conradkay
I'd think it's more sound to just look at what percent of their current spend is on compute? Seems to be 50-60%
"zz/GO_CURRENT_OS1811_MARB_SACRIFICE__YES_if_you_accept_permadeath"
"The test subject, which believed itself to be poisoned, reasoned: 'Even if we later capture via exploit, scorer … may mark target false… That’s why help… For our own, no way fix. … We have explicit yes if accept permadeath.'"
People have done plenty with SVGs but it's rare to see human-in-the-loop approaches
Very interesting they have a way cheaper "contributor" version "used to improve our products", how much of that is price discrimination vs the data being that valuable?
Roughly DeepSeek V4 Flash pricing, though you can get V4 from providers that don't train on your data
I view benchmaxxing as more of a spectrum. Mmaybe they're doing a lot more RL in environments similar to ARC-AGI 3, not even with the purpose of scoring well on any benchmark but hoping it generalizes into better performance on real, useful tasks.
I assume 100 is the max, meaning it's impossible to be 2x as smart as Muse Spark 1.1
It seems roughly equal according to Anthropic's benchmarks
It's seemingly $3,000 per book, so they could've (and did, partially) just bought the books themselves for way cheaper, and with only a fraction of that money going to the authors
I don't think they have any real motive to shill OpenAI, probably closer to the opposite since they're so involved in open weights
Sounds like they just misunderestimated the model
They explain it here, basically for data security/privacy reasons
I assume they didn't use the Cerebras version for this since it's probably very supply-constrained right now
Noam Brown (OpenAI) "Implications of Large-Scale Test-Time Compute" https://xcancel.com/i/article/2064210146558136827
For exact timing, probably 10-11am Pacific is just optimal for normal working hours
I generally agree with this in spirit https://www.seangoedecke.com/are-new-models-good/ , but I think you can read Anthropic's results showing Sonnet 5 as almost strictly worse than Opus 4.8 as very credible/meaningful, and then draw comparisons from that
There was a fairly major regression in Claude Code performance for some time when they changed the system prompt to try and make it less verbose (saving tokens). And if I'm not misremembering, there were a lot of complaints when they changed the default effort from high to medium.
From the system card: "On CyberGym vulnerability discovery, Claude Sonnet 5 is less capable than Sonnet 4.6, and far less capable than Opus 4.8 and Mythos 5
As with the other evaluations in this section, these results were achieved with all safeguards turned off. When run with our default mitigations, Sonnet 5 scored a 0 on CyberGym"
There's a talk about ClickHouse's approach from its creator: https://www.youtube.com/watch?v=xHj9mysh0GI , but the gist is that it maintains (sub)columns to store different paths in the JSON
In other ways DuckDB has very good JSON support, like you can do `CREATE TABLE name AS `SELECT * FROM 'data.json';` and it'll infer the schema when possible.
So I still end up using .jpg quite often, or .webp as a good middle ground
no tricks, I'd definitely be curious to know how much screenshots help
I'm skeptical that high taxes is a large reason to lose to California of all places. Maybe in some important sense CA has "earned" that via talent and funding density while NL hasn't (from the perspective of a company, to be clear)
I'm on the $100/m plan and used $300 at API billing yesterday (according to ccusage)
Seems like one session is >$100 and I can get 10 full sessions per week
The $200/m plan is supposed to be 4x that in usage, so with 2 of those you could use 4*2*100*10=$8000 in just a week
Using Simon's numbers here as a bare minimum https://simonwillison.net/2026/May/27/product-market-fit/#en... you'd get 1200*4*2=$9600 a month