vervez

There's also recent Bayesian Optimal Experimental Design methods that allow you to directly design experiments using gradient ascent. Not sure how it compares with BayesOpt on your problem, though.

github copilot won't pay for itself.

Their paper is more for the case of "we can't gather more data, so what to do?" but your solution is in line with optimal experimental design and choosing a utility function to distinguish between models using as little data as possible.

min{d(problem)/d(theta)} is essentially what LLMs are doing with a prompt. Every session that a chatgpt user has leads to either a resolution or not, which is the loss function given the prompt used to reach that point. It's getting better at not hallucinating in my experience over just the past 4 months.

Thanks for sharing and i'm happy to be your friend :)

I see your point and agree it can be frustrating if DEI initiatives are used as a political tool or blindly without considering the pitfalls of a naive implementation.

Indeed, I didn't cite any sources but I'm fairly sure there's literature studying this phenomenon. I agree that DEI when used as a political tool is a distraction for a company.

I think many of the things you mentioned w.r.t. diversity is actually forward-thinking and potentially yields more returns later. If there's a brilliant swe that feels uncomfortable at a company because of how they name their branches, they won't work there. Same goes for hiring a "less-qualified" individual. How many decisions or projects can be improved because of a diversity of input? Again, a non-linear outcome that is harder to measure than immediate profitability. Those are some arguments for those programs focusing on your productivity point.

Equivariance is growing in popularity in machine learning, so these tricks will be helpful if one wants to study, or publish, in that area (I'm thinking about it). I'd recommend for any ML-related folks to look into this area and save this thread whenever they're trying to implement an equivariant neural network.

So you `pip install`ed them at a new company? I'll see myself out.

Couldn't read the article but yea, if it's a small molecule, most likely it's inhibiting some protein specific to cancerous cells. In this case, it sounds like it's blocking some protein that blocks human cells' innate ability to produce antigens, which signal to T-cells that they are defective and need to be destroyed.

Sometimes we understand the biology after we discover a treatment.

Briefly, no. Why: they're most likely not training your average ML models on this dataset. Instead, they are likely taking a model of some physics and seeing how it performs in these simulations.

You can think of this as a form of causal inference - "if this model is true, how well does it work with our current understanding (simulations) of physics?" type of questions. There are measures of error and bias that come with evaluation of these models.