Deepseek - Chinese opensource ChatGPT rival
Discussion
I bet this has got the sphincter of some American AI company investors twitching!
Article at The Register about it.
Apparently if you ask it about the Tiananmen Square massacre or I daresay the Uighur Muslims it doesn't answer.
An interesting development though.
Article at The Register about it.
Apparently if you ask it about the Tiananmen Square massacre or I daresay the Uighur Muslims it doesn't answer.
An interesting development though.
It's quite amazing, considering it costs £3.99 to develop and ongoing projects on it are going to cost peanuts compared to the billions quoted by american companies.
NVidias growth is largely driven by AI and this was developed on 2000 GPUs not the huge amount of computers the US AI models use.
A game changer.
NVidias growth is largely driven by AI and this was developed on 2000 GPUs not the huge amount of computers the US AI models use.
A game changer.
BigMon said:
Apparently if you ask it about the Tiananmen Square massacre or I daresay the Uighur Muslims it doesn't answer.
Interestingly it starts to answer and then as it gets to the keyword it censors itself, but it does it live so you can see what happens. Quite a shonky experience, and not one that I think will inspire confidence when there's obvious government involvement in the answers. As much as you can criticise Altman for being an abrasive weirdo, OpenAI is tied to US tech companies, and comes with an expectation of transparency. Gemini is the same, where (unlike in China) a lot of their reputation relies upon trust.
I'm still unsure on AI, having a software background and working in a tech company that is looking to use it, I'm still not sold on it being more than a re-attempt of calling machine learning something different to market it better. All we have now is an interface and way of presenting results that makes it much prettier to people.
jimmsy said:
I'm still unsure on AI, having a software background and working in a tech company that is looking to use it, I'm still not sold on it being more than a re-attempt of calling machine learning something different to market it better. All we have now is an interface and way of presenting results that makes it much prettier to people.
I know what you mean but I think it is going to, in time, render a lot of current jobs unviable.jimmsy said:
Interestingly it starts to answer and then as it gets to the keyword it censors itself, but it does it live so you can see what happens. Quite a shonky experience, and not one that I think will inspire confidence when there's obvious government involvement in the answers.
As much as you can criticise Altman for being an abrasive weirdo, OpenAI is tied to US tech companies, and comes with an expectation of transparency. Gemini is the same, where (unlike in China) a lot of their reputation relies upon trust.
I'm still unsure on AI, having a software background and working in a tech company that is looking to use it, I'm still not sold on it being more than a re-attempt of calling machine learning something different to market it better. All we have now is an interface and way of presenting results that makes it much prettier to people.
Conversely Open AI don’t allow their models to show their working to avoid giving information inadvertently. As much as you can criticise Altman for being an abrasive weirdo, OpenAI is tied to US tech companies, and comes with an expectation of transparency. Gemini is the same, where (unlike in China) a lot of their reputation relies upon trust.
I'm still unsure on AI, having a software background and working in a tech company that is looking to use it, I'm still not sold on it being more than a re-attempt of calling machine learning something different to market it better. All we have now is an interface and way of presenting results that makes it much prettier to people.
Things like dangerous material etc so there is similar processing of information, they just don’t let you see that it happens.
Nvidia stock is down ~11.5% pre-market and I presume it's going to go down more when they open.
I've got a decent amount of money in AI stocks, including Oracle. All of them are at least 4-5% down pre-market.
Could be the beginnings of a black swan event for any/all AI-adjacent US stocks unless Deepseek is somehow revealed to be smoke and mirrors (which seems unlikely as it is apparently 100% open source?)
I've got a decent amount of money in AI stocks, including Oracle. All of them are at least 4-5% down pre-market.
Could be the beginnings of a black swan event for any/all AI-adjacent US stocks unless Deepseek is somehow revealed to be smoke and mirrors (which seems unlikely as it is apparently 100% open source?)
Durzel said:
Nvidia stock is down ~11.5% pre-market and I presume it's going to go down more when they open.
I've got a decent amount of money in AI stocks, including Oracle. All of them are at least 4-5% down pre-market.
Could be the beginnings of a black swan event for any/all AI-adjacent US stocks unless Deepseek is somehow revealed to be smoke and mirrors (which seems unlikely as it is apparently 100% open source?)
It's answers, based on my testing, seem very similar to ChatGPT.I've got a decent amount of money in AI stocks, including Oracle. All of them are at least 4-5% down pre-market.
Could be the beginnings of a black swan event for any/all AI-adjacent US stocks unless Deepseek is somehow revealed to be smoke and mirrors (which seems unlikely as it is apparently 100% open source?)
I'd imagine stocks will drop like a stone, I was watching Trumps speech this weekend saying how they are going to be a world leader in AI and his executive order to build loads of power stations as AI need more power than they produce.
Then the Chinese do it on 2000 RTX cards.
That's gonna make him look a bit silly.
Then the Chinese do it on 2000 RTX cards.
That's gonna make him look a bit silly.
There is something in what Sam Altman was saying about it being easier to catch up than make the breakthrough. The issue the US have is that China are doing it on smaller numbers of vastly cheaper kit (because of US blocking the best kit) and so able to under cut the US companies on price massively. So when companies are looking to purchase products they can get 90-95% of the capability for a much lower cost.
And whilst there will be questions about what the models can’t talk about and whether that has an impact, which may limit which countries China can sell to, it gives their companies a huge advantage over others on just a cost basis alone.
And whilst there will be questions about what the models can’t talk about and whether that has an impact, which may limit which countries China can sell to, it gives their companies a huge advantage over others on just a cost basis alone.
Sheets Tabuer said:
I'd imagine stocks will drop like a stone, I was watching Trumps speech this weekend saying how they are going to be a world leader in AI and his executive order to build loads of power stations as AI need more power than they produce.
Then the Chinese do it on 2000 RTX cards.
That's gonna make him look a bit silly.
It's worth remembering that the human brain is the size of a grapefruit and runs all day on a couple of mars bars, and most of that isn't even contributing to the general problem solving skillz. The route to efficient ANNs isn't going to involve conventional CPUs or GPUs, and the economics of chucking resources at doing ANNs really inefficiently is going to eventually force some better engineering.Then the Chinese do it on 2000 RTX cards.
That's gonna make him look a bit silly.
BigMon said:
I asked it about the treatment of Uighur Muslims in China and it gave me a full answer (mentioning genocide several times), then deleted it and said it was beyond it's current scope.
Interesting.
I cant shake the feeling that this isn't powered by a bunch of Uighurs doing the typing anywayInteresting.
BigMon said:
I know what you mean but I think it is going to, in time, render a lot of current jobs unviable.
I expect it'll make a lot of "non-jobs" redundant.Generative AI has a few niche uses, but most of it is user facing stuff, being a better/easier way of searching information, as you can RAG it against your own docs, then ask it questions about said docs. Likewise help bots, helpdesk triage etc.
There's also the "writing" words type stuff, and "making images" type stuff, but as ever that'll just remove people who were bad at it to start with. The people who were good at it, might find this gives them a quicker start, but they'll still need to add their own stuff to the mix.
It's the same sort of thing with coding, it can write you little snippets, or a whole game you didn't need in Python... but it won't turn a customers vague request to change a complex system that's been modified outside it's design pattern over decades and whack out all the code needed. So like the above people who weren't very good will be out of a job, people who didn't write noddy code for greenfield projects will continue to be needed. And anytime you ask it for any facts, it needs checking, so again not something you'd use if you knew what you'd doing, as you already knew.
If anything I think the big deal about generative AI, isn't all the stuff that's being pushed at the moment, but the way it makes searching so much more natural for the end user. I think once the hype has died down this, is where we'll see more of it, as the user interface to systems and sites to make interacting with them, finding things etc much more natural and easy.
But that's just my guess.
Dingu said:
There is something in what Sam Altman was saying about it being easier to catch up than make the breakthrough. The issue the US have is that China are doing it on smaller numbers of vastly cheaper kit (because of US blocking the best kit) and so able to under cut the US companies on price massively. So when companies are looking to purchase products they can get 90-95% of the capability for a much lower cost.
And whilst there will be questions about what the models can’t talk about and whether that has an impact, which may limit which countries China can sell to, it gives their companies a huge advantage over others on just a cost basis alone.
That's the devil in the detail imo.And whilst there will be questions about what the models can’t talk about and whether that has an impact, which may limit which countries China can sell to, it gives their companies a huge advantage over others on just a cost basis alone.
America will ultimately jump ahead because they have more compute, and the latest Nvidia kit. But plenty of companies will look at Deepseek and think "that's good enough for us". Instead of having to pay ChatGPT licensing costs, they might just go with the Chinese version instead. It sounds like it's broadly equivalent.
It pretty refreshing to see a project where the developers have exploited the limited hardware they've got, rather than complacency where developers just aim at a "good enough" level on the basis that hardware will get better. That's a oversimplication of course, but China have definitely brought into question whether all of the billions these US tech companies have spent is legitimate.
phil4 said:
I expect it'll make a lot of "non-jobs" redundant.
Generative AI has a few niche uses, but most of it is user facing stuff, being a better/easier way of searching information, as you can RAG it against your own docs, then ask it questions about said docs. Likewise help bots, helpdesk triage etc.
There's also the "writing" words type stuff, and "making images" type stuff, but as ever that'll just remove people who were bad at it to start with. The people who were good at it, might find this gives them a quicker start, but they'll still need to add their own stuff to the mix.
It's the same sort of thing with coding, it can write you little snippets, or a whole game you didn't need in Python... but it won't turn a customers vague request to change a complex system that's been modified outside it's design pattern over decades and whack out all the code needed. So like the above people who weren't very good will be out of a job, people who didn't write noddy code for greenfield projects will continue to be needed. And anytime you ask it for any facts, it needs checking, so again not something you'd use if you knew what you'd doing, as you already knew.
If anything I think the big deal about generative AI, isn't all the stuff that's being pushed at the moment, but the way it makes searching so much more natural for the end user. I think once the hype has died down this, is where we'll see more of it, as the user interface to systems and sites to make interacting with them, finding things etc much more natural and easy.
But that's just my guess.
I think it's effect will be far reaching and deep. Very deep.Generative AI has a few niche uses, but most of it is user facing stuff, being a better/easier way of searching information, as you can RAG it against your own docs, then ask it questions about said docs. Likewise help bots, helpdesk triage etc.
There's also the "writing" words type stuff, and "making images" type stuff, but as ever that'll just remove people who were bad at it to start with. The people who were good at it, might find this gives them a quicker start, but they'll still need to add their own stuff to the mix.
It's the same sort of thing with coding, it can write you little snippets, or a whole game you didn't need in Python... but it won't turn a customers vague request to change a complex system that's been modified outside it's design pattern over decades and whack out all the code needed. So like the above people who weren't very good will be out of a job, people who didn't write noddy code for greenfield projects will continue to be needed. And anytime you ask it for any facts, it needs checking, so again not something you'd use if you knew what you'd doing, as you already knew.
If anything I think the big deal about generative AI, isn't all the stuff that's being pushed at the moment, but the way it makes searching so much more natural for the end user. I think once the hype has died down this, is where we'll see more of it, as the user interface to systems and sites to make interacting with them, finding things etc much more natural and easy.
But that's just my guess.
Everything from accountants to architects, copywriters to back office admin. I You can kiss goodbye to 20% of the workforce in the not too distant future.
Sim75 said:
Timing feels pretty coincidental.
China taking a swipe at the US?
On the back of Trump's AI announcement? China taking a swipe at the US?
That's very much what I think.
Is this a 'thing' (I hate that phrase), or is it something which has caught the headlines and will recede back into nothingness by the end of the week?
Let's wait until the end of the week to find out.
Collectingbrass said:
BigMon said:
I asked it about the treatment of Uighur Muslims in China and it gave me a full answer (mentioning genocide several times), then deleted it and said it was beyond it's current scope.
Interesting.
I cant shake the feeling that this isn't powered by a bunch of Uighurs doing the typing anywayInteresting.
anywayTim Cognito said:
Necessity is the mother of invention.
Quite ironic that the chip ban, intended to impede Chinese AI progress has probably led to the engineers making massive efficiency improvements.
Although this - https://fortune.com/2025/01/27/china-deepseek-nvid...Quite ironic that the chip ban, intended to impede Chinese AI progress has probably led to the engineers making massive efficiency improvements.
- includes a claim that DeepSeek actually used 50,000 nVidia H100 GPUs.
So perhaps was trained on more than just the H800s they've admitted using.
Gassing Station | Computers, Gadgets & Stuff | Top of Page | What's New | My Stuff