saeranv

Interior ceilings are often built as modular panels for this reason, and sometimes even floors. But walls, especially exterior walls, are engineered to serve multiple functions, such as acting as a moisture, thermal, and acoustic barrier, in addition to carrying utilities and structural loads. All of this is best served by arranging, and sealing different material layers carefully within the enclosure. Within the ruthless logic of cost optimization, incorporating modularity conflicts with building-code-regulated neccessities, and will only be used by some small percent of people, for some small percent of time.

I used to work at an research/architecture (buildings, not software) firm that was a leader in carbon emission estimation. To estimate the carbon emissions from building operational energy, what we did was find the rate of carbon emission at the local grid[1] and then multiply it by the predicted operational energy from an energy model. For example, for a single family home drawing from a blended grid mix of coal, gas, and nuclear, you might have a carbon emission rate of 150 kgC02/MBtu for 0.04 kBtu/ft2 of operational energy, over the course of an hour. This works out to:

    150 kgC02/MBtu * 4E-5 MBtu/ft2 = 0.006 kgC02/ft2

You then integrate this over the course of the year. This gets complicated as there are variations in the resolution of carbon emissions from different grids, the grid emission rates, and which grid the building uses.

To estimate server carbon emission rate, you just have to swap the building energy use intensity for server energy use, and everything else should be the same.

[1] Grid Emission Rate Sources EPA: https://www.epa.gov/egrid/data-explorer WattTime: https://www.watttime.org/

Ah, fair enough, I misunderstood. Interesting point.

Not sure what you're talking about... I'm the person who submitted this, and I have no idea what archive.today is.

I found this referenced in Politico's Global Insider columns[1] and thought the use of tech to aid public opinion manipulation would be relevant to the HN crowd (i.e. the use of StyleGAN to create authentic user profiles).

[1] https://www.politico.com/newsletters/global-insider

Honestly, Andreesen's understanding of the underlying cause for new housing is really poor, I'm not surprised that he is also a hypocrite, or his ignorance has lead him, perhaps unintentionally, to a hypocritical act.

This probably sounds harsh, but if anyone doubts me, I urge you to actually read Andreessen's “It’s Time to Build”. It's shockingly naive. An entire essay written to convince us of "the need to build", without the awareness that the obstacle to building is not the lack of a SV visionary pointing out the obvious, but the tangle of policy and political incentives. Read Ezra Klein's rebuttal for a more comprehensive review of why Andreesen's essay misses the mark completely[1].

I do want to highlight one positive attribute about him though. He actually had the courage, and intellectual sensibility to write out his argument and publish it in a public forum for others to read. That's the first step towards getting your ideas challenged, and, in the back-and-forth of criticism, strengthening your ideas. As far as I know he didn't engage with the more robust criticism of his ideas (i.e. Ezra Klein) and so that second step is missing, but I at least admire that he took the first step. It's frustrating to me how much society idolizes wealthy "visionaries" who don't actively publish or debate their ideas, like Musk, or Thiel (with regards to his political activity), giving them the benefit of doubt that there's nuance and intentionality behind their actions without any evidence of such.

[1] https://www.vox.com/2020/4/22/21228469/marc-andreessen-build...

That makes sense. Since you seem to have expertise in this domain, may I ask what is the current state-of-the-art for generative models for modeling learning/behaviour, in the neuroscience community? If any...

Thank you, that's a very clever explaination. Even though I can't grok the mathematics off the top of my head (I would have to look up the conjugate gaussian-gaussian formula and work this out) I think converting this to a posterior probability calculation is much more intuitive, then the article's explaination.

So, my crude understanding is that therefore the signal and noise is the mean and variance of the Gaussian, respectively. The posterior uncertainty in this is case is thus a function of a weighted average of the sample data, weighted by the priors, normalized by the data, which results in a higher signal-to-noise ratio, or lower uncertainty.

> This latter property is particularly useful for enhancing the precision of information processing. For example, information represented by an individual neuron may be noisy (say, with a precision of 1 in 100). By taking the average of input from 100 neurons carrying the same information, the common downstream partner neuron can represent the information with much higher precision (about 1 in 1,000 in this case).

What does precision mean here, mathematically? Precision as in precision/recall doesn't seem to work, as precision in that context means TP / (TP + FP), and going from 1/100 to 1/1000 would mean worse precision.

Does it mean something like floating point precision, where increasing bits can increase information represented, i.e. 2 bits represent max of 11=3, 3 bits represent maximum of 111=7 and so on? I can't quite this to work...

> Another salient property of the brain ... is that the connection strengths between neurons can be modified in response to activity and experience—a process that is widely believed by neuroscientists to be the basis for learning and memory. Repetitive training enables the neuronal circuits to become better configured for the tasks being performed, resulting in greatly improved speed and precision.

This seems like a clear description of a bayesian process. The new connections are a posterior updated from the likelihood informed from input sensory data. But the fact that they don't describe it as such makes me wonder if there's other reasonable ways to model this learning behaviour.

I am familiar with Karl Friston's work (as a layperson) of animal behaviour and learning as a form of bayesian update via variational inference. Does anyone know if his ideas, or the general idea of learning as a bayesian process well accepted in the neuroscience industry?

That's an interesting idea. Do we know what is different about the brains of geniuses that could be replicated by changing how the brain grows or consumes fuel in a normal pereson? I know descriptions of "high-IQ" and "normal" are extreme simplifications of a complex, multidimensional concepts, but I have to imagine someone like John Von Neumann who apparently never forgot anything he read, has had to have had something physically different in his brain.

For example, do we know if high IQ/abilities are associated with more neural connections in the same brain volume?

Other people have answered, but I think this is slightly more intuitive way to put it: Newton's law of cooling states that the rate of temperature reduction (heat loss) of a body is proportional to the temperature delta between it and its ambient surroundings. We can write this as:

dT/dt = kT

Where T is temperature delta, t is time, and k is some proportionality constant. So the rate of cooling is changing as the delta temperature reduces, specifically, it is getting exponentially reducing as it gets closer to thermodynamic equilibrium (where no heat is exchanged):

T(t) = Ce^kt C = T(0)

Which means there is a larger cool down with the hot water at first, but under this simple model, once the T(t) hits the same temperature of the cool water it's being compared to, it's cooling rate should be equivalent (and a lot slower).

Agreed, I think your points get at the right point. I would go further and say - we consider meaningful "intelligence" as having to do with "consciousness", and its absense causes the disconnect between Wolfram's definition and our intuition about human intelligence.

Defining consciousness is its own challenge, but one definition that I like (as a layperson) is that it consists of an approximate model of our self, that emerged as the brain recursively modeled itself while attempting to model other's behaviours, a critical trait once group co-operation and adversity became key to our survival.

This seems correct to me, a hurricane exhibits computational complexity, but it does not have a sense of self, and does not use that sense of self to make decisions.

Of course, that sense of self can also be broken down into manual computation, it's just a very complicated computation. For example, a bayesian interpretation of that "approximate model of our self" might define it as a generative model or joint distribution of actions and states, embedded within a markov decision process (MDP). Modeling, or speculating about your self would therefore consist of sequentially sampling from the joint distribution over several timesteps of the MDP.

The worst part about reading a Wolfram article on HN is having to read for the millionth time all the people complaining about Wolfram's ego, or his rambling. I understand, and agree (to a layerperson's extent) the criticism of his refusal to work within conventions for novel research.

Anyway, does anyone have a good rebuttal, elaboration on this idea:

> But actually I think there’s no lack of “alien intelligence”; indeed it’s all around us. But the point is that it really is alien. At an abstract computational level it’s like our intelligence. But in its details it’s not aligned with our intelligence. Abstractly it’s intelligence, but it’s not human-like intelligence. It’s alien intelligence.

There are a couple of comments making the point that this is reductionism, and is equivalent to saying any computational in theory exists in a dust cloud, but I'm not sure why we can't classify human intelligence/sentience/consciousness as simply a type of emergent computation in our neurons.

I've recently been fascinated by Karl Friston's work on the free energy principle, which models intelligent behaviour as a function of us trying to minimize the divergence between the entropy of our perception and a generative, probabilistic model of the world, and the entropy of the likelihood of evidence we collect about the world. From that perspective, at a high-level consiousness just reduces to a really, really, high-dimensional modeling of the world embedded in our neurons that is driven to minimize entropy.

Once a species evolves to have its own theory of its mind (which is how I think of sentience/consciousness), the entropy minimization gets really complicated, but ultimately still fundamentally a computational process.

I myself don't by this argument, completely, I feel like its missing something, but can't quite figure it out.

This definition clears up it up nicely for me:

"In signal processing, sampling is the reduction of a continuous-time signal to a discrete-time signal... A sample is a value of the signal at a point in time and/or space; this definition differs from the usage in statistics, which refers to a set of such values."

So it's a sample in the sense a set of (x, y, z, w) values from discrete timesteps represents a continuous timestep, but doesn't require specifying and using random variables to determine those timesteps.

I'm a little confused about why the curves are "sampled" to update the joint transformations: ``` So if we are rendering 60 frames per second, we are basically doing the following 60 times a second:

Sample all the curves you see on the graphs to get the orientation of each joint... ```

When the curves are "sampled", does that mean in a 60 fps animation, there are 60 timestep values t={t1, t2, ... t60} sampled from a uniform distributions, and evaluated at x(t), y(t), z(t), and w(t), and then those values are stepped through sequentially? If that is the case, why not just set t={1/60, 2/60, ... 1}?

Or does "sampling" here just mean function is evaluated at an appropriate t value, and not refering to statistical sampling?

I wouldn't reccomend Strang's "Introduction to Linear Algebra" textbook to a beginner. Strang has a very odd, dense way of writing, often with references to material that has yet to be introduced. I think this is a consequence of it's intended use as an aid to his lectures, and can't really stand on its own. The goodreads reviews on the textbook seems to share my opinion: https://www.goodreads.com/book/show/179700.Introduction_to_L...

I think it's great for an intermediate student, or someone who's also watching Strang's lectures.

Looking at the wikipedia derivation of the Kelly criterion, taking the logarithm of the expected rate of return E[r] = (1 + fb)^p (1 - fa)^(1-p) is just to make finding the derivative easier (because logs are monotonic). Nothing else.

What other antipatterns are there? I can understand the criticism of having configuration code with business logic, and think there should be better commenting/docstrings.

But other then that, thought it was fine. The use of type hints is pretty awesome, in particular. Do you guys just not like higher-order functions?

The argument actually understates the work Shaq put into becoming an elite player. Shaq's physical gifts obviously give him a huge advantage, but there are plenty of 7-footers in the NBA that don't do as well as him. Shaq learned actual skills (shot-making, agility, footwork) that other big men tend to ignore because they can get by on their size.

I'm always frustrated to see the pervasivness of this type of political interpretation, which constructs a false equivalence between two parties and therefore places the blame on both.

I think this kind of view is the default interpretation for a majority of people, since it is a tempting conclusion to draw for people who aren't well versed in the actual details they're evaluating: to the ignorant, it does look like both sides are just slinging mud at eachother, and it's inherent cynicism seems "realistic".

I think to solve this, we need to educate people on the idea of "false equivalance" (that just because two parties are engaged in conflict, does not mean they are equally to blame) and to force them to state the details of what exactly both parties are doing in order to hopefully push them to recognize there is a meaningful difference between actions.

Yes, you're correct. Jane Jacobs[1] in "The Death and Life of Great American Cities" wrote about how dense, walkable environments with storefronts, residents on porches contribute to the concept of friendly strangers that make areas safer. The isolation of the suburbs creates danger.

[1] Who unfortunately then proceeded to lock others out of also having this kind of quality of life by becoming a NIMBY...

Awesome, thanks for the calculations! It's cool that even with the massive transmission loss, the server itself is so low-energy it can happily survive with just 36 Wh a day.

One of the more compelling framings for democratic socialism to me, is the idea of taking profits from companies and using it to radically raise wages of people with "undesirable" jobs: janitors, garbage-men, warehouse workers. There's no market pressure to do so (not without unions!) but can be regulated.

Very cool project. I'm really intrigued by this idea of having websites reflecting geographic location, and even local weather conditions (with some lag). What's the peak and average power (energy/second) you get from your solar panel, and what's the power consumption of your server? Also what's the physical size of your solar panels?

My guess is that the server might use around the same power as a laptop (~50 W). Assuming a square meter patch of earth gets around 100 W (dependent on latitude/time of day), and assuming something like a 10% efficiency that works out to around 80 W (on average during sunlight hours)? So you're maybe getting 8 hours of server power per day from 10 hours of sunlight? Is that a reasonable ballpark estimate?

Yes, although I suppose even in languages where this isn't mitigated (i.e. Clojure example below), you can offset that inefficiency with the parallelism enabled by functional programming.

I've never understood how someone had the nerve to trademark "Bollinger bands", when all they are is just a measure of 2 standard deviations from a rolling mean of share price. It's like renaming 1+1=2 after yourself and and trademarking it's use.

Personally, I feel like the biggest problems with Python for math is the lack of a native vector datatype and our subsequent reliance on Numpy which really disrupts the elegance/terseness of working with vectors/matrices.

First, there's the constant inelegance/clutter/inefficiency of having to cast into and out of arrays and lists, even when doing basic list comprehension. R, Julia, and Matlab are all vector-based languages (I think), so you avoid having this casting as much.

Secondly, having a native vector type means you don't have to worry about the performance penalty of operating directly on arrays if an existing prebuilt method exists. Since the efficiency of Numpy comes from calling it's underlying C library, you're forced to memorize and use prebuilt Numpy functions rather then just use the more obvious and elegant array manipulations. For example rather than calculating the the cumulative sum like this:

  cumsum = reduce(lambda a, b: a+b, [1,2,3,4]) 

 We have do this: 

  cumsum = np.cumsum([1,2,3,4])

(There are better examples, but this is all I can think of right now).

And once you add something like Pytorch tensors on top of this, we now have an additionally layer of casting/redunduncy/memorization of prebuilt functions!

Can someone break this down for me? val speed = 5.4 m/s // Type = Float[m/s]

I understand that the speed variable is automatically assigned a type of Float[m/s] based on the m/s there, but I'm confused about how the units are just placed at the end of the value assignment: val speed = {value} {unit}

Is this just a F# property that units can be added at the end of value assignment, and F# interprets them as units correctly?

Also, anyone know of methods for incorporating units as types in Python? It would be great to have an equivalent method as demonstrated here, where dividing types automatically creates a new type unit, and perhaps even associates with a related type, i.e: type(kg / (m * m * m)) == density

Waste heat can be recovered by a heat pump in various ways, then recirculated either within a building (this is just typical HVAC heat recovery) or between buildings (district heating system) as supply system heating.

Are transformers competitive with (for example) CNNs on vision-related tasks when there's less data available? I'm not that familiar with "injecting inductive bias" via positional encodings, but it sounds really interesting. My crude understanding is that the positional encodings were used in the original Transformer architecture to encode the ordering of words for NLP. Are they more flexible then that? For example, can they be used to replicate the image-related inductive bias of CNNs and match CNN performance on small datasets (1000 - 10,000)?

If not, then to me, it seems like only industries where it's possible to get access to a large amount of representative data (i.e. greater than a million?) benefit from transformers. In industries where there are bottlenecks to data generation, there's a clear benefit in leveraging the inductive bias in other architectures, such as the various ways CNNs have biases towards image recognition.

I'm in an industry (building energy consumption prediction) where we can only generate around 10,000 to 100,000 datapoints (from simulation engines) for DL. Are transformers ever used with that scale of data?