> Your best employees are the most likely to leave via attrition, because they have the most opportunity elsewhere.
But this remains true after a layoff and the layoff often acts a motivator for your best employees to start looking even if they weren't previously.
Usually they aren't thinking "well, glad I survived that layoff and now my job is safe forever", they are thinking "huh, is this a sinking ship? Maybe I should look around and see what else is out there..."
...speaking as someone that has been at several companies during layoffs...
yeah, it seems like it would have to be accompanied by a pay bump for the ones you really want to retain... which is challenging from an optics perspective.
Perhaps, but there are optics only when someone sees. It's not unusual to fund retention increases using some of the budget freed up by a layoff, but that won't be explicitly stated in a public announcement.
I think this article is a bad writeup about it. The attack is academically interesting but not practical or worth worrying about.
The claim that it provides a 'near-perfect attack success' is misleading; for the majority of datapoints it had no success at all. However for a minority of datapoints (anywhere from 0.001% to 10% depending on the model) it was able to state with >95% confidence that they were in the training data.
They are also not simply 'tricking' it into revealing your data. The attacker needs to already have your data in order to check if it is in the model. It also required retraining the model 200 times on different subsets of the training data and comparing the differences.
That does not sound stupid, but safe? I giant truck that has no chance of stopping, controlled by a computer? Just build railways and then there are no issues, no fancy AI to control them.
The trouble is: there is no deterministic algorithm that can do the things neural networks can do.
For many of these problems, I think it is likely that no deterministic algorithm can exist because the problems are fundamentally underspecified. E.g. a common task in computer vision is generating a 3D depth map from a 2D image. This is inverting a lossy projection, so any solution must be a least partially a hallucination.
I think we just have to accept this. It's a different type of algorithm, built out of statistics instead of logic, with different strengths and weaknesses compared to traditional software.
I feel the same way. Analogy: we’ve been geologists this whole time, building our dynamic and interesting mechanical planet.
Now, biology exists. It’s wet and messy and impossible to understand (we haven’t invented the microscope yet). That doesn’t mean biological study is not worth doing.
GPUs are not really the ideal architecture for running neural networks; they are heavily bottlenecked by memory bandwidth and struggle to keep all their tensor cores supplied with data.
There is significant room to make more specialized neural network accelerators with new compute-in-memory architectures.
If the brain can run 86 billion neurons on 30W it must be possible.
There are already some companies doing specialised inference hardware, Cerebras Systems for example. Such designs are still early days and I wouldn't be surprised to see more innovation there. Though because custom silicon design takes time I expect a multi-year cycle.
For training, not sure. But even if training runs on GPUs, once you have the model the main cost is inference.
Whole cryptography often time feels like snakes oil. You can't tell if algorithm you are using is not actually fundamentally broken. Or we know that algorithm is fundamentally broken (like RSA) and we are just one mathematical discovery from a digital catastrophe.
Actually this is reality of cryptography. Is AES secure? Well lot of people tried to break it and failed. Which does not answer the question because we don't have any mathematical tools to say yes this cryptography algorithm is formally verified to be secure and unbreakable. It can turn out that tomorrow we will find a flaw in AES which will make it leak key after capturing few blocks.
Only thing which we know is unbreakable is OTP and that's because it is easy to prove it.
MidJourney has better style than most other image generators, especially for artistic images.
Unfortunately it hasn't kept up on image quality, detail, or text rendering. I think this is because they don't have the $ to keep up in the scaling game.
There have also been proposals to use flash memory in inference accelerators instead of DRAM. You can make high bandwidth flash using the same stacking technique used for HBM DRAM.
It is obviously unsuitable for training because of limited write cycles. But the read bandwidth is decent, and the density/$ is much better.
In theory, a small layoff can target the least productive employees.
reply