AI coding is gambling

some_random

How often do you have to win before it's no longer gambling?

apitman

See also https://www.fast.ai/posts/2026-01-28-dark-flow/

minimaxir

The gambling metaphor often applied to vibecoding implies that the outcome cannot be fully controlled or influenced, such as a slot machine. Opus 4.5 and beyond show that it not only can be very much can be influenced, but also it can give better results more consistently with the proper checks and balances.

samschooler

I think there are levels to this. - One shot or "spray and pray" prompt only vibe coding: gambling. - Spec driven TDD AI vibe coding: more akin to poker. - Normal coding (maybe with tab auto complete): eating veggies/work. Notably though gambling has the massive downside of losing your entire life and life savings. Being in the "vibe coding" bucket's worse case is being insufferable to your friends and family, wasting your time, and spending $200/month on a max plan.

xnx

...and the payouts are fantastic.

Retr0id

> But now either the AI can handle it or it can pretend to handle it. Frankly it's pretending both times, but often it's enough to get the result we need. This has been how I think about it, too. The success rates are going up, but I still view the AI as an adversary that is trying to trick me into thinking it's being useful. Often the act is good enough to be actually useful, too.

simonw

Assigning work to an intern is gambling: they're inherently non-deterministic and it's a roll of the dice whether the work they do will be good enough or you'll have to give them feedback in order to get to what you need.

rvz

It is indeed gambling. You are spending more tokens hoping that the agent aligns with your desired output from your prompt. Sometimes it works, sometimes it doesn't. Watching vibe gamblers hooked onto coding agents who can't solve fizz buzz in Rust are given promotional offers by Anthropic [0] for free token allowances that are the equivalent in the casino of free $20 bets or free spins at the casino to win until March 27, 2026. The house (Anthropic) always wins. [0] https://support.claude.com/en/articles/14063676-claude-march...

halotrope

idk it works for me it build stuff that would have taken weeks in hours ymmv

dwa3592

few thoughts on this- it's not gambling if the most expected outcome actually occurs. It also depends on what you're coding with; - If you're coding with opus4.6, then it's not gambling for a while. - If you'r coding with gemini3-flash, then yeah. One thing I have noticed though is- you have to spend a lot of tokens to keep the error/hallucination rate low as your codebase increases in size. The math of this problem makes sense; as the code base has increased, there's physically more surface where something could go wrong. To avoid that you have to consistently and efficiently make the surface and all it's features visible to the model. If you have coded with a model for a week and it has produced some code, the model is not more intelligent after that week- it still has the same layers and parameters, so keeping the context relevant is a moving target as the codebase increases (and that's why it probably feels like gambling to some people).

bensyverson

This "slot machine" metaphor is played out. If you're just entering a coin's worth of information and nudging it over and over in the hopes of getting something good, that's a you problem, not a Claude problem. If, on the other hand, you treat it like a hyper-competent collaborator, and follow good project management and development practices, you're golden.

rustyhancock

Life is full of variable reward schemes. Probably why we evolved to be so enamoured by them. Sometimes I think we put the Carr before the horse. We gamble because evolution promotes that approach. Yes I could go for the reliable option. But taking a punt is worth a shot if the cost is low. The cost of AI is low. What is a problem is people getting wrapped up in just one more pull of the slot machine handle. I use AI often. But fairly often I simply bin its reponse and get to work on my own. A decent amount of the time I can work with the response given to make a decent result. Sometimes, rarely, it gives me what I need right off the bat.

jsLavaGoat

Everything is "fast, cheap, good--pick two." This is no different.

CraftingLinks

I see whole teams pushed by c- level going full in with spec driven + tdd development. The devs hate it because they are literally forbidden to touch a single line if code. but the results speak for themselves, it just works and the pressure has shifted to the product people to keep up. The whole tooling to enable this had to be worked out first. All Cursor and extreme use of a tool called Speckit, connected to Notion to pump documentation and Jira.

CodingJeebus

For me, the feedback loop accelerating the way that AI now permits is so addictive in my day-to-day flows. I've had a really hard time stepping away from work at a reasonable hour because I get dopamine hits seeing Claude build things so fast. Addiction and recovery is part of my story, so I've done quite a bit of work around that part of my life. I don't gamble, but I can confidently say that using LLMs has been an incredible boost in my productivity while completely destroying my good habits around setting boundaries, not working until 2AM, etc. In that sense, it feels very much like gambling.

aderix

Sometimes I feel that subsidising these packages (vs cost via API) is meant to make more and more people increasingly addicted

Terr_

I'd emphasize that prompting LLMs to generate code isn't just metaphorical gambling in the sense of "taking a risk", the scary part is the more-literal gambling involving addictive behaviors and how those affect the way the user interacts with the machine and the world. Heck, this technology also offers a parasocial relationship at the same time! Plopping tokens into a slot-machine which also projects a holographic "best friend" that gives you "encouragement" would fit fine in any cyberpunk dystopia.

lasgawe

haha.. I agree with the points mentioned in the article. Literally every model does this. It feels like this even with skills and other buzzword files

amw-zero

So is human coding.

thisisbrians

It is and will always be about: 1) properly defining the spec 2) ensuring the implementation satisfies said spec