Of course not, all these sloppers are doing is training the models so at the eyes of management they are good enough for a replacement. The ones who stay will have 10x more work.
How surprised people will be when they learn that their prompts and skills and etc are being saved for ai training even though they said it would not be
They select for people who are beholden to AI, probably to eventually have the model do the job of prompting if the model is expected to be doing the doing anyhow. Anthropic job posts I've seen have explicitly said you should use claude to claude-ify your resume before submitting. I'm guessing it's an auto reject if you don't. If they are asking for you to use their ai tool for step 0 before you even work there, they are going to want you to use it for all your job functions and communications. And all of that will be logged, used as training data, and will justify not hiring to fill your seat when you leave or get canned.
Indeed this will likely happen in the future, but not today. I was experimeting with SSD streaming in DwarfStar for DeepSeek v4 PRO inference in 128GB systems (and Flash inference iwth 32/64). GPT 5.5 ran the whole night, I checked what it had accomplished regardless of all the hints I provided in the specification document. After reasoning on the problem I gave him the design fixes and the tokens/sec were 4x after 10 minutes. And this is true for every domain where the human babysitting the AI know a few things in that domain. However this is a moving target, and at the current rate, soon or later, indeed AIs will do much better than us in many domains.
The article is quite general. Here's some notes on how AI is being used to do AI research at frontier labs specifically. It's not the singularity (yet?) but it's heading in that direction.
Most training is now actually inference, not directly gradient descent. Reinforcement learning requires the generation of lots of 'rollouts' that are then compared with each other via an algorithm like GRPO. Or they might be compared using a critic model - AI judging AI and causing it to self improve. Generating a rollout means inference. And there's lots of data cleaning by older models. This has been called in the past 'textbook' or 'curriculum' learning, not sure what it's called now. But AI is also used for things like data/document labelling, transcription of videos, detection of images/videos with watermarks or subtitles, elimination of content that shouldn't be in the dataset, creation of new content that should and so on.
AI has proven capable of some routine work, like brute-force optimizing GPU kernels or doing hyperparameter sweeps.
Obviously, researchers are all using coding agents too.
So that's a few ways AI is self-improving. But there are lots of other ways in which even frontier models are still beaten by human researchers. Experiments in closing the loop have failed. For instance, people have tried giving the latest models access to some GPUs and an old version of an AI codebase that was recently optimized by human researchers (a NanoChat speed run goal, I believe). Could the models match the performance of the AI researchers? Nope. They only got 10% as far as the humans did, mostly because their approach was uninspired. They wasted a lot of time and budget doing low-IQ stuff like hyperparameter tuning. The humans had many other tactics like studying the research literature and inventing new algorithms that the models didn't even attempt.
The bottleneck is therefore currently the level of insight and inspiration the models are capable of. I've also seen this in my own work. I come up with an idea I think is novel and see if I can get a frontier model to reach the same idea. It never works without questions so leading it's more or less pointless.
It's very unclear why AI struggles so much with innovation yet can invent new songs, poems etc without apparent difficulty. Obvious answers like "it's not in the training set" don't feel right to me, the issue is deeper.
I had the same thought about ML engineering a while back, when Google released the AutoML suite, that was banned from Kaggle competitions. At the time, it seemed obvious to me that the closer you were to the models, the easier it was for them to replace your work, since most of the work on models was itself hill-climbing, grind searching and mutation search. So, the more your work is an explicit, measurable search loop, the more automatable it is.
Same with prompts, most attempts seem to be fidgeting with the models till they get your intend right, which is also a matter of hill-climbing, subtle mutation, and so on.
If I were to clarify anything from the article, I'd probably say that I'd rather do the factorisation of programming roles by how long they already existed. If someone is an AI engineer and his work only became relevant a month ago, very probably it will be obsolete in another month. If they do the same thing for the past 10 years, changes are that their skills would be useful for another 10 years to come.
I agree with you, although you could argue that history teaches the opposite, for example the many jobs that disappeared after medieval times, even if they existed for centuries. Maybe you will still find a cobbler or a blacksmith, but now it's something very niche; which is more or less the point of the post.
With human productivity already fully unleashed, the natural next step was to initiate self-evolution of both the model and the organization. M2.7 is our first model deeply participating in its own evolution.
25 comments
[ 5.2 ms ] story [ 36.8 ms ] threadOf course not, all these sloppers are doing is training the models so at the eyes of management they are good enough for a replacement. The ones who stay will have 10x more work.
Although this may be more relevant to replacing AI researchers, not AI engineers...
(Submitted as https://news.ycombinator.com/item?id=48380643 )
Most training is now actually inference, not directly gradient descent. Reinforcement learning requires the generation of lots of 'rollouts' that are then compared with each other via an algorithm like GRPO. Or they might be compared using a critic model - AI judging AI and causing it to self improve. Generating a rollout means inference. And there's lots of data cleaning by older models. This has been called in the past 'textbook' or 'curriculum' learning, not sure what it's called now. But AI is also used for things like data/document labelling, transcription of videos, detection of images/videos with watermarks or subtitles, elimination of content that shouldn't be in the dataset, creation of new content that should and so on.
AI has proven capable of some routine work, like brute-force optimizing GPU kernels or doing hyperparameter sweeps.
Obviously, researchers are all using coding agents too.
So that's a few ways AI is self-improving. But there are lots of other ways in which even frontier models are still beaten by human researchers. Experiments in closing the loop have failed. For instance, people have tried giving the latest models access to some GPUs and an old version of an AI codebase that was recently optimized by human researchers (a NanoChat speed run goal, I believe). Could the models match the performance of the AI researchers? Nope. They only got 10% as far as the humans did, mostly because their approach was uninspired. They wasted a lot of time and budget doing low-IQ stuff like hyperparameter tuning. The humans had many other tactics like studying the research literature and inventing new algorithms that the models didn't even attempt.
The bottleneck is therefore currently the level of insight and inspiration the models are capable of. I've also seen this in my own work. I come up with an idea I think is novel and see if I can get a frontier model to reach the same idea. It never works without questions so leading it's more or less pointless.
It's very unclear why AI struggles so much with innovation yet can invent new songs, poems etc without apparent difficulty. Obvious answers like "it's not in the training set" don't feel right to me, the issue is deeper.
Same with prompts, most attempts seem to be fidgeting with the models till they get your intend right, which is also a matter of hill-climbing, subtle mutation, and so on.
If I were to clarify anything from the article, I'd probably say that I'd rather do the factorisation of programming roles by how long they already existed. If someone is an AI engineer and his work only became relevant a month ago, very probably it will be obsolete in another month. If they do the same thing for the past 10 years, changes are that their skills would be useful for another 10 years to come.
Drag it out for a couple of years and you'll be set.
https://www.minimax.io/news/minimax-m27-en
With human productivity already fully unleashed, the natural next step was to initiate self-evolution of both the model and the organization. M2.7 is our first model deeply participating in its own evolution.