echen

10 ms·

61.
by echen 4y ago
Note that it's often difficult to tell if these signals are good or bad. For example: 1. Clicking on a link is often a negative signal. If you're searching for "when was barack obama born?", hopefully you get the answer
62.
Chronicles of Opt-175B Training(github.com)
1 points by echen 4y ago|0 comments
63.
Anthropic Raises Series B to Build Steerable, Interpretable, Robust AI Systems(anthropic.com)
3 points by echen 4y ago|1 comments
64.
Nat (Unit)(en.wikipedia.org)
1 points by echen 4y ago|0 comments
65.
Risks from Learned Optimization in Advanced Machine Learning Systems(arxiv.org)
2 points by echen 4y ago|0 comments
66.
by echen 5y ago
(Post author here.) Agree with both you and the parent here! We work a lot in the NLP and Trust & Safety space, and many of the models and datasets we see do ignore context -- and so real-world "toxicity models often end up simply
67.
Lessons Learned on Language Model Safety and Misuse(openai.com)
11 points by echen 5y ago|1 comments
68.
Fine-Tuning OpenAI's GPT-3 (Weights and Biases)(youtube.com)
2 points by echen 5y ago|0 comments
69.
by echen 5y ago
I used to work at Facebook, YouTube, and Twitter. One of the big questions I focused on: what was the right objective function to align our AI systems towards? When we started optimizing for watch time at YouTube, for example, our algorithm
70.
What if social media optimized for human values? A Facebook case study(surgehq.ai)
12 points by echen 5y ago|1 comments
71.
by echen 5y ago
> For starters, you can't unless you actually include the context in the training set. 100%. It's crazy how often context is missing from datasets! We wrote a separate blog post recently about this too: https://www.s
72.
by echen 5y ago
Yeah, Google and a couple other companies often hire "Analytical Linguists" as their labelers, or to help write their guidelines and manage labeling projects. Although -- and I say this having done a lot of my graduate coursework
73.
by echen 5y ago
Exactly. This is why it's important not just to have language skills when creating these kinds of datasets, but also cultural knowledge and context. For example, to pick a bit on the Google Emotions dataset again, it's difficult t
74.
by echen 5y ago
Yeah, we love what Jigsaw's building! This is all with the hope of improvement and collaboration. We deal with these hairy problems a lot too. Even before ML proper, getting the definitions right is very tricky. Should a comment that&#
75.
by echen 5y ago
Yeah, I think this is part of the problem. Is large-scale, low-quality data good? Sometimes it is (depending on the tradeoff), but from a model performance perspective, it's often more effective to get smaller amounts of higher-quality
76.
by echen 5y ago
One of the problems with real world machine learning is that engineers often treat models as pure black boxes to be optimized, ignoring the datasets behind them. I've often worked with ML engineers who can't give you any examples
77.
Are popular toxicity models simply profanity detectors?(surgehq.ai)
183 points by echen 5y ago|211 comments
78.
by echen 5y ago
Putting "series b" in quotes doesn't actually help! https://imgur.com/a/zNNTK3d And "series b" is a common enough phrase that I wouldn't guess you should have to.
79.
by echen 5y ago
Yeah, when I was at Google, the SERP (and caring about users) was a shrine at all levels. Here's an example of one of the weekly Search meetings AlbertCory is talking about: https://search.googleblog.com/2012/03&#x
80.
by echen 5y ago
There's been a lot of HN discussion [1][2][3] about Google Search recently, and whether it’s gone downhill. I used to work on Search and Search Measurement at YouTube, Twitter, and Microsoft, so I thought it would be fun to move beyond
81.
Is Google Search Deteriorating? Measuring Google's Search Quality in 2022(surgehq.ai)
470 points by echen 5y ago|414 comments
82.
by echen 14y ago
It was a tongue-in-cheek comment :) -- sad from the perspective of me trying to find a cute pattern for Berkeley.