And you actually traveled in Scotland and Wales but England was synonymous with UK for you. Technically incorrect of course.
B) "I travelled around Britain"
And in actuality you only traveled in England.
C) "I went to Europe"
As an American (yes, a US-ian) and you start in Poland at the Baltic sea and then via some nice Eastern European countries with other attractions like Kroatia, all the way down to Greece. But without further explanation and context most people would've assumed you probably traveled Spain, France, Italy etc.
Everquest looked sooooo good back then. Man! No more Meridian 59!
Why did you have to destroy my memories? :)
It's like looking at the first DivX encoded movie that was soooooo good vs. VideoCD/LaserDisc. And then you look at it today and it's so damn blocky and also in that tiny resolution. Ugh!
I just did the LoC measure that isn't "official" but actually closer to how AI would read it: I.e. simply actual lines. Never mind "context window".
Just in the actual source folder there are 1.9 million lines. That's not tokens. Lines. Not even LoC as per some official canonicalized definition. Various languages. ~15 year old code base.
We work with Claude just fine.
For variable measures of "fine" depending on how much devs just trust whatever Claude spits out ;)
Yes, a 15 year old code base written by hand obviously, not by a coding agent from the start.
At the same size I only heard of AI generated projects reimplementing compilers and browser engines which come with amazing tests written by hand, but even these projects generally don't become reliable enough to use.
Well we also have an AI first code base, new product 160k "LoC" by the same measure. Not even a year old.
That 1.9m code base, within the same time frame as the 160k was created? <10k by the same measure.
10k I can comfortably throw at a local agent on my system (see recent posts in my history). Whether I can keep it that small is another question I suppose lol!
It was actually great. I have like a non-AI box so to speak 8GB VRAM, co-incidentally from a gaming PC ...
All the previous models that were "frontier level, just try it!" but wouldn't run at all in agentic mode, including previous Qwens, just disappointed, period.
Then I ran then Qwen 3.8 27b and while it was super slow (4t/s) it literally one-shotted creating a usable "web search/pull" skill for `pi.dev`. while any other model previously just entirely failed to create anything usable even with actual guidance.
Since then I have actually gotten a gemma-4 12B qat 4bit quantized with a ~250MB MTP from unsloth to work with a 32k context "working" on this setup at 80-120 t/s. That's usable for private stuff on a co-incidental box!
It's still only 32k context and it's entirely dumb vs. our API paid at-work Claude Opus. But for entirely private local stuff it's totally workable without breaking the bank even after all these AI price hikes!. I bought this rig literally just for gaming a month ago.
Have you tried a mixture of experts model? Dense models have been quite slow for me, as I have only 6 GB VRAM. But with llama.cpp and --cpu-moe I get 200 t/s input and almost 30 t/s output with Gemma 4 26B A3B, which feels ok to use. Would be interested about your mileage there.
If I use the 12B Unified (dense) model I mentioned without MTP, then I get 37t/s, input ~700t/s.
It's all still quite frustrating in the end, like a Claude from a very long time ago by now but usable. If I want 64k context, I can't use MTP. I still haven't decided whether I'd rather have 37t/s but it's "less dumb" or I want MTP speed but it's going off the rails more. All of this is also with `-ctv q4_0 -ctk q4_0)`, which is not ideal. I'm actually right now contending with 35k context but using q8_0 KV quantization. More like 35t/s coz with those settings I can't use MTP.
But I'm not ready to go back to 4t/s. It's not interactive enough for me. That said, I had tried to use the Gemma E4B for example to have it build itself that websearch/fetch skill. It utterly failed, as did previous qwens.
I don't see a Gemma 4 26B A3B GGUF for download, but there is a gemma-4-26B-A4B-it-MXFP4_MOE.gguf that should fit into my overall RAM and then use lots of CPU like the Qwen 3.8. I guess I'll give it a try just to see the difference in speed though I don't expect anything "usable" out of that tbh.
Haha, OK, so it processes input tokens at 55t/s if started like the 12B. That's so slow that I don't even have the patience to wait for the context of the session I aborted mid-processing to test this, to load. Heck loading the system prompt alone is not fun.
That said, after that it seems to run at the same ~35t/s speed but, I was able to start it up with a 128k context (and no KV quantization), humm humm. I am now playing more with all the parameters like I did to get the 12B going e.g. I'm already up to ~125t/s for inputs now and with an 8bit MTP it's running about 45-85t/s output when its coding.
I run with `--spec-type draft-mtp --spec-draft-p-min 0.8 --spec-draft-n-max 8` and when it was that fast (during actual code generation) got some `draft acceptance = 0.91801 ( 3482 accepted / 3793 generated), mean len = 7.92`. I also tried 16 with no real improvement for those phases but it seemed slower, wasting too much with low 50% acceptances during other phases like thinking.
EDIT: Oh look, I can now use larger batch sizes again and that gets prompt_eval reads back to 700+
Regarding how good it is: Probably, maybe, better than the 12B? It immediately noticed an error the 12B made, where it hadn't followed a skill's format description. But it's also annoyingly not following instructions the other one was and instead just gives me its opinion lol! And it couldn't figure out how to read a simple pi SKILL.md and run an included script. It complained that it wasn't registered as a tool. But like all other model changes, will get used to it and adjust I suppose :)
Thank you, very interesting to read. I haven't considered batch sizes at all up to now. Seems I should look into it.
How did you get the prediction that good? The default spec-draft-max-n is 3 and if I go beyond, the output rate stays the same or gets lower any magic spell you could share?
Regarding the MTP, if you haven't found out why yet, the `--spec-draft-p-min 0.8`. Something to play with and of course it's non-deterministic and depends on both the model itself and whatever you're actually throwing at it.
I'm having the same constraint, 8GB VRAM, and was trying the past months to get a local model running to drive my Home Assistant setup via voice, so many tool calls, little room for errors. I've tried multiple models, different temperature and top p values, different context windows... turns out that a quantized gemma4 is the best pick so far but qwen3.5:9b isn't far off. Surprisingly newer models are not necessarily performing better.
18C is more like the maximum OK temperature. But then I also gotta sleep without PJs when its that hot. Still a blanket for comfort.
Best sleep from my experience? It's usually about 5C at the thermostat, which is farthest from the window, by the time I wake up. And that's likely because it's in the basement, which is a typical NA basement, so you have the uninsulated and below the frost line basement wall there. The cold air coming in will be 0C and colder and I sleep next to the window. Winter is my happy sleep time.
For you, and that's kind of my point. For me, my preferred temperature would be about 28C with no blanket or a light sheet. Something like 18C can be tolerable with blankets for me, but it's certainly not "maximum comfort", even with blankets. There is too much variation and too many things that wouldn't be controlled for, like number/thickness of coverings, and other such things. Having some "best" universal temperature is just not something generalizable like this.
If you are wearing clothing or any covering, clearly it isn’t your optimal temperature. I’m not trying to be glib here; I understand the appeal of being warm in a cool or even chilly room with air that keeps you from overheating. But that’s not sleeping at the optimal temperature; it’s choosing the sleeping setting you like.
And there are huge swaths of the world where the temperature basically never gets that low (18 C is the lowest temperature ever recorded in Kingston, Jamaica; 16 along the coast of west Africa, 18.9 in Jakarta). Even here in the southeastern US it is normal for temperatures to stay above 20 C at all times for three full months of the year (late June through early September).
Nothing wrong with liking to sleep in cold air, but it’s a preference.
Your feeling might be different. But for me, body temperature and feeling is way different from the air I'm breathing.
I can be, actually want to be toasty warm under the blanket. But I want the air I am breathing to be fresh and cold. That is when I feel comfortable and "at peace".
I will often stick out my feet from under the blanket because it's way too toasty but the rest of my body needs to be warm.
Sure, when it's 25C I'll essentially just lay there without anything but a sheet across the chest to "feel safe" but anything else will just make me way too warm and the air I'm breathing will feel restricting.
What is really great in 25C+ temps is to be outside in a hammock without an underquilt, which cools things down considerably. But I've also slept comfortably at 0F (-18C) in a hammock. With two underquilts and a 0F sleeping bag as well as a winter tarp with doors and snow cover all around, weighing down the tarp sides (wind does suck at that temp ;)).
I was really responding to the article’s odd claim of 18 C being the perfect sleeping temp, and your comment was so extreme in the make-it-cold direction that it seemed to be a good spot to reply. If we’re talking about “optimum” sleep temp as though such a thing has meaning beyond personal preference, then clearly it must not be cluttered with cultural detritus like depth of blankets, etc.
My wife wants it to be 20 C or lower and sleep under two blankets. I pointed out that 23 with a fan and a sheet is less energy-intensive and gives equivalent results, and she said stuff it. Not a battle worth fighting at home, but on the internet…
(Of note, the one of our dogs that loves me but worships her still sleeps next me every night, under covers, because she’s a Chihuahua and gets cold easily, and unlike my wife I sleep like a dead log.)
And "From a technology standpoint, Tesla will have a car that can do full autonomy in about three years, maybe a bit sooner.". Uttered by Musk in 2015. So ... yeeeeaaah.
Musk has 10+ years of history we can go by that says: Nobody has to follow anything closely to know that he will just say whatever he wants to say and you should give a rats ass about it timeline wise.
Now, do not get me wrong: I got frickin' goose bumps when I saw the first Falcon touchdown. Goose bumps again when they did the first landing out in the open ocean. Of course I still love you! The pincers catching a starship booster were super awesome. Did I mention goose bumps? I am getting them as I write this, thinking back to those moments and now I want to re-watch all of those on youtube after I finish typing this.
But believing a single timeline prediction by Musk? No, forget it.
EDIT: haha, and yeah I did go re-watch and this is so super awesome when I got reminded of it: https://www.youtube.com/watch?v=bvim4rsNHkQ&t=1s No this is not meant as a jab btw. I love that they have this up. This is what many of us do in software development. Just we don't blow up actual frickin' spaceships to do it "It's just a rapid unscheduled disassembly!" ;)
Can you tell me what the difference is between "technologically" and "safely" standpoint?
"Technologically we can do self-driving, we'll just be killing any person that happens to stand in our way" does not qualify even as "technologically" for me ;)
The technology has to take into consideration the safety part. Legal, sure I'll give leeway. If you've seen some of the stuff Teslas have done past that predicted timeline, then no, they really did not qualify.
In all fairness, it may not have actually been bullet proof. It was very thick, so I made a bit of an assumption based on what I know about it (which is mostly watching super slowotion videos on YouTube). Perhaps it was only hammer or knife proof.
It was definitely meant to protect employees in a high crime neighborhood, at the very least.
I'm not sure why you all have issues with CC commenting too much. My rules in the CLAUDE.md specify that comments are evil, never comment unless there is an actual need to explain a WHY and since I do read what CC writes, if I spot it still adding such WHY comments and they make no sense, I'll have it adjust, in many cases by removing them.
Given the code base has a minimal amount of such comments, it's also less likely to go "copy what the rest of the codebase does".
Of course I've now jinxed it and some update will cause it to ignore the instructions coz I didn't write them in the new model's style or something.
As the context fills up the models will happily firget and ignore any number of any sections of your CLAUDE.md/AGENTS.md.
Edit:
I've had explicit instructions for communication style in CLAUDE.md, in Claude's project "memory", in global "memory", in "skills": it couldn't care less where it was. It would just ignore it.
When I would point this out it would just say "Yes, I violated communication guidelines, I won't do that again". Only to do that again in the next session.
This applies to everything: code guidelines, communication guidelines, preferences, decisions etc.
reply