pull down to refresh
@optimism
453,155 sats stacked
stacking since: #879734longest cowboy streak: 97npub13wvyk...hhes6rk47y
0 sats \ 0 replies \ @optimism 32m \ on: Desktop Linux Longform Nostr blog client nostr
Cool.
Do you understand the code you vibe?
this thing is clearly trained via RL to think and solve tasks for specific reasoning benchmarks. nothing else.
Perfectly aligns with the perceived villain arc of the CEO. I made a small comment yesterday about how it's apparently okay in AI to do what gave VW massive reputation problems: build inferior products that only perform well on the benchmarks and safety tests.
Re:
electrs
on a phone.Unless someone improved blockstream's implementation or mempool's implementation completely... you need a bitcoin core node, for either, the readme says:
Install Rust, Bitcoin Core (no txindex needed)
So unless you hang something like a 1TB Samsung T7 disk to your android phone, this is not really feasible.
I think that you'd probably want LND with Neutrino if you're on an armv8 android phone, which is exactly what Blixt does on android, but then you get a much better management UI than
termux
, so you'd probably use that?You can also just use ollama + goose on a mac. Since
v1.2.0
of goose, most problems that were plaguing it seem to be resolved and it works rather stable again.If you’re in scenario 3b, you could, of course, just “stick it out” and grow your business and net worth by so much that you move up into scenario 4, i.e. that you have enough money to pay fancy tax advisors to set up a Liechtenstein trust for you, which would enable you to leave the country without paying exit tax again. But this feels both shady and wrong.
Funny how the author feels that evading
your exit tax is calculated by taking the average of the past 3 years of earnings of that company, multiplied by 13.75, and then taking 60% of that which is taxed at your personal income tax rate [of 42%]
feels both shady and wrong.
No, because it doesn't add anything. Blockchain anti-counterfeit measures only work for blockchain native assets where the asset itself is part of validation. This is why there are persistent doubts even about something as simple as Tether.
If you need an audit trail for your medicine but not everyone else's, an auditable central database with the manufacturer (linked to a sealed-in identifier) is probably a more useful solution.
How would you do that on an unmoderated territory? I just downzap and looking at the stats there's more than just I.
I've always felt that "justice" has taken a strange form in the face of (world) order. For example: have victims of war or their families ever received justice? Will those of the current wars? Will those of the future? I'm pessimistic, but I'm also undereducated.
I discussed this with a friend not too long ago and at their recommendation I have Foucault's Discipline and Punish: The Birth of the Prison sitting on the ebook shelf waiting for me to read it. I guess your question is a sign that I should not procrastinate this and simply spend time reading this.
Going to be honest:
- On desktop, both implementations are a bit distracting when I'm working on a longer reply. I often quote source articles, pictures, and so on and I spend a longer time with the write screen open. Also I'm a slow af boomer, sorry
- On PWA, I don't mind as much because I anyway don't do long replies from mobile, so there its cool.
I was wondering: can I turn it off on desktop, to retain focus?
I was looking at these
k
chartsand wondered whether .38 for higher complexity and .56 for lower complexity a great result, if human experts reach .78 and .81 among themselves?
I knew I'd seen a paper about this: https://arxiv.org/abs/2501.08167, but it's kinda stone age:
Comparisons Percentage Agreement Cohen’s Kappa Human vs Claude 2.1 Ratings 79% 0.41 Human vs Titan Express Ratings 78% 0.35 Human vs Sonnet 3.5 Ratings 76% 0.44 Human vs Llama 3.3 70b Ratings 79% 0.39 Human vs Nova Pro 76% 0.34
Looks awesome if we realize that Google's results were with a 3.25B model, but the evaluation data provided in the paper was "a mockup", so we don't know if this is apples-to-apples. Nevertheless, I'm a big fan of "less junk in".
There's another note in #1071526 regarding "Instruction Hierarchy Evaluation"
We note regressions in performance for gpt-5-main. We will follow up with a fix to improve these behaviors.
that may explain
Im hearing its routing to dumber faster model and getting some things wrong then if you ask "think" it gets it right
This explains why these models are not really usable.
Any small online community for people who run local models is at least 50% perverts.
lmao!
PS: apparently pulling a VW is okay in AI.