The smallest image models I know of are the Stable Diffusion Models, ~1B parameters for something like ~3 GB when including the VAE/text encoder. I'm surprised there hasn't been significant progress on how small a coherent image model can be since then.
Welcome to the part of Bluesky that takes AI seriously! Here are a couple starter packs for serious AI discussion on Bluesky if you want to find more:
bsky.app/starter-pack...
bsky.app/starter-pack...
Her piece does talk about how it loses meaning at long lengths:
The circle packing in circle page dates to 1999.
This site has been around a very long time, I remember first finding it over a decade ago.
1T seems possible a few years hence, the largest companies by revenue today are Walmart and Amazon each at 2/3 of a trillion. 1T is also less than 1% of world GDP. But this requires speculating beyond three years which is problematic.
yeah, surprisingly weak relationship between water % and how wet they seem
Same! I remember learning watermelons were ~92% water as a kid and assuming that other fruits must be so much lower.
Many fruits and veg are surprisingly high % water
I agreed with Dario at the time, and still do. It's good to be cautious on account of the known unknown of future growth. But the unknown unknowns like the supply chain risk thing makes not overextending yourself look even better.
yup
A couple reasons:
-It's a good indicator of capabilities progress, more useful than benchmarks IMO, and I'm more worried about all types of risk in a world where AI capabilities progress is very rapid.
-More revenue is itself one of the drivers of infrastructure spend which drives further progress.
At their current 10x/year growth rate, they should automate all labor in about 3.5 years.
Anthropic revenue growth is terrifyingly fast www.bloomberg.com/news/article...
These are four particularly good long reads on the topic of military software provisioning from the past few days.
www.hyperdimensional.co/p/clawed
thezvi.substack.com/p/secretary-...
jessicatillipman.com/what-rights-...
www.astralcodexten.com/p/all-lawful...
TBH I'm not sure and it's possible it's speculation. Lin was an open source champion, and many people on X are suggesting that the firing is a shift toward further commercialization. I'm not so sure myself though.
Targeting daily active users as the key metric you want more of
Some more context on Lin, via SCMP's Vincent Chow: x.com/vince_chow1/...
TBC I'm not sure about all the claims in this post, like the claim that it's about DAUmaxxing.
Some drama at Qwen today. Junyang Lin has been replaced, and at least one Qwen employee claims to be quitting in response. I've seen several employees tweet something of the form "Qwen is nothing without its people". x.com/Xinyu2ML/sta...
One thing I can say though is that I'm glad this discussion is happening, and that it's happening out in the open. I have lots of respect for the folks at OpenAI who have been sharing their perspectives the last few days.
I find the OpenAI/DoD stuff hard to reason about. It activates my political tribalism and my feelings about the labs and their leaders. And the specifics of contract language seem crucial and aren't something I'm used to thinking about.
Great shots! the only waxwing I've seen is the cedar waxwing, glad to see there are other waxwings halfway around the world.
Most Captain Beefheart is too weird for me but I really like this song: www.youtube.com/watch?v=MRlW...
Interesting album choices scaruffi.com/music/best10...
Note that this is just API spend, not all enterprise, and Ramp users may not be representative. But my impression is indeed that Anthropic does have more enterprise revenue than OpenAI.
Fortunately you can choose to label humans as part of the anthropogenic mass
One important bit is that there's precedent for a preliminary injunction
