johndough

8 ms·

61.
by johndough 4mo ago
I see that OVH offers Qwen3.5-397B-A17B, which is a bit surprising to me. I thought that EU providers had to comply with the AI act where you have to provide opt-out and information about the training data once the model is sufficiently lar
62.
by johndough 4mo ago
Probably not that important in practice. Firefox allows 2^20 - 4 and Chrome allows 2100000 characters. Also, 8000 characters already allows for an unreasonable amount of SQL and could be extended even further with compression. And if that s
63.
by johndough 4mo ago
In academia, this is called "planar embedding" and can be computed in O(V) where V is the number of vertices of the graph. However, there are graphs that do not allow planar embeddings (e.g. K_5 or K_3,3, see https://en
64.
by johndough 4mo ago
Feel free to replace "racism" with "discrimination" if you prefer. English is not my first language and the minutiae elude me.
65.
by johndough 4mo ago
The recent MiMo-V2.5-Pro-UltraSpeed can be served from 8 GPUs, which is certainly within the reach of sophisticated on-prem setups. https://mimo.xiaomi.com/blog/mimo-tilert-1000tps
66.
by johndough 4mo ago
MiniMax M3 weights have already been released: https://huggingface.co/MiniMaxAI/MiniMax-M3
67.
by johndough 4mo ago
As a foreigner, the casual racism against red-haired people in the US and UK always baffled me.
68.
by johndough 4mo ago
Cerebras got lucky that they IPOed last month instead of now.
69.
by johndough 5mo ago
I recall that Facebook asked me to identify faces of my friends in order to "verify" myself about 15 years ago. Do you know whether Facebook still stores that data?
70.
by johndough 5mo ago
Likely just a coincidence. There is a huge number of node-based image editors: https://www.google.com/search?q=node-based+image+editor&udm=...
71.
by johndough 5mo ago
Goomba fallacy https://en.wiktionary.org/wiki/Goomba_fallacy
72.
by johndough 5mo ago
Fair enough. I am optimistic that 2 and 3 will be doable eventually.
73.
by johndough 5mo ago
I've got a few. Pick whichever you like. Factor 1350664108659952233496032162788059699388814756056670275244851438515265106048595338339402871505719094417982072821644715513736804197039641917430464965892742562393410208643832021103729587257
74.
by johndough 5mo ago
It looks like the issue was that, in Italy, copyright expires 70 years after the death of the author or the first translator of a work.
75.
by johndough 5mo ago
Yes, single RTX 3090 with this model https://huggingface.co/llmfan46/Qwen3.6-35B-A3B-uncensored-h... following these https://huggingface.co/havenoammo/Qwen3.6-35B-A3B-MTP-GGUF instructions (should
76.
by johndough 5mo ago
> it gets caught in circles more often then say a 1t parameter model does. I've found that the Q5+ quants are less loopy than Q4. Still not perfect, but noticeably better. > reasonable enough tokens per second The speed has been
77.
by johndough 5mo ago
While Qwen3.6 27B and 35B-A3B are very good, I am skeptical about them being that good. I think another factor is at play here. The Qwen3.6 models have memorized some common games. For example, if you ask it to create an index.html with a
78.
by johndough 5mo ago
API pricing for Claude is about an order of magnitude more expensive than subscriptions (numbers: https://she-llac.com/claude-limits ). But it may be worth it with DeepSeek V4 Pro, which is currently on discount.
79.
by johndough 5mo ago
At my university, everyone had to pay their AI subscriptions out of their own pocket, until a communal AI service was introduced recently. It took 2 years to set up and only serves gpt-oss-120b, so everyone is still using other services. Bu
80.
by johndough 6mo ago
> I'm running this locally. Impressive! What is your setup? Are you running the full DeepSeek V4 Pro, or V4 Flash?
81.
by johndough 6mo ago
You mentioned "pareto-competitive", and EXAONE certainly was that. The statement that the "above countries have never had a model that is even close to being so" is simply too broad.
82.
by johndough 6mo ago
> Korea EXAONE from LG AI Research https://huggingface.co/LGAI-EXAONE They had one of the best small models a few months ago and they released a new model just last week. There's also HyperCLOVA X (haven't test
83.
by johndough 6mo ago
> Sure they aren't as good as the big providers If you haven't done so already, finetune the model on all your company's code that you can get your hands on. This is one of the great advantages that you get when running lo
84.
by johndough 6mo ago
> Gemini on-prem boxes (recent release announced last week) Do you have any more information on this? I only found this article about it: https://venturebeat.com/technology/googles-gemini-can-now-ru... It mentions t
85.
by johndough 6mo ago
Yes, to be honest, that graph could use some improvements as well. I should probably just link to the blog post with actual numbers: https://github.blog/news-insights/octoverse/octoverse-a-new-...
86.
by johndough 6mo ago
> You don't need to know the bottom left axis number. We very much do. The graph suggests an insane growth in PRs from almost zero to 90M. Now compare this misleading graph with this much clearer one, which shows that the growth ove
87.
by johndough 6mo ago
Unfortunately, Opus was removed from the student plan in March. So far, I had been happy with GPT-5.3-Codex, but that model seems to have been removed this morning.
88.
by johndough 6mo ago
Is there somewhere I can look up whether a certain provider is VC-funded?
89.
by johndough 6mo ago
> More powerful models can do a lot more work (consuming a lot more tokens) per prompt. That is not my experience. Each model since at least GPT-4 can fill up an entire context window. In fact, more powerful models can solve tasks faster
90.
by johndough 6mo ago
That was not my experience. When I tried to use Opus for longer tasks with Copilot, it would fill up the context completely and then crash without any output, while still consuming premium requests. (At least from September 2025 to January

More ›