daemonologist

There is no "coder" version of Qwen 3.6; I think they just mean it's a coding-focused model of similar size and performance (to Qwen 3.6 35B-A3B).

Regular Qwen 3.6 benchmarks slightly better and has much wider software support though, so this is probably of interest only to organizations which disallow models trained in China.

The allegation here is that it's not actually a fine-tune of Qwen, but instead an undisclosed mashup (merge) of someone else's fine-tune of Qwen and the original model. Rio subsequently said that the model was in fact a merge, that they did additional fine-tuning after the merge, and that they accidentally uploaded the base merge instead of the version with additional fine-tuning. But this seems like quite an oversight...

There are also significant economies of scale (namely: utilization and batching), which tend to make inference on a shared server more economical even after the operator takes a cut.

You can bounce the ball up slightly (presumably the spin from rolling is modeled or approximated, and gives lift when hitting a bumper), which might be enough to skip from the tee to near the end of the course. Not sure that should be considered for "par" though. Took me 14.

Opus 4.7 and 4.8 are also rather "proactive" - several times I've seen them try to inspect compiled binaries before there's even a problem, just to check that their changes are included (and if I let them do so they often get stuck down that rabbithole).

I admit I snorted when that was mentioned. It's frequently ranked as the most desirable place to live on earth.

Not to say the message of the article is completely without merit - there are things to see and do almost everywhere. But if I just get in the car and start driving I will 95% of the time find only strip malls and cornfields. Perhaps a suburban park with some trees.

Unfortunately Radxa and Milk-V are almost completely out of stock and not much cheaper. If you need more than a microcontroller there's no circumventing the memory shortage at this point.

Kicking myself for not buying the Q6A at the beginning of the year (I wanted three and arace would only sell one per customer, but one would've been better than none).

In the US, 99th percentile household wealth is ~$14M, which at historical rates of return is enough to live opulently indefinitely. (Of course although we're discussing a scenario where capital holds most of the cards, who knows if those returns would be dependable.)

lol

    > can it be slower than without speculative decoding in worst case then?

Yes - running the draft model costs compute and memory bandwidth, and running the drafted futures through the main model costs compute. If the draft model were really inaccurate or you're already compute-limited (usually: running large batches) you would expect some slowdown.

In practice, for single-user (non-batched) inference with a working configuration, you pretty much always get some speedup. For non-coding tasks I've seen it be nearly a wash for some people, in which case you might want to avoid it due to the extra memory usage (you'd rather use that memory to run a bigger quant/model, even at a slightly lower speed).

The "library" UI has also gotten radically worse over time (in my family there is a 3G, an early Paperwhite, and a relatively recent base model, and each has a worse and sparser UI than the last). The pages turn faster though, due to improved display/display driver tech.

The tiles are not supposed to ablate - they're supposed to be ~fully reusable. That said I think it's plausible that the much higher iteration speed and lack of a need for human-rating (at least during reentry, for now) will allow for more success than the space shuttle saw with its similar approach.

uv has a lot of great features, but the dependency resolution is why I'm a fanboy. It can resolve trees that pip gives up on, and it does it 20x faster than poetry (100x faster than pip) - saves me half an hour on some big projects. All the python resolution and environment management and stuff is just gravy.

They have a subsequent post (from Monday) about what they've been working on: https://rosmine.ai/2026/05/18/fixing-llm-writing-with-distri...

(I would assume they haven't made a lot of $ off of this, if nothing else because they've only just put out that post and demo. They do seem to have produced a model that doesn't sound very LLM-y to my ear, though it also seems rather weak for its size.)

I wonder about this when I see someone post their own work without the Show HN prefix - is it always supposed to be a Show? (Enforcement/community objection to the lack thereof doesn't seem to be very strenuous, if so. Or, maybe it gets fixed after a little while and I haven't noticed.)

If this is accurate it raises the question: why is this model so expensive? DeepSeek v4 Flash is 284B total/13B active, FP4/FP8 mixed, and only costs $0.14/$0.28 - even less from OpenRouter. Of course Gemini 3.5 Flash is most likely a better product, and therefore it can command a higher price from an economics perspective, but does this imply Google is taking roughly a 90% profit margin on inference? If so they're either very compute-limited or confident in the model and wanting to recoup training/fixed costs (or both).

Looks like Flash 3.5 is GA ("stable"): https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flas...

This is a problem with all of devops imo - everything is a magic yaml config file and they're very difficult to debug or reason about unless you _just know things_.

Yes - I was thinking about starting my own business but am staying put instead and saving as much as possible.

And in my experience this means you usually have to go to both the DMV and then across town to the tag agent.

The consumption aspect is perhaps similar, but the crowds at Costco are much, much worse (in quantity mainly) than any other grocery or big-box store I've ever been to.

I also refuse to go to Costco these days. Every once in a while my memory fades and I agree to accompany a family member or friend, and am quickly reminded why I should stick to Aldi.

We quoted that book for years (probably because the accompanying audio version had a somewhat amusing cadence, but I do also think it was a lot more beneficial to learning than trudging through classical texts with a dictionary).

I theorize that this kind of thing - verification approaches that do already exist but which a human would not usually reach for - is a result of RLVR. I've noticed Claude Opus a couple of times trying to use `nm` to check that its changes made it into the binary (and often tying itself into knots in the process), and have stuffed some extra coaxings along the lines of "don't try to be clever" into its prompt.

"Special tool" is not used in the actual regulation; the requirement is that replacement must be possible with basic tools, defined:

> (50) 'basic tools' means a screwdriver for slotted heads, a screwdriver for cross recess screws, a screwdriver for hexalobular recess heads [Torx], a hexagon socket key, a combination wrench, combination pliers, combination pliers for wire stripping and terminal crimping, half round nose pliers, diagonal cutters, multigrip pliers, locking pliers, a prying lever, tweezers, magnifying glass, a spudger and a pick;

(Excepted devices can require "commercially available tools" which is defined exactly as you'd expect.)

https://eur-lex.europa.eu/eli/reg/2023/1670/oj

I think you're right, but journalists have gotta stop calling them ebikes. We already have a widely used term that fits them perfectly and is legally accurate - moped.

To retain and recruit talent, and protect from lawsuits (you usually have to sign a release to receive severance/a buyout). Generally only difficult to recruit and highly in-demand employees get these packages - most people over here would also kill for something like this.

Problem is that it's been heavily contaminated with people speculating about who the author is. It would probably be difficult to get an unbiased answer out of it (although who knows - it's crazy that it can do this at all).

Probably because it appears to have been written by an LLM (the post too, but I assume comments are more easily killed by flags (?)). (To be clear I did not flag it myself.)

This might actually be quite nice - the Blender Python API is currently very useful and very touchy. Lots of differences in behavior in headless mode which are hard to debug (because you can't open the GUI to see what's happening, because that changes the behavior).

You can use Copilot Chat* with basically any API provider, and if you switch to the VS Code Insiders build you can configure it to use literally any OpenAI API-compatible endpoint.

Other than that Zed has a similar experience which is pretty decent.

* By which I mean the good one, whatever it's called now - the part of Copilot that used to be a plugin and is now part of VS Code, not the thing that has always been part of VS Code.