"Considering your Country decided not to give me the Nobel Peace Prize for having stopped 8 Wars PLUS, I no longer feel an obligation to think purely of Peace, although it will always be predominant, but can now think about what is good and proper for the United States of America."
Still gonna take 2-3 years to get DeepSeek V4.1 Flash quality at decent speeds on reasonably priced hardware.
Hardware update cycles are 2-3 years even on the high end, so it's still a ways away before "good enough" and "local" belong in the same sentence for the average person.
And by then, DeepSeek V6 Flash will be too cheap to meter, 5x faster, and 10x better, so... You'd still need to go out of your way.
Most people are spending most of their time on their phones anyway. ..
Flash Next is a basically there. It really depends on what you are doing. This model is great. People forget that they felt Opus 4.6 was a great model and now you have it at home.
I do use DS 4.1 flash a good bit. It's only real fault is that it is a very eager model and I have to system prompt it down to a more reasonable level and it will still take initiative, a little more than I would prefer.
I have ZDR enforced and see only compatible models and providers, yet am able to use it. DeepSeek as a provider may not be ZDR, but the models are available from ZDR and no training providers on EU/US servers.
Yeah, they all do it though. And now that models are being natively trained at FP8/NVFP4, I’m not sure it matters.
Only way you can really know you’re getting the full model is to host it yourself. Every inference provider has every reason to lie and it’s impossible to find out the degree to which they are.
It’s a pretty common requirement in the enterprise world. If you’re processing data for enterprise customers, it’s a lot easier to retain nothing than to deal with all the compliance issues that arise if you’re retaining data.
When 98.5% of my requests are cache hits (according to Pi for the last week), the cache miss price isn’t that important to me, and $0.003-0.006 per 1M input tokens is shockingly cheap.
It’s also the major difference between using DeepSeek directly vs other providers also serving it, though I have not looked lately: it’s possible other providers have matched its cache hit pricing better?
Read the tech report and see how laser-focused they've been on compressing the disk KV footprint specifically. 890 bytes/token is absurd, and is probably the reason I regularly get ~0.5 Mtok request full cache hits after an hour. On their end that's just yanking a 414 MiB file off an SSD, then doing a little bit of compute (bounded SWA replay). Nobody else seems to be serving their model as well as they do.
It will of course depend on what you’re doing with it, but right now my session at work has a 99.8% cache hit rate, and I’ve been running this session for hours with 23M tokens read and 713K tokens written (Opus 5.5 in this case though)
I've heard that certain inference providers may have different quality of caching implementations, so even if the listed numbers are as you say, the practical cache hit % you get might be significantly different/incur significantly different costs.
There’s a big difference in speed & quality between using DeepSeek API directly with DSH vs. DeepSeek in Opencode Go with Opencode CLI. Can’t tell if it’s the provider or the harness - but worth to give it a try.
This isn't true at least on the API. If you read their privacy policy you'll see the training clause is scoped specifically to the consumer terms i.e. for the chat product. No such clause exists for the API service, and it would absolutely be required under Chinese law if it was taking place.
Contrary to popular belief, DeepSeek really aren't interested in your prompts.
The standard terms includes clause 4.3 which grants them the right to retain inputs and outputs for training purposes, and this is missing from their API terms. The standard terms also cover the right to opt out (which you can do from your user settings). No such opt-out exists on the API because it isn't applicable.
Why does liking a product mean you have to give them all of your data? People are so outraged at LG because they make the best TVs and people wanted their expensive product, yet some MBA convinced them they could make more money by spying on your entire household all the time.
It actually was awesome in the early Facebook days where you could have your entire phone contacts and other apps filled out with a profile picture and Birthday by connecting them together. But that relationship has been completely abused, privacy has been invaded, and my data has been sold to multiple companies.
The goal going forward is to keep that data private. If your company can't survive without it then I hope your company goes out of business
All these products, western or Chinese, are built on a hell of a lot of running rough-shod over licensing or IP laws in general.
I'm writing a program I personally need, but I would be happy if there existed something like it already, if someone else vibecoded a better version of it than mine, or if DS got better at vibing this kind of thing.
Just pin your config to a single provider, or several providers with the params `order` and `allow_fallbacks: false`. I regularly get ~98-99% cache hit rates with OpenCode. And some providers are much faster than DeepSeek; I was getting 200-300 tokens/second the other day with Together as my provider.
It's regrettable that OpenRouter doesn't even try to pin you to a single provider per session, but once you know about it, it's a problem that's easily solved.
Yes, but that's why you pin them. If you specify more than one provider in `order`, OpenRouter will use the first one unless it's down, so that's the only time it would switch providers on you. And personally, I'd pay a few cents instead of waiting for the API to come back.
Yes, I send it to China, and they use it to improve open-weight models. I'm ok with this arrangement. At least, I'm happier with this than with companies using my open-source work to improve proprietary models without my consent.
Now, you're going to continue attracting LESS of the top tier talent as you make it clearer and clearer that you don't want them. This is the talent we want most.
Even if you're a laborer, it's quite hard to argue that the absolute pinnacle of the labor pool is dragging down your wages and not increasing total jobs.
If other countries are giving you better alternatives, why bring your talents to the US?
LOL. It's not "top talent" it's "cheap talent" or even just "more talent so the supply is larger and wages go down".
If you're EG a PhD+ and working in a frontier AI lab on cutting edge problems, then fair enough. If you're developer x of 100 writing Java CRUD apps in the finance department of some corporate, you are not.
> Now, you're going to continue attracting LESS of the top tier talent as you make it clearer and clearer that you don't want them. This is the talent we want most.
No. It’s not.
> other companies that would be suspended […] included Cognizant Technology Solutions Corp., Infosys Ltd., Tata, Wipro Ltd., HCL Technologies Ltd., and Capgemini SE.
Top talent exists domestically, this is to hold wages down. It’s fine to impair the import conveyor and see what happens. If and when there is a labor shortage or wages spike horrendously (highly unlikely), certainly, we can revisit.
Neoliberalism ignored domestic worker concerns for shareholder returns and corporatism, and an alternative was offered, here we are. It’s a natural experiment we might not otherwise had had the opportunity to run.
If the US can bully countries into releasing 100M barrels of their own petroleum reserves, what makes you think the US can’t solve for that too? Other countries!
Everyone has a leverage point, everyone has a point at which the pain is worse than the ask. Companies and executives are easy, and usually of weak constitution. They fold fast, based on the evidence so far. US Big Tech fell in line super fast, right? “Fall in line or else.” If trillions of dollars in enterprise value and the resources it commands won’t save you, I’m unsure if anything else would.
It’s not right or wrong, it just is. We are living through a “might makes right” era. It’s not the first, won’t be the last (if history is our guide). This too shall pass.
Frankly talent doesn’t matter any more with AI - the skill floor for software has dropped to zero so an average dev now competes with the best devs. Marketing and capital matter way more.
It still matters a bit for architectural direction - but those roles are so few I wouldn’t be concerned as an American.
Yes -> every 18 months they've gotten 90% more efficient for the same level of quality for about 5 years. There's little sign that trend is slowing. If anything, there's reason to believe that System 1 models (plus potentially 1-2-3 workflows) may increase that over the next 3-5 years.
You'll know when the trend stops -> when the intelligence differential between smaller models like 7B starts to grow instead of shrink from 32B models -> that means 7B is getting about as smart as it can get. Then, 32B will follow next, then 70B, etc etc.
Super intelligence that doesn't have to deal with the real world, maybe.
I wouldn't be surprised if less than 1B param equivalent of our brain deals with solving math and writing computer programs and physics and all the things we tend to associate with "intelligence" - especially if you ultra optimized for that, I doubt our brain works like that.
Dealing with the real world, I highly highly doubt it.
How about if we get away from written text as the input, to something more fundamental, that then also is able to produce text (among other things)?
Given that humans learn to talk while having encountered a measly number of word instances, and, given enough time, we should always be able to improve on the lottery that is biology, it does seems fairly likely.
It would be interesting if running ends up being a more complex task than advanced math. And our brains are just 95% allocated to dealing with the real world.
IMO, this is better. Luna is super cheap, but it's not that capable. At higher levels of reasoning, it's not that fast.
This is more expensive, but it also looks like it's better enough that it's far more useful.
I also won't be surprised if you look at cost per completed task + wall clock time that it comes out ahead for the majority of what you'd want to actually use it for.
Luna will still be a great option for doing non-engineering tasks super cheaply.
> just move crucial bits in another language and use webassembly. so good section of your code is in compiled binaries hence blocking anyone stealing your IP.
Can't you just steal the entire WASM code just as easily? I mean, I guess if the ads are stuffed in WASM that becomes a problem. But LLMs are pretty good at reverse engineering. I can't imagine it would be too much effort to get them to take the ads out, or to replace your ads with their ads.
Vision is just processing a still image (why it's by far the smallest). Text requires dealing with the entropy of human language. Audio is meaningless without time.
reply