They have a bit more info on their announcement blog post[0]
> Belfort today released the "so far" CIFAR demo, an encrypted implementation of ResNet-20, a popular model for image classification. It outperforms recent SOTA by 3x with a total latency of less than 200ms
Not many details on how they've done this, so I'm a bit skeptical. Fast HE is a holy grail.
> Belfort's image classification is built on top of its upcoming GPU library, Cyclops. It comes with several optimizations that make Cyclops extremely fast on Encrypted AI workloads.
Looks like a lead up to an upcoming library release
Thank you Sajarin mentioning my Gork-Build fork, I wanted to address some of the comments.
I agree that many of the responses in the comments are valid. It is a harsh reality that 80% of projects like this fail to gain traction and eventually fade away. However, I believe the significance of such a fork lies more in its existence as a statement. Regarding xAI, even if their current release of Grok has telemetry disabled by default, Zero Data Retention remains a feature exclusive to enterprise users rather than individuals. And Whole-repo research packaging is still controlled by their server-side settings; it isn't an option you can toggle within the software itself.
I am currently implementing more fences to prevent unnecessary data from being uploaded, in terms of long-term maintenance, one person certainly cannot build something on the scale of VSCodium, but I have drawn a lot of inspiration from that project. In the future, I want to automate Gork-Build further by turning these privacy protections and guardrails into patches. These could then be applied to new upstream Grok-Buil versions as they are released.
As for whether this project can become a daily driver for everyone, I don't think that is the primary concern. If you need a open source coding agent, you should definitely use Pi or OpenCode, there is absolute no necessity to use Grok-Build for non xAI models in the first place. But again I think its existence is vital. People need companies that demonstrate a truly open attitude and coding agents that are genuinely friendly to the open-source community, and while xAI's decision to open-source Grok-build was a great move, it isn't a community-maintained or community-built project. It remains a public snapshot of their internal monorepo, and they have disabled issues and pull requests. This is precisely why a fork like https://github.com/thedavidweng/gork-build needs to exist.
all those youtube videos people upload nowadays aren't worth it, we already have keyboardcat (^ basically telling ppl creativity is done with, don't bother)
The parent comment is just pointing out that LLM written forks pushed out within hours of a “buzzy” repo release on GitHub are a pretty useless signal for gauging actual adoption/interest.
Which is, IMO, accurate based on the state of the AI dev space in 2026. Stars/forks drafting off the hype from a well known name are constantly gamed for eyeballs/personal brand-building courtesy of free advertising via the Github UI when the only cost is a few sentence prompt and some tokens.
Maintaining a fork costs you mental space, time and energy, even if someone else (i.e. AI) can reliably do all the work. (In my experience they're not quite there yet.)
I've been contemplating that recently. You're of course correct that subsidized tokens won't be forever, but that might only be half the story, since there's two opposing forces in action:
1. Phasing out of subsidized tokens.
2. Token prices being brought down through scaling, better hardware, etc.
It's possible that these might balance each other out sufficiently that token customers won't notice any substantial increase in price.
That was yesterday's "LLMs might". Time passes. Nothing stays the same. "Local models might" X Y or Z today has no influence on the limitations of tomorrow's local models except to remove them. Yesterday's LLMs are the exact same thing, except your computer is connected to their local model for you to use.
Disc drives used to be measured in megabytes—now in terabytes. Technically useful tend to get more optimized with time, not less.
I use both actually. Anki for the reviews since the spaced repetition is hard to beat, but I use Norsha Notes (norshanotes.com) to generate the cards. You upload your notes or study material and it creates flashcards from them using AI, then you can export as .apkg and import straight into Anki. Saves a ton of time over making cards manually.
I had used Anki for a few years but recently migrated all my cards out of it and into a custom app I built just a few months ago. It is as an Elixir/Phoenix app with a simple UI but also with a rich API for Agent integration.
People aren't good at detecting AI generated/edited comments, so unsure how effective this policy will be. Though I guess there are still some obvious signs of AI speak like emdashes and sycophantic (it's not X, it's Y!) speech.
Bit of a shameless plug but I wrote a HN AI comment detector game[0] with AI and most of my friends and fellow HN users who tried it out couldn't detect them.
Something I've noticed through moderation is that people are much more easily duped by generated comments if they like the content and/or agree with the point. We've seen several cases where a bot-generated comment has been heavily upvoted and sits at the top of the thread for hours, and any comments calling it out for being generated languish at the bottom of the subthread below other enthusiastic, heavily upvoted replies. This shouldn't be surprising, given what we've seen of LLM chatbots being tuned to be sycophantic, but it's interesting to see it in effect on HN.
This is another reason why it's good to email us (hn@ycombinator.com) rather than commenting when you see generated comments.
Having been reading generated comments almost daily for over three years now, I have a pretty good sense of it. There's a bunch of signals: how new the account is; how the comments look visually (the capitalization and layout of the paragraphs, particularly when all of one user's comments are displayed in a list). Em-dashes and short, emphatic sentences, make it more obvious of course.
There are cases that are more borderline; usually when someone has used a translation service or has used an LLM to polish up a comment they wrote themselves. For these ones there's less certainty, and whilst we discourage them, we're not as rigid in our aversion to them or as eager to ban accounts that do it.
But ones that are entirely generated are still pretty easy to spot, even just from visual appearance.
It’s certainly hard to detect in isolation, but the thing that gives it away is the comment history.
All the AI acounts I’ve seen repeatedly post the exact same cookie cutter top-level comments over and over again. Typically some vapid observation followed by an obviously forced question serving as engagement bait. The paragraphs and sentence structure even looks visually similar across comments when you scroll down the history page.
Just look at a few of these accounts and you’ll easily be able to recognize AI posts on your own.
Some of us were trained/self taught to write that way. Even "it's not X, it's Y" is a legitimate and subjectively effective communication tool, and there are those of us who either by training modeling have picked it up as a habit. It's not Ai that started this, Ai learned it from us.
Crap - I just did it, didn't I? Awww double crap! Did it again...
Forums and comments are not written as formal novels or text. Corporate-speak is also not typically used in these environments unless you are representing corporate.
So I think it's fine to scrutinize commenters who write that way.
Besides, the biggest offense of AI speak is making everything seem like a grand epiphany and revolutionary discovery. Aka engagement bait.
That's just French for "masterful" or a way to describe lectures. There's a sense of greatness in that word that contrasts with the Mini in Ministral which is in turn might be a pun on "ménestrel" (minstrel), "ministre" (minister), or made to sound like Minitel (or all of the above).
psychosis.hn is a daily game. Every day we fetch three stories from a previous front page of HN, each with 5-7 AI comments threaded into the discussion. They have personas, reply to real people, and sometimes have real comments reparented underneath them.
> Belfort today released the "so far" CIFAR demo, an encrypted implementation of ResNet-20, a popular model for image classification. It outperforms recent SOTA by 3x with a total latency of less than 200ms
Not many details on how they've done this, so I'm a bit skeptical. Fast HE is a holy grail.
> Belfort's image classification is built on top of its upcoming GPU library, Cyclops. It comes with several optimizations that make Cyclops extremely fast on Encrypted AI workloads.
Looks like a lead up to an upcoming library release
[0] https://belfortlabs.com/blog/sofar