I had a similar experience with tado that tried to paywall a basic temperature control of a smart thermostat. Their ratings got trashed and they pulled back.
It’s a feeling that I got from reading his stuff over the years. Asking the AI just now comes up with a number of examples where he himself admitted to it. I also found him increasingly disingenuous.
My interpretation of this is something like: if LLMs are to be mega useful token counts need to increase by many orders of magnitude -> broad adoption (and spending) would require token costs to drop by many orders of magnitude -> before the common folk get mega useful tools existing GPUs will be worthless
OpenAI were initially competing with Anthropic in $ value of tokens, and then number of tokens, and then the 5h limit removal meant you could buy multiple plans for huge continuous usage. This gravy train was never sustainable from a margin perspective and must have cost them billions.
It’s now going to be interesting how anthropic react - 5h limit and retain 15-20x $ value, or go the same was as OpenAI.
Anthropic have the revenue to subsidise longer. OpenAI is lagging behind Anthropic on the API revenue which is where Anthropic is making the majority of its money. I expect Anthropic will smell blood in the water and double down on their subsidies.
Yes very possibly. If you assume your competition will always exist then I think it makes sense to increase prices as they do even if you have the best business model.
I think on the app you should have an auto sort for your local supermarket. I have a list of aisles that I go through on my route around and I ask an ai to sort my shopping list by that route through the supermarket.
You can sort the aisles manually in the app, and there's a (hacky) route to use AI to categorise new items into an aisle.
It shouldn't be too hard to use AI to sort the aisles I'd guess
ive heard but not confirmed that some store chains have apis to tell you what asile to find something on, but i assume that would probably require upc info which would be an annoying extra level of specificity to need to track and manage. you just wanna write "corn chips" not... "mission brand corn chips in the 64oz bag"
LVT and things like market value based taxes are all far more complicated than is what already works, and what Singapore does (and a couple other countries like China): all land owned by the state, sold to private individuals with [99] year leases.
So the revenue comes through in resales once leases end.
Super simple: no valuation (only market sales); no annual expenses to consider (all upfront); development issues are simply a matter for leases to expire; speculation is capped given the capped lease length (although it can still get to stupid valuations).
Implementation is also extremely simple if you’re already onboard with taxation as expropriation: the state just legislates that all property is now a 99 year lease and they own it at that point (alternatively they compensate you 2-5% of market value, and retain existing annual taxes, or cut existing annual taxes to compensate, or they just do it without compensation!)
Unless the state is renewing the lease every time you do non-trivial development work or actually undertaking the development work itself as in Singapore, short leases are a rather significant deterrent from making optimal use of the land. If you think developers hold back from developing land because it might be worth more developed differently in ten years' time, wait to see what incentive "the state will repossess it and everything you've built on it in ten years' time" has...
Well yes, in reality you want a system where you don’t wait until the lease expires because that would leave many short lease properties without investment. It is trivial to value (competitive real estate market, rational negotiation, market interest rates) a lease extension from 30 to 99 years (enabling development) and the state shouldn’t really care about it unless it wanted to take ownership early as part of a wider land development plan.
The state is technically repossessing it when you own a lease, but it’s not something you had rights to anyway, or had paid for. The terms of what you own are up to the lease end date, under the terms of the lease.
I agree this seems like the most realistic way to implement an lvt. A pure lvt would need to have tons of loopholes to convince people to build industry in low value places. If a builder thinks their taxes will skyrocket in the future they will go overseas. The yearly evaluation is ripe for corruption(my city has tons of corruption in property taxes evaluations).
What would be a revolution in excel is if we could have probability distributions in cells. Right now all spreadsheets are fairly deterministic and we think that tells us something about the real world. Having an ability to make cells distributions (and then all resultant cells would be distributions too) would represent the uncertainty in the world better (and expose lots of interesting things like distributions of tails)
Guesstimate (https://getguesstimate.com) has been around for years. I assumed Excel and Google would incorporate its features quickly, but it never seems to have taken off.
Features:
- All cells are named (auto named if you don't set them)
- Cells can have single values, many types of probability distribution, or sample data
- Formulas also produce distributions as first class outputs, because calculation is a Monte Carlo sim
- Each cell also has a pop-up box that invites you to explain your reasoning -- built in documentation
It got so much right. Sample data particularly useful because you can build a probabilistic bottom-up forecast and then update it with real-world data.
This is really amazing, and it reminds me of an Excel modification I did 24 years ago for an animation studio. They had around 3,000 digital artists and then about 200 support staff, and around 8 different productions moving through the studio at the same time. Each production was at some different stage of completion. The studio's production management staff spent a huge amount of time negotiating what staff was assigned to what productions, and the reasonings were all hazy and non-fact based. So I wrote what people would now call a deep learning system that mined all the accounting and production tracking records the studio had, and created probability distributions and confidence intervals for every individual and every task, and every combination of people and their combined tasks, and that all composited together into a giant predictive engine of real time whatif scenarios for the entire studio in different configurations.
It was used for a series of feature films you all probably saw: the Chronicles of Narnia series, the Scooby Doo live action CG films, the Garfield films with the CG Garfield, and Life of PI.
People both loved and hated it. They loved the insights into the studio's capabilities, and they hated thinking through how the thing operated, and they hated the thought process of constructing a new whatif scenario, because they had to THINK.
I'd say a broader "people has to WORK" about various accounting systems I developed. E.g. to make something warehousable, people need to enter various details about their products, and they hate it. And Amazon shows us that marketplace can be successful without even basic weight and dimensions.
Right, and that's why it doesn't really matter what innovative things don't get added to Excel because people wouldn't be using them.
Spreadsheets at their core are a visual intetaction model; a table with cells that can affect other cells. Beyond that core, you can do whatever you want - including with Excel, so the 10% that don't want to treated as a calculator can do so with a little bit of bespoke programming.
I've implemented several business applications where Excel was the user agent against specialized backend services. This was before the browser took over the front end in developer's mindset.
First thing that came to mind is: can this model correlation between each cell, or does it assume that all cells are independent?
For example, in this https://www.getguesstimate.com/models/2205 hours of sleep per night is perhaps correlated with hours per weekday commuting. You might not have the data, but if you do, it could improve the probability distribution in the "weekday hours occupied" cell
What would this look like? Whenever I need something like that, I just stick each parameter in its own cell and use functions from there.
I would love an easier way to communicate to tech/math-illiterate CEOs that worst case, average case, and best case are different things, but must all be planned for.
When I did Process Improvement we used MiniTab for this.
I honestly was not a fan of that program and would have loved to be able to use Excel, but there are so many statistical things you just can't do in it.
As a number of the comments have suggested there are a number of Monte Carlo methods (including this addin) but I was hoping that someone would make one which calculated based on random variables and output a resultant random variable, over brute force monte carlo. Monte Carlo is also somewhat slow, but doing the maths analytically should be fast.
MC is used because there is no analytic solution. Only some likelihood functions have a conjugate prior. You have to rework the analytic solution for every model change (where most changes would result it in being non-analytic anyway), and they don't work well for high dimensional problems. Restricting the tool to analytic would mostly result in a toy app that just does Gaussians.
Check out @Risk and crystal ball - both have been around for a long time.
I should also say what’s really nice is you can define the shape of the probability distribution and parameters for random variables, as well as the correlation of different random variables. Both are effective but also feel dated these days.
> This library solves that problem. It registers xarray data structures as custom JAX PyTrees. This allows JAX to seamlessly flatten xarray objects into their raw arrays for accelerated computation and then unflatten the results back into fully labeled xarray objects, preserving critical metadata like dimension names and coordinates.
flatten and unflatten with datatypes is necessary for unrolling loops for performance.
Which is the correct logic for probabilistic logic, for expressions with frequentist or symbolic distributions as values? Are quantum logic and quantum statistical mechanics the appropriate or useful tools for all probabilistic logic?
dist_a1 <operator> dist_a2
uncertainties does mean±dev in Python with numpy types.
> W3C CSVW supports per-column schema. ( with URIs for datatypes )
> Serialize a dict containing a value with uncertainties and/or Pint (or astropy.units) and complex values to JSON, then read it from JSON back to the same types. Handle datetimes, complex values, and categoricals
IEEE-754 specifies NaN (null), ±0, three infinities (positive, negative, and unsigned), but IEEE-754 does not specify a representation for categoricals, datetimes (like ISO8601), or complex numbers.
XSD (XML Schema Datatypes), which RDFS vocabularies often use to specify the rdfs:range of an rdfs:Property, does not specify how to specify abstract complex numbers; but OpenMath RDF, and QUDT (Quantities, Units, Dimensions, and Types) and OM Ontology all have a way to save complex numbers to disk, too.
The matrix product is not something specific to matrices, but it is a particular case of the contracted tensor product.
(While now almost everybody uses the names "tensor product" and "contracted tensor product", historically these names are wrong. Better names, which do not misuse the word "tensor", would have been "Zehfuss product" and "Cayley product", after the names of the mathematicians who have introduced them.)
The word "tensor" has become increasingly corrupted over the years. Originally, it had a meaning that matched its etymology, i.e. "tensor" meant "symmetric matrix", because a matrix could be decomposed through the polar decomposition into a tensor and a versor (i.e. a symmetric matrix and a rotation matrix).
Then what is now called tensor theory was created by Ricci, but without ever using the word "tensor". The meaning of the word "tensor" was changed by Einstein in his theory of the general relativity, where for unknown reasons he chose this word to refer to arrays with an arbitrary number of dimensions that satisfy a certain kind of transformation relationships when the systems of coordinates is changed.
After WWII, due to the huge popularity of Einstein, this weird usage of the word has become widespread.
Then for many decades tensors were just a subset of arrays, those that are changed in a certain specific way when the system of coordinates is changed.
Eventually, there was the second corruption of the word "tensor" which is completely unjustified, when "tensor" has become just a synonym for "array".
Vectors are 1-dimensional arrays (though originally, like tensors, vectors were only a subset of 1-dimensional arrays), matrices are 2-dimensional arrays and nowadays "tensors" means just "arrays", i.e. of an arbitrary dimension.
The tensor product and the contracted tensor product are defined for arrays of arbitrary dimensions. For 2-dimensional arrays the contracted tensor product is a.k.a. matrix product, and for 1-dimensional arrays it is a.k.a. scalar product.
TIL MLIR supports sparse tensors with sparse_tensor.encoding = {
dense, compressed, specialized structures for hypersparse regions, and hardware-specific sparse constraints, such as NVIDIA's 2:4 structured sparsity layout }
But it looks like [MLIR and all other implementations of] SIMD only accept vectors; so there can't be Zero-Copy there because the tensor must (?) be copied to a vector to pass to a SIMD e.g. matmul routine, and then the resultant vector must be copied back into a tensor only if there are subsequent references to the complete tensor instead of just a slice?
FWIU, AFAICS, GPUs are designed for 3x3 tensors (and affine transformation to 2D) but for greater degrees like for 4x4 tensors (e.g. for SQG) you must implement shaders?
From https://news.ycombinator.com/item?id=49855122 re: gravity from QED without any GR spacetime curvature, using the ~amplituhedron to solve n-body gravity with scattering amplitudes, and SQG/DDF:
> normed tensor gaussian splatters work well as a simulation primitive for this too because they do conservation.
This has been my dream for decades. I even pestered the excel team when I worked at MS. Everyone thought it would be cool but not worth the effort and complexity.
The main one I was thinking about was in finance when doing lending to corporates. You build a financial model of the company, and then you “stress test” it to see where the company struggles to pay you back by changing a few variables up or down. If it was distributions you’d simply assign distributions to those variables and look at the distribution of outcomes for your loan.
There is not that much difference between the state and a mafia. The fact that the mafia has successfully managed to keep out fentanyl shows just how much democratically elected states are failing their citizens. This outcome is not surprising.
The government can’t regulate a black market, but they could stop the charade of illegal drugs, convert the black market into a regular market, and regulate that. Suddenly you get less fentanyl overdoses, better outcomes for addicts (they're not criminals anymore), AND you remove the power of drug cartels (they'll still exist, but will compete with each other through less murder-y means, like any other capitalist company). Plus, stabilize a lot of third world countries where illegal drug production and export is a big part of the economy, which is what allows the cartels to fight and corrupt the state.
Or, just keep doing what they're doing. It's bound to work eventually, right? Drug cartels are definitely growing less powerful every year, the war of drugs is being won, rate of addiction is going down.
Switzerland has prescription heroin to provide a safe supply to people addicted, as Diaphin (diacetylmorphine) 200mg. This dosage is huge and lethal for for someone without a tolerance.
This is a bit of a straw man “do the ends justify the means” statement, with hints of anarchism read between the lines.
Western democracies haven’t really prioritized “war on drugs” (for lack of a better term) over the last probably 10ish years. Yeah there’s still funding and initiatives and some lip service.
Instead the focus has been largely infighting and major political factions fighting for political power in the next generation (“tech bros” political class for one)
So it’s no surprise state-sponsored (and just demand driven) drug trafficking into the US hasn’t received much action.
There is some hope though. With rare earth mineral access being a major national security priority, and many cartels controlling those areas in SA, the converging interests may produce some net positive. Background: https://youtu.be/q_n-74ifBZs?is=nw5g-mmPmKphGH2W
A war on drugs can only be won through heavy penalties on the users and the political classes refuse to come down hard on the users (themselves, their children) but rather come down hard on the dealers (poor people) or traffickers (foreigners). For this reason we never really had a war on drugs, just lip service towards one, hence a failure of government.
This is a false equivalence, which means equating two things based on surface similarities while ignoring key differences.
The Mafia keeping fentanyl out isn't a public service. It's a cartel protecting its own drug market, which says nothing about whether democratic states ( globally ) are failing their citizens.
https://www.trustpilot.com/review/www.withings.com
I had a similar experience with tado that tried to paywall a basic temperature control of a smart thermostat. Their ratings got trashed and they pulled back.
reply