dannyw

4 ms·

31.
by dannyw 1mo ago
Precisely. BIS does not consider cloud compute services to be exports. There are some rules around when you knowingly know that a customer is military or WMD, and you obviously can’t supply services to someone on the entity list; but KYC-th
32.
by dannyw 1mo ago
Are you using direct or via OpenRouter? I think OpenRouter Luna always uses the `flex` tier, which is quite a bit slower.
33.
by dannyw 1mo ago
That’s incompatible with the file not leaving your browser, which you can trivially verify with the Network tab in chrome (or wireshack, etc). This is just a C2PA metadata checker.
34.
by dannyw 1mo ago
The tests measure Fable 5.1 (with fallbacks). The increased cost can come from Fable 5.1 triggering fallbacks less; which means less (cheaper) Opus when AA ran it.
35.
by dannyw 1mo ago
The new cloth is different. It's softer and thinner, but the same size. If you want to spend 12 minutes on what changed, AppleInsider made a video: https://www.youtube.com/watch?v=iL-4kCkvEvQ This is actually the 4th A
36.
by dannyw 1mo ago
The narrative — ie being associated and seen as an AI-winner — is worth billions of market capitalisation at Apple scale. I don’t think there’s a direct guerilla marketing division, but I would absolutely bet money that Apple’s marketing&#x
37.
by dannyw 1mo ago
Comments that the HN community find interesting are surfaced higher. Just tap on the [-], and upvote what you find more interesting :)
38.
by dannyw 1mo ago
I agree, and I think what you’re describing is only scratching the surface of what’s possible today. It’s even more powerful with large data and knowledge sources connected. Takes a lot of work to set up effectively, but when connected to S
39.
by dannyw 1mo ago
I consider publicly enforced to be one where a government agency (e.g. the FDA), or prosecutors make judgements in what cases they file, make the arguments, etc. Otherwise, it's just a standard case between two private parties resolved
40.
by dannyw 1mo ago
You can scale with more memory channels. Workstation/server platforms go up to 12 or 16 channels if I remember correctly. Consumer platforms have been stuck at dual channel for decades; most of it I attribute to intentional product seg
41.
by dannyw 1mo ago
What do you do for long-running goals / tasks where you are not interactively prompting?
42.
by dannyw 1mo ago
Companies are well within their rights to choose who they sell to. I don't see how this is 'privately enforced copyright'. Also, copyright has always been privately enforced anyway?
43.
by dannyw 1mo ago
FP8 experts not BF16.
44.
by dannyw 1mo ago
Try Parallel.ai (no affiliation). Instead of keywords, the model writes objectives.
45.
by dannyw 1mo ago
I’ve found a very similar creative spark with Gemma4 base models. You’ll have to do old school prompting, but it’s kinda fun too.
46.
by dannyw 1mo ago
Thank you for making and maintaining rclone. It is truly a blessing.
47.
by dannyw 1mo ago
I mean, if you don't care about your inputs being trained on, you can just use one of the many free models on openrouter.
48.
by dannyw 1mo ago
I just used Qwen3.8 & Deepseek v4 flash to make it for me.
49.
by dannyw 1mo ago
The M3 Ultra was a lot cheaper for most of its lifetime, and really the main reason for buying it is if you want a lot of unified RAM, to run bigger models than 27B models. The new M5 Ultra should deliver ~50% faster token generation (1.2T
50.
by dannyw 1mo ago
Yes, Kilmar Abrego Garcia is back in Maryland, as one example.
51.
by dannyw 1mo ago
There's a lot of platforms that are more actively developed and maintained, for example Prolific (which also offers discounts on service fees for academia).
52.
by dannyw 1mo ago
OpenAI has been doing wonky stuff with subagents, including encrypting the prompts sent to subagents in Codex. Who knows what’s really going on.
53.
by dannyw 1mo ago
A lot of llama.cpp contributions come from the community and ecosystem, like Unsloth. If something goes awry, I fully expect lots of forks.
54.
by dannyw 1mo ago
The Intel Arc Pro B60 Dual has 48GB of VRAM; and the B70 goes up to 32GB.
55.
by dannyw 1mo ago
It's really excellent. I've been using Tailscale for a number of years, just for my home network and with a single user, and the service is excellent, and the free tier is incredibly generous. I'm hoping that my Mullvad subsc
56.
by dannyw 1mo ago
If you ever connect over a mobile hotspot, then you need NAT traversal. And if you want something that "just works" across various network environments, you need NAT traversal.
57.
by dannyw 1mo ago
Yep. It's like someone asking "How can we make cycling in [city] safer?" and answering it with "Drive a car."
58.
by dannyw 2mo ago
NVIDIA is one of the most open labs. They even share many of their pre-training and even post-training datasets for Nemotron on HuggingFace; for example: https://huggingface.co/datasets/nvidia/Nemotron-Post-Trainin
59.
by dannyw 2mo ago
I generally agree — go local for the hobby/tinkering, privacy, and control (ie not getting refused by an AI to defend and secure your own network and codebase; as HuggingFace has seen). But whether you make a loss or not depends on how
60.
by dannyw 2mo ago
I will never use or verify biometrically with Persona. I don’t care how out of the way I have to go, but I will never do it. Working well so far.

More ›