3 ms·
Ignoring the effects of volume is a common and fatal mistake. "This effect already exists" is not incompatible with "Multiplying the impact of this effect by or
by banannaise 4y ago
Ignoring the effects of volume is a common and fatal mistake. "This effect already exists" is not incompatible with "Multiplying the impact of this effect by orders of magnitude will cause it to become a massive problem."
Or, put another way: A slightly degraded signal-to-noise ratio isn't really a problem, but once it degrades far enough, the whole stream becomes impossible to decode, and worthless.
- falcor84 4y agoMy Information Theory is rusty (and wasn't too strong to begin with), but didn't Shannon prove that we can decode a signal given arbitrarily bad noise, but that the more noise, the lower the channel capacity?
- banannaise 4y agoYep, pretty much. But therein lies the problem: filtering out garbage writing is something done by humans, whose limited time, attention, and effort all make for a pretty tight bottleneck in the first place. Degrade that channel capacity, and our ability to find texts of value becomes pretty close to nil. edit: Also, when the noise comes in the form of competing signals rather than random noise, that becomes much, much more difficult, but we're also outside my area of expertise at this point.
- version_five 4y agoAs I understand it, LLMs use temperature scaling to optimize the variability in the tokens they generate so the output appears "good", somewhere between getting stuck always saying the same thing and saying nonsense. [0] I wonder if more output being used in the training data will result in a shift to higher temperatures over time to keep variety in the output. [0] Discussion the other day: https://news.ycombinator.com/item?id=35131112 https://news.ycombinator.com/item?id=35131112