alphazard

There's a whole lot of information on this page that doesn't tell me anything about what this actually is. Can anyone spell out what the architecture is here?

They claim it's not an LLM, which I read as "not an auto-regressive token generator". I assume they are still using a transformer, otherwise they would be talking about the thing that's not a transformer, instead of all the fluff on the linked page. But they emphasize parallel generation, so is it like a text diffusion model?

> Democratization of access has its victims.

Democracies require conserved stake to function. How individuals initially get that stake is a degree of freedom, but if the stake can be printed freely, the system is unstable.

For working implementations of democracy, look at: invite trees, shareholder corporations, proof-of-stake distributed ledgers, etc.

This framing of security as something that belongs in the harness is completely wrong, and I hope no one is relying on a correct harness to keep their agents isolated.

VM, or even just a container will do. The agent should be able to run as root in its environment and do whatever it wants. If you can't give it that, you aren't sandboxing correctly.

> Wouldn't the democratic way be to petition...

This is also the democratic way, arguably more democratic, it's just probabilistic. Juries are supposed to be randomly sampled from the jurisdiction's population (a kind of democratic process). The ability of a jury to judge the law, and not just the particular crime is a feature, not a bug of the justice system. This is how the system is set up, and it is as legitimate as action through a legislative process.

EDIT: the parent asked a legitimate question in good faith, there is no need to downvote them because you disagree with their implied current understanding of an issue.

Not sure why your comment is being downvote. That is in fact the best explanation for what is going on here, and the outcome has little to do with the competence of the prosecutor, as is perhaps alluded to by the original comment.

It's good to see individuals on a jury acting according to rational self-interest.

Humans have been inserting benign noise into their writing without affecting the signal for thousands of years. They called it "style".

Apparently superfluous descriptions, unnecessary words, and paragraph footnotes are all fine, but gumbel softmax (or whatever's going on here) isn't.

I should have expected this level of discourse from 2026 HN.

Here's another way of framing things that might encourage better discussion: Suppose in a few years, it was announced that an LLM did discover a cure for a cancer. What approach would you expect it to have taken?

1. Developing an intervention for every possible mutation that leads to loss of cellular identity, and uncontrolled growth.

2. Preserving immune function later into life.

3. Preserving genetic and epigenetic accuracy later into life.

Maybe it will be 1. But I would bet on 2 or 3. Which both assume that by reducing the difference between rates, the phenomenon of cancer could be reduced or eliminated.

"Cancer isn't a single disease" is often shouted from the Dunning-Kruger peak of cancer research. I think it was popularized by an explainer video on Youtube, and now it's being memed. The people who think this way want to legitimize funding a particular research niche that they benefit from.

A more sophisticated view is that cancer (one idea) is an problem of rates (like most things in biology) and there is a tumor burden per unit time from whatever causes vs. the immune system's capacity to detect and destroy them per unit time. As humans age, the balance tips in one direction. If the balance is too far off for too long, tumors accumulate, doctors take notice, and assign whatever labels to the condition.

Looking into particular causes of tumors (and calling each cause its own cancer) can create a vast research program that employs a lot of people, since there are so many mutations that lead to tumors. But interrupting the development of a particular mutation is not going to have the same society-changing effect that solving the fundamental issue of rates would.

The answer, of course, is that the taste-making apparatus inside the typical tech company is now entirely imposters (non-technical, non-power-users), who continue to LARP as visionaries.

The people who notice broken or degraded behavior (devs and power users) and conceive of good designs are not in charge of what goes into the product. So you shouldn't expect it to get better, only to continuously change so that the imposters can point to a difference that they were responsible for, knowing that upper management also has no idea whether that difference was actually an improvement.

The solution is to gut your product org, replace management with leading engineers, and hire some of your most fanatic users. They don't need to do anything other than give their opinions, and use the product every day. Put them in a room with your top engineers, and the software will mysteriously get better.

All of this was obvious decades ago, but now everyone wants to have a do-nothing job at a tech company, and to help their friends get one too. The result is imposters everywhere; "imposter syndrome" was brought into the vernacular to normalize a lack of expertise. "No one knows what they are doing, it's okay to not know how to do your job". That's what your (product) manager tells themself in the mirror each morning. Turns out: skill predicts quality, and an org ships quality proportional to the skill of the decision makers.

I agree with everything you said. Most open source projects are limited by contributor labor. Generative AI does help with that problem, but it introduces a sea of vibe coded slop as a side effect. Truly a Genie/Jinn.

> And no, HN is not social media in any normal sense of the word. The pedantry involved in that comparison is extremely tiresome.

Whatever you want to call it, HN has followed a similar trend. It's rare to see the authors of small, but interesting/innovative, projects show up in the comments, surprised to see their work on the front page. That used to be common, even the default, if you look far enough back.

Now the front page is current events and marketing campaigns. I don't think I've seen a single software project here in the last year that wasn't already extremely popular, or being pushed by a company with a marketing budget.

In theory AI should have helped. I know people are still making cool stuff, faster now with AI, but it's harder and harder to find it.