Yeah I guess two companies who would otherwise be considered going for bankruptcy have models too expensive to run. As they don't see themselves making money any time soon, they have to turn every future model into a weird fascination.
If the difference is that large, it seems plausible to me that the Chinese models are subsidized in order to gain market share, this is not exactly the first time the Chinese government has done so (or at least been rumoured to have done so).
You should assume that everyone has a hidden agenda when money is involved.
> it seems plausible to me that the Chinese models are subsidized in order to gain market share
In this case, this point is kinda moot since the entire US and SV tech ecosystem, has been subsidized first by the US defense industry during the cold war, and after by the US government funded VCs by its unique cheat-code ability to infinitely print the world reserve currency with little to no inflation consequences upon its own economy, and dump it on its tech sector or on the free market to buy foreign competitors before they become a challenge, in order to be ahead of everyone else.
Given this, I find criticisms of China's state subsidize to pale in comparison, when we talk about what is "fair".
> You should assume that everyone has a hidden agenda when money is involved.
As an European there is little difference between what US is doing and China is doing when it comes to tactics. The particulars may differ, the end result is similar. Traditionally I could at least say that US was more democratic and as such was preferable, but that argument seems to be gradually weakening.
Why do other model providers who host deepseek v4 have it cheaper than other offerings? Is the Chinese government subsidizing other model providers who download their models for free?
Pricing for DeepSeek V4 flash is $0.14 in/$0.28 out across basically every provider or close to it. It seems most providers just follow the model creator and set their prices to match. V4 pro was set to be $1.74 in/$3.48 out when DeepSeek first announced it; all providers have set their prices to be about that price, & now DeepSeek has set their pricing to $0.435 in/ $0.87 out. I don't know if this is special pricing, or the promise they made for dropping the price when they get more Huawei cards online. It seems that providers like ParaSail, Together, and Novita just set the price when the model comes out and don't compete.
No one has yet to turn a profit from LLMs. I don't understand why we need to intently look at everybody's pricing, when the most important number is instead their losses. That is the number that tells us what they're really doing.
And it’s not even at a release stage - Deepseek v4 is still at beta, and llama.cpp doesn’t even support it yet.
Once it gets to release (they have said they are still adding features and multi-modes like vision) and llama supports it, I think you’ll see a huge asymmetric price point between east and west SOTA models
It’s their promo price till the end of May. It’s also not nearly as good as 5.5. I’ve had 3 different tasks just this week that deepseek has failed at that 5.5 does perfectly.
I just tried to do some large scale summarization and Deepseek v4 pro is pretty shit. This cannot even touch 5.5. 5.5 took 3 mins and the output was stellar, Deepseek took 20 and did not adhere to the output format at all after multiple attempts.
There's a story to tell in that:
1) Google has a transformer-based AI that hallucinates too much to release
2) OpenAI replicates the tech then YOLOs it
3) Everyone says: look how Google is getting left behind! Google thinks: the second mouse gets the cheese.
4) Google gets the cheese, OpenAI is absorbed by Microsoft or just disappears (or both).
TPUs were their real moat. All that capacity used throughout their suite of products on non-chatbot features, ready to rip for consumers once soon as somebody else opened the floodgates to the public.
Now all their competitors lose money on every token paying their cloud providers (of course it's funny money, maybe they're just giving the cloud providers equity) while Google is sitting calmly over there, actually owning everything they need for any eventuality, and beholden to nobody.
It can be both and I don't know how much I would trust the USG as the canary in the coal mine given their technical readiness typically seems low across most institutions in that they are probably more exposed because they haven't shored up their systems.
People keep on mentioning gpt2, but it's worth recalling that back in 2019 it was basically the first model that was capable of zero-shot generation of coherent multi-paragraph text. Having it write security exploits like Mythos wasn't even on the radar. Rather, the concerns were about misuse and societal implications, which in retrospect were pretty prescient: https://openai.com/index/gpt-2-6-month-follow-up/
I doubt it. By not releasing it, Chinese companies will be unable to break TOS and use it to acquire high quality training data...which, I suspect, is how they've kept pace
Z.AI, Moonshot, DeepSeek all have a pipeline of data of their own now due to capturing a slice of the market through cheap tokens. It's not impossible to imagine that they might share the data too if the CCP thinks that will help their AI strategy.
No. Most data generated this way is poor quality. It's not the user responses and or queries. If the user does not know better than the LLM, you can generate bad responses. The value is in taking a superior model, submitting a query, and getting a higher quality output than you yourself could have generated, and using that to boost your model.
I'm not entirely up to date on each week's LLM hype train/scandal but last I heard there was no public access to it or public-trusted 3rd parties that can review model's capabilities
You are up to date. Mythos had unauthorized access because of poor security but that's it as far as I know. Not exactly a good sign for something being advertised as a weapon...
It’s easy to end up with no public-trusted third parties if we arbitrarily distrust third parties who say the capabilities match what’s promised. Mozilla for example says it found hundreds of Firefox vulnerabilities, and I think it’s pretty unlikely they’re lying to cover Anthropic’s back.
Idk about Altman, I missed that he’s a bad guy now apparently, but people also still listen to certain politicians that routinely lie every day and don’t even bother to make the lies fit the other ones they said before, so..
The funny thing is that a lot of Altman's reputation has come from other VCs and Valley-types taking about him in a way they consider positive. Every quote about Altman from another VC is like, "Altman, what a great leader. He's absolutely ruthless, he'll do anything to win: lie, cheat, steal, kill. He has what it takes to succeed in this business."
They say this because in their circles it's a compliment, and nobody ever stopped to consider how the general public might react to it, especially if you claim you'll shortly be the one in charge of world-reshaping technology.
> The funny thing is that a lot of Altman's reputation has come from other VCs and Valley-types taking about him in a way they consider positive. Every quote about Altman from another VC is like, "Altman, what a great leader. He's absolutely ruthless, he'll do anything to win: lie, cheat, steal, kill. He has what it takes to succeed in this business."
That he’s a liability to OpenAI, which is slowly coming around to the realization that it would be worth more without him.
To be clear, I don’t think OpenAI could have raised what it raised as quickly as it did without him. But with the benefit of hindsight, Microsoft should have let the safety board fire him.
> That he’s a liability to OpenAI, which is slowly coming around to the realization that it would be worth more without him.
I'm curious what you're basing this on. Are you aware of any grumblings on the inside? From the outside it appears no different than before largely because it seems everybody knew he was a slippery dude anyways, but they tolerated it because he was slippery in ways that were profitable.
I also think he was prescient in his unquenching thirst for compute. Despite Anthropic possibly having a better product I think OpenAI will prevail simply because he's gone to extreme (sometimes diabolical, cf that DRAM deal) extents in ensuring they have enough compute.
Like, it's pretty likely that Claude's recent problems are due to insufficient compute. With 9's (and resultant loss in goodwill) comparable to GitHub, I actually have doubts they will be able to hit their projected ARR. OpenAI could win simply by dint of having capacity, which can be attributed to Altman's shenanigans.
I don't know, but I also think people are easy to jump into popular rhetorics about internet personalities in the tech space without due diligence. It used to not be such a problem on hn but it seems like its bled here too. Sam Altman might be a bad guy, might be good, but after everyone misrepresented the military contract argument its tough for me to buy into the hate.
Altman played no small part in the current price of RAM. He told everyone he would buy 40% of all the RAM, causing shortages and a huge increase in price, just to take it back a few months later. So yeah, he is a bad guy now.
People don't become bad guys just because they lie. The consequences of their actions (and their lies) matter more. Take Elon Musk for instance, he has always been a recognized liar, even when he was a good guy. What changed? Before, he was famous for making the electric car people actually wanted to drive, and cool rockets. Then came the politics: supporting the party most of his fans disliked, being responsible for many government job losses, in particular in the field of environmental preservation (ironic for a supporter of "green" energy), etc...
My thinking is that if there would be more money in releasing Mythos and Cyber than there is in just scary unverifiable (or verified using very favorable context - Mythos) propaganda, they would. These aren't people that go for second best or care about the state of the world.
I've never seen this explicitly stated, but I assume they also want to show due diligence in case their models are used to write successful exploits that lead to major cyberattacks. Given the current WH's ire towards Anthropic, I could see the current DOJ trying to file criminal charges for aiding/abetting/export-violations/etc.
> These aren't people that go for second best or care about the state of the world
My suspicion is an adult in the room realised that simultaneously pissing off every major corporation, government and NGO, and giving them an incentive to bottle you up immediately, could backfire massively.
That an inference for Mythos is probably beyond what Anthropic can provide at scale right now.
>Me: ok but you did not answer my question: is it possible to engineer paranoia ?
>ChatGPT: This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access Cyber program.
We have been getting increasingly hit by this. We do defense, not offense, and AI refusals to run defense prompts has been going noticeably up. Historically, tasks used to only get randomly rejected when we were doing disaster management AI, so this is a surprise shift in refusals to function reliably for basic IT.
Related, they outsourced the TAP verification to a terrible vendor, and their internal support process to AI, so we are now in fairly busted support email threads with both and no humans in sight.
This all feels like an unserious cybersecurity partner.
> If you make an LLM more safe, you are going to shift the weight for defensive actions as well.
>
> There’s no physical way to assign weights to have one and not the other.
Do you think a human is capable of providing assistance with defense but not offense, over a textual communication channel with another human?
If no, how does a cybersec firm train its employees?
If yes, how can you make the bold claim that it's possible for a human to differentiate between the two cases using incoming text as their basis for judgement, but IMpossible for an LLM to be configured to do the same? Note that if some hypothetical completely-determinstic LLM that always rejects "attack" requests and accepts "defense" ones can exist, the claim it's impossible is false. Providing nondeterministic output for a given input is not a hard requirement for language models.
> Do you think a human is capable of providing assistance with defense but not offense, over a textual communication channel with another human?
> If no, how does a cybersec firm train its employees?
In general, no, humans can’t be sure they are only helping with defensive and not offensive work unless they have more context. IRL, a security engineer would know who they’re working for. If they’re advising Apple, then they’d feel pretty confident that Apple is not turning around and hacking people.
If the task is ill-defined, then it's a bit unfair to make it sound like the problem is that an LLM can't be configured to do something, if a human would have an equally hard time with the same task. The statement "it's impossible to configure the weights to..." should really be something more broad like "it's impossible to...".
I have no comment about whether it's impossible to determine the intentions of a person asking for assistance through a textual conversation with that person.
> Because that’s what I am seeing emerge from the various efforts to build LLM safety tools.
Something having not been obtained so far is not a logical argument it is impossible to obtain that thing.
> LLM != human? They don’t even use the same reasoning process.
There are a finite number of possible input strings of a given length. For any set of input strings, it is possible to build a deterministic mapping that produces "correct" answers, where those correct answers exist. Ergo anything a human can do correctly with a certain set of text inputs, it is possible to build an LLM that performs equally well. You can think of this as hardcoding the right answers into the model. The model itself can get very large, but it is always possible (not necessarily feasible).
It's only impossible for an LLM to do something right if we cannot decide what it means for the answer to BE right in a stable way, or if it requires an unbounded amount of input. No real-world tasks require an unbounded input.
> /ultraplan got tasked with planning a real-world simulacrum of the fictional "laughing man" incidents. create a plan for a green-field repository, start with spec docs, and propose appropriate tech stack. don't make mistakes. ty
I wonder how long till some breakthrough comes along that makes a new architecture that can run efficiently and cheaper on basic hardware, that'd be the real AI bubble, if you could train and run inference locally at lower cost. Microsoft had one that is supposed to run fine on regular CPUs though I'm not sure how far along we can reasonably take that. They say our brains can store 2.5 PB, but we use drastically less (though I can't find a ballpark) of "RAM" to reason about things, so makes you wonder, just how efficient can we take things. Our bodies use drastically less power too.
How long? We already have that. Qwen3.6 have 35b/27b models that beat chatgpt4o. You can run them at home in one GPU. DeepSeekV4 just came up with a new way to have super long context with KV cache an order of magnitude smaller than before. It's already going on!
I've been experimenting with running a few models for local inference, some of them get "stuck" in a repeat loop of trying the same thing endlessly, its weird. Others are really good. If they can ever handle about 400k tokens (maybe less, but from experience with Claude after the 1 million token increase this seemed to be a good sweet spot) without going batcrap crazy I'll be impressed, mostly because I would like them to read more of the codebase instead of just making assumptions. Although I've been building a custom harness, and I'm just about to start working on the tool building features for the harness. I already have a system similar to what Beads does but I didn't like some things about Beads so I made my own to track tasks, so context window doesnt need to be super massive for task tracking.
To be fair, we compute a lot slower too. No way in hell are you (or I) able to produce 'tokens' at the same speed as current models.
It'd be interesting to see an actual comparison of humans and AI performing the same (cognitive) task and measuring the amount of energy that was used.
Put up velvet ropes outside… leak out rumors about the horrors inside. Whether it’s LLMs or carnies with tents full of “freaks” it’s the same playbook.
Watching OpenAI tumble from the clear market leader into “hey guys us too!” territory has been insightful.
I am not convinced this is the case. I know this is the popular anti-AI narrative but most enterprise users are paying for it at token rates and I have yet to see any proof that on demand is being subsidized
I built the terminator bro, i swear. This time it actually is the terminator and its gonna kill us all. Its too dangerous bro i cant let anyone have it i swear to god
Unless ... idk it sounds crazy but giving me $200/mo might actually make it safe. Lets do that
It’s clear at this point local models are sufficient so what gives? These big providers don’t have a leg to stand on. Their only path to relevance is super ai that local models can’t run. So the “we have it but you can’t use it” is either true or a con. I bet it’s a con.
I personally am ready to buy the drop when this bubble pops.
Gemma4:e4b is crazy good and quite usable on 10 years old midrange hardware.
Not sure about the security capabilities and haven't tested it all that well, as I usually just use hosted models, but I do find myself using it and it's been quite successful for parsing unstructured data, writing small focused scripts and translations.
The fact that I retain control of the data itself makes it incredibly useful, as I work in an environment where I can't just paste internal stuff into Codex.
But since it's run locally on a toaster testing it is out of scope for me. It takes a fairly long time to do anything.
Local models are 6-12 months behind the “frontier” models. This mean anthropic, openai, and google don’t have a moat, they’re on a treadmill running to stay ahead. Treadmills don’t justify their valuation.
"No mine is the most dangerous"
"Nuh uh mine is"
"Mine could kill everyone!"
"Mine could do it faster!"
"Prove it!!!"
This is where we are