "Glad to be of service" . All the doors chose to do this in a slow servile voice that soon got irritating.
A fully automated space craft, it's AI refused to authorise lift off as the passengers snacks hadn't been delivered ... It apologised to passengers for the delay. It was still apologising when the fossilised space ship was re discovered 300 years later, pax mummified.
You get the picture. PacMan wasn't even around in 1970s, , but already Silicon Valley was high on hype.
Adams .. no means writer .. skewered them.
42
Shouldn't a good agentic reasoning model figure out for itself that it can't do a good job and just go on the web to grab the first plausible-seeming Pac-Man clone? Plagiarism is fully expected, not to mention the explicit forms of cheating reported recently. LLMs often say they're sorry, but they don't really mean it.
It makes you wonder why we (devs + PMs) spend hours understanding and formalizing a business feature request. In the end they usually are close to what the business actually wants but far from what goes into the code. I think it will require some pseudo code/LLM language to succinctly describe what AI should create.
o3-mini-high just solved it? I now have a perfectly fine pac man running. I don't think dario meant that literally anybody could code up anything (e.g., a networked air traffic control system). if you put in some effort into your prompt, one should not downplay what already can be achieved w/o any coding experience imho. I do not believe all coders will be replaced soon, but to be fair a lot of coders for the longest time could get away with pretending to be more skilled than they actually were. I'm convinced only people who are great at it will be coding in 2030, and it will be 80% less than are today.
To be fair, Dario also says we'll be immortal by 2030 so you'll have to forgive me for not giving a shit what he says
"o3-mini-high just solved it? I now have a perfectly fine pac man running"
What do you mean? Was this one-shot or multi-prompt?
The world is awash in frustrating code that often makes it necessary to call customer support. Banks, phone companies, others where actually talking to someone is hard to figure out. What possible use will the next gen of autocomplete software do to improve functionality, UX? What will AI produce but more frustrating software.
Trillion dollar company software sucks pre-AI and it will be like quicksand post-AI