XCSme

6 ms·

61.
by XCSme 1mo ago
In my tests 3.8 Flash is considerably more expensive[0]/less token efficient than 3.7 or 3.6, and not necessarily much smarter. I assume it is faster in tps, but hard to tell because ot also outputs more tokens, so response time is slo
62.
by XCSme 1mo ago
Yeah, most people keep bashing my aibenchy.com benchmarks, because Gemini is on top, but that's because questions are not coding only...
63.
by XCSme 1mo ago
We can't even predict/understand our current world, yet alone the world in 20 years. Even if we get to one GPU pe person, it's not like that's the end state, the world constantly changes and will constantly evolve. If it
64.
by XCSme 1mo ago
I tried benchmarking it, but it keeps timing out/rate limiting, so the current provider(s) are unusable.
65.
by XCSme 1mo ago
> Autistic people pretty reliably score high on a set of diagnostics test Any sources? In my experience, most people that I know are on the limit with their diagnostic, and the majority discovered it late in life (30+, so a lot of maskin
66.
by XCSme 1mo ago
But that's the thing, sight and blind can be easily defined. "Autism" can't, most of the people live undiagnosed, many mask the symptoms unknowingly, specialists can't even tell exactly if someone is autistic or not
67.
by XCSme 1mo ago
Yup, exactly my thoughts. When I first considered my condition, I was looking around and was trying to mentally label people as neurotypical/neurodivergent. After a few years, I started labeling more and more people as neurodivergent,
68.
by XCSme 1mo ago
That's interesting. Are there studies on this? I.e. 1000 participants, "count" the symptoms and make a graph of autistic symptomps/person.
69.
by XCSme 1mo ago
The problem is that the set of problems is very broadly defined, changes over time, varies from person to person, it's often impossible to obiectively test. Can't it be that a neurotypical person is bothered by some autistic trigg
70.
by XCSme 1mo ago
That's a good example: they hate certain things, and are ok with others. Depending on what you "hate" your life and disability can vastly drastically. Be it eye contact, clothing, certain foods, or the inability to change you
71.
by XCSme 1mo ago
Ha, that's a great analogy, I like it. Even more, what is considered "positive impact" changes over time, so what was "crazy" 50 years ago, might be normal or even recommend now.
72.
by XCSme 1mo ago
I said that if everyone is neurodivergent, then no one is. Same with IQ, if everyone is 150IQ, then everyone is still 100IQ, as per the definition, so "normal". Divergence can only be mentioned when talking a base value to diverge
73.
by XCSme 1mo ago
But different autistic people struggle with different things, in different ways. What's even the point of having a general diagnosis umbrella, if the symptoms are different?
74.
by XCSme 1mo ago
As a potential autist, I think everyone is autistic in one way or another (which would make everyone neuroytpical, lol). In the end, maybe we are just all unique individuals with our character sliders set in various forms.
75.
by XCSme 1mo ago
I don't think taxes are the main issue, but people losing their jobs. We need UBI. Taxes are still there anyway, company income/profit tax, VAT, and maybe we can have other taxes too for automatically-generate income, same as with
76.
by XCSme 1mo ago
Nice, finally they fixed the huge reasoning tokens count. Now it's similar cost to DeepSeek v4 flash, but smarter. My tests: https://aibenchy.com/compare/z-ai-glm-5-3-flash-max/deepseek...
77.
by XCSme 2mo ago
Why would they infer faster with new shoes?
78.
by XCSme 2mo ago
I am also on the pro $200 plan, the limit is high enough to do what I want for a week, and binge run Ultra Fast last day to use the remaining credits.
79.
by XCSme 2mo ago
I am still pretty happy with some UX decisions I took: - the spotlight search, you can do ctrl+k and search "gpt vs claude vs google" and it automatically shows the most relevant comparison - drag to reorder compared models - filt
80.
by XCSme 2mo ago
I do try to guide it, I don't think strictly telling LLMs what to do works best. I like letting the LLM do what it thinks it's best (supposedly it is a lot smarter than me), and then guide it in the directions I think it should go
81.
by XCSme 2mo ago
Maybe, but only it's marginally off. This is what I pay extra per kWh,in Netherlands: € 0,25162 ($0.291), excluding the connection costs.
82.
by XCSme 2mo ago
Thanks for the feedback! That might be my fault, I am not a designer, and I asked for most UI decisions. I tried to use colors to diferentiate models, the site is very data dense and it's hard to make everything readable. I've spe
83.
by XCSme 2mo ago
Also, note that the Luna was run and costs were calculated through OpenRouter, via API. With a ChatGPT subscription it's likely even cheaper. Also, Qwen 3.7 27B is actually Terra level.
84.
by XCSme 2mo ago
I removed the extra links to sources for the electricity prices, but that's the average cost in EU, where I live. https://ec.europa.eu/eurostat/web/products-eurostat-news/w/d...
85.
by XCSme 2mo ago
My comparison of its reasoning efforts[0] seems to show that it only really supports 3 modes: none, low, xhigh. Low and medium are basically the same. Also, the electricity it costs to run on a 3090 is not negligible, so that it's che
86.
by XCSme 2mo ago
It's an amazing model, it's GLM-5.2 level[0], running locally... I tested it on my 3090, took like 8 hours to benchmark it and my room became a furnace (35+ deg outside temp), but it's really good. Now, in theory, you can tal
87.
by XCSme 2mo ago
Msi afterburner, you go around 900mv curve editor, raise it up to normal clock frequency, and it uses like 260w instead 300w for same performance. There's some YouTube guides for it. I also undervolted my new 5070ti, same tdp, around 2
88.
by XCSme 2mo ago
5900x, 3090 24gb (slightly undervolted), 128gb ddr4, running via Ollama. I am benchmarking it now locally, will put the results and speed/tps on aibenchy.com
89.
by XCSme 2mo ago
Oh, ok, that's like the average tps for most AI providers
90.
by XCSme 2mo ago
Why slow? I see ~50tps on a single 3090

More ›