> Finally, we use our dataset and LRE-estimating method to build a visualization tool we call an attribute lens. Instead of showing the next token distribution like Logit Lens (nostalgebraist, 2020) the attribute lens shows the object-token distribution at each layer for a given relation. This lets us visualize where and when the LM finishes retrieving knowledge about a specific relation, and can reveal the presence of knowledge about attributes even when that knowledge does not reach the output.
They're just looking at what lights up in the embedding when they feed something in, and whatever lights up is "knowing" about that topic. The function is an approximation they added on top of the model. It's important to not conflate this with the actual weights of the model.
You can't separate the hallucinations from the model -- they exist precisely because of the lossy compression.
Did the author even read the abstract let alone the paper?
> We also found that while risk for the six outcomes in our analysis combined was minimized at 0 g unprocessed red meat intake per day, the 95% uncertainty interval that incorporated between-study heterogeneity was very wide: from 0–200 g d−1. While there is some evidence that eating unprocessed red meat is associated with increased risk of disease incidence and mortality, it is weak and insufficient to make stronger or more conclusive recommendations.
Doesn't sound like it particularly slams years of shoddy research.
I guess you could say this, but given that the planet is essentially burning it seems like we're getting way ahead of ourselves with this type of "solution" when we could be working on more important problems.
Tablespoon of peanut butter really isn't any more work. It's probably less.
> How many beans equal a serving of meat?
> In general, 1 ounce of meat, poultry or fish, ¼ cup cooked beans, 1 egg, 1 tablespoon of peanut butter, or ½ ounce of nuts or seeds can be considered as 1 ounce-equivalent from the Protein Foods Group.[1]
People eat more meat than they need. Most people would be healthier if they consumed less of it or none and substituted vegetables/beans/nuts for those calories while still getting proper amounts of protein from non-meat sources.
Reducing meat consumption reduces a bunch of other consumption of resources letting people get their calories more efficiently.
> Meat is considered one of the prime factors contributing to the current biodiversity loss crisis.
> the livestock sector is a major stressor on many ecosystems and on the planet as a whole. Globally it is one of the largest sources of greenhouse gases (GHG) and one of the leading causal factors in the loss of biodiversity, and in developed and emerging countries it is perhaps the leading source of water pollution.
> farmers would reduce their land use of feed crops; currently representing about 75% of US land use, and would reduce the use of fertilizer due to the lower land areas and crop yields needed. A transition to a more plant based diet is also projected to improve health, which can lead to reductions in healthcare GHG emissions, currently standing at 8% of US emissions [1]
What's not to understand? Stop eating meat. We don't need lab-grown meat. We don't need meat substitutes.
Most americans would not meet suggested exercise levels for "relatively active" even if that was at the 30 min. per day level.
16% is 60% more than 10%. Most americans are consuming meat for at least 2 meals a day, possibly 3 for those who are eating sausage or bacon for breakfast.
The general point is that americans consume way more meat than what they actually need. This has been promoted heavily in the high protein diet fads recently. Educating people on what they actually need for a healthy diet would go a long way and cost a lot less than alternatives like lab-grown meat.
Most americans are consuming 100+ grams of protein a day (its very likely all of that coming from meat - I don't know anyone who consumes a lot of beans).[1] Average person needs about .36 grams of protein per pound of body weight.[2]
I wonder if lab-grown meat is also carcinogenic.[1]
Funds directed at educating the public on the amount of protein they actually need instead of high-protein diets would go a long way towards reducing energy consumption and emissions to help the climate. Lab-grown meat seems like the wrong direction when there are numerous good protein sources already.
> Getting people to stop beef in countries where it's not a problem is a waste
It's a problem. Most americans consume way more meat than what they could possibly need and they do it because they've heard they need lots of protein which could come from many other sources besides dead animals.
> If cultured protein is going to be even 10 percent of the world’s meat supply by 2030
Why do we need protein from meat? The money to engineer cultivated protein could be directed to educating simple americans on the fact that they're consuming way too much protein already. Most americans consume twice the daily protein they need.[1][2]
Plenty of brown rice, soy, nuts, beans etc for people to stop eating meat altogether if they actually needed more protein (they probably don't). ¯\_(ツ)_/¯
Maybe we should direct funds towards advertising meat-free and the benefits of not consuming carcinogenic meat instead of all of this meat replacement. [1]
Do AMP clicks count as no click since the user didn't leave google?
And how many are no click because google has presented semi-correct but not really correct information at the top of the results so the user finds what they think they're looking for without actually clicking?
Seems like google is doing just fine so I'd imagine both of the above scenarios account for a significant portion.
> Finally, we use our dataset and LRE-estimating method to build a visualization tool we call an attribute lens. Instead of showing the next token distribution like Logit Lens (nostalgebraist, 2020) the attribute lens shows the object-token distribution at each layer for a given relation. This lets us visualize where and when the LM finishes retrieving knowledge about a specific relation, and can reveal the presence of knowledge about attributes even when that knowledge does not reach the output.
They're just looking at what lights up in the embedding when they feed something in, and whatever lights up is "knowing" about that topic. The function is an approximation they added on top of the model. It's important to not conflate this with the actual weights of the model.
You can't separate the hallucinations from the model -- they exist precisely because of the lossy compression.