3s

Confidential computing as a technology will become (and should be) commoditized, so the value add comes down to security and UX. We don’t want to be a confidential computing company, we want to use the right tool for the job of building private & verifiable AI. If that becomes FHE in a few years, then we will use that. We are starting with easy-to-use inference, but our goal of having any AI application be provably private

this is really cool! congrats on the launch

Usually the unnecessary use of em dash and dramatic prose is an indication, which is what I think the comment may have been referring to. but many humans write that way too

I agree. My point was more in terms of surface area. Sending data to a party with a reputation to uphold and resources to do proper data handling is still better than sending data to 10+ parties with little or no reputation and/or resources to implement security measures.

The problem too is that while google or the “big players” can be somewhat “trusted” with security, I don’t trust a random third party startup to properly handle this extremely sensitive data. The sheer number of places AI related data is being sent for processing is staggering and unprecedented.

For instance, when I program with Cursor, that data is sent to Cursor severs, then to their tooling and analytics third parties, then to the foundation model providers, then gets trained on, etc.

The problem with all anonymous communication systems is that they suffer from network effects even more than traditional communication systems. By having few people use it you immediately expose yourself by being part of a small group of people using it, becoming an easy target. Unfortunately the reality is that it’s often easier to hide as a needle in a haystack on something like WhatsApp than it is to use a theoretically anonymous communication system that lacks the “haystack” altogether

This is an interesting idea, and indeed such an approach to providing privacy has been formalized to different degrees and varying levels of success (eg. [1][2]).

[1] https://arxiv.org/abs/1204.2136 [2] https://arxiv.org/abs/2210.03458

Unfortunately, as described, such a solution would only satisfy a somewhat meaningless notion of privacy. Specifically, the embeddings by definition contain potentially private information about the user, revealing things like "I'm asking about birds" to use your example. Even though it might "compress" the query in a slightly lossy way, it would still reveal a great deal of information about the query.

A true solution to this problem would require something like differential privacy and adding noise to the embeddings. However, the noise required would (likely) end up destroying too much information from the embedding to preserve accuracy of the LLM.

Yeah but as one of the replies points out the resulting tokens/second would be unusable in production environments

For a lot of tasks like frontend development I’ve found that a tool like cursor can get you pretty far without much prior knowledge. IMO (and experience) many tasks that previously required to hiring a programmer or designer with knowledge of the latest frameworks can now be replaced by one motivated “prompt engineer” and some patience

Exactly! While I have llama running locally on RTX and it’s fun to tinker with, I can’t use it for my workflows and don’t want to invest 20k+ to run a decent model locally

> How are HN users handling this? I’m working on a startup for end-to-end confidential AI using secure enclaves in the cloud (think of it like extending a local+private setup to the cloud with verifiable security guarantees). Live demo with DeepSeek 70B: chat.tinfoil.sh

We just applied and we were wondering the same thing too! As others mentioned here, it looks like the applications aren’t posted anywhere even if you get accepted.

A really important application of reproducible builds is running code inside Secure Enclaves that has been committed to on a public transparency log. A client can connect to a remote secure enclave that can then prove to the client that it’s running the commit code via a process known as remote attestation. It’s pretty cool stuff. However it’s only possible if the build inside the enclave is reproducible (deterministic) and always identical to the build on the transparency log

Are people concerned about the privacy implications of computer use at all? This is why I haven’t been using Claude computer use personally. Somehow the idea of sending everything I do on my computer to a random third party seems creepy. There are a lot of applications of AI (rewind comes to mind) that I simply cannot accept the idea of sharing my screen with

Curious to know why browser fingerprinting has a “half life” of 24h? I always assumed it would be more static compared to an IP address and therefore a more powerful tracking mechanism (albeit more shady / less legally understood compared to using IPs).

SHE vs FHE has nothing to do with security. Instead, it’s about how many operations (eg homomorphic multiplications and additions) can be performed before the correctness of the scheme fails due to too much noise accumulating in the ciphertext. Indeed all FHE schemes are also SHE schemes.

What typically makes FHE expensive computationally is a “bootstrapping” step for removing the noise that accumulated after X operations and threatening correctness. After bootstrapping you can do another X operations. Rinse and repeat until you finish the computation to you want to perform.

While not exactly software per-se, I created a system of multiple text files to manage todos, long term goals, and various reminders (eg, IOUs, deadlines, etc). This was inspired initially by Jeff Huang’s blog post [1] but then grew to a complex collection of different files. A problem I ran into was building an interface for displaying and editing these text files (each file has a different width and for some files I want to have different heights when editing them). Ultimately I settled on multiple vim tabs in a terminal window. Been using this for close to five years now and I couldn’t be happier with it. However, at this point the system of files (and the terminal “user interface”) is completely customized to my life and would likely never fit someone else’s requirements.

[1] https://jeffhuang.com/productivity_text_file/

I disagree. Avoiding PKI and post-quantum security correlate very much. Even under plausibly post-quantum assumptions we only have a couple of assumptions from which we can build public key encryption. In contrast, here they avoid all use of public key cryptography which makes it provably post-quantum secure. It’s not using a buzzword for the sole sake of selling the paper. In general, using “minimal cryptography” (like random oracles / one-way functions) translates to real-world efficiency because you can instantiate these from a plethora of different concrete candidates.

Paywalled. Here’s the archive link: https://archive.is/aSU79

Check out Googles Topics API [1] (available in chrome). It spits out a list of advertising topics based on your browsing history. [1] https://developer.chrome.com/docs/privacy-sandbox/topics/

In the article: “The odours defied the expected effects of dilution since workers in the laboratory did not find the odours intolerable.”

I think there has to be a component of “jadedness” in this attitude, for lack of a better term. It feels as though there are so many problems in the world (income inequality, climate change, covid, homelessness, political polarization, and now the very real possibility of a nuclear war). The world doesn’t feel like a place we can save anymore, which I think can be a driver of the “bring it on” attitude you mention, not unlike the Monty python scene with the defeated knight…

You're absolutely right. It is not enough to use anonymity tools, you also have to make sure everything else around you doesn't compromise your anonymity. Made me think of a Harvard bomb threat incident where the student posting a fake bomb threat (through Tor) to avoid final exams was the only person using Tor on campus at the time, which trivially identified him.

https://theprivacyblog.com/blog/anonymity/why-tor-failed-to-...

I got into computer science by making video games for the first couple generations of the iPhone; now I’m a phd student writing papers. Things were fun and simple back then and my games were quite successful. There was something so magical about it. However, the magic was that people were actually playing my games and I could focus on making the game rather than dealing with all the crud that one has to do to ship an app now. While I still get the urge to make a mobile game sometimes, there is little real life “reward” in it beyond making something beautiful and maybe showing it off to a couple friends. Gaining visibility on the AppStore is next to impossible without turning it into a full time job (which was never my end goal). I think if there was even a small but non negligible chance that say 50k people would download and play my game (as I had with some of my old games) I would get the motivation to turn some of my ideas into reality. However, the way things are now, I would be lucky to get 1000 downloads and that would require me developing for all the different interfaces, versions, etc. making it really difficult to focus on “just the game.”

Unfortunately, this is standard practice across schools (and most big employers). This is definitely not a problem unique to Brown, or even elite schools.

There was a scandal back in 2010 where a high school in Philadelphia was accused of spying on kids through school-issued laptop webcams [1], which is arguably more egregious.

https://www.computerworld.com/article/2521075/pennsylvania-s...

The NYTimes [1] did an excellent expose on this a year ago. They purchased location data on 12 million phones across the country.

[1] https://www.nytimes.com/interactive/2019/12/19/opinion/locat...

This reminds me of #SmashTheStone from a few years back when the 4chan community came together track down and prevent 9gag from burying a meme-carved rock. Video from the internet historian: https://www.youtube.com/watch?v=nzFykQv6Q08

Very cool idea and beautiful design execution!

One nitpick: it would be helpful to only have to type out the words on the page rather than different symbols (maybe having it be a toggle). For example, in The Art of War, typing out all the bullet points and numbering can be tedious and not really what I'm looking for from a typing exercise.

WebTorrent [1] can be used to run BitTorrent on the web. It's actually integrated into the Brave browser which is a nice feature. Other browsers might catch on soon.

[1] https://webtorrent.io/