dannyw

22 ms·

241.
by dannyw 3mo ago
Reminds me of this great read: https://lantian.pub/en/article/fun/ai-agent-bankrupted-their... (HN: https://news.ycombinator.com/item?id=48500012 ) My view: I consider it ethical if you disclo
242.
by dannyw 3mo ago
We almost need a proof of human check. Before submitting a PR, you must find and paste the actual lyrics of a random song as challenge-response. Lyrics seem to have the strongest safeguards out of everything. (Try it! On some APIs, you migh
243.
by dannyw 3mo ago
You'd be surprised at what Luna can do, especially on xhigh or max. It's capable of working overnight, usually productively, just like Sol. Haiku 4.5, on the other hand, is comparable to performance to Gemma4 31B (with working too
244.
by dannyw 3mo ago
Deep research is basically a LLM with web search, and a "work really hard" goal-orientated prompt, and some output formatting suggestions.
245.
by dannyw 3mo ago
OpenAI's APIs are extremely reliable for sure. I don't even remember when the last incident or downtime was.
246.
by dannyw 3mo ago
Been using OpenAI models since ada/babbage/curie/davinci and at least from my own experience, their APIs feel the same. If you use Codex it's different, the harness has a lot to do with it and there's definitely bee
247.
by dannyw 3mo ago
If you decide to give your agent the ability to autonomously send bitcoin, I'd rather it go to the creator of this website.
248.
by dannyw 3mo ago
I feel like everyone does it though. Like who really runs frontier LLMs at `max`? It feels like a benchmark-only setting, just for ArtificialAnalysis really.
249.
by dannyw 3mo ago
If you train on any form of HTML, or any code that includes URLs, your model is going to learn them somehow. Stripping URLs would certainly break some code, and also train the model with unhelpful behavior. You want your coding model to wri
250.
by dannyw 3mo ago
Definitely annoying, and YMMV, but a lot of jurisdictions don't treat credits (that cannot be redeemed back to cash) as income. In the same way signing up for $App and getting $100 in API credits isn't income; paying a bill early
251.
by dannyw 3mo ago
At scale, a decent number of commercial/business agreements have Net Discount provisions (or provisions to that effect); basically get a % discount if you pay within X days; or pay the full amount with no discount in Y days, so the gen
252.
by dannyw 3mo ago
I do appreciate you doing this btw, I find it nice and clever. I like how it democratizes a mechanism that enterprises have (net discounts, similar effective mechanisms) and makes it accessible to everyone. I was replying to the parent comm
253.
by dannyw 3mo ago
You absolutely don't need to 'retrain' to reduce your context window. In vLLM it is an inference parameter. Smaller context window, smaller KV, less RAM needed to serve the same volume of requests. With a lot of architectures
254.
by dannyw 3mo ago
There's the infra cost of having multiple SKUs. If you create 3 buckets of inference pods, say, 256k, 512k, and 1M, then you have to worry about filling/dynamically-scaling all of them. And my guess is there's probably not a
255.
by dannyw 3mo ago
How many small hobbyists really want to front load $150, instead of paying $5 a month, and would only be swayed if they got interest/credits? At 5% simple interest, your $150 would give you 62.5c per month. So you'd need closer to
256.
by dannyw 3mo ago
I just mean in terms of incremental inference cost. If you already committed to your hardware, huge models at single or sub-digit TPS are still useful. LLM-as-judge is a good use case. If your machine is gonna be idle overnight (and the pow
257.
by dannyw 3mo ago
The OpenAI and Anthropic batch APIs give you a 50% discount. This is like 98%.
258.
by dannyw 3mo ago
A lot of API customers don't want their chatbot or agent to identify as Claude or GPT, they want it to identify as <customisable marketingname>.
259.
by dannyw 3mo ago
The architectures are high level concepts and the mechanics usually have enough detail for you to try and implement. Transformers are very "mendable" in that you can permute the architecture in crazy or random ways, and still basi
260.
by dannyw 3mo ago
I suspect a lot of the Python and TypeScript code is getting prompted by folks who don't know any difference between dynamically typed and statically typed.
261.
by dannyw 3mo ago
A license and presumably revshare is required for large scale inference as a service of Kimi K3. There's clearly something going on right now with every router at the same or higher price as the Moonshot list price. So I wouldn't
262.
by dannyw 3mo ago
You can get basically any part from Shenzhen with an Alibaba account, if you know what you're looking for. For reaching out to a seller/merchant, I found it helps significantly if you google translate into Chinese, and just commun
263.
by dannyw 3mo ago
Do you run your coding harness in a completely segregated sandbox? Even folder names of what projects you have, what projects you actually work on, etc are very useful for targeted advertising.
264.
by dannyw 3mo ago
You can even use /model to use Opus 4.5 if you want. Broadly though I find Opus 5 and most of their point releases (4.7 being the exception) to be excellent and upgrades.
265.
by dannyw 3mo ago
What's wrong with it? Licenses like the MIT license already requires you to attribute, the only difference is the "prominent" part.
266.
by dannyw 3mo ago
The western AI crawlers are generally very-well behaved. GPTBot and ClaudeBot are nice crawlers and I see them respecting my robots.txt. There are of course plenty of mystery, obfuscated/camouflaged scrapers/crawlers from god know
267.
by dannyw 3mo ago
Even companies like OpenAI segregate their training crawlers (GPTBot for general corpus training; OAI-SearchBot for their search index inclusion; and ChatGPT-User for user-driven agentic browser use, etc). When even OpenAI is more respectfu
268.
by dannyw 3mo ago
The career prosecutors at the DoJ are not the same as a couple years ago. I hope this case ends the same way as the sub sandwich assault.
269.
by dannyw 3mo ago
Learning how to make a cuckoo wristwatch and everything that involves helps you acquire skills, even if what you build is fundamentally useless.
270.
by dannyw 3mo ago
It would be lovely to see NVIDIA accidentally delay some allocations to frontier AI labs calling for banning open weight models. Oops, sorry, production issue, chip shortage… now about that open model ban you’re lobbying for…

More ›