dcl

thank you for this laugh

Wouldn't have even been a believable April fools joke even 5 years ago.

Search engines are still required for me. LLM's still get lots of very important things wrong.

Last night, I asked Claude 3.7 Sonnet to obtain historical gold prices in AUD and the ASX200 TR index values and plot the ratio of them, it got all of the tickers wrong - I had to google (it then got a bunch of other stuff wrong in the code).

Also yesterday, I was preparing a brief summary of forecasting metrics/measures for a stakeholder and it incorrectly described the properties of SMAPE (easily validated by checking Wikipedia).

I constantly have issues with my direct reports writing code using LLM's. They constantly hallucinate things for some of the SDK's we use.

But how good is it at Pokemon?

I like this bit:

> Personally, when I want to get a sense of capability improvements in the future, I'm going to be looking almost exclusively at benchmarks like Claude Plays Pokemon.

Definitely interested to see how the best models from Anthropics competitors do at this.,

6 months?

I always thought it was interesting that my modern CPU takes ages to plot 100,000 or so points in R or Python (ggplot2, seaborn, plotnine, etc) and yet somehow my 486DX 50Mhz could pump out all those pixels to play Doom interactively and smoothly.

He kept more of his MSFT stock.

I didn't realize how bad I was at writing and also how incomplete a lot of thoughts and ideas were until I tried blogging about them.

Definitely made me respect the bloggers I read even more.

Running/managing the firm, doing pitches, sales, schmoozing, etc.

Not seeing ads is going is going to be a very luxury premium in the future.

But for what purposes?

I watched this just the other day (as a joke Christmas film) after having not seen it for about 10-12 years. It's waaaay better than I remember and I thoroughly enjoyed it. Definitely hits a bit different after Epstein.

The Very Bad Wizards podcast on it is interesting/fun too. https://podcasts.apple.com/us/podcast/very-bad-orgies-kubric...

That just makes it more likely to sample less likely outcomes from the same distribution. No real novelty.

That sounds pretty promising.

Is there any benchmarks for these chips doing like regular 'data-sciency' CPU grunt work? Dataframe-wrangling, inverting matrices, doing large matrix, factorisations, fitting decision tree's, etc?

I'm very keen on one of these, but I simply have no idea how good they are at my day to day tasks in R or Python.

Yes. I've seen lots of twitter/X posts lately about how Gravity is not actually a force. But how can that be true if there is a force carrying "gravity" particle? Or is the word 'force' being used loosely here?

Radiologists will not be replaced. They will just have better tools.

What part about the genuine statistical arguments made in the article would make you believe it is satire?

I've found the reaction to this article can be pretty intense. We read this in a journal club many years ago and one of the mathematicians who was kind of new to the idea that research papers (in other fields) didn't more or less represent 'truth' said this article was _dangerous_.

I'd say Moneyball made statistics kind of cool for a bit. It would have inspired a lot of people to get in to sports analytics.

Llama and its variants are popular for language tasks, https://huggingface.co/meta-llama/Meta-Llama-3.1-8B

However, as far as I can tell, it's never actually clear what the hardware requirements are to get these to run without fussing around. Am I wrong about this?

Knowing your longitude is very difficult without a clock... Can't do it with just stars iirc.

Why would that stop if they got got broken up?

Because then you couldn't use a pretrained LLM to give you the embeddings. If you added these numerics as extra dimensions, you would need to train a new model that somehow learns the meaning of those extra dimensions based on some measure.

Well hopefully this means they will begin taking security seriously.

Silicon Valley/SF wouldn't be what it is today with all the defense tech investments in the 1950s and onwards, paving the way for more "consumer" tech

I had heavily pondered investing in Nvidia in 2014 when I thought deep learning was really becoming interesting (90+% of my other investments are in index funds) in some sense, I had some expert knowledge of the field. However, I thought that margins would compress because AMD, Intel, and perhaps other firms would get involved... Boy was I wrong.

Why would you expect anything to work well for economic forecasting :p

Understood, thank you. There are certainly applications in demand sensing/demand forecasting where things like recent order information, recent sales, CRM inputs are quite predictive of near-term outcomes, but become useless for longer horizon forecasts. In my experience, when information like this is available, no time-series technique that is unable to leverage this information would beat even simple regressions for short term horizon forecasts.

What does 'any-variate' forecasting mean? Can you use this pre-trained model to produce forecasts when there exists useful covariates/features/predictors? Is this something the other TS foundation models can/cannot do?