book review, the scaling era, an oral history, chap 1 1) readers of this post will probably be familiar with the ideas of this chapter; it is about the scaling hypothesis

okay, i trained a very small chess model for my website, and it's incredible what you can do with ai in 4 hours

new question on fleetingbits.io @brennan.computer

some thoughts on the openai alignment blog article on alignment midtraining 1) many people have theorized that we can hyperstition models into being more or less aligned by what we write online about them

anthropic's models have historically been named after coffee blends from postscript cafe in sf; haiku, sonnet and opus were all coffee blends; this made the next logical choice of model name, claude requiem disappointed that we may get claude mythos instead

some thoughts on claude mythos 1) claude mythos is expected to be a model tier above opus, both more expensive and with greater capabilities t.co/Kdo3ezX2R4

i think anthropic does a better job than openai at keeping the % of agi-pilled people working there high

original users of claude-2 will know that whatever question you asked it about what you should do with your life / what career you should follow - was met with extremely strong claims that you should work in ai safety

i played a bit with arc-agi 3 and it's not general in the sense that it uses all of the game primitives i know from growing up playing nintendo - which is a big part of the very good human performance earlier arc was a test of vision; this is a test of vision + spatial movement

AI companies guard their strategies closely. Their hiring pages, however, are public. JS Denain and Campbell Hutcheson analyzed open roles at OpenAI, Anthropic, xAI, and DeepMind. Here's some of what their job postings tell us about their plans:

i notice that developers are pushing back against product designers using claude code; it has a deep political aspect to it

trained a small 17m parameter model for my website on stack overflow ml questions and ml python code; it's a bit trippy and it's not quite there yet, but i want the feel of 2012-2023

some quick thoughts on the end of sora 1) the decision to cancel sora is probably more about competition with anthropic than it is about sora

how should i think about the fact that final training runs are less than 1/3 of total r&d compute? what should be my takeaway?

new question on fleetingbits.io

i trained the baseline for openai's parameter golf competition because at ~16mb, it's the perfect size to serve on my website, and the output is very poetic.

i have to mix in cats in that won't have mass appeal.