It's described at a high level in media often. Here's more detail though:
Person buys a product, or downloads some software - that unknowing to them, was reaching out to a central server elsewhere to establish a network tunnel. And over that tunnel, the other server routes traffic onto that person's home network. Which can then of course go onto the internet, and it appears like the traffic originated from their home internet connection.
Other people rent access to this network. The goal being, you have surreptitious access to appear as any one of tens of thousands of home network IP addresses on the public internet. Bad guys can use this to rotate through many different IP addresses when their access gets blocked because of malicious (or benign but unwanted) behavior on the internet. Residential IP addresses are more 'valuable' because they don't have the negative connotation that datacentre and commercial IPs can have, and they can be spread geographically or across many ISPs.
A very common way for people being a part of these residential proxy networks is by buying dodgy streaming boxes, and also installing shady apps on TV app marketplaces.
This isn't a theoretical problem either. My employer's website was being hammered by residential proxies a month ago. It wasn't until we setup WAF captcha rules that they mostly subsided. It wasn't even the typical exploit scanner. I don't know what the purpose was but it was effectively a DDoS.
It's definitely not all VPNs (typically the ones you pay for will buy their own hardware and internet connections in the places you can route your traffic through), but it is a common situation for the shadier (especially 'free') ones, and in that case it's usually not just a peer-to-peer thing, they make money by selling your connection to people who pay because they want a large pool of relatively reputable and disconnected IPs to send traffic through.
You're right, a lot of 'free VPN' software also uses your own internet as a residential proxy. It's often mentioned in the ToS but the users don't read that.
Honestly, I expect any of those free VPN services that require an app install to be doing precisely this. At this point, I just assume anything with "free" in it will cost me more than if I just paid for something instead.
But most likely these residential proxy services are more valuable for scraping, DDoS and manipulation plays on social media so they might not actually be used for things that draw direct attention. Using them to download illegal materials would be a waste of resources when a VPN does the job just as well.
It means some networked devices in there are being remotely piloted to tunnel network traffic to somewhere else. This scenario usually happens with more "unprofessional" ISPs, which sell bandwidth from their networked clients for people in need of such services.
Such as ignoring web scraping countermeasures, and DDoS.
There are companies that provide proxy services for scraping services, and residential proxies route through consumer networks (like Comcast) which is flagged less than a datacenter IP.
Many questionable TV streaming boxes pretend to deliver you video, but in reality they are also piggy backing on your personal ISP to power those companies.
Some good answers in this thread. I had the same question but I'm still a tad confused:
* I initially thought this meant someone in the WH network was using commercial proxies to appear as if they were exiting from a different network (which would align with "some user is firing up a proxy on their phone to exit in a different location")
* But then the description of residential proxies seemed to align more with traffic hitting an endpoint in the White House, and _exiting_ from the network there. But this didn't explain the timing (why isn't it 24/7).
* Finally it sounds like some person, legitimately on the WH network, has a device (cell phone, TV, whatever) that, when activated, is proxying outside traffic into the WH network.
There are services you can pay for such that you can route internet traffic thru a specific IP in a specific part of the world. People (usually unwillingly and unknowingly) run software that effectively makes them part of a giant botnet. A computer running in Romania might have some kind of weather application or gambling/casino app that I could use to make an HTTP request as if I were on their network. So if you need access to 1,000 or 10,000 IP's to circumvent rate limits or perform crawling tasks, this is how you could do it.
> Residential proxy services sell access to everyday devices, letting customers send traffic through them so it appears to come from that device's network.
So what? Packets flowing through machines is just how the internet works. Stop believing all this propaganda that other peoples papers are EVIL and if you route them you are getting in BIG BIG trouble. This is harmful messaging that does not reflect reality.
> So they've invested around $2.5M-3M into hardware + monthly operating costs (electricity, updates, cooling).
Are you assuming that they bought the hardware? because they could also be renting the metal off a provider, in which case they either pay by the hour or have a yearly commitment to save on costs.
yeah, seems like they're renting (according to the info from the video), but doesn't it make the bill worse? I mean the provider makes money somehow, so it should be more expensive for revolut. I guess it's a lot anyway
Maintenance, capacity management and spares all add up, not to mention labour. If they have a requirement to maintain geographical diversity then that moves the needle significantly if they’re primarily based on the other side of the country.
I've quickly checked and guestimate of the rental would be ~250-300k per month for this kind of usage they claim to have. It's all based on open prices from together.ai. So probably it worth the switch for them afterall, I can imagine that 56 engineers can burn tokens for more than 300k per month
I believe the reason why they have to point at science is because they are often discredited because the knowledge they present is not backed up by papers but rather generational wisdom (they explain in the text how they do it).
But you never heard someone say "you should still code in assembly so you can understand what your compiler is doing".
That's true but a similar advice that is given to junior developers in their early career is to actually reinvent the wheel, so that they can learn from the pain and then rip the profits of using reusable components. That's unvaluable. You have to read through thousands of lines of code and write some more yourself to get good at programming.
Finally, I have a bit of a grudge against the naming Mistral gives to their models.
This is Voxtral 3, right? Instead of appending the variations at the end consistently, they just do this:
Voxtral Mini Transcribe 3 (batch), Voxtral Mini Transcribe Realtime 3, Voxtral Small Transcribe 3, Voxtral Speech 3 (batch and realtime, same API)
Why not Voxtral 3 Transcribe Mini (batch), Voxtral 3 Transcribe Realtime, Voxtral 3 Transcribe Small, Voxtral 3 Speech. Then it reads more like it's layered, like it should be.
> And then just do a cleanup pass with a cheap LLM
I am especially interested in this part. Could you please share the prompt you are using to instruct the LLM to clean up the dictation? Thanks in advance!
I do this with a summary prompt + a JSON config. The config keys are the full name of the thing that gets commonly mistranscribed, and the values include a phonetic spelling of the name and a n "entity id" which is just the name of a markdown file that provides information about the entity, which the LLM can use to infer what should have been referenced if it's not clear.
My prompt is pretty simple:
- Intro of the purpose (faithful, high-quality summaries of meeting transcripts useful for readers who did not attend the meeting)
- Some basic rules
- Summarize only what is actually said in the transcript.
- If a speaker's name isn't clear from the transcript, describe them generically (e.g. "a participant") rather than guessing
- If the transcript is a fragment or cuts off, say so rather than inventing a resolution
- Only state a causal or explanatory link between two facts (e.g. why someone is absent) if that link is clearly stated in the transcript
- Preserve relative time references precisely rather than flattening them. If a speaker says something will happen "in two hours" don't reword it as "today" or "same day"
- Maintain the original voice and framing to keep original intent
- Attempt to infer meeting attendance based on who actually spoke or was spoken to
- Prefer plain, direct sentences over compound or clever phrasing
- Respond directly with the summary. Do not preface it with phrases like "Here is a summary" or "Based on the transcript".
- A preferred structure for the summary, with examples
- An instruction to attempt correction of mis-transcribed words based on the included JSON configuration, but only if the certainty is high. If unsure, leave as-is, but add (spelling?) after the word to flag it.
- The JSON config as a string.
I invoke this with a skill that can also update the JSON config if I identify new mistranscriptions, so it gets updated regularly.
I told Kimi to set up a local Matrix server for me and to write a trivial little Claw-like.
Basically it's just a loop with bubblewrap based sandboxing for the tools (not really enough if you're going to expose it to a lot of untrusted data - I intend to move the entire thing into a vm) and the "identity", "soul" and "memory" split that the original openclaw used to give the agents persistent identity.
The e-mail access is read-only simply by virtue of it having no direct access to credentials but access to tools to run read only operations on my mailbox.
They also have tools to read and send Matrix messages, but at the end of the turn any new messages also get injected automatically if any to start a new turn. In additional there's heartbeat that triggers occasional requests to clean ensure anything they "want to remember" is consolidated in the memory files, and to clean up, or advance any tasks, and a very basic reminder system.
I'm using a mix of Kimi, GLM5 to run the actual agents at the moment.
I might release parts of the code at some point. It's nothing special, but it's small and easily auditable unlike OpenClaw.
I picked Matrix because it's trivial to self-host with federation turned off so the agents are reasonably isolated (with the caveat that they do have web access but at least they'll have to work a bit more to leak my data - but be careful). I run it on in my wireguard network and can access it via a Matrix client on my phone and/or a web app on my home machine.
I can also access it via Claude, Codex, Kimi CLI via a trivial little mcp server that wraps the same Matrix client code the agents use.
I have gone down this route but I'm afraid that 16K of context is nothing for agentic coding tasks. Even if you use a smarter model directly from an inference provider to plan all the work and then use your own 16GB GPU to execute the plan, managing those 16k tokens becomes unusable at some point even with a minimalistic harness like pi.
For other agentic tasks it's nice, and more than just the budget, it's the data sovereignty that you gain (imagine analysing tax forms, contracts, other legal documents, etc.) in my opinion.
Qwen3.8-Flash-Next+strix halo+halogen server = TAM
Im convinced AI is local now. Halogen loads QFN in 30 GB at 4bits. Its token gen is slightly faster than readable. You can follow along and leave it alone based on its thinking traces.
It gets 256k context but using dynamic context pruning, i got it to loop between 64k and 128k to maximize its speed vs prefill.
Its real. If the memory cartel is busted, the moat is busted.
Not sure if it's the same but the concept was the same: the ability to combine two emojis together.
reply