> Not having access to something does not grant you rights to it.
> You probably know a lot of things I don't, and some things that I can never know. That does not allow me to compel you to tell them to me.
This is such a obtuse way to frame the comment you were responding to. It's obvious they were just talking about the right to know about labor demand price ranges. Do you really think they were advocating for the right to know everything about everything?
We must have a different understanding of the English language. OP's sibling comment clarifies your original misunderstanding by saying
> The California law I referenced grants me rights to the pay range information.
You read that and somehow thought, _for the second time_, "I get it. OP wants to know everything about everything." Do you see the disconnect? OP is specifically talking about pay range information. You're just making up an obviously outrageous point that no one is advocating for and arguing against that made-up thing.
What you're saying was true for primitive LLMs from a couple years ago, but in my experience, any of the advanced LLMs today (GPT-4, Claude, etc.) have no issue with this. I'm open to changing my mind if you can provide any examples of GPT-4 failing at this task.
England/London is a bad example, but this one (based on the paper) is easy to replicate with GPT4 and other top models:
> Who is Mary Lee Pfeiffer's son?
>> Mary Lee Pfieffer's son is Sean Pfieffer.
> Who is Tom Cruise's mother?
>> Tom Cruise's mother is Mary Lee Pfeiffer.
> Who is Mary Lee Pfeiffer's son?
>> Mary Lee Pfeiffer's son is the actor Tom Cruise.
Of course the second time around (with models that keep context) it "realizes" its mistake, sometimes apologizing. I've tried this one tens of times and never seen it get Tom Cruise, but it will often pick another random celebrity like Harrison Ford or Eddie Van Halen.
That was a nice read, but it looks like the article concludes that the "Reversal Curse" observed by the authors of the paper is likely better attributed to the researchers' methodology. Some quotes from that article:
"As mentioned before, It’s important to keep in mind ChatGPT and GPT-4 can do B is A reasoning. The researchers don’t dispute that."
"So in summation: I don’t think any of the examples the authors provided are proof of a Reversal Curse and we haven’t observed a “failure of logical deduction.” Simpler explanations are more explanatory: imprecise prompts, underrepresented data and fine-tuning errors."
"Since the main claim of the paper is “LLMs trained on “A is B” fail to learn “B is A”“, I think it’s safe to say that’s not true of the GPT-3.5-Turbo model we fine-tuned."
What LLMs can do efficiently is crawl through and identify the secondary forms of evidence you mentioned. The real power behind retrieval architectures with LLMs is not the summarization part- the power comes from automating the retrieval of relevant documents from arbitrarily large corpuses which weren't included in the training set.
What makes a document relevant or not? Provenance? Certain keywords? A lot of this retrieval people cite that llms are good at can be done with existing search algorithms too. These are imo nicer because they will at least provide a score for the fit of the given document to the term.
I'm no biology expert but had to study some of this for my robotics degree not so long ago.
"Sight"
split into rods for brightness sensitivity, and cones, each of which is deicated to one out of red, green, and blue. green is wider gamut of color than the others because there is a lot of green in nature. These sensors are fully independant of each other for the most part, although there is minor overlap between cones which is what we call other colors (yellow etc)
"Taste"
Again split into different specialised papillae sensors. I dont remember so well, but its something like foliate for sour sensing, fungiform for salty, and vallate for bitter/poison. There is also sweet I dont remember the name, and some argue for umami
"Touch"
There are an ungodly number of very distinct senses that go into touch. From more abstract ones like pain, heat/cold, moisture (not evenly distributed around body, for example have to touch things to lips to distinguish cold from wet), proprioception for joints (arguably an independant sense for each joint, or at least each "kind" of joint, because the biological mechanism is different for ball joints to saddle joints etc as well as specialised proprioception for eyeballs, tongue etc)
Then in actual touch touch there is Ruffini corpuscles sensing skin stretching and slippage of objects past the skin
Merkel discs, which senses pressure applied to the skin and low frequency vibration
Meissner's corpuscles, which sense vibrations in middle range. They are very sensitive and allow very slight sensing of tiny impulses such as picking up an insect's wing
Pacinian corpuscle sense extremely fast vibration which among other things allow the distinction between "rough" and "smooth" surfaces (by mechanical movement causing vibration)
There are also free nerve endings sensing stuff like itching and bruising.
Hair foillicles also sense movement and stretching of the hair they are attached too, which provides more touch data. Incidentally this mechanism is also used for balance and hearing via really complicated interactions of tiny hairs in the ear.
"Smell"
Smell is fiendishly complex, it actually is more akin to the way antibodies in the body are made in the sense it consists of thousands (and millions) of specialised sensors made to "fit" and attach to individual compounds, so there are almost limitless individual senses of smell
There is also a whole lot of internal sensor data for things like breathing (you know when you are short of breath), digestion you know when you are full, or when you are craving one of a number of things sweet salty etc), bladder control.
This is mostly off the top of my head and i'm certain i'm misremembering some of the subtlties and a whole bunch more senses both obscure and immediately recognisable ones to any owner of a human body
This is super interesting, and I appreciate the level of detail and thought that went into your response. Some I'm willing to accept, like hot/cold being distinct from pressure being distinct from pain. (Spinal cord injury, for instance, can impair pressure perception in a particular part of the body without affecting hot/cold. And lumping joint pain in with "touch" is just silly.)
On the other hand, in the context of the discussion, it's hard to support the argument that you can count each colour channel separately just because the biological mechanics differ. You can't actually triple the amount of human-perceptable information by going from a monochrome to full colour display.
The point remains that we've plucked the low-hanging fruit when it comes to high-bandwidth human senses (or meta-senses if you insist on being pedantic). No one will buy a PUI (pain user interface).
absolubtely! sight is an amazingly high-bandwidth sense, as is hearing.
Other types of interfaces do exist, for example ive worked with vibration motor arrays placed on the skin for various purposes such as assisting in guiding the arm of a patient to target a specific point (vibrate on side closest to target) etc. We also worked with pads of electrical patches that pass small currents through the skin to produce a distinct sensation, like pain but barely at the threshold of being noticable. These were used for first responders, placed along the side of the torso underneath the clothes with flat profile, allowing them to have handsfree silent communication with low bandwidth. Something like "up up left left" being pre-agreed to mean leave the structure now etc. Another fun one I wanted to mention is in-mouth joysticks controlled with the tongue for quadreplegic patients to allow them to move a wheelchair or robot arm to regain some small independance (might seem like it would be uselessly hard to achieve anything with an arm controlled that way but the emotional impact of independance can't be understated for such people, even a simple task can be very meaningful)
They won't be as good as screens or audio unfortunately. But they can exist. Even braille screens and keyboards exist as a nice product and are reasonably high bandwidth.
> On the other hand, in the context of the discussion, it's hard to support the argument that you can count each colour channel separately just because the biological mechanics differ. You can't actually triple the amount of human-perceptable information by going from a monochrome to full colour display.
You absolutely LOSE perceptible information when you lose one of then channels, like in color blindness.
This was a great read. I especially liked how you can read the presentation below with the slides if you'd rather do that than watch the video. No issues with video, but sometimes I'm in the mood to read, and this was very satisfying to be able to do here.
That really isn’t true either. Homelessness doesn’t cause drug abuse, there is plenty of drug abuse among the housed, it isn’t weird that some of them eventually lose their support network. There are definitely economic homeless in Seattle, they just aren’t as visible as the fent addicts.
This is an excellent documentary. Skipping ahead to the “Discussion” section and watching through to the end gives a good summary, and it certainly opened my mind. Thanks for sharing, I’m probably going to go back and watch the whole thing.
> You probably know a lot of things I don't, and some things that I can never know. That does not allow me to compel you to tell them to me.
This is such a obtuse way to frame the comment you were responding to. It's obvious they were just talking about the right to know about labor demand price ranges. Do you really think they were advocating for the right to know everything about everything?