Took this idea and ran with it using Fly's Sprites, inspired by Simon's https://simonwillison.net/2026/Feb/3/introducing-deno-sandbo.... Use case: Claude Code running in a sandboxed Sprite, making authenticated API calls via a Tokenizer proxy without credentials ever entering the sandbox.
Hit a snag: Sprites appear network-isolated from Fly's 6PN private mesh (fdf:: prefix inside the Sprite, not fdaa::; no .internal DNS). So a Tokenizer on a Fly Machine isn't directly reachable without public internet.
@tptacek's point upthread about controlling not just hosts but request structure is well taken - for AI agent sandboxing you'd want tight scoping on what the proxy will forward.
Physical systems that can learn by themselves
Andrea Liu
Brains learn and perform an enormous variety of tasks on their own, using relatively little energy. Brains are able to accomplish this without an external computer because their analog constituent parts (neurons) update their connections without knowing what all the other neurons are doing using local rules. We have developed an approach to learning that shares the property that analog constituent parts update their properties via a local rule, but does not otherwise emulate the brain. Instead, we exploit physics to learn in a far simpler way. Our collaborators have implemented this approach in the lab, developing physical systems that learn and perform machine learning tasks on their own with little energy cost. These systems should open up the opportunity to study how many more is different within a new paradigm for scalable learning.
The fallout from the hack of a little-known but pivotal health-care company is inflicting pain on hospitals, doctor offices, pharmacies and millions of patients across the nation, with government and industry officials calling it one of the most serious attacks on the health-care system in U.S. history.
[....]
Change Healthcare is a juggernaut in the health-care world, processing 15 billion claims totaling more than $1.5 trillion a year, the company says. It operates the largest electronic “clearinghouse” in the business, acting as a pipeline that connects health-care providers with insurance companies who pay for their services and determine what patients owe. It supported tens of thousands of physicians, dentists, pharmacies and hospitals, handling 50 percent of all medical claims in the United States, the Justice Department wrote in a 2022 lawsuit that unsuccessfully tried to block UnitedHealth from acquiring the company.
Wednesday, January 31, 2024
The Future of Artificial Intelligence
AI is all around us—recognizing our faces in photos, transcribing our speech, constructing our news feeds, navigating our driving routes, answering our search queries, and much more. But rapidly improving AI is poised to play a much bigger role in all of our lives. In this lecture, AI expert Melanie Mitchell will demystify how current-day AI works, how “intelligent” it really is, and what our expectations—and concerns—about its near-term and long-term prospects should be.
I'm surprised by the results (from what I've heard so far on the podcast) and a bit skeptical given the involvement of Meta. Looking forward to reactions from various scientific and technical communities.
> The AI Text Classifier is a fine-tuned GPT model that predicts how likely it is that a piece of text was generated by AI from a variety of sources, such as ChatGPT.
> This classifier is available as a free tool to spark discussions on AI literacy.
> Universal Summarizer is a tool developed by Kagi that is able to quickly and accurately summarize articles, videos, and other texts. It is able to pick out unique and interesting details from the text, such as Ratz having a prosthetic arm in the bar in the novel Catch 22. It is also able to summarize technical documentation and Moby Dick, as well as provide a summary of the themes in Walden. It is able to summarize videos as well, and is a useful tool for developers to familiarize themselves with academic papers. It is also able to provide summaries in multiple languages, and is able to detect jokes and humorous elements in the text. It is an impressive tool that is sure to be a hit with developers and researchers alike.
> Moby-Dick is a novel by Herman Melville about the obsessive quest of Ahab, captain of the whaling ship Pequod, for revenge on the white sperm whale Moby Dick, which crippled him on the ship's previous voyage. The novel follows Ishmael, a sailor on the Pequod, as he narrates the story of Ahab's monomaniacal pursuit of Moby Dick. Along the way, Ishmael and his fellow crew members encounter a variety of characters and situations, including the mysterious Fedallah, a harpooner who prophesies Ahab's death; Pip, a young African American cabin-boy who jumps in panic from Stubb's whale boat; and the Delight, a whaling ship that is badly damaged and with five of her crew left dead by Moby Dick. The novel also contains a variety of themes, including the nature of evil, the power of fate, and the search for identity. Ultimately, the novel suggests that while revenge may be a powerful motivator, it can also lead to destruction and death.
Not sure what to make of its summary of Moby Dick (as presented in the Project Gutenberg text version) -- https://labs.kagi.com/ai/sum?url=https://www.gutenberg.org/f.... The summary is full of platitudes. I'm wondering how they were generated. Did they actually arise directly from the text itself or from outside commentary of Moby Dick that Kagi had ingested?
> The story follows the adventures of a whaling crew led by Captain Ahab, who is obsessed with hunting down a white whale named Moby Dick. The crew includes Queequeg, a harpooner from a distant island, and Starbuck, the first mate. Along the way, they encounter many dangers, including storms, sea monsters, and the wrath of Moby Dick himself. The story is a meditation on the power of nature, the dangers of obsession, and the importance of friendship and loyalty. The message of the story is that even in the face of great danger, it is possible to find courage and strength in friendship and loyalty.
> A new study reveals how unconscious bias leads us to neglect negative externalities of driving. You may call it ‘car brain,’ but this research team calls it ‘motornormativity.’
Thanks for providing the detailed comparison among the many password managers. I think it's more accurate to describe 1Password's CLI as "yes" rather than "yes?poor" and submitted a PR for consideration: https://github.com/Soft-wa-re/password-manager-comparer/pull...
I submitted this article because I didn't know about the new(?) type of fraud around gift cards. From the article:
> Walmart's gift cards are worthless until customers load them with cash. Once loaded, the company requires shoppers to input a card's hidden security code when using it to make purchases online, but not at self-checkout.
> Wilson says a fraudster could easily take photos of a bunch of the cards' bar codes at Walmart, and then try to buy goods with them at self-checkout at a later date — in the hopes the cards have since been loaded with cash.
> "It's sort of, like, egregious," he said. "All the cards in Walmart are on bulk display. The bar codes are in plain sight."
The podcast episode struck me (as someone with a superficial understanding of crypto) as a devastating critique of crypto. This episode is a companion to an episode of the Ezra Klein Show featuring Katie Haun (partner, a16z), who is also a crypto booster: https://www.nytimes.com/2021/10/15/opinion/ezra-klein-podcas...
I'm hoping that this episode attract comments in the HN community.
Hit a snag: Sprites appear network-isolated from Fly's 6PN private mesh (fdf:: prefix inside the Sprite, not fdaa::; no .internal DNS). So a Tokenizer on a Fly Machine isn't directly reachable without public internet.
Asked on the Fly forum: https://community.fly.io/t/can-sprites-reach-internal-fly-se...
@tptacek's point upthread about controlling not just hosts but request structure is well taken - for AI agent sandboxing you'd want tight scoping on what the proxy will forward.