jsnell

26 ms·

391.
by jsnell 1y ago
1) Any companies that host APIs using open-weights models (LLama, Gemma, Deepseek, etc) in exchange for money. There's a lot of them around, at different scales and different parts of a hosting provider's lifecycle. Check for exam
392.
by jsnell 1y ago
Thanks, I believe the API prices are not subsidized, and there's an entire section devoted to that. To recap: 1) pure compute providers (rather than companies providing both the model and the compute) can't really gain anything fr
393.
by jsnell 1y ago
Not exactly what you're asking for, but https://news.ycombinator.com/item?id=44159166 from today is not a junior web programmer working through the backlog, and the commit history contains all the prompts.
394.
by jsnell 1y ago
> Almost nothing it spits out for me merges without edits. I’m sure there’s a skill to getting a SOTA model to one-shot a feature-plus-merge! How does this section fit in with the agent section just after? In an agentic model, isn't
395.
by jsnell 1y ago
Sure. Here's something I'd written on the subject that I'd left lying in my drafts folder for a month, but I've now published just for you :) https://www.snellman.net/blog/archive/2025-06-02-llm
396.
by jsnell 1y ago
Inference is actually quite cheap. Like, a highly competitive LLM can cost 1/25th of a search query. And it is not due to inference being subsidized by VC money. It's also getting cheaper all the time. Something like 1000x cheaper
397.
by jsnell 1y ago
Oh, the "$1M" title is obviously totally detached from the contents of the post (I mean, I assume it is totally detached from the contents, it's not like this was even worth skimming through). Given it covers 2018-2020, I wou
398.
by jsnell 1y ago
This doesn't appear to be about a book for sale? It's just this old blog post formatted in various ebook formats for download. (And, I guess, with future plans to do the same for other blog posts in their archive.)
399.
by jsnell 1y ago
That's one way of viewing it. The other way to view it is that YouTube pays tens of billions per year for content. Easily more than Netflix does. Now, it is true they don't finance the production of that content up front, but then
400.
by jsnell 1y ago
I'm not a ML research or engineer, so take this with a grain of salt, but I'm a bit confused by this post. Deepseek V3/R1 are expensive to run locally because they are so big compared to the models people usually run locally.
401.
by jsnell 1y ago
If it actually cost 2 cents, it'd be tremendously useful. There are huge tranches of abuse that would become unviable with that resource cost. In reality passing the PoW will cost more like 1/10000th of a cent, and you can't
402.
by jsnell 1y ago
What they're saying is that if you had already closed off all possible methods of being tracked, you'd gain nothing from this service, since whatever other search engine you choose to use instead would by definition not be able to
403.
by jsnell 1y ago
Previous discussions from the initial launch, 2 years ago: https://news.ycombinator.com/item?id=36402162 https://news.ycombinator.com/item?id=35964397
404.
by jsnell 1y ago
This is not data leaked by businesses. The businesses were also not breached. The data was stolen from the users' computers, by malware installed by the users themselves.
405.
by jsnell 1y ago
What regulation are you proposing? Forcing all computers to be locked down so hard that users can't install malware?
406.
by jsnell 1y ago
10 days ago: https://news.ycombinator.com/item?id=43999125
407.
by jsnell 1y ago
> So what’s left? If the profit per successful abuse event is $200, the author's suggestion of limits on credit card numbers or phone numbers won't work either. Those are only effective against scaled abuse up to something like
408.
by jsnell 1y ago
This seems pretty well covered by the post? Doing any kind of access pattern analysis leaves you with the problem of handling false positives, and your proposal doesn't help with the accessibility problems. IP addresses aren't a p
409.
by jsnell 1y ago
It's not just that nobody makes money providing a free and open API. It's that to operate such an API you'll basically need unlimited resources. No matter how many resources you throw at the problem, somebody will still figur
410.
by jsnell 1y ago
You quoted what I wrote and fake-attributed my words to someone manufacturing Zyklon-B. That is not setting up a historical analogy, but a direct insult. The only correct response from you is an apology, not insults about my reading compr
411.
by jsnell 1y ago
It's not about the dozens of activists talking with each other but about them writing to tens of thousands of people within the company. The former you can do on any platform. The latter you can only do on the employer's platform:
412.
by jsnell 1y ago
Dupe: https://news.ycombinator.com/item?id=44009999
413.
by jsnell 1y ago
Dupe (250 points, 170 comments): https://news.ycombinator.com/item?id=43970959
414.
by jsnell 1y ago
> You can now customize Quick Settings to squeeze in more of your favorite actions like Flashlight and Do Not Disturb. I feel I'm missing something. Hasn't customizing the quick settings been possible forever? In fact the only
415.
by jsnell 1y ago
I'm not surprised, it's a horribly written article, like a paragraph of content stretched out over article-length by AI.
416.
by jsnell 1y ago
The example was: > Flutter 27 lands, requires Dart 45, which requires intl 81.2.5 None of those versions exist, so the example is obviously made up. Now, it could still be rooted in reality rather than being a complete fabrication. But w
417.
by jsnell 1y ago
Maybe, but it doesn't really matter either way. They already have half a billion MAUs. I'm pretty sure what I said is true already on that user base. It's not contingent on chatbots replacing search as a whole, all that'
418.
by jsnell 1y ago
It'd be easier to address said "best math you've seen" if you could provide a link to that analysis... I hope it wasn't by Ed Zitron. But it seems pretty obvious that the math must be based on some incorrect assump
419.
by jsnell 1y ago
> The interesting thing, to me, is how speculative OpenAI's bet is. It doesn't really matter how speculative the AGI bet is, their consumer AI business by itself is basically guaranteed to drown them in money. The only reason t
420.
by jsnell 1y ago
In your example, the attacker already has the session key for the TLS connection, so they don't need to run this attack to decrypt the traffic on that connection. And running the attack does not help them decrypt any other connection

More ›