Often (most?) of the time "solving" a hard problem is partly taking it apart into separate pieces and phrasing those pieces in various novel ways that allow one to tackle them individually and then at the end it all magically falls into place (in this case after 900 pages).
I think that's really the key here right? It doesn't really "beat" Qwen3.8-27B.. more of like it's on par / a few points ahead on some, behind on others. But contrary to the Qwen (which is dense) it actually does it as an MoE - so less active parameters.
And the fact that they trained it as a looped model in 4-bit native is also kind of impressive to me as well, because it combines recurrence, routing, and narrow-format - all of which adversely impact numeric stability during training.
It beats Qwen on effectively every coding benchmark, which I think are the most relevant ones to HN :)
On the other benchmarks it trades blows or beats Qwen; as Thinky points out, looping helps reasoning but doesn't scale memorization, so tasks requiring broad world knowledge without tools/search won't be helped by looping.
Still super impressive, and probably replaces Qwen 27B in pretty much any workload!
you can also just make qwen looped / carry forward its reasoning from pass to pass. kind of does the same thing and you dont need to change the weights.
Also, for us Apple Silicon users, MoE models fit the hardware better: I have a 64G Mac, and non-dense models like qwen 3.8 next flash, Poolside.ai Laguna XS, and deepseek 4.1 flash 1017 are practical workhorses on hardware that I can afford. I liked inkling-2-mini when I was using it yesterday with pi as a coding harness.
I mostly use oMLX (with KV cache settings set to use external SSD for very large KV caches), Sushi for Qwen 3.8 next flash, and Local Coder for deepseek 4.1 flash 1017 (streams layers from SSD so it runs a little slow.)
It is a real nuisance having to use multiple runtimes. I run Inkling 2 Mini with oMLX that had an update 2 days ago for this model.
the moe-ness is orthogonal to whether something can loop as least as far as i can tell, or as least as far as i have tried. i have local dense and moe models looped in the way that this trained one does, qwen3.8-27b does it fine and thats good, but the model in question here misses the link that would make the looping kind of neither here nor there, which is you could avoid doing all that fancy loopy machinery altogether and get several orders of magnitude more benefit just by carrying the state all the way through decode between forward passes. i mean you may as well do both i just mean the bigger and lower hanging fruit is on the table here.
when you do the carry forward link which is just a single save-and-write qwen goes from being the guy in memento without any notes to actually pretty brilliant, i havent looked into what it takes to get a clean benchmark on it but i assume itd be as easy as cloning a repo and hitting go, so maybe i'll do that and see what the numbers actually say.
It's anti-ai because people didn't understand code in many (most?) companies already before agents wrote code. Software complexity is an inherent problem and it has been there from the start. Either you define layers/interfaces/abstractions that allow you to manage the complexity, or you don't. Yes, it gets worse if you don't, but blaming the tool for that is an excuse.
I find physical exercise, not smoking, and talking to people very enjoyable. These things in particular keep me sane and reassured.
When I was an obese smoker who avoided people, I was stressed out by a lot of other things going on. As soon as those stressors were gone, the bad habits also went away without so much effort.
It's worth considering how many of these perceptions of "discomfort" are caused by other aspects of culture, not nature or nationality or era. All my stress was coming from trying way too hard as an insecure younger person.
That doesn't mean you can hit it too. The only case which I remember was well documented was for a turret gun "controlled by satellite" hit the passenger of a standing vehicle from the back of a truck in front of it (nuclear scientist Mohsen Fakhrizadeh, killed in Tehran in 2020 while his wife, sitting immediately beside him, was unharmed). Also those reports were pretty questionable at the time, there were multiple competing narratives, and I'm not sure the satellite gun story was ever independently verified.
I'm pretty sure that is a mis-characterization of what actually happened: a 3rd party app collected user profile data from FB _through somebody's friend connection_ [1].
The Zuck then apologized for this undocumented API access leak and FB subsequently patched it out of their platform. It had nothing to do with running ads. We shouldn't let time paper over the details here.
I know someone at FB who worked on a project whose entire goal was to use compiler/type checker technology to prevent intentional and unintentional leaking of personal data in FB back-end code. If you can believe that. The project had a fair amount of resources and high-level support.
I don't know if it was a psy-op or deception campaign or what.
I think you're thinking of something else, but they also added "capabilities" to Hack (their main frontend language; it's a PHP fork) to make leaking data between requests or in logs painful.
> But when reflecting on the biggest mistakes of his career, Zuckerberg said his largest one was a “political miscalculation” that he described as a “20-year mistake.” Specifically, he said, he’d taken too much ownership for problems allegedly out of Facebook’s control. “Some of the things they were asserting that we were doing or were responsible for, I don’t actually think we were,” said Zuckerberg. “When it’s a political problem… there are people operating in good faith who are identifying a problem and want something to be fixed, and there are people who are just looking for someone to blame.”
TechCrunch then goes on to blame Zuckerberg by insinuation. As someone who dislikes Facebook and Zuckerberg, I think Big Tech is a rather convenient boogeyman for fallout from the 2016 election & beyond.
There's also the fact that all parties involved benefits from overstating the story. Facebook doesn't hate the reputation that their impact is so powerful as to sway elections. Cambridge Analytica and whatever they turned into has a reference customer like any other. Those who lost the election has an explanation that doesn't involve having to change.
Zuckerberg sounds a bit entitled in that interview. Yes, it was a chore having to appear in the senate and all, but he would probably not hesitate to take that same chance again if he could. As they say, all attention is good attention, and that probably goes double if you personally control the biggest social network on the planet.
reply