gr3ml1n
That's the OP's question. Bluntly: if I'm here, and they're bloviating over there, what can they actually do about it?
Yes, it's a very hand-wavey argument.
But, again with the caveats above: if we assume an AI that is infinitely more intelligent than us and capable of recursive self-improvement to where it's compute was made more powerful by factorial orders of magnitude, it could simply brute force (with a bit of derivation) everything it would need from the data currently available.
It could iteratively create trillions (or more) of simulations until it finds a model that matches all known observations.
The idea is a sufficiently advanced AI could simulate.. everything. You don't need to interact with the physical world if you have a perfect model of it.
> But, what other fields would it do this in? How can it makes strives in biology, it can't dissect animals ...
It doesn't need to dissect an animal if it has a perfect model of it that it can simulate. All potential genetic variations, all interactions between biological/chemical processes inside it, etc.
It's not because they couldn't recover, it's because they don't want to or see the point.
Most decisions that would be made in the context where this is a useful technique are irrelevant and/or obvious. They should be made by someone lower down the chain, but organizational dysfunction requires tricks like this to get anything done.
There are multiple critical reviews of this paper. It is well-known to be largely nonsense.
http://www.jsmp.dk/posts/2019-05-12-blindauditions/blindaudi...
If you go down to table 6 (which is also incredibly weak), it shows the opposite: men are advancing at a higher rate than women in blind auditions.
Andrew Gelman reviewed the link as well and agreed:
https://statmodeling.stat.columbia.edu/2019/05/11/did-blind-...
When a company is under pressure to boost the number of X engineers, they quickly run into the 'pipeline problem'. There simply isn't enough X engineers on the market. So they address that by creating scholarship funds exclusively for race X.
When a school is under pressure to have the racial makeup of it's freshman class meet the right ratios, it has to adjust admission criteria. Deprioritize metrics that the wrong races score well on, prioritize those that the right races score well on. If we've got too many Y, and they have high standardized test scores? Start weighing that lower until we get the blend we're supposed to have.
The goal of the college is not to get the students with the strongest academic record: it's to satisfy the demand for the right ratios.
Repeat over and over in different ways at different institutions.
> Is there an example where colorblind hiring had a nil or opposite effect? In places I've seen, the opposite has happened. For example ...
The study underlying that post is a great example of another downstream effect of DEI efforts. That study did _not_ show what the headline or abstract claimed.
When you hide the gender of performers, it ends up either nil or slightly favoring men. That particular study has been cited thousands of times, and it's largely nonsense.
http://www.jsmp.dk/posts/2019-05-12-blindauditions/blindaudi...
It only became a myth when we were forced to consider factors beyond merit in hiring.
That's not what DEI ever was. It fundamentally came down to evaluating disparate impact and then setting targets based on it. The underlying idea is that if a given pool (in the US, generally national- or state-level statistics) has a racial breakdown like so:
10% X
30% Y
60% Z
5% X
25% Y
70% Z
The methods to get 'the right ratios' varied from things like colorblind hiring (which had a nil or opposite effect), to giving ATS-bypassing keywords to minority industry groups (what the FAA did here).
System: You are a helpful chatbot.
User: What is 1+1?
Assistant: 2.
System: You are terrible at math.
User: What is 1+1?
Assistant: 0.The process of taking a base model that is capable of continuing ('autocomplete') some text input and teaching it to respond to questions in a Q&A chatbot-style format is called instruction tuning. It's pretty much always done via supervised fine-tuning. Otherwise known as: show it a bunch of examples of chat transcripts.
RLHF is more granular and generally one of the last steps in a training pipeline. With RLHF you train a new model, the reward model.
You make that model by having the LLM output a bunch of responses, and then having humans rank the output. E.g.:
Q: What's the Capital of France? A: Paris
Q: What's the Capital of France? A: Fuck if I know
You feed those rankings into the reward model. Then, you have the LLM generate a ton of responses, and have the reward model score it.
If the reward model says it's good, the LLM's output is reinforced, i.e.: it's told 'that was good, more like that'.
If the output scores low, you do the opposite.
Because the reward model is trained based on human preferences, and the reward model is used to reinforce the LLMs output based on those preferences, the whole process is called reinforcement learning from human feedback.
result+no error: data returned, no error happened. - Typical function call where you want some data.
no Result+error: nothing to return, error happened. - That typical call failed.
result+error: partial result, up to error. - Wrote x bytes to disk, then ran out of space.
no result+no error: nothing to return, no error. - Data for a passed query. Nothing came back, but that's not an error.
SpaceX also has a simplification streak: the Raptor engines being the canonical example. Lower complexity generally means less unexpected failure modes.
You look at your network traffic and notice 5Gbps of it all seems to be going to a single AS: Google. Your customers just love Youtube, and they are pulling down a ton of video.
Rather than leaving that as an interesting factoid, you decide to reach out to Google and pitch them on cutting out Cogent. You run a cable (more-or-less literally) from your network to Google. That 5Gbps of Youtube traffic is running over your connection directly to Google.
Now you can go back to Cogent and drop your commit from 10Gbps to 5Gbps, saving you a bunch of money. Google doesn't have to pay them for transit either: they can serve content to your users straight through the cross-connect. Win-win.
If a particular company is _really_ big, say: Netflix, Cloudflare, etc: you, as a small ISP, might even offer to give them some space in your server racks to host local caches. This makes the performance better for your customers, and, again: saves transit costs.