scellus

lol

Here's a rough breakdown from Claude:

"[...] psilocybin converts to psilocin in the body at roughly a 1:1 ratio by active effect [...]

Psilocybe cubensis (most common): Contains about 0.5-1.0% psilocybin by dry weight. Since psilocybin converts to psilocin in the body at roughly a 1:1 ratio by active effect, 30mg of psilocin would be equivalent to roughly 3-6 grams of dried P. cubensis.

Psilocybe semilanceata (liberty caps): Much more potent at 1-2% psilocybin content, so you'd need only about 1.5-3 grams dried.

Psilocybe azurescens: Even more potent at 1.5-2.5% psilocybin, requiring roughly 1-2 grams dried.

Important caveats:

- Individual mushrooms within the same species can vary by 3-5x in potency Growing conditions, harvesting time, and drying/storage methods all affect potency

- The caps are typically more potent than stems

- Fresh vs. dried makes a huge difference (fresh mushrooms are ~90% water)"

Have to note that the paper is from 2016; for those really interested, it's good to read recent review papers.

Yes, price elasticity increases the demand for software work in total, including LLMs and humans. But to me at least, it is not clear that humans will increase their total amount of work as LLMs obviously do. Is it possible that LLM coding grows faster than the total, so that the human piece of cake actually shrinks?

I agree. AI has made even mundane coding fun again, at least for a while. AI does a lot of the tedious work, but finding ways to make it maximally do it is challenging in a new way. New landscape of possibilities, innovation, tools, processes.

Google's is initialized with a gridded dataset, ERA5, from ECMWF. Using ERA5 is the current standard here, and ECMWF themselves build on that mostly now. Meanwhile, Aardvark tries to do the same directly from observations.

It's lower than many other medium-range AI forecasts, but note that those other models get state-of-the-art with pretty coarse grids, 0.5° or so. The point is that upper atmosphere and broad patterns are smooth, so with ML/AI they don't require high resolution (while simulating them with physical models does require). And at the forecast lag of say 5-10 days, all local detail is lost anyway, so what skill remains comes from broad patterns, in all models. (Some extra skill can be gained by running local models initialized with the broad patterns, for there are clear cases like mountains where fine resolution is useful.)

No, they say end-to-end, meaning they use raw obsevations. Most or all other medium-range models start with ERA5.

There's a paper from Norway that tried end-to-end, but their results were not spectacular. That's the aim of many though, including ECMWF. Note that ECMWF already has their AIFS in production, so AI weather prediction is pretty mainstream nowadays.

Google has a local nowcast model that uses raw observations, in production, but that's a different genre of forecasting than the medium-range models of Aardvark.

I'm 57, a data scientist and just can't keep my hands off concrete problems, which means I need to write code as well. Although I enjoy good modeling most, right now AI makes even mundane parts of the work fun again.

But ECMWF itself runs a diffusion model that is practically on par with ENS in accuracy. They also seem to collaborate closely.

ECMWF runs many such models at their site, a run two or four times per day, and they have verification statistics too, no need to doubt the accuracy.

The Google model is probably the best so far but ECMWF's own diffusion model was already on par with ENS and many point-forecast models (graph transformers, not diffusion) outperform state-of-the-art physical models.

What is missing is initialization directly from observations. All the best-performing models initialize from ERA5 or other reconstruction.

It's a medium-range global model, while Google Weather (which I don't have) is mostly about local short-range weather? But Google Weather is already based on an AI prediction on most cities: https://research.google/blog/metnet-3-a-state-of-the-art-neu...

Google says GenCast forecasts will later be available from them too.

Also ECMWF runs a very similar diffusion model, it's not operational but run a couple of times a day with results available on their graph site (and as data files too I guess): https://charts.ecmwf.int/

For someone living in Helsinki and currently working in steel industry, the title of the post was particularly hard to parse.

Yes, except that I'm not so sure there is a clear distinction between following general instructions and generating new heuristics. It's just a difference in the level of abstraction there, and probably not even that one in any discrete sense, more like a continuum.

(Current) models may of course lack sufficient training data to act on a metalevel enough ("be creative problem solvers"), or they may lack deep enough representations to efficiently act in a more creative way. (And those two may be more or less the same thing or not.)

In Finland, i get 5Mbps LTE uplink for EUR 4 per month, for a trailcam, with unlimited use (at least in principle). So $20 per month sounds expensive, but obviously there are places where one has no earthly LTE and then it could be justified.

In general, having low-bandwidth Starlink IoT connections globally accessible would be just great, I can see lots of usage.

Priors on parameters are not an issue. On models of scale, priors are just some computationally convenient shrinkage, and what works is found empirically and canonized into the practice; projecting prior knowledge of the problem at hand by parameter priors does not really happen except in some vague sense ("I think most predictors are irrelevant, so make it sparse by Cauchy/horseshoe/whatever").

The important thing in bayesian (statistical, ML) modelling in general is the ability to gain in flexibility and do model structures that otherwise would be hard or impossible: latent states, hierarchies, etc.

In bayesian NNs the main advantages would be around uncertainty quantification (UQ) and in finding good optima and partly to avoid overfitting. These do apply in some cases of simple NNs.

Mostly however, especially with larger conventional models (not speaking of normalizing flows and such here), using explicit bayes is not feasible. Instead, people use approximate point estimates with tricks:

(1) UQ has been taken care of by post-calibration. (2) Stochastic gradient actually searches for large posterior masses like a variational approximation would do, so it is kind of bayes. (3) And those priors: using dropout is commonplace, it has a bayesian interpretation, and L2 regularization aka gaussian priors are frequent too.

So bayes is there in practice, just not in a neat, pure form but as a collection of practical hacks.

Although not probable in this case as far as I know: if vegetable oils seem to cause problems for you, you should be aware of sitosterolemia (phytosterolemia), caused by rare mutations in genes ABCG5 and ABCG8. It needs to be homozygous for symptoms.

(I have it, discovered by a full-genome sequencing by myself, accidentally around the age of 50.)

Semantic leakage could be just weakness of the model, and related to claims that they don't _really_ reason. Maybe more training could help.

Or maybe it's a more fundamental weakness of the attention mechanism? (There are alternatives to that now.)

I generate or modify R and Python, and slightly prefer Claude currently. I haven't tested the o1 models properly though. By looking at evals, o1-mini should be the best coding model available. On the other hand most (but not all) of my use is close to googling, so not worth using a reasoning model.

That's interesting, zaps are a withdrawal syndrome of SSRIs (and duloxetine).

Same here, occasionally I find both the case and the ear pieces with empty batteries while they haven’t been nowhere near empty when I left them, usually previous day or evening.

No, Graphcast is a graph transformer trained on ERA5 weather reconstructions of the atmosphere, not a general time series prediction model. It by the way outperforms all traditional global point forecasts (non-ensembles), at least on predicting large-scale global patterns (Z500 and such, on the lag of 3–10 days or so). ECMWF has AIFS that is a derivate of Graphcast, they'll probably get it or something similar to production in a couple of years.

GraphCast, Pangu-Weather from Huawei, FourCastNet and EC's own AIFS are available on the ECMWF chart website https://charts.ecmwf.int, click "Machine learning models" on the left tab. (Clicking anything makes the URL very long.)

Some of these forecasts are also downloadable as data, but I don't know whether GraphCast is. Alternatively, if forecasts have a big economic value to you, loading latest ERA5 and the model code, and running it yourself should be relatively trivial? (I'm no expert on this, but I think that is ECMWF's aim, to distribute some of the models and initial states as easily runnable.)

Not yet, sorry. :) But see the Github repo and source code therein.

Price data is sparse, often from a small number of sales, and censored when number of sales is smaller than six. So the local price levels, and especially trends are not at all clear from the raw data.

The model has three-level geographical hierarchy from zip-code prefixes, and the local population density as a (hierarchical) covariate. There are also covariances for price level, trend and trend change, and residual model takes number of sales into account and adapts to outliers.

All this helps in getting a better idea of what really happens behind the more or less random individual sales.

http://louhos.github.io/figs/2015-05-07-asuntohintojen-muuto...

Thanks to the Stan team for making the modelling feasible! http://mc-stan.org