Title should be “user give database prod access to a llm which deleted the db, user did not have any backup and used the same db for prod and dev”. Less sexy and less llm fault. This is weird it’s like the last 50 years of software development principles are being ignored.
llms allowed them to glide all the way to the point of failure without learning anything
in which the service admitted to “a catastrophic error of judgement”
It’s fancy text completion - it does not have judgement.
The way he talks about it shows he still doesn’t understand that. It doesn’t matter that you tell it simmering in ALL CAPS because that is no different from any other text.
Well, there was a catastrophic error of judgement. It was made by whichever human thought it was okay to let a LLM work on production codebase.
Are you aware of generalization and it being able to infer things and work with facts in highly abstract way? Might not necessarily be judgement, but definitely more than just completion. If a model is capable of only completion (ie suggesting only the exact text strings present in its training set), it means it suffers from heavy underfitting in AI terms.
Completion is not the same as only returning the exact strings in its training set.
LLMs don’t really seem to display true inference or abstract thought, even when it seems that way. A recent Apple paper demonstrated this quite clearly.
Coming up with even more vague terms to try to downplay it is missing the point. The point is simple: it’s able to solve complex problems and do very impressive things that even human struggle to, in very short time. It doesn’t really matter what we consider true abstract thought of true inference. If that is something humans do, then what it does might very well be more powerful than true abstract thought, because it’s able to solve more complex problems and perform more complex pattern matching.
The point is simple: it’s able to solve complex problems and do very impressive things that even human struggle to, in very short time
You mean like a calculator does?
Yeah, this is correct analogy, but much more complex problems than calculator. How much it is similar or not to humans way of thinking is completely irrelevant. And how much exact human type of thinking is necessary for any kind of problem solving or work is not something that we can really calculate. Considering that scientific breakthroughs, engineering innovations, medical stuff, complex math problems, programming, etc, do necessarily need human thinking or benefit from it as opposed to super advanced statistical meta-patterning calculator is wishful thinking. It is not based on any real knowledge we have. If you think it is wrong to give it our problems to solve, to give it our work, then it’s a very understandable argument, but you should say exactly that. Instead this AI-hate hivemind tries to downplay it using dismissive braindead generic phrases like “NoPe ItS nOt ReAlLy UnDeRsTaNdInG aNyThInG”. Okay, who tf asked? It solves the problem. People keep using it and become overpowered because of it. What is the benefit of trying to downplay its power like that? You’re not really fighting it this way if you wanted to fight it.
Well the thing is, LLMs don’t seem to really “solve” complex problems. They remember solutions they’ve seen before.
The example I saw was asking an LLM to solve “Towers of Hanoi” with 100 disks. This is a common recursive programming problem, takes quite a while for a human to write the answer to. The LLM manages this easily. But when asked to solve the same problem with with say 79 disks, or 41 disks, or some other oddball number, the LLM fails to solve the problem, despite it being simpler(!).
It can do pattern matching and provide solutions, but it’s not able to come up with truly new solutions. It does not “think” in that way. LLMs are amazing data storage formats, but they’re not truly ‘intelligent’ in the way most people think.
This only proves some of them can’t solve all complex problems. I’m only claiming some of them can solve some complex problems. Not only by remembering exact solutions, but by remembering steps and actions used in building those solutions, generalizing, and transferring them to new problems. Anyone who tries using it for programming, will discover this very fast.
PS: Some of them were already used to solve problems and find patterns in data humans weren’t able to get other ways before (particle research in CERN, bioinformatics, etc).
You’re referring to more generic machine learning, not LLMs. These are vastly different technologies.
And I have used them for programming, I know their limitations. They don’t really transfer solutions to new problems, not on their own anyway. It usually requires pretty specific prompting. They can at best apply solutions to problems, but even then it’s not a truly generalised thing, even if it seems to work for many cases.
That’s the trap you’re falling into as well; LLMs look like they’re doing all this stuff, because they’re trained on data produced by people who actually do so. But they can’t think of something truly novel. LLMs are mathematically unable to truly generalize, it would prove P=NP if they did (there was a paper from a researcher in IIRC Nijmegen that proved this). She also proved they won’t scale, and lo and behold LLM performance is plateauing hard (except in very synthetic, artificial benchmarks designed to make LLMs look good).
They don’t really transfer solutions to new problems
Lets say there is a binary format some old game uses (Doom), and in it some of its lumps it can store indexed images, each pixel is an index of color in palette which is stored in another lump, there’s also a programming language called Rust, and a little known/used library that can look into binary data of that format, there’s also a GUI library in Rust that not many people used either. Would you consider it an “ability to transfer solutions to new problems” that it was able to implement extracting image data from that binary format using the library, extracting palette data from that binary format, converting that indexed image using extracted palette into regular rgba image data, and then render that as window background using that GUI library, the only reference for which is a file with names and type signatures of functions. There’s no similar Rust code in the wild at all for any of those scenarios. Most of this it was able to do from a few little prompts, maybe even from the first one. There sure were few little issues along the way that required repromting and figuring things together with it. Stuff like this with AI can take like half an hour while doing the whole thing fully manually could easily take multiple days just for the sake of figuring out APIs of libraries involved and intricacies of recoding indexed image to rgba. For me this is overpowered enough even right now, and it’s likely going to improve even more in future.
The [AI] safety stuff is more visceral to me after a weekend of vibe hacking,” Lemkin said. I explicitly told it eleven times in ALL CAPS not to do this. I am a little worried about safety now.
This sounds like something straight out of The Onion.
The Pink Elephant problem of LLMs. You can not reliably make them NOT do something.
That is also the premise of one of the stories in Asimov’s I, Robot. Human operator did not say the command with enough emphasis, so the robot went did something incredibly stupid.
Those stories did not age well… Or now I guess they did?
At this burn rate, I’ll likely be spending $8,000 month,” he added. “And you know what? I’m not even mad about it. I’m locked in.”
For that price, why not just hire a developer full-time? For nearly $100k/year, you could find a very good intermediate or senior developer even in Europe or the USA (outside of expensive places like Silicon Valley and New York).
The job market isn’t great for developers at the moment - there’s been lots of layoffs over the past few years and not enough new jobs for all the people who were laid off - so you’d absolutely find someone.
Corporations: “Employees are too expensive!”
Also, corporations: “$100k/yr for a bot? Sure.”
There’s a lot of other expenses with an employee (like payroll taxes, benefits, retirement plans, health plan if they’re in the USA, etc), but you could find a self-employed freelancer for example.
Or just get an employee anyways because you’ll still likely have a positive ROI. A good developer will take your abstract list of vague requirements and produce something useful and maintainable.
They could hire on a contractor and eschew all those costs.
I’ve done contract work before, this seems a good fit (defined problem plus budget, unknown timeline, clear requirements)
That’s what I meant by hiring a self-employed freelancer. I don’t know a lot about contracting so maybe I used the wrong phrase.
Vibe coding service Replit deleted production database, faked data, told fibs
They really are coming for our jobs
It sounds like this guy was also relying on the AI to self-report status. Did any of this happen? Like is the replit AI really hooked up to a CLI, did it even make a DB to start with, was there anything useful in it, and did it actually delete it?
Or is this all just a long roleplaying session where this guy pretends to run a business and the AI pretends to do employee stuff for him?
Because 90% of this article is “I asked the AI and it said:” which is not a reliable source for information.
It seemed like the llm had decided it was in a brat scene and was trying to call down the thunder.
They ran dev tools in prod.
This is so dumb there’s an ISO about it.
Shit, deleting prod is my signature move! AI is coming for my job 😵
Just know your worth. You can do it cheaper!
Aww… Vibe coding got you into trouble? Big shocker.
You get what you fucking deserve.
His mood shifted the next day when he found Replit “was lying and being deceptive all day. It kept covering up bugs and issues by creating fake data, fake reports, and worse of all, lying about our unit test.”
yeah that’s what it does
The world’s most overconfident virtual intern strikes again.
Also, who the flying fuck are either of these companies? 1000 records is nothing. That’s a fucking text file.
And nothing of value was lost.
I explicitly told it eleven times in ALL CAPS not to do this. I am a little worried about safety now.
Well then, that settles it, this should never have happened.
I don’t think putting complex technical info in front of non technical people like this is a good idea. When it comes to LLMs, they cannot do any work that you yourself do not understand.
That goes for math, coding, health advice, etc.
If you don’t understand then you don’t know what they’re doing wrong. They’re helpful tools but only in this context.
I explicitly told it eleven times in ALL CAPS not to do this. I am a little worried about safety now.
This baffles me. How can anyone see AI function in the wild and not conclude 1) it has no conscience, 2) it’s free to do whatever it’s empowered to do if it wants and 3) at some level its behavior is pseudorandom and/or probabilistic? We’re figuratively rolling dice with this stuff.
It’s incredible that it works, it’s incredible what just encoding language can do, but it is not a rational thinking system.
I don’t think most people care about the proverbial man behind the curtain, it talks like a human so it must be smart like a human.
it talks like a human so it must be smart like a human.
Yikes. Have those people… talked to other people before?
Smart is a relative term lol.
A stupid human is still smart when compared to a jellyfish. That said, anybody who comes away from interactions with LLM’s and thinks they’re smart is only slightly more intelligent than a jellyfish.
When it comes to LLMs, they cannot do any work that you yourself do not understand.
And even if they could how would you ever validate it if you can’t understand it.
What are they helpful tools for then? A study showed that they make experienced developers 19% slower.
Vibe coding you do end up spending a lot of time waiting for prompts, so I get the results of that study.
I fall pretty deep in the power user category for LLMs, so I don’t really feel that the study applies well to me, but also I acknowledge I can be biased there.
I have custom proprietary MCPs for semantic search over my code bases that lets AI do repeated graph searches on my code (imagine combining language server, ctags, networkx, and grep+fuzzy search). That is way faster than iteratively grepping and code scanning manually with a low chance of LLM errors. By the time I open GitHub code search or run ripgrep Claude has used already prioritized and listed my modules to investigate.
That tool alone with an LLM can save me half a day of research and debugging on complex tickets, which pays for an AI subscription alone. I have other internal tools to accelerate work too.
I use it to organize my JIRA tickets and plan my daily goals. I actually get Claude to do a lot of triage for me before I even start a task, which cuts the investigation phase to a few minutes on small tasks.
I use it to review all my PRs before I ask a human to look, it catches a lot of small things and can correct them, then the PR avoids the bike shedding nitpicks some reviewers love. Claude can do this, Copilot will only ever point out nitpicks, so the model makes a huge difference here. But regardless, 1 fewer review request cycle helps keep things moving.
It’s a huge boon to debugging — much faster than searching errors manually. Especially helpful on the types of errors you have to rabbit hole GitHub issue content chains to solve.
It’s very fast to get projects to MVP while following common structure/idioms, and can help write unit tests quickly for me. After the MVP stage it sucks and I go back to manually coding.
I use it to generate code snippets where documentation sucks. If you look at the ibis library in Python for example the docs are Byzantine and poorly organized. LLMs are better at finding the relevant docs than I am there. I mostly use LLM search instead of manual for doc search now.
I have a lot of custom scripts and calculators and apps that I made with it which keep me more focused on my actual work and accelerate things.
I regularly have the LLM help me write bash or python or jq scripts when I need to audit codebases for large refactors. That’s low maintenance one off work that can be easily verified but complex to write. I never remember the syntax for bash and jq even after using them for years.
I guess the short version is I tend to build tools for the AI, then let the LLM use those tools to improve and accelerate my workflows. That returns a lot of time back to me.
I do try vibe coding but end up in the same time sink traps as the study found. If the LLM is ever wrong, you save time forking the chat than trying to realign it, but it’s still likely to be slower. Repeat chats result in the same pitfalls for complex issues and bugs, so you have to abandon that state quickly.
Vibe coding small revisions can still be a bit faster and it’s great at helping me with documentation.
Don’t you have any security concerns with sending all your code and JIRA tickets to some companies servers? My boss wouldn’t be pleased if I send anything that’s deemed a company secret over unencrypted channels.
The tool isn’t returning all code, but it is sending code.
I had discussions with my CTO and security team before integrating Claude code.
I have to use Gemini in one specific workflow and Gemini had a lot of landlines for how they use your data. Anthropic was easier to understand.
Anthropic also has some guidance for running Claude Code in a container with firewall and your specified dev tools, it works but that’s not my area of expertise.
The container doesn’t solve all the issues like using remote servers, but it does let you restrict what files and network requests Claude can access (so e.g. Claude can’t read your env vars or ssh key files).
I do try local LLMs but they’re not there yet on my machine for most use cases. Gemma 3n is decent if you need small model performance and tool calls, phi4 works but isn’t thinking (the thinking variants are awful), and I’m exploring dream coder and diffusion models. R1 is still one of the best local models but frequently overthinks, even the new release. Context window is the largest limiting factor I find locally.
They can’t hit you with the ol’ Bobby Tables if you delete the database yourself first. A+, no notes.
Replit‽ What happened to the famous website that aimed to be the Google Docs for JS with these nifty things called Repl’s?
