All about energy, and energy cost plays a role here, DeepSeek can go cheaper than western models…
No wonder. Since deepseek has open license, they have to compete with 3rd party providers, and in case of smallest models with local generation.
“Permanently” lol it’s a subscription and the terms say they can change the price at any time. How is it legal for them to advertise with the word “permanent”?
lol it’s a subscription
It’s actually API access price, and it’s charged per input + output tokens. $0.87 per million tokens is damn cheap.
They probably have super cheap electricity and it’s possible they use cheap Chinese Ai chips for inference.
I think it’s meant to convey that it’s not a temporary deal on the old price, but a permanent new price point.
What is the effective difference? It’s not like they’re offering long term contracts.
60% of the time it works every time
Don’t use any of them much and from my limited experience they all seem to be pretty much the same. In fact DeepSeek probably has been a little better than ChatGPT.
Permanent under the current pricing model, subject to change.
Prices are funny. My last job we were changing clients extra for doing a thing that didn’t cost us anything and was fast to do. How much we charged was completely arbitrary and depended on the partners mood. It’s all made up folks.
Yeah, which is why the “if minimum wage increases, so will prices” aregument is BS. They were going to charge the highest price they thought they could either way, the difference is that they are forced to increase the amount that goes to the people they are trying to pay the least.
It’s not BS, it’s just not as direct of an impact as they are implying. If payroll is 10% of their expenses (assuming EVERYONE makes minimum wage) then doubling the minimum wage will increase costs by 10%.
Which could be (partially) absorbed from profits, could cause a 10% price hike… or a 50% price hike and fat bonuses for the executives.
It could also affect the cost of any other inputs that are labour-heavy.
There is an element of minimum wage increasing, increasing prices because now there are more people that can afford to pay for things.
But yes it isn’t because costs go up, and it really only applies to things people on minimum wage can afford and it’s always less than the increase in wages.
This would impact the companies pnl though, so shareholders and c suite will get less money. That’s why they’re scaring people into not wanting to increase wage.
The hilarious irony is that is not even conclusive. There are plenty of studies, both real-world and contrived, that indicate that employers paying more, in broad, yields returns in excess of the added payroll costs.
Not only are there more customers, but increasing pay increases the quality and quantity of labor output.
All numbers in AI are made up it’s wild to see tankies glaze DeepSeek’s fake numbers while being skeptical of Western corporations’ numbers
Not glazing when its simply enjoying watching China beat the US at its own game
But the numbers are fake, so it really doesn’t mean much to reduce a fake number by 75%, it isn’t an indicator that DeepSeek is beating anyone at anything.
Cost to end consumer is not a fake number.
What does that number meaningful represent as DeepSeek doing well?
They can afford to lose more money on this? They have lower operating costs? They have a better way to make money of their users?
It could indicate any/all/none of theses
What does that number meaningful represent as DeepSeek doing well? I don’t understand the question
They can afford to lose more money on this? Yes They have lower operating costs? Yes They have a better way to make money of their users? They are not as profit motivated as their competition
I don’t think you understand Deepseek’s role in the market. It’s to intentionally undercut US providers.
What do you mean by the numbers are fake? Are you saying the worth is over inflated? If that’s the case, of course it is, none too different than virtually any other commodity.
What does that number meaningful represent as DeepSeek doing well?
They can afford to lose more money on this? They have lower operating costs? They have a better way to make money of their users?
It could indicate any/all/none of theses
The lower prices could be aimed at undercutting the competition.
Mobster voice: Sure would be a pity if the monetization potential of those 2 huge IPOs (3 if you count SpaceX with xAI deadweight rolled in) went boom when that’s all that’s holding your economy out of recession (depression depending on how they cook the books).
The way SpaceX IPO got crammed into index, it’s invulnerable to anything but an immediate incarceration of everybody involved.
Index funds will be required to buy the stocks at a listing price before market can decide how much they are worth exactly.
Afterwards, “economy in a recession” is synonymous to “free buffet” to those at the reins.
Yeah, the whole plan is to have every US citizen’s 401k’s autobuy into the SpaceX IPO.
Your retirement fund is Elon’s exit liquidity.
Its a truly fantastic fraud.
Because… the Nasdaq… well a few weeks ago it changed its rules on the delay time between an IPO and it being part of the index, the index that everyone’s 401k’s buy into.
I guess you could say its going to be ‘epic’ when this all blows up.
See this is basically how the us economy works:
Poors roll over negative equity into their next car loan.
The ever diminishing ‘middle class’ basically does the same with homes, helocs, etc.
The owners roll over debt via corporate amalgamations.
But because the rich have a magical legal barrier of ‘all the bad and dumb things i do are a legal fiction doing them, not me personally’, well, the legal fiction gets what its due and/or evaporates when it can’t pay what it owes… and the rich remain on top.
Yeehaw!
Yep this will be the fourth or fifth record breaking upward transfers of wealth I’ve lived through. I really don’t want to live through another.
At this point many people won’t. Hard to squeeze blood out of a dried, overworked, malnutritioned poisoned and diseased husk of a laid off worker.
Can’t wait until enough people are unemployed that nothing is stopping them from marching in the streets.
Remember what happened during covid layoffs? People brought tents to the protests and camped out for months…
Wait, I don’t remember that happening during Covid. I remember some years before, that happened as part of the 99% anti Wall Street protests… which did nothing.
Well, they leaked the Panama papers, which showed exactly how and how much they are stealing/defrauding… which did nothing.
You’ve already forgotten about the BLM protests?
I think I’m following what you’re saying but I haven’t read anything about this. Where can I read more?
https://youtube.com/watch?v=-X6YzlY_8tM presents a good summary.
https://youtube.com/watch?v=IHD8BDFYyGI is a longer analysis coming at it from a slightly different angle.
Updoot for Patrick Boyle.
Afterwards, “economy in a recession” is synonymous to “free buffet” to those at the reins.
Not at all wrong, but there’s only so much blood parasites can suck before the host dies (and with luck kills the parasites, and / or sends a strong signal to everybody else to get their infestations eradicated, or at the very least under control), and that host is already hurting bad.
Perhaps I’m being optimistic, but a collapse of the likely magnitude could be that straw, or maybe it’ll just be the back of US influence that breaks.
Still doesn’t know what happened at Tiananmen Square, but can tell in detail how protests were brutally ended a few years later in South Africa…
It’s not what you would like to hear though I get it
It’s not what you would like to hear though I get it
If you run it locally there’s no censorship…
And no price per token. So I assume we are discussing the cloud version.
Really? You sure, this is still true?
I have never run that one locally, but qwen doesn’t “know” about specific Chinese historic events either when executed locally.
It was a couple of months ago at least. Also, if you use the deepthink mode online you can actually see the reply ( really criticising the Chinese regime) for a couple of seconds before it disappears. I’ve manage to screeshot it and also to trick him once about a “fictional” regime so I could have the answer.
Interesting. I tried their R1 model when it was all the rage and that in its thought process mentioned something about being a Chinese AI and having to provide the user with safe information. It then responded it doesn’t know about any specific historic events.
It does, and it’ll tell you about it. But it’s their interface that censores the output, and it’s not perfect. Ask it in English or Chinese and it’ll censor it. But ask in Spanish or other languages and it doesn’t get caught.
Huh. Apparently in Korean it’s censored. It’s also not their interface. Am using their API and still getting rejected.
It will happily give you details if search is involved(via Searxng in my case) though, so that’s something.
Other than that, amazing model. I’m not having political conversation with LLMs, let alone Chinese ones.
“permanently” means nothing when it comes to technology
Okay I still won’t use it.
Its safety rails are far worse than any in the West. But to your point, fuck AI.
There are plenty of rails, they’re just different ones. Like criticizing dear leader or Tiananmen square.
Blah blah blah blah. China bad. Invading Iran, Venezuela, supporting the genocide in Palestine is all good, getting out of the WHO, removing USAID which had let to thousand of deaths around the world, threathing Greenland and Cuba. Causing global inflation to rise due to oil cost. Tariffs . Destroyed relationships with organizations and countries that took 80 years to build.
And that is just this year for the US.
What catastrophic action on this scale has China in let’s say the last decade?
Whataboutwhataboutwhatabout
Two superpowers can be shit at the same time. Three, even. Four is phisically imposible.
First of all, no one said “The US is good actually.”
Second of all, aren’t you the same type of person who used to cry about USAID being a front for US imperialism? Now you want to complain that it’s gone, because oh wait they were actually helping to contain infectious diseases, provide clean drinking water, and feed people in food scarce areas…
That got shut down last year, by the way, not “just this year.”
As for catastrophic actions China has engaged in, I suppose if you ignore forced assimilations in Tibet and Urumqi, the hostile police takeover of Hong Kong, and the aggressive and environmentally destructive colonization of the South China Sea, then it would start to be somewhat difficult to come up with any examples…
Its rather revealing that your response to “China has severe restrictions on political speech” is to accuse that person of bootlicking for American empire.
You claim someone is biased against China, because they said one true thing, then you bring up all kinds of unrelated things to attack another nation, which wasn’t being discussed. I wonder if you might possess the bias your against, But for China and against the US? Both nations do good and bad things. If you’re so angry about bias, maybe you should check your own. Or maybe bias isn’t the issue, and you just support the Chinese leadership?
“See, China did a bad thing in 1989, while listed here are multiple worse things the US did this year alone. These are equal”
I’m unfamiliar with AI chatbots that you pay for. What is a token?
A token is basically just a word. Know how your phone’s auto suggest tries to anticipate the words you want to use as you type? In this case, your phone is using an extremely small token amount (typically only the previous two or three words you have typed) to try and predict your next word, which would also be a token. Your phone only uses a few tokens at a time, because as token count rises, processing requirements also quickly balloon.
And AI chat is basically the same concept, but with a massively inflated token limit. Instead of looking at your previous two or three words, it looks at entire conversations. And it also uses tokens to generate responses, the same way your phone is using one token at a time to predict your next word.
So when you pay for tokens, you’re essentially paying for a word count. As you continue a conversation, the token requirement for each subsequent request will increase, because it is attempting to look at the entire context of the conversation you have had.
Models have built-in token limits, to put a cap on how much memory is required to run the model. As conversations stretch on and you reach the model’s token limits, it will begin losing context for things that happened earlier. It will try to summarize earlier parts of the conversation to shorten them but keep relevant pieces in memory, or it will just outright drop old parts of the conversation and “forget” that context, the same way my phone has already forgotten the start of this sentence.
It’s a little more complicated that “each word is a token”, because the chatbot will combine your prompts with its own internal systems. Especially as conversations stretch on, and it begins to summarize old parts to keep them in memory. But that’s the most straightforward way to explain it.
My understanding is that tokens are basically words, and that when you ask a question it charges for all the tokens it consumes, produces, or processes. There’s a lot of internal processing for each request, where the input text is summarized in different ways and combined with previous parts of the conversation, so it’s not as straightforward as “word count of what you say plus what it says”.
Worth noting that a token is not necessarily a word, though can be. One word could also take multiple tokens. It can also vary from LLM to LLM and their tokenization methods.
There’s a lot of internal processing for each request, where the input text is summarized in different ways and combined with previous parts of the conversation, so it’s not as straightforward as “word count of what you say plus what it says”.
In other words obfuscation so they can charge whatever they want using some obscure formula that only they know.
Not really, there are ways to count tokens before running an inference. Some providers make tokenizers public, so they even work offline. APIs also usually return rolling costs per response and have budget limits - though some could have more fine-grained limits.
Users who are surprised by the bill are usually not paying attention to each call, or using autonomous subagents, or a setup where they have little or no control to what is sent to the provider.
So the problem isn’t really the API provider, as much as it’s the tooling around it, which makes it too easy to overspend.
In very simple terms, a token is more or less a word. You pay per input and output tokens (your prompts and the answers) as they correlate the most closely to the energy expended by the LLM to process your request.
Damn, fire sale already?
Thank you daddy Xi
They have to cut the price because its behind the frontier models. No one would buy it otherwise
The ones paying attention and on a budget would still use them. “The best” of anything is usually not cost effective.
Even before reducing the prices, they were already 2 to 3 times cheaper than equivalent alternatives from Anthropic’s ($3in, $15out) and OpenAI’s ($1.75in, $14out) at $1.74in and $3.48out. Now they’re around 10x cheaper.
How fast do you burn through tokens that $4 for a million of them was a lot of money?
If you use it for Q&A, that’s a lot of tokens. If you use it to write software somewhat autonomously, it’s easy to go through a million tokens every few hours. Do that every day and you’ll be paying over $100 a month at that rate.
Still gonna self host it instead (maybe)
FYI the flash model is ~158 GB
How are they running it? Doesn’t the model have to fit in (V)RAM? Does Nvidia have such huge memories in the H cards?
For self hosting it essentially needs to fit in VRAM + RAM but it’ll take a lot of CPU for the part in RAM
Deepseek probably uses those big fancy H cards and not one but several together to increase VRAM.
The destiled models?
