just quick retraction of my (deleted) post that I've not been using AI. after being traumatized by GPT 5.5 fucking up, Fable being withdrawn, and Opus 4.8 being depressingly dumb (relatively), I spent some days coding manually and I made a lot of progress. that said, I over-compensated. pure manual coding is not faster than using AI. it is faster than using AI poorly. like everything, there is a balance where you use it responsibly: audits, well specified refactors, sanity checks... where you can extract a lot of value. my mistake was letting codex edit a lot of code unsupervised without really reading its output. that's clearly my own fault for using it poorly and Bend's delay is attributable to it. it is a new powerful tech. I think not using it is a mistake, but using it too aggressively is a mistake too. I've been using AI carefully and doing things way slower now, taking the time needed to assert each added functionality is correct, well written, and well understood, before moving to the next, and, by doing so, things are actually moving faster, because progress only moves forward, never backward - which happens a lot when you let AI unsupervised. I'm halfway through. Bend2 has 50% of the pre-rewrite features, except the code is now beautiful and extremely robust. this is not adding anything it didn't have, I'm just ensuring the codebase mets my own quality standards
Victor Taelin, Bend creator, retracts his vow to abandon AI coding, finding manual programming slower than highly disciplined AI use
Story Overview
Victor Taelin had publicly stepped away from AI coding tools after frustrating runs with models that injected bugs and slowed his work, yet several days of pure manual effort on Bend2 convinced him the opposite approach was even slower, prompting a return to tightly controlled AI assistance.
Manual effort fell short of careful AI pairing
Taelin now limits AI strictly to well-scoped refactors and audits, reading every generated line, running sanity checks, and refusing any change that risks regressions before moving forward.
Bend2 sits at a robust halfway point
The language project is described as about fifty percent complete with a codebase that meets unusually high personal standards, though no further details on timelines or remaining scope have been shared.
Positive users endorse Taelin's balanced AI coding approach as relatable and effective when every line is reviewed, while negative users report bugs, wasted time, and hidden errors from over-reliance on AI-generated code.
No Digg Deeper questions have been answered for this story yet.
Most Activity
People don't take me literally enough when I say you need to read every single line of code yourself!
just quick retraction of my (deleted) post that I've not been using AI. after being traumatized by GPT 5.5 fucking up, Fable being withdrawn, and Opus 4.8 being depressingly dumb (relatively), I spent some days coding manually and I made a lot of progress. that said, I over-compensated. pure manual coding is not faster than using AI. it is faster than using AI poorly. like everything, there is a balance where you use it responsibly: audits, well specified refactors, sanity checks... where you can extract a lot of value. my mistake was letting codex edit a lot of code unsupervised without really reading its output. that's clearly my own fault for using it poorly and Bend's delay is attributable to it. it is a new powerful tech. I think not using it is a mistake, but using it too aggressively is a mistake too. I've been using AI carefully and doing things way slower now, taking the time needed to assert each added functionality is correct, well written, and well understood, before moving to the next, and, by doing so, things are actually moving faster, because progress only moves forward, never backward - which happens a lot when you let AI unsupervised. I'm halfway through. Bend2 has 50% of the pre-rewrite features, except the code is now beautiful and extremely robust. this is not adding anything it didn't have, I'm just ensuring the codebase mets my own quality standards
Example: GPT 5.5 is continuously scanning holes in Bend as a proof language. This is no substitute to a formal proof, but works great as a stopgap, and for finding bugs. Since GPT isn't editing my code, its inputs are zero risk. Either it adds insight, or does no harm.
just quick retraction of my (deleted) post that I've not been using AI. after being traumatized by GPT 5.5 fucking up, Fable being withdrawn, and Opus 4.8 being depressingly dumb (relatively), I spent some days coding manually and I made a lot of progress. that said, I over-compensated. pure manual coding is not faster than using AI. it is faster than using AI poorly. like everything, there is a balance where you use it responsibly: audits, well specified refactors, sanity checks... where you can extract a lot of value. my mistake was letting codex edit a lot of code unsupervised without really reading its output. that's clearly my own fault for using it poorly and Bend's delay is attributable to it. it is a new powerful tech. I think not using it is a mistake, but using it too aggressively is a mistake too. I've been using AI carefully and doing things way slower now, taking the time needed to assert each added functionality is correct, well written, and well understood, before moving to the next, and, by doing so, things are actually moving faster, because progress only moves forward, never backward - which happens a lot when you let AI unsupervised. I'm halfway through. Bend2 has 50% of the pre-rewrite features, except the code is now beautiful and extremely robust. this is not adding anything it didn't have, I'm just ensuring the codebase mets my own quality standards
@VictorTaelin bro you are so bipolar with AI (just like me fr)
@puhlkit just like everyone who still updates their priors !
> I've been using AI carefully and doing things way slower now, taking the time needed to assert each added functionality is correct, well written, and well understood, before moving to the next, and, by doing so, things are actually moving faster, because progress only moves forward, never backward - which happens a lot when you let AI unsupervised.
This sounds like the recovery phase from AI psychosis, smth 😛
@VictorTaelin Imo eventually we need to transition to writing type specifications so we can then be sure the code the AI writes is correct (because it typechecks)
@VictorTaelin Maybe you still did not learned how to use agents properly
It also depends on the classes of problems too are tackling. Web sites, CRMs, simple saas, mostly glue already know domains models. AFAIK, you are creating new languages to solve hard problems like automatic verification. I suspect your intuition is right, moving slower, also building very specific harnesses
@VictorTaelin Highly relatable post.
@VictorTaelin it's why i really don't get the hype about loops, like are these people building something so mediocre that they can just leave a bunch of agents working unsupervised on their codebases?
@VictorTaelin looks like your AI already forgot the context and made the feeling of bipolar
no worry, it happens with me sometimes as well
Yeah. Totally agree. Lots of people do the long run loops and goals, but I never had a good experience with letting senior developers code big features without a fairly tight review process, unless every element was already spec’d and had full mockups with interactions mapped out.
I don’t see why we think AI is mind reader.
I used to use issues and keep them near atomic, but now it’s focused instructions and tight check.
My sense is that developers who have never run teams will eventually just settle back down to the best practices that senior managers of good size dev teams have been using for a couple decades.
Ai is an army of junior devs but can be made to to behave like an army of junior devs with senior devs cleaning up the mess and pushing back before going to main.
@VictorTaelin Exactly my experience - it looks good and then lingering bugs and eventually I discovered absolutely epic concept level f ups that I have to fix manually
@VictorTaelin I recently made the same realization.
@VictorTaelin “which happens a lot when you let AI unsupervised.” You mean even if letting AI unsupervised?
@yacineMTB code doesn't lie, but AI sure can
btw, this was tongue-in-cheek but there is a serious hypothetical observation here. I have seen a few of my engineering friends kind of "wake up" faster when forced to go back to 4.8/some weaker model from Fable.
Because it "seemed so good" before but now it seems dumb, "why did it seem so good before?".
@VictorTaelin AI is for boys, typing is for men, be a man gd!
Did i do that right? 😂
@VictorTaelin claude codex struggled for weeks over a simple 6 table db WHERE I DELIBERATELY DID NOT USE SQL, these models know SQL but perplexity inevitably blows out as u refine model, starts conflating and renaming, cannot tell things apart
@VictorTaelin Yep. This is the way.
Tomorrow may be different.