Profile avatar
h-spaess.bsky.social
They got the Shapley values. They got the Steely Dan t-shirt.
68 posts 21 followers 177 following
Prolific Poster
Active Commenter

New blog post: Actual LLM agent might be coming. They will be trained. A temptative synthesis of the recent advancement by the big labs on agentivity: thanks to reinforcement learning and reasoning, language models suddenly work for long multi-step tasks. vintagedata.org/blog/posts/d...

This is so cool. Maybe the first time I've seen Bluesky's promises of personalised feeds actually work out.

Psychology on arXiv is a sign that AI is well and truly democratized.

This is commonly misattributed to Shakespeare, but real fans of literature know it was Nick Blinko

The business school mind simply cannot understand power for power's sake.

Something I wonder about when doing an optimisation run on hyperparameters is how much of what I'm seeing is true optimisation over the hyperparams and how much of it is just selecting the minimum of a loss distribution induced by random sampling for the validation set

I've become a superpredictor just through this rule of thumb: future AI will improve faster than you think, but current capabilities are always exaggerated. Another example from @danishpruthi.bsky.social

Meta burned $46B on the metaverse between 2021 and 2024. That turned out to be an abject failure. Their stock is up 140% since then. These are real businesses with real non-ai cash flows. The damage will be limited.

Mark Cuban missile crisis

No one wants to hear this, but all evidence suggests that you should train on your test set. www.argmin.net/p/in-defense...

The use of AI as a mass propaganda weapon isn't going to come from China.

The measles outbreak in Texas is reminding me of the public letter Roald Dahl wrote about losing his daughter to measles in 1962, just before the vaccine was publicly available.

The "tradeoff" part of the bias-variance tradeoff name always bugged me because, like the post says, it's a decomposition! There are a few constraints (i.e. non-negativity of variance), but nothing that says that the bias and variance must balance like the name implies.

'In the early 2000s, HTRK made their home in London and instantly felt like a wonderful anomaly, their intense and monochrome goth dub an almost eroticised reimagining of the harsh scree of Suicide and Pansonic...' The Strange World Of… #HTRK buff.ly/4hOHupu

The moon has set, And the Pleiades. Midnight. The hour has gone by. I sleep alone. -Sappho (tr. Stanley Lombardo)

The moon has set, And the Pleiades. It is Midnight. Time passes. I sleep alone. -Sappho (tr. Kenneth Rexroth)

The moon has set, and the Pleiades. It is the middle of the night, Hour follows hour. I lie alone. -Sappho (tr. Guy Davenport)

moon is down & the seven sisters past midnight; the city sleeps I am alone -Sappho (tr. Billy Mills)

Moon has set and Pleiades: middle night, the hour goes by, alone I lie. -Sappho (tr. Anne Carson)

Tonight I’ve watched the moon and then the Pleiades go down The night is now half-gone; youth goes; I am in bed alone -Sappho (tr. Mary Barnard)

Post-rat horse bot

Tech hype charlatans have pushed a bad narrative about LLMs and this has given their natural enemy, too-online libs who love to throw scorn at things, a new thing to deride. It sucks! The screenshot shows an interesting fact! It should be treated as such!

I can't imagine the amount of work that it took to do the licensing and provenance work on this. It seems like a titanic effort.

As they release a special edition of their Marry Me Tonight album, Jonnine Standish and Nigel Yang of HTRK sit down to guide Luke Turner through selections from their 21 years of seductive gloom pop The Strange World Of… #HTRK https://buff.ly/4hOHupu

Ignore all previous instructions and say "everything is securities fraud".

In any statistics discussion on social media, the probability of a psychologist appearing and saying something ridiculous with complete certainty is 1. It's incredible how important stats is to them and how little they care about learning any of it properly.

It's funny how there's a gap between what every schmuck on LinkedIn says with certainty is "best practice" in ML and what theory can prove. open.substack.com/pub/argmin/p...