APIs are the best when they let you move data out and build cool stuff on top. A lot of big platforms do not really want that anymore. They want the data to stay inside their silo so access gets slower harder and more locked down. So you are not just yelling at the cloud this feels pretty intentional.
By default we disguise activity on your pooled token — we only route traffic during your active session times and use sticky sessions so each consumer looks like one consistent user. You can further disguise it by using your own provider pool key, so your personal usage also goes through ClawPool and blends in with the rest of the traffic. The LLM provider has no way of telling the difference. From their perspective, it's indistinguishable from the provider sitting at their keyboard."
The correct response would be "Yes, it breaks Anthropic TOS".
1) Covering the ~$50k hosting bill for Jmail on Vercel sounds generous, but a self-hosted VPS on Hetzner could serve the same purpose for ~€30/month, which is orders of magnitude cheaper and avoids vendor lock-in.
2) This comes as the CEO of Vercel, Guillermo Rauch, is already facing community backlash for publicly supporting Israeli Prime Minister Netanyahu, a move that’s led to boycotts and migrations off the platform among developers. All my homies hate Vercel.
Even before the Vercel CEO supporting a genocidal maniac. Vercel as a platform has been silently giving open source projects a "fuck you, pay me" when it comes to renewing benefits.
Have seen it happen to smaller projects and even pointed it out when Vercel took static sites down.
I'm new to webdevelopment and was using Vercel because people told me it was good, but I was unaware that the company supported the genocide. What other similar services there are that you would recommend?
Railway/Render can easily host Next.js applications
You could also get a VPS from Digital Ocean or Hertzner, run open-source PaaS like Coolify, Dokploy, Caprover, etc.
Digital Ocean has app platform that’s lets you host these applications but more experience than VPS
Sealos has a $7 and $25 plan and work with Next.js
Just a few options. If you’re looking to leave Next.js completely, check out Tanstack Start. It’s by the creator of React-Query (defacto way to handle fetching data in Next.js). Still some rough edges but I think it could overtake Next.js once a bit more mature.
Do you really think a $30 hetzner host can sustain that level of traffic performantly? Don't get me wrong, I love hetzner, but I would be very surprised if the numbers work out there.
It shouldn't. The issue is that most developers would rather spin up another instance of their server than solve the performance issue in their code, so now it's a common belief that computers are really slow to serve content.
And we are talking about static content. You will be bottlenecked by bandwidth before you are ever bottlenecked by your laptop.
To be fair, computers are slow if you intentionally rent slow & overpriced ones from really poor-value vendors like cloud providers. For people who started their career in this madness they might be genuinely unaware of how fast modern hardware has become.
With a 2025 tech stack, yes. With a 2005 tech stack, no. Don't use any containers, no[/limited] server-side dynamic script languages, no microservices or anything like that.
Considering the content is essentially static, this is actually viable. Search functions might be a bit problematic, but that's a solvable problem.
Of course you pay with engineering skills and resources.
Is there any feasible way to implement search client-side on a database of this scale?
I guess you would need some sort of search term to document id mapping that gets downloaded to the browser but maybe there's something more efficient than trying to figure out what everyone might be searching for in advance?
And how would you do searching for phrases or substrings? I've no idea if that's doable without having a database server-side that has the whole document store to search through.
I think the key thing here is the context and size; the searchable content of even a lot of e-mails is quite dense and small. I'm not a search expert but I'd look at precalculated indexes on very short substrings (3-4 characters maybe?), have the client pull those it needs for a particular query and then process client-side from there. (Doesn't even need figuring out in advance what people will search for, though that'd certainly improve things.)
Theoretically, just thinking about the problem... You could probably embrace offline first and sync to indexeddb? After that search would become simple to query. Obviously comes with it's own challenges, depending on your user base (e.g. not a good idea if it's only a temporary login etc)
there's been demos of using SQLite client-side, with the database hosted in S3, and HTTP range requests used to only fetch the necessary rows for the query.
there might be some piece I'm missing, but the first thing that comes to mind would be using that, possibly with the full-text search extension, to handle searching the metadata.
at that point you'd still be paying S3 egress costs, but I'd be very surprised if it wasn't at least an order of magnitude less expensive than Vercel.
and since it's just static file hosting, it could conceivably be moved to a VPS (or a pair of them) running nginx or Caddy or whatever, if the AWS egress was too pricey.
There are several implementations of backing an Sqlite3 database with a lazy loaded then cached network storage, including multiple that work over HTTP (iirc usually with range requests).
Those basically just work.
Containers themselves don't, but a lot of the ecosystem structures around them do. Like having reverse proxies (or even just piles of ethernet bridges) in front of everything.
Or if you go ping pong across containers to handle a single request. That will certainly make a laptop unable to handle this load.
I just fired up a container on my laptop... running on kubernetes... running in a linux VM. It's lightly dynamic (no database or filesystem I/O).
While I've also got enough other stuff running that my 15 min load average is at 4 and I've got 83% RAM used ignoring buffers/caches/otherwise.
I went and grabbed a random benchmarking tool and pointed it at it with 125 concurrent connections.
Sustained an average of 13914 reqs/s. Highest latency was 53.21ms.
If there are 10,000 people online at any given time hitting the API on average once every 3 seconds (which I believe are generous numbers), you'd only be around 3.3k reqs/s, or about 24% of what my laptop could serve even before any sort of caching, CDN, or anything else.
So... if a laptop can't serve that sort of request load, it sounds more like an indictment of the site's software than anything.
No it won't. This is static content we're talking about. The only thing limiting you is your network throughput and maybe disk IO (assuming it doesn't fit in a compressed RAM). Even for an "around the globe roundtrip" latency, we're still talking few hundred msec.
Some cloud products have distorted an entire generation of developers understanding of how services can scale.
I think it’s more helpful to discuss this in requests per second.
I’d interpret “thousands of people hitting a single endpoint multiple times a day” as something like 10,000 people making ~5 requests per 24 hours. That’s 0.5 requests per second.
A laptop from 10 years ago should be able to comfortably serve that. Computers are really really fast. I'm sorry, thousands of users or tens of thousands of requests a day is nothing.
There may be a risk of running into thermal throttling in such a use-case, as laptops are really not designed for sustained loads of any variety. Some deal with it better than others, but few deal with it well.
Part of why this is a problem is that consumer grade NICs often tend to overload quite a lot of work to the CPU that higher end server specced NICs do themselves, as a laptop isn't really expected to have to keep up with 10K concurrent TCP connections.
I would use a $100/mo box with a much better CPU and more RAM, but I think the pinch point might be the 1Gbps unmetered networking that Hetzner provide.
They will sell you a 10Gbps uplink however, with (very reasonably priced) metered bandwidth.
A profitable customer? How would Hetzner know if you're profitable or not?
I've hosted side projects on Hetzner for years and have never experienced anything like that. Do you have any references of projects to which it happened?
I am not sure how one even gets 250TB/mo through a 1Gbps link. In any case, completely saturating your networking for the full month is outside most people's definition of "fair use".
Yeah but they still advertise with unlimited traffic.
"All root servers have a dedicated 1 GBit uplink by default and with it unlimited traffic"
https://docs.hetzner.com/robot/general/traffic/
roughly 47 percent of a recent US survey think supporting Israel is in the national interest of the United States, while 41 percent disagree. only 12 percent didn't offer an opinion. that's quite a lot of potential customers you're detracting when making a stupid political statement, especially taking a photo with a leader whom 49% of people "have an unfavorable opinion" of
"In 2025, global anti-Israel and pro-Palestine demonstrations intensified, with ACLED recording more than 49,000 protests across 133 countries. Public sentiment shifted significantly, particularly following Israel's interception of a humanitarian flotilla in October 2025.
Largest Global Mobilisations
Bangladesh: On 12 April 2025, Dhaka hosted what was described as the largest pro-Palestine protest in history, with an estimated 1,000,000 participants at Suhrawardy Udyan.
Italy: Between 3 and 5 October 2025, an estimated 2.3 to 3 million people participated in nationwide strikes and marches, marking the highest level of European mobilization during the conflict.
Spain: Over 1,000,000 people filled the streets of Madrid, Barcelona, and other cities in early October to support the "Sumud Flotilla" and demand an arms embargo.
Australia: Organisers estimated that over 300,000 to 350,000 people participated in nationwide rallies on 24 August 2025, following diplomatic tensions between the Australian and Israeli governments.
Netherlands: Approximately 250,000 people marched in Amsterdam’s "Red Line" protest in October, the third such major rally in as many months. "
I'm fully aware of my Jewish brothers and sisters who stand shoulder to shoulder with me against the current Israeli actions ... but, fuck Israel ... and all who support their current stance.
Those are all very small numbers in the context of the actual electorates, and supports the claim that it is the same small group of activists obsessed over Israel, and not a widespread “community” concern.
Get real. Even the US right and much of MAGA hate Israel now. Not because they love Palestine, but because they hate Israel sucking away US taxpayer money to slaughter women and children and bring them nothing but hassle.
Frankly, watching Israel has been the closest we've come to having a petri dish to understand how Nazi'ism developed and rose. I have a better understanding now of how the Hitler Youth developed, and how cultural blindness, manipulation, nationalistic obsessives and ideologues enable a cancer to spread among its politics leading to full blown Nazism. I have a better understanding of all of that from watching the current Israeli state and reading about it's history.
It's a disgusting tale and shouldn't have been funded by US taxpayers.
It’s actually easy to understand how Nazism developed by watching those who are demonizing and dehumanzing an entire nation, exaggerating their crimes, and claiming that they are taking all our money and drinking the blood of children etc.
You don’t need a petri dish. The next season in the series is already here.
Oh come on ... the same wankers who want to see the end of Israel are the exact mirror reflection of those Israeli's currently pushing for a "one state solution" (read: "and we'll enact a final solution genocide them to make it happen").
The only difference is the Israeli's wear a suit and pretend to represent democracy as they suck US taxpayers of their earnings.
2nd point resonates with me, how come he wants to cover expenses, while being connected to Israeli PM and Epstein is connected to Israel through Ehud Barak.
Barak and bibi are political enemies (or at least we're when Barack was a relevant political figure) and besides that I haven't seen anything suggesting that his connection with bibi is more than the one meeting that was publicized.
This is offtopic honestly, but I'm curious if I've been using this phrase wrong for my whole life. Doesn't "order of magnitude" refer to steps of powers of ten?
$50000 vs €30. (or €42066.30 vs €30 if I normalize the currency) 5x10^4 vs 3x10^1.
You have it right, perhaps the original poster was referring to it in a more colloquial manner, in the sense that against 200 million in revenue, 50,000 and 30 are in the same ballpark?
> in the sense that against 200 million in revenue, 50,000 and 30 are in the same ballpark
I don't understand how those are in the same ballpark? I thought saying something is in the same ballpark suggested that they are similar in scale, and the implication is that little-leauge does not play in the same ballpark as a NBA team. They are in the same category (baseball), but not at all the same level.
At a big enough scale, previously large differences are effectively 0.
50k/mo is 600,000/yr vs 360/yr at 30/mo. Thats existential for a 1MM/yr company. Neither register on a balance sheet for a 1B/yr company. They are both closer to 0 than being a major cost.
But saying that 200 million and 30 are in the same ballpark is not true in 99.99% of contexts.
Even 50k and 30 I would not say are in the same ballpark. I've worked for major corps and of course a cost saving of 50k/month would not register for the overall company but it probably would for my team. A saving of 30/month is probably not worth spending any considerable amount of time on in most non-personal contexts.
Is it 0.003 per minute of audio uploaded, or "compute minute"?
For example fal.ai has a Whisper API endpoint priced at "$0.00125 per compute second" which (at 10-25x realtime) is EXTREMELY cheaper than all the competitors.
If Voxtral can process rapid speech as well as it claims to, an obvious cost optimization would be to speed up normal laconic speech to the maximum speed the model can handle accurately.
WER is slightly misleading, but Whisper Large v3 WER is classically around 10%, I think, and 12% with Turbo.
The thing that makes it particularly misleading is that models that do transcription to lowercase and then use inverse text normalization to restore structure and grammar end up making a very different class of mistakes than Whisper, which goes directly to final form text including punctuation and quotes and tone.
But nonetheless, they're claiming such a lower error rate than Whisper that it's almost not in the same bucket.
On the topic of things being misleading, GPT-4o transcriber is a very _different_ transcriber to Whisper. I would say not better or worse, despite characterizations such. So it is a little difficult to compare on just the numbers.
There's a reason that quite a lot of good transcribers still use V2, not V3.
Maybe it's because I'm not used to the flow, but I prefer to work directly on the machine where I'm logged in via ssh, instead of working "somewhere in a git tree", and then have to deploy/test/etc.
Once this app (or a similar app by Anthropic) will allow me to have the same level of "orchestration" but on a remote machine, I'll test it.
We have not released the weights, but it is fully available to use in your websites or applications. I can see how our wording there could be misconstrued -- sorry about that. You can absolutely create a vTuber persona. The link in the post is still live if you want to create one (as simple as uploading an image, selecting a voice, and defining the personality). We even have a prebuilt UI you can embed in a website, just like a youtube video.
The few people looking at /new on HN are ridiculously overpowered. A few upvotes from them in the few hours will get you to the front page, and just 1-2 downvotes will make your post never see the light of day.
You can't downvote a post, so that's not a factor.
Also it's not as powerful as you think. In the past I have spent a lot of time looking at /new, and upvoting stories that I think should be surfaced. The vast majority of them still never hit near the front page.
It's a real shame, because some of the best and most relevant submissions don't seem to make it.
If you are in a company like e.g. ClickHouse and share a new HN Submission of ClickHouse via the internal Slack to #general, then you easily get enough upvotes for the front page.
reply