The reply has been flagged/marked dead by HN since, and should at minimum auto-collapse (thus hiding even worse replies to it) regardless of your account's "showdead" settings. I'm not going to link to it, and it's not worth reading regardless.
One of the options: money go to a 2/3 multisig address, 1 key is controlled by the customer, 1 key is controlled by the service provider, 1 key is controlled by an escrow service.
Of course US has a huge head start, but if AI keeps growing, what matters is how the market's gonna look like years from now.
Most of my clients using AI in the business workflows (in products) use Chinese LLMs, because after benchmarking for a specific use case you nearly always end up finding that you pay half or a tenth.
That's not a new phenomenon. I've adapted Gemini Flash 2.5 years and years ago when people were dissing it as "crap", yet it was the best budget and quality fit for the task I had at hand back then (translating and summarizing tons of documents). It was both faster and around 100 times cheaper than the best GPT 4 model available.
Needless to say, medium-sized Chinese models are far better than those LLMs and a perfect fit for countless applications.
Even with all the Qwens and Kimis and GLMs etc, the latest Gemini Flash models are still an insane value! I recently settled on 3.1 Flash Lite after testing basically everything on offer on OpenRouter, and it was not a close call - cheaper, faster, and better (for translation and visual understanding tasks).
Yes, they were uglier but they had advantages that got lost such as an easy to control rendering life cycle.
Just opening dev tools on MIT-written big tech corps and startups confirms nobody is able to write good websites or applications that arent filled with performance and memory leaks.
Yes, Redis scope got bigger, but not at the expense of the core functionality.
It's not like using it as a val key store got worse and more complex.
Redis itself is more of a data service that can bend to many needs, and those all of those things well.
Not sure why supporting more data structures would be bad.
reply