Whenever someone does such a huge drastic change like this, it's by ignoring a large chunk of code that most people were afraid to touch for good reasons. Now, that code is gone, AI is celebrated, things will break, people will work very hard in the background to fix it, with no fanfare.
Key statement:
> But at the rate we were going, we were still roughly five years from finishing.
Everyone has seen how this sort of thing comes about: it's meaningful work for engineering but never business/product critical so it just drags along.
Seems like this time around someone just went "I wonder if we could do it this way" and it worked. Perfect example of ditching sunk-cost and starting from scratch. Great outcome for them.
Because OpenAI is burning $15 billion/year, and outrageous stories like those get parroted in the media. It's free marketing for a company desperate to get middle management to believe that a $500/mo subscription is absolutely crucial for every single employee.
ugh just recently I saw a news article gushing about how AI had helped with some health/medicine study, and various patient organizations were all like "oh yeah this is a Good use of AI" and then I click the link to read the study and it's decision trees and clustering on a tiny dataset that you could analyze with a ten year old laptop.
I mean, sure at some point decision trees and clustering were called "AI", but the way the article was written you'd think OpenAI and Anthropic were responsible for the advancement of medicine.
This appears entirely unreasonable.
Normally the reason is not good. The reason is that unit testing is missing or that downstream effects are not entirely mapped out.
Exactly activities that traditional software developers are loathing because they are boring and mentally straining.
Seems a bit strong.
Unit testing can get you some of the way, but its not a full-on all-case guarantee. Sometimes the code is encapsulating some particularly complex system/behaviour. Sometimes the reason is interop/compatilibity issues or some kind of politics.
P.S. you managed to introduce a typo in your quote (were –> where)
This comes directly after they wrote "statements we can't verify".
Why is it that we can not verify the statements, but we can believe the reasons that people will not touch the code base to be "good"?
This. It looks like AI companies have managed to use this for their advantage. Can't really blame them.
― Woody Allen
You see: normally one reads for recreation or enlightenment.
If that same one would approach their job as a recreational activity, or solely for enlightenment, you might not have that job for very long.
2. It's a joke! Relax.
Latently they say that it is a quality of itself to do the work.
That is a reasonable statement, however, I am sure they also extend it to harvesting, baking etc. As to not use mechanical help.
Heck, if they actually believed their own statement then they'd not even be commenting here, as that require immense amounts of mechanised help.
all work is stack ranked against other opportunities. no matter how many teams help to parallelize, leadership basically comes down to ranking the stack correctly. Valuable work will be at the top by any means necessary. “mvp”, probes, task forces, code yellows, and so on.
edit: it’s even a backhanded compliment to agentic coding. Asana a public company considers a $12k outlay for something they’d never spend real resources on worth a case study =|
I did subagent based removal tasks a few times. These were the ones that required the least amount of input or thinking from me, because the requirements are abundantly clear. "Remove this part of the code without breaking any other part or by porting the tests done with it onto another system."
In these situations the code acts as the bookkeeping ledger itself, and coordination complexity is almost a no-brainer.
It's like every time they make a plan, there's something about things taking "a week or two", "month of focused work", or whatever.
This is something that would've been RL'd out a long time ago if it wasn't great for business.
What I was going after with "aware", is that the actual people working at the companies, training the models, are aware that people aren't mostly going to be implementing the plan by hand, if they've already made the plan in Claude Code or Codex. As for a specific Claude / GPT instance, "aware" would definitely be the wrong word choice there, but the instance does have its stats and environment information in its context window, unless you specifically remove it.
Either way: Training does include estimates on working with the model, and adjustments of the model itself based on that. That's literally what RLHF is.
It's straightforward to have a portion in post training that aims specifically at the model being able to give better estimates on how long that model takes to complete a certain type of task.
When the post training run is nearing its cutoff point, there's a massive amount of data on how long coding tasks take to complete by that model in the golden format of "task -> time task took to complete", separable to whatever amount of subtasks, in the same format. With the parts from the end of the dataset being useful for evaluating the finished model's capabilites, whether that data is then fed back into another step in post training or not.
Completely separate from even the actual training: If you have a model proactively giving estimates that are an order of magnitude wrong, you can already fix the worst of it as of this moment by just changing the system prompt. It's a dirty fix, but it's the type of fix that has been used by Anthropic and OpenAI since forever when a model is dishing out blatantly wrong outputs.
Not sure if I'm misunderstanding your point?
(And also prompts will need to be refined, etc., unless you have amazing prompt skills it won't immediately deliver what you wanted. Even if it did, the work to put together the AI inputs also needs to be included.)
⇒ I don't think it's easily possible right now to give a time estimate for **the full picture of** a project to be implemented with AI assistance.
> For comparison: the previous plan was expected to take at least five years and estimated to cost roughly $6M.
If that's an estimate from an LLM, those have always been way off for me. I'm constantly amazed whenever something that an agent estimates would take weeks ends up being completed in an hour.
I suspect the labs could improve the models such that they are estimating these sorts of things but they don't prioritize doing so (or perhaps RLHF selects it away) because, as you say, it feels amazing to do a week's worth of work in an hour.
The headlines feel similar
Non prod code, good use case.
Hopefully a human did give a quick look to make sure it didn’t just delete the tests.
This kind of projects are where AI is most helpful. Long tedious and highly testable projects like ports or legacy system replacements where humans have to grind through millions of lines of code without really thinking are the perfect target for AI.
I once did a C to C# port of several math libraries, and while I was able to automate most of it, it still required a ton of manual work. I bet if I had Claude, I could do what I did in 3 months in a day or two.
Although also agree, that even if human's did it, they probably would try to automate as much as possible and speed up development, and maybe that original 5 year estimate was if it was done completely manually.
We engineers, had the power to dictate how our days will look, and we were able to give ballooned timelines to give us room to braeth.
This has changed. The power is no longer in our hands, for good or bad.
I suspect it as a two month project without agents. Max.
It doesn’t sound like that much changed. If one engineer can really review all the changes in under two weeks, the original estimate of 5 years of engineering work has to be waaaaaaay off.
The impressive thing here is that Asana has such a poorly run engineering org that replacing their testing framework was estimated to take 5 years and $6 million.
Oh dear, poor Asana customers...
Summary: The Takeaway Did an AI generate 5 years' worth of highly creative, novel software features from scratch in two weeks? No.
Did an AI complete 5 years' worth of tedious, widespread code-migration technical debt in two weeks? Yes.