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.
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.
AI research is not over yet. If OpenAI and Anthropic fail to bring on the Singularity, what are we going to call the continuing research on AI? Are we going to call it something else than "AI" because that was taken by LLMs? Does that make any sense at all?
I kind of agree, though I find "AI" to be an almost useless term, in particular these days. I dislike the current state of discourse where AI is used to mean inhumanly large models, so I tend to use a specific term like LLMs instead of AI. But the point here wasn't whether this or that form of ML is in fact AI. The term "AI" was almost never used in mainstream media until a few years ago, when it suddenly it was all over the place https://trends.google.com/trends/explore?date=all&q=AI,Biden... and nearly without exception referring to very large generative text/image models. So when a news article in 2026 talks about AI without specifying that it's nothing to do with ChatGPT etc. (and in fact using a very different method requiring no large pre-trained model and no dependence on large companies), I find it highly misleading.
I don't agree that it's a useless term. AI, as a research area, is well-defined by its journals and conferences and the grants that support that research. It's not the fault of the AI research community that OpenAI's app went viral and now there's so much confusion in mainstream discourse about the term. I try to set the record straight whenever I can but it does feel like fighting a pointless rearguard action, even in this community that is better informed than most.
this allows said middle managers to do the analysis themselves and feel mostly confident in the results.
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.
> Back in 2022, we set out to migrate Asana's frontend test suite off Enzyme, our aging testing library, and onto React Testing Library (RTL).
Your telling me they had a full team of engineers at Asana, doing nothing but rewriting tests for 4 years until AI came along and did the last year in a couple days? I'm extremely doubtful. I don't doubt for a minute they were one-track to take 5 years, but not because it was 5 years of engineering effort for humans.
Far more likely they finally cleared up some tech debt they had been plugging at off and on for 4 years, then a PR flack got ahold of it and it became a breathless "AI did 5 years of work in a couple days".
With enough budget this seems a rather reasonable and fun task.
But yes, LLMs are great at tests. I'm not doubting that an LLM migrated some legacy tests, or that it was a task taking a long time. I am doubting the way it was presented in the article, that this was taking a full team of engineers dedicated to nothing but this, 5 years to accomplish (and presumably already spent 4 million working on this, since the total estimate was 6 million, and they've been at it for 4 year).
I think some PR flack got ahold of the fact that an LLM wrapped up migrating some legacy tests that a team had been slowly chipping away at for 4 years, between their normal feature work, and were on track to finish in 5, then wrote it up like it was that teams entire focus instead of a piece of tech debt.
That makes far more sense to me than spending millions for a team of engineers dedicated to nothing but rewriting an existing test suite.
I don't think that's true. All you see is all the test pass. You don't know that the tests still cover everything that they used to.
LLMs excel where there is an excellent test suite, and you ask them to modify the thing that the test suite tests, and forbid them from changing the tests.
The ultimate bad faith interpretation. "Unless I can verify the results that contradict my worldview, I don't acknowledge them."
> https://www.mikekasberg.com/blog/2026/08/19/hacking-with-cla...
"I haven't done this, so this doesn't prove it."
> https://www.bbc.com/news/articles/clyq011414eo
"I haven't seen the paper trail, so this doesn't prove it."
on and on...
> But you _can_ verify if you had bothered
I think pointing out I can't is indeed fair.
We absolutely should be skeptical when an AI company makes big claims. The fact that the company the AI company is talking about also claims the same thing doesn't change that. OpenAI is getting awareness and marketing out of this, and I'm sure Asana is getting something out of it too.
And I'm not even saying OpenAI or Asana are necessarily lying. Asana might not find out for months or years that some tests had been rewritten poorly, and don't sufficiently test the thing they were supposed to test anymore. For example. If they truly had 5 years of work, then I find it hard to believe that in two weeks of the LLM churning, they had the time to review all the new tests. They spot-checked, at best.
Maybe everything is great. Maybe the LLM did a wonderful job, and this was awesome for Asana. But we have no idea, and we're unlikely to ever find out. Unless, of course, it's in Asana's interest from a marketing perspective to tell us.
(Not sure what the URLs you posted in your comment are supposed to prove. They're unrelated to the issue at hand.)
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"?
Because "we all" is a bubble, and many people do not have AI at work, or at least not the level of usage that many people here have the budget for at their company.
This. It looks like AI companies have managed to use this for their advantage. Can't really blame them.
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.
either it was a hail mary 5 year change-the-market offering, or it was something that staff can "get around to" whenever they have the cycles, and would, at current rates, take like 3-5 years to do.
was the tool worth 12k? if it's not a line in capex or opex budgets it's value is $0
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.
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.
But here’s the rub: this is, basically by definition, low-priority engineering work. Those fixed papercuts are nice to have fixed but they do not necessarily add a lot of value.
And it’s all too easy to lose value by doing this. For example, current LLMs really really like adding test cases, and a lot of those test cases have basically no value, and carrying them around is not free.
I also realized that review generated code sucks, I can generate a whole app quickly, but I have no understanding of small decisions. It is extremely difficult to wrap your hand around hundreds of lines of code written by someone else.
codex-rs, popularly known as Codex or maybe the Codex CLI, is a fine choice.
https://learn.chatgpt.com/docs/codex/cli
https://github.com/openai/codex
Looks like it's still a more-or-less-open-source product called Codex CLI.
FWIW, there have been issues with various versions of ChatGPT not knowing what Codex is -- at least in the (IIRC) GPT-5.3 timeframe, there were serious issues with the knowledge cutoff.
First, the app was called Codex, then they combined them and now the app is called chatGPT, but recently inside the app you can switch between ChatGPT and Codex. lol.
press X to doubt [x]
Non prod code, good use case.
Hopefully a human did give a quick look to make sure it didn’t just delete the tests.
I tried it in AI and virtually did it all in a few hours! Magic?
Well I had gone through it to explore the problem space, and we had found a working solution which just petered out. But I still remembered the learnt lessons, and I was just eager to see what it would be like to use DuckDB on it.
My point being these second passes on projects often skip the hard won knowledge part.
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.
The headlines feel similar
Oh dear, poor Asana customers...
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's entirely possible for a 2-3 month project if one senior person was left the hell alone to do it to take a dozen people five years.
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.
"Don't be snarky."
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.
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.