johndough

7 ms·

121.
by johndough 7mo ago
Could you point at some more public info about active parameter count? You said: > and while an exact number is hard to compute, let me tell you, it is not 17B or anywhere in that particular OOM :) I can see ~100B, but that would near th
122.
by johndough 7mo ago
Unfortunately, that comment can not be edited anymore. Maybe @dang can change it or remove the comment chain. I am fine with both.
123.
by johndough 7mo ago
I dug a bit deeper. The cause was that navigator.storage.getDirectory() is not available in Firefox private browsing mode. The performance is pretty amazing. fib(35) runs in 60ms, compared to 65ms in NodeJS on Desktop. But I can't find
124.
by johndough 7mo ago
I see. I thought they were the same and https://vitedemo.browserpod.io/ failed to load, so I was not able to test there.
125.
by johndough 7mo ago
Do you have any clues to guess the total model size? I do not see any limitations to making models ridiculously large (besides training), and the Scaling Law paper showed that more parameters = more better, so it would be a safe bet for com
126.
by johndough 7mo ago
I ran two experiments: ~20x slower for a naive recursive Fibonacci implementation in Python (1300 ms for fib(30) in this VM vs 65ms on bare metal. For comparison, CPython directly compiled to WASM without VM overhead does it in 140ms.) ~250
127.
by johndough 7mo ago
Opus 4.6 likely has in the order of 100B active parameters. OpenRouter lists the following throughput for Google Vertex: 42 tps for Claude Opus 4.6 https://openrouter.ai/anthropic/claude-opus-4.6 143 tps for GL
128.
by johndough 7mo ago
Oh, you are right, I forgot that "major version" is a technical term and incorrectly read it as "For Python, 0.1 increases make a big difference". My bad!
129.
by johndough 7mo ago
That is true, and it is also a huge pet peeve of mine. If more library maintainers showed some restraint it using the newest and hottest features, we'd have much less update churn. But on the other hand, this is what keeps half of us e
130.
by johndough 7mo ago
AI-written text is not necessarily incorrect, but if the author did not take their time to remove the AI slop, they probably did not put much effort into it elsewhere. In addition, the text is often over two times longer than without the sl
131.
by johndough 7mo ago
I don't understand. Could you elaborate?
132.
by johndough 7mo ago
That seems reasonably active to me. You can't really expect more from an open source project without paid full-time developers.
133.
by johndough 7mo ago
I don't know about that. For me, f-strings were the last great quality-of-life improvement that I wouldn't want to live without, and those landed in Python 3.6. Everything after that has not really made much of a difference to me.
134.
by johndough 8mo ago
> Works on: RTX 3060 Ti 16GB This seems to be a hallucination. There is no RTX 3060 or RTX 3060 Ti GPU with 16GB memory.
135.
by johndough 8mo ago
Maybe related to this: > Opus 4.6 now defaults to medium effort for Max and Team subscribers. Medium effort works well for most tasks — it's the sweet spot between speed and thoroughness. You can change this anytime with &#x2F
136.
by johndough 8mo ago
Anthropic has never been able to send emails to my outlook email address (since 2023). Maybe changing your email address helps.
137.
by johndough 8mo ago
I've heard that Google keeps Google Drive data around for up to two years if your subscription expired and your account is over quota. They could certainly do the same with other cloud storage.
138.
by johndough 8mo ago
Again, your claims are impossible to verify or falsify, because they are too unspecific. > Meanwhile we have several industry experts pointing not only towards DeepSeek ridiculous claims of efficiency, but also the lies from other labs.
139.
by johndough 8mo ago
Only if chip-to-chip communication was a bottleneck. Which it isn't. If a layer completely fits in SRAM (as is probably the case for Cerebras), you only have to communicate the hidden states between chips for each token. The hidden sta
140.
by johndough 8mo ago
> why are you so defensive? Because you made stuff up, did not show any proof, and ignored my proof to the contrary. You made the claim: > Deepseek lies about costs systematically. DeepSeek broke down their cost in great deta
141.
by johndough 8mo ago
> So the size of GPT-5.3-Codex-Spark isn't limited by the memory of a single Cerebras chip, but the number of such chips that you can chain together and still hit the 1000 tokens per second target. Chaining chips does not decrease t
142.
by johndough 8mo ago
Did Amazon really not turn a profit, or apply a bunch of tricks to make it appear like they didn't in order to avoid taxes? Given their history, I'd assume the later: https://en.wikipedia.org/wiki/Amazon_tax_a
143.
by johndough 8mo ago
What do you base your accusations on? Is there a specific number from the link above that you claim is a lie? And how are 7 other providers able to offer DeepSeek API access at roughly the same price as DeepSeek? https://openrout
144.
by johndough 8mo ago
I am using piper-tts ( https://github.com/OHF-Voice/piper1-gpl ) with these voice files for GLaDOS: https://huggingface.co/rokeya71/VITS-Piper-GlaDOS-en-onnx/tr... It is not perfect, but quit
145.
by johndough 8mo ago
DeepSeek had a theoretical profit margin of 545 % [1] with much inferior GPUs at 1/60th the API price. Anthropic's Opus 4.6 is a bit bigger, but they'd have to be insanely incompetent to not make a profit on inference. [1] h
146.
by johndough 8mo ago
Nutika uses CPython as fallback. From the Limitations section: > Standalone binaries built using the --standalone command line option include an embedded CPython interpreter to handle aspects of the language that are not determined
147.
by johndough 8mo ago
Oh, so you are looking for a real compiler. I do not think that it is possible to compile Python, since the language is just too dynamic. You'd have to compile every function for every possible combination of types, since the types of
148.
by johndough 8mo ago
Not sure if this is what you are looking for, but here is Python compiled to WASM: https://pyodide.org/en/stable/ Web demo: https://pyodide.org/en/stable/console.html
149.
by johndough 9mo ago
> Edit: Python has no JIT There are quite a few JITs: JIT-compiler for Python https://pypy.org/ Python enhancement proposal for JIT in CPython https://peps.python.org/pep-0744/ And there a
150.
by johndough 9mo ago
I was wondering whether multiple GPUs make it go appreciably faster when limited by VRAM. Do you have some tokens/sec numbers for text generation?

More ›