Balinares

Albert Einstein is the world-famous physicist. This article was originally published in the first issue of Monthly Review (May 1949). It was subsequently published in May 1998 to commemorate the first issue of MR‘s fiftieth year.

Whenever I see something in OCaml I assume it's Jane Street and that ends up correct a surprising amount of the time.

What's Stroustrup up to these days though?

So, if I'm reading this correctly, Chat Control [2.0, implied] is bound to become law?

Nope. This is bad, but not THAT bad.

This is an extension of the existing Chat Control 1.0, which was set to expire (or maybe already has, I didn't keep track). AIUI it gives chat companies permission to scan user chats for illicit content, but does not mandate it.

This is bad, but it's not the much worse still Chat Control 2.0 that was defeated several times already.

A hundred thousand furries consuming unfathomable amounts of porn without shooting up anyone kind of cast doubt about that point.

I have a hunch that the Epstein class is getting increasingly upset about the kids encountering ideas about what ought to be done about the Epstein class, and mostly are keen to see the next generation molded back into good little subservient laborers. It really isn't about the well-being of the kids.

Mainly I disliked the walrus operator because Python already has a keyword that means "capture the output of this clause into this variable": as.

E.g.,

    try:
        with open("file.txt") as f:
            do_something(f)
    except OSError as e:
        print(e)

    L = [abs(x) as x_abs for x in some_list if x_abs < 10]

Ok, I gave the 35B MoE weights a shot at Q6. It failed my go-to code reasoning benchmark (local models usually do), but delivered the most advanced output I've seen so far for my (non-agentic) coding execution benchmark. It was also full of bugs, though more in the form of coding mistakes than the deep logic bugs I usually see in local models. The code was also a bit haphazard, and some of those bugs could have been prevented entirely with a better code structure.

It was fairly good at diagnosing the bugs once informed of their symptoms. However, if I mischaracterized the symptoms, it would weigh my input too heavily and reject its own (correct) hunch about the root cause.

So it's an interesting one. There's definitely some latent capability in there that arguably exceeds Qwen 3.6, which is absolutely no small feat. But that capability seems to come in a somewhat erratic package.

It's probably worth benchmarking it unquantized if you can. I've grown to suspect that quantization damages small models more than perplexity and KL divergence accurately reflect.

I'll also give the 9B weights a shot when I can.

A 9B model with the capability of the 35B SOTA from last February is too good to be true, and a wild claim to make IMO even if there's a newer 35B SOTA. I'll need to make time to take it on a test drive and see how it holds up.

I'd have expected this to get more HN attention. Qwen 3.6 35B capability in a 9B model is a bonkers claim.

The interesting thing here is that Amodei is implicitly admitting he foresees open-weight models (and not "open-source", please...) reaching Mythos-equivalent capability before much longer.

Which, I mean, d'uh; if you've been paying attention it's not a wild forecast.

But it's interesting that he's admitting it.

According to US AI labs, training on other people's output is fair use. So that's how.

In fairness, goroutine is far catchier than >>=<%>.

Thank you for helpfully de-darmoking that for the less pony-attuned of us.

I mean, there's such a thing as reality and there's such a thing as being wrong.

By construction, the closer you approach reality, the narrower your band of uncertainty becomes, and the less overlap you have with viewpoints that are just plain wrong. Doubtless their proponents will call that biased.

Burning even more energy in response to the consequence of burning too much energy is seen by some as counterproductive.

Do also note that AIUI in the US only deaths that were officially ruled from heat are counted as such by the CDC (source: https://www.nytimes.com/2024/08/23/us/extreme-heat-deaths.ht...), whereas the EU usually performs estimates based on overall deaths in excess of the expected baseline. So it's not apples to apples.

I do agree that avoidable deaths should be avoided, though, so I'm with you on that.

I can't help wondering what kind of models we'll see coming out of China once it gets its own chip fabs up and running. Right now it sounds like the US's export ban is not slowing them down a whole lot.

Presumably they're from one of the happy countries that you believe don't exist. https://en.wikipedia.org/wiki/World_Happiness_Report

In the original implementation, SREs have the authority to block further releases until the SLOs are green again -- and, if it comes to that, to hand the pager back to SWE until such a time the product can pass a production readiness review again. That requires a lot of trust on both sides, though, the building of which takes both time and deliberate effort, and I don't know how other shops do it.

After using i3 for a heck of a long time, I'm finding myself liking KDE quite a lot. The amount of obvious care and polish that went into it is delightful. I just wish it had support for tabbed window layout!

Whenever someone wants "small government" I assume they mean "a government that does not prevent me from shackling you" and hoo boy does that end up correct a lot.

Given that the "proprietary" model is built on stolen work at an unprecedented scale, it's at the very least hypocritical to a degree that would not be possible without a fundamentally amoral mindset.

AdGuard MV3 works fine. Still switch to FF if you can, more diversity in the ecosystem benefits everyone.

code is just a means to an end?

Exactly. Which is the reason why code that increases maintenance friction and costs, or decreases the product's operability or fitness for purpose, does not properly serve the end of advancing the business. Not in competitive industries anyway.

There's this whole manufactured narrative going on geared toward reframing the quality of code design and structure as an aesthetics problem. If you think code design and structure is an aesthetics problem, you will not understand the pushback.

A much more interesting question, I think, is why bad AI code is perceived as a worse problem than bad human code.

My observation is that bad human code ultimately slows down its own maintenance: the speed at which the human can add bad code decreases as the human adds more bad code. So it's a self-limiting problem. Whereas AI will quickly and efficiently add extra layers of bad code to work around the issues introduced by its own earlier bad code. If you let that go on, you can end up in runaway complexity scenarios where only AI can maintain the code at all, until even it can't. There have been a lot of arguments on HN that such scenarios are not a problem because AI will keep getting better, but to me this amounts to a Ponzi scheme for tech debt, which I find undesirable.

And that's not to even mention the externalities.

So, basically, depending on whether you care about all of the above, or not, the machine that supercharges your LoC-per-day metric will be seen as a good thing or a bad thing, and it's not a divide that can be reconciled because it boils down to what individuals care about. We can all agree about the end of writing code, but that doesn't mean you can't be wrong about the ways to get to that end, I guess.

Meanwhile down here in the real world the US double-tapped a primary school.

They (correctly IMO) seem to think that the hard thing in sourcing the top candidates by the sort of metric they care about, is getting them to even know about you as a potential employer. That means figuring out where those candidates might be and finding ways to get your name in front of their eyeballs.

So, cool nerdy blog posts on HN, cool nerdy puzzles on Cracking the Cryptic, etc.

Having been sorta involved in hiring in the same sorta market, I can respect how astute their approach is.

We're a GitLab shop. The GitLab UI lets you read a PR commit by commit.

Overheard recently: "Thanks to AI we're producing more code and more MRs, faster than ever, but the milestones aren't getting hit any sooner. Actually the opposite, if anything."

I wonder how widespread that phenomenon is. Perhaps it's no wonder the prominent actors are trying to rush to IPO...

So one of the prominent LLM advocates known for testing every model shared a prompt intended to exhibit Opus 4.7 capabilities, and Qwen 3.6 sorted it out okay? Interesting.

Not saying they're equivalent, local models still decohere much quicker as the context grows in my experience. But... Interesting.

Some do, but we should not be level-setting at mediocrity.