The whole thing reads like great literature. There should be a scientific magazine for papers that can be understood by the motivated layman. I accidentally read the whole thing in one sitting.
For more of excellent writing read Derek Lowe's https://www.science.org/blogs/pipeline His "Things I Won't Work With" series is particularly entertaining whilst remaining firmly scientific.
You misspelled “terrifying” there. I’ve never practiced chemistry, just got my B.S. and went to med school, but simply reading the names of most of those compounds made my blood run cold.
We stop pretending that universal time exists and add location to every time stamp relative to some predefined clock at a known location. The issues of when something happened dissapear.
If you stop pretending that universal time exists, then each timezone would need its own NTP stratum 0 atomic clock., Higher strata NTP servers would have to be told which timezone you wanted, and only return you the time sourced from whatever country's national labs count for that timezone, and those labs would not be allowed to compare each others times.
Each clock would run at a slightly different rate, based on imperfections in the equipment, the altitude of the laboratory and the amount of nearby mass in its vicinity. National clocks would drift, relative to each other, and so would everybody's timestamps. CET would be perhaps +1:00:00.000002 ahead of GMT rather than being exactly +1:00.
You'd have to continually measure and publish this drift in some kind of timezone-to-timezone comparison service, so that people who make network connections across the world don't end up finding that packets appear to arrive before they've been sent (due to National Clock drifts)
It's an interesting thought experiment. But I prefer where all the worlds' labs work together to produce a consensus universal time, and we just add fixed political offsets to it.
I like the odea of reference frames. But it isn't just geographic. The unix clock on a starlink is doing double digit mach and goes horizon to horizon in minutes. GPS satellites have always had to account for relativity, so we could look to that system... Hmm, seems gps time doesn't add leap seconds, only using those already added when it started ;)
Sure, you can define the reference frames to allow any orbit on any body. Mars landers & orbiters have slightly different time due to general & special relativity than Earth's surface, and also don't need Earth days/months/years, for example.
The issue without outsourcing is that the benefits are widespread (lower prices!) but the drawbacks are concentrated (factory town is now a hellhole). And our political system is incapable of redistributing correctly even though the net effect is highly positive.
The seminal study on the topic is the "China shock" paper from Autor et Al.:
Strange, because put this way, it should be entirely positive - widespread benefits and concentrated drawbacks are what we want to happen, as it benefits more people and concentrated problems are much easier to manage. What's very bad is when benefits are concentrated (often in the hands of a small group), and drawbacks are widespread, and therefore near-impossible to manage. See e.g. pollution, emissions...
... and outsourcing. The benefits are concentrated: profits captured by the companies doing the outsourcing. Sure, they may sometimes trickle down to the consumer, but the costs - the distributed drawbacks - are inferior quality of goods, elimination of local jobs, high ecological footprint, abusive business practices, lack of effective customer support. And the extra magic here is, it spreads direct responsibility over national borders, so it's near-impossible to hold anyone to account.
I don’t know or understand this multiplier effect you’re referring to. If you’d like to persuade me (and I assume other readers) explaining your argument might be more effective. Instead I get a sense of “don’t argue against me” as opposed to “this is why I’m right”
Still have not answered my question. Why is lower prices better? Why would vastly higher consumption coupled with vastly decreased production be beneficial?
Similar to OP if information is low density like a legal contract I can do 1200wpm after a few hours of getting used to it. Daily normal is 600wpm, if the text is heavy going enough I have to drop it down to 100 wpm and put it on loop.
Like usual the limit isn't how fast human io is but how fast human processing works.
Yeah 600wpm is passive listening. 900-1200wpm is listening lecture on youtube at 3-4x speed. Skim listening for content I'm familiar with. Active listening for things I just want to speed through. It's context dependent, I find I can ramp up 600-1200 and get into flow state of listening.
>text is heavy going enough I have to drop it down to 100 wpm
What is heavy text for you? Like very dense technical text?
>put it on loop
I find this very helpful as well, but for content I consume, not very technical, I listen at ~600wpm and loop it multipe times. It's like listening a song to death. Engrain it on a vocal / story telling level.
E: semi related comment to a deleted comment about processing speed that I can no longer reply to. Posting here because related.
Some speech synthesis are much more intelligible at higher speeds, and aids processing at higher wpms. What I've been trying to find is the most intelligible speech synthesis voice for upper limit of concentrated/burst listening which for me is around 1200wpm / 4x speed, i.e. many have wierd audio artefacts past 3x. There's synthesis engines whose high speed intelligbility improves if text is processed with SSML markup to add longer pauses after punctuation. Just little tweaks that makes processing easier. Doesn't apply to all content, all contexts, but I think some consumption are suitable for that, and it's something that can be trained like many mental tasks, and dedicated speech synthesis like fancy sport equipments improve top end performance.
IMO also something neural model can be tuned for. There are some podcasters/audiobook narrators who are "easy" listening at 3x speed vs others because they just have better enunciation/cadence at same word density. Most voices out there from traditional SAPI models to neural are... very mid fast "narrators". Think need to bundle speech sythensis with content awareness - AI to filter content then synthesis speech that emphasis/slow on significant information, breeze past filler - just present information more efficiently for consumption.
>This is of course true, but there's something very broken with lisp: metaprogramming and programming aren't the same thing,
Metaprogramming and programming are the same thing. It's just that no language, including all lisp, (but hilariously not m4) get quotation wrong. Lisp gets around this with macros which let you ignore quotation and deal with meta language statements expressed as object language statements when they clearly should not be.
This issue stems from the fact space in the object and meta language is treated as the termination of an atom without distinction between the two.
>Cognition is different in that it uses an antisyntax that is fully postfix. This has similarities with concatenative programming languages
Postfix languages are a dual of prefix languages and suffer from the same issue. You either need to define the arity of all symbols ahead of time and not use higher order functions or you need a pair of delimiters which can serialise a tree. Relying on an implicit zeroth order stack solved the problem in the same way a lobotomy solves depression.
Thanks for your feedback! I'm not accusing you of not reading the full article, but you should if you haven't already. Also, we don't know the extent to which we made anything new; if you think you can do what we're doing in lisp in some way, you're free to prove us wrong.
> Relying on an implicit zeroth order stack solved the problem in the same way a lobotomy solves depression.
I mean, fair and flowery, but an implicit stack has been a thing ever since proto computers/calculators. Just like a lobotomy reduces the availability of higher order processing, so too does going back to the utmost primitives in calculating a string of instructions
Competent companies tend to put a lot of effort into building data analysis tools. There will often be A/B or QRT frameworks in place allowing deployment of two models, for example, the new deep learning model, and the old rule based system. By using the results from these experiments in conjunction with typical offline and online evaluation metrics one can begin to make statements about the impact of model performance on revenue. Naturally model performance is tracked through many offline and online metrics. So people can and do say things like "if this model is x% more accurate then that translates to $y million dollars in monthly revenue" with great confidence.
Lets call someone working at such a company Bob.
A restatement of your claim is that Bob decided to launch a model to live because of hype rather than because he could justify his promotion by pointing to the millions of dollars in increased revenue his switch produced. Bob of course did not make his decision based on hype. He made his decision because there were evaluation criteria in place for the launch. He was literally not allowed to launch things that didn't improve the system according to the evaluation criteria. As Bob didn't want to be fired for not doing anything at the company, he was forced to use a tool that worked to improve the evaluation according to the criteria that was specified. So he used the tool that worked. Hype might provide motivation to experiment, but it doesn't justify a launch.
I say this as someone whose literally seen transitions from decision trees to deep learning models on < 100 feature models which had multi-million dollar monthly revenue impacts.
Granted, however this approach does not require that constant-one input either.
> There isn't much difference between weights of a linear sum and coefficients of a function.
Yes, the trained function coefficients of this approach are the equivalent to the trained weights of MLP. Still this approach does not require the globally uniform activation function of MLP.
