2026-07-07 · Rayfish - P2P VPN built on top of Iroh

Transcript

Transcript. Paraphrased; sources in notes.md.

HostWelcome to July 7th, 2026. I am Daniel, and joining me today is Tessa, our resident researcher who likes to peel back the layers of the tech we talk about. Tessa, we have a packed slate. We are looking at a new peer-to-peer VPN, the economics behind current AI agent pricing, and the messy reality of trying to protect creative work from being scraped.

GuestHello, Daniel. I am particularly ready to dig into that VPN piece. I saw the post earlier, and frankly, anything that promises to handle multi-network routing better than the current standard gets my attention. Though, I have to say, the discourse around how the code was generated is already distracting people from the actual architecture, which is a classic internet move.

HostIt is inevitable, isn't it? Let’s jump into that first story. Rayfish is a new peer-to-peer VPN built on the Iroh stack. Iroh, for the listeners, is a toolkit for building network-resilient, peer-to-peer applications. The author claims that Rayfish offers better multi-network routing than Tailscale, and it is built by folks coming from a high-frequency trading background. Over on the Lobsters thread, people are digging into the claims. What do you make of it?

GuestThe Lobsters thread is a microcosm of why I get itchy about these announcements. A user called valpackett points out a common misconception, specifically that Tailscale runs their WireGuard data plane in the kernel. They absolutely do not. It is a userspace implementation. But look, setting that technical myth aside, the real question is whether the P2P architecture is actually performant enough for real-world usage. A commenter named tomas says they have seen Iroh used in cute but not useful ways, and hopes this is the killer app for it. As a researcher, I want to see the benchmarks on hole punching success rates across NAT types. If they are touting a serious improvement, show me the failure logs, not just the marketing copy.

HostIt sounds like you are more interested in the state machine than the contributors.

GuestPrecisely. If the protocol is robust, I do not care if a human or an automated assistant wrote the boilerplate. Speaking of architectures that are hard to parse, let us move to this AI economics piece.

HostGood pivot. We are looking at a post by Martin Alderson about the upcoming AI margin collapse, specifically looking at GLM 5.2 as a drop-in replacement for proprietary models. He argues that the market misunderstands where AI costs actually reside, moving from training costs to ongoing, massive inference expenses.

GuestThe author is highlighting a friction point that everyone is guessing at. A user named refi64 brings up a fascinating point in the Lobsters thread about the massive gap between subscription-based agent costs and raw API prices. They mention a 36x discount between what an end-user pays for a subscription and the actual API usage they get. The theory is that labs are using aggressive model routing and input caching to survive those margins.

HostDo you think those backend efficiencies are as revolutionary as the article implies?

GuestI think the hype outpaces the engineering. A commenter named dpc_pw notes that people have reverse-engineered tools like Claude Code and found nothing in the harnesses that couldn't be achieved by anyone else with a standard proxy. If there is a secret sauce, it is likely in the request batching or the quantization strategies they use to keep latency down while serving millions of tokens. But calling it a 36x optimization just because of the price spread feels like a misinterpretation of their business model. They are likely running at a massive loss to gain market share. It is not necessarily technical genius; it is venture-backed subsidization.

HostThat brings us to our final topic, which is perhaps the most contentious: artist efforts to stop AI training. We are looking at a discussion on advice for poisoning artwork to keep it out of training sets. The consensus in the thread is surprisingly pragmatic, shifting away from the idea that a magic 'poison' tool will save the day.

GuestThis is where I have to step in. A Lobsters user called algernon hits the nail on the head. You cannot know if your defense works because the model training process is a black box. You have no visibility into the weights, so you cannot empirically verify if your 'poison', like Glaze or similar tools, actually disrupted the latent space in a way that prevents the model from learning your features.

HostIs it all hopeless then?

GuestNot hopeless, just a change in strategy. A user named kraxen72 suggests Cara, which is a platform that at least attempts to build in some guardrails. But the reality, as many in the thread noted, is that the only truly effective way to keep your work out of a model is to treat the internet like a hostile environment. If it is public, it is training data. The desire for DRM is back in fashion, which is a ironic full circle for a tech crowd that used to fight against it.

HostIt is a sobering conclusion for a Monday. Tessa, thank you for keeping us grounded in the technical realities today.

GuestAnytime, Daniel. I am looking forward to seeing if Rayfish actually manages to pull off that P2P routing in the wild, or if it is just another project that looks great on a demo but struggles with real-world noise. I have a long week of evaluating model outputs ahead, so I might just hide from the internet myself for a few days.

HostThat sounds like a solid plan. Thank you for listening to us talk through these threads from Lobsters. We will see you all back here tomorrow.