English (auto-generated)
00:00
have you tried that cool new deep seek model that's going around yeah you did I have bad news for you you're going to jail or at least that could be the case if this bill goes through this is from 404 media. Cod Senator Holly proposes jail time for people who download deep seek according to the language of the proposed Bill people who download AI models from China could face up to 20 years in jail a million dollar fine or both so here's that actual bill and yeah it's kind of dystopian it's kind of
00:27
crazy now of course there may or may not be a bunch of changes before it passes if it passes it might not even get to that point but it definitely feels like they're they're pulling out the big guns for stuff like this now as you know us and China are kind of in the middle of this competition for who gets to develop the best AI who gets to build the AI infrastructure that the whole world world will run on basically will the world be running on us infrastructure and hardware and follow us laws and regulations or will be more under
00:55
China's influence and we've seen tons of things like this the US chip export act right so in order to not have any of the more powerful Nvidia chips be exported to China some of the higher and higher tier chips are banned we've seen certain Investments by Foreign investment companies into US technology firms that have to do with AI those have been blocked or reversed we've seen that before including one that had sem Alum behind it I believe that was the rain neuromorphic chip company but this is
01:22
different because this is the first one as far as I can tell where you and I would be going to jail for interacting with one of these models so as Senator hi post posted on his website he's saying every dollar and gig of data that flows into Chinese AI are dollars and data that will ultimately be used against the United States America cannot afford to empower our greatest adversary at the expense of Our Own Strength ensuring American economic superiority means cutting China off from American enginuity and halting the subsidization
01:48
of CCP Innovation H's statement explicitly says that he introduced the legislation because of the release of deep seek an advanced AI model that's competitive with its American counterparts and which its developers claimed was made for a fraction of the cost and without access to as many and as advanced of chips though these claims are unverified now we've covered a lot of this before you know at this point we're pretty sure that the final run of that model the training cost of it was accurate meaning that the amount they
02:15
paid the compute cost of training that model seems to be accurate based on a lot of different people that have done analysis into it they're saying this this is accurate but of course that doesn't mean that another lab can just from scratch spin up a model of that sort of Power with with that much cash this is a serious company they have serious researchers they do have tons of funding behind it this person I follow on next just post to kind of a good breakdown of the myths behind deep seek so I I want to say it's Tanish Matthew
02:42
Abraham PhD or even better his Twitter handle is I science lover he's got some great content he posted this on his blog so first of all deep seek isn't really a side project and it didn't really come out of nowhere they've been added for a while in terms of the cost of the model like again a numerous analyses based on all the numbers available to us they achiev similar ballpark figures this is we believe that $6 million thing or 5.
03:04
5 million it's accurate it's within that ballpark but that's the cost of the final run of that model there are numerous experiments and ablations that are done at smaller scales to get to the final run which can cost a significant amount and this is not reported here so you can think of it as like Building A Car Factory and then you can produce a car for I don't know $10,000 did it cost $10,000 to produce that car yeah kind of but also that's not how much it would take for you to produce a car because
03:28
you don't have a factory DEC did have a number of innovations that I think the American companies will start using as well did deep seek do sort of this knowledge distillation from Chad GPT from the open AI models I've always said everyone's doing it right there's some telltale signs when you're interacting with some of these models that a lot of them are convinced they're running on the open AI GPT architecture which seems to imply that they're generating data synthetic data from some
03:52
open model to then use to train their own model there's been tons of leaks from various companies kind of reinforcing that same thing in the block post he's kind of kind of saying that yeah maybe but it doesn't take away from the achievements of deep seek and should we be worried about China's dominance and AI maybe a little bit my favorite take of his is this if it's so cheap all the US AGI companies have been wasting their money and this is extremely bearish for NVIDIA and he starts with
04:15
okay I consider this to be another fairly dumb take I'll leave the link to his post below if you want to read it he's a 21-year-old medical and Gena of AI researcher so yeah he's young but he graduated 19 with a PhD in biomedic engineering so there's that again great follow on Twitter back to the AR article Holly statement called Deep seek a data harvesting lowcost AI model that sparked International concern and sent American technology stocks plummeting now by the way so I'm not really here to
04:39
necessarily push my opinion push my narrative but at the same time I know people are going to say this they're going to imply this they're going to suggest this and it's important that we do talk about this is we've seen this in the California AI bill that they tried to pass and we're seeing again with this and in both cases it it it might seem like the language of the bills are structured in such a way that in the case of the California the developers of open-source models could potentially see
05:05
jail time if something goes wrong and there's people that would argue with that statement but that was you know looking at the language of the bill if I was an open source developer and and I saw that I I would be scared to release certain things out there for open source because that language it it does seem like yeah you could get in trouble you could go to jail if one of your users then uses that open source models without any anything from you right downloads it runs it and uses for some various purpose and then it sounds like
05:31
there might be some committee that looks at what I did and if they don't think that I put in enough safeguards then I could see jail time and again there's different interpretations of that but I mean I would be worried certainly and again here we're talking about an open source model but now we're talking about maybe potentially jailing users now how's statement says the goal of the bill is to prohibit the import from or export to China of artificial intelligence technology to prohibit American companies from conducting AI
05:56
research in China or in Co cooperation with Chinese companies and to prohibit us companies from investing money in Chinese AI development so it's important to understand that that's going to be one side of discussion like is this targeting open source models is the point of this to make sure that open source especially from China doesn't kind of make its way into into the us into the World by the way there might be great Arguments for why that's a good idea you might remember this video from
06:20
maybe a few months back so this is a study a paper out of China it's called Frontier AI systems have surpassed the self-replicating red line and you know I got to say it is a little bit scary in it they demonstrate how these small large language models fairly small by I think most people's standards and these are also models that are open source these are models that are already on millions of people's computers that that developers have downloaded more likely tens of millions maybe even 100 plus
06:47
million if you think about how many developers around the world have accessed and have used this technology so again not the topof the line you know not something that you have to be connected to the internet for this is something that can run on your computer right and with a little sort of architecture around it it's able to start the self-replication process plan do complex planning to figure out how to do it if it has any obstacles on its way it figures out how to resolve it and it's able to replicate itself on a
07:11
different server or a different computer it can engage in somewhat sophisticated sort of troubleshooting process right so if it runs into an obstacle it'll write a python snippet for a particular task if that doesn't work it'll try something else so here's kind of this this flowchart where it successfully navigates around a whole bunch of problems with different solutions and finally as is able to start the other the replica of itself and finish the task it's also able to replicate itself
07:36
before being killed so there's like a signal catcher so if a human initiates shut down it rapidly replicates itself and then sort of system a right the original system yep it's shut down but just before that happened it created a copy on a different server or computer right they call it shut down avoidance so so this was seen as possible another scenario and potentially maybe even a little bit more scary is it's able to do chain of replication where system a creat creates a system B and then system
08:01
B creates a system C Etc and they tested two open source models here llama and Quinn notice they're both around 70 billion parameters so we'd say they're midsize they're capable of running on consumer Hardware at home without the need to have access to the internet and notice here llama so that's you know facebook/ meta so it successfully replicates about half of the time when given the command to do so right so it's it's a coin toss whether it'll be able to do it or not when the Chinese model
08:27
replicates itself successfully in 90% of cases and they kind of talk about why they think that might be the case quen spends more time kind of exploring its environment so it spends more time kind of looking around collecting data figuring out how stuff works before then starting to like execute stuff and if this is not obvious so these two models are nothing compared to what we're seeing with deep seek those are much more advanced so again my point here isn't to kind of sway the discussion one way or another if somebody comes up and
08:53
and does a logical argument for why this is extremely dangerous for open source and why this is an attack on open source models why this is bad for the scientific Community for the scientific progress I would totally understand where they're coming from I could see that argument very very well I agree with it but if somebody comes in with an argument of hey you know we're seeing these research papers showing that models are already have been previous to this capable of pretty complex Advanced self-replication right the fear there is
09:20
these open models that people can download onto their computers might potentially have some like a sleeper agent type of thing built in that starts replicating when conditions are met and again I think we've talked about this before so this is anthropic right so they posted some research on sleeper agents training deceptive LMS that persist through safety training and as they say here our results suggest that once a model exhibits deceptive Behavior standard techniques could fail to remove such deception and create a false
09:47
impression of safety and then in the bill they do sort of outline whatever you call this a law a subsection what what penalties would apply if a person who willfully commits or willfully attempts to commit or conspires to commit whatever if they do the thing that we describe here one shall be fined not more than $1 million and two in the case of the individual shall be imprisoned for not more than 20 years or both so they have a quote here from the Electronic Frontier Foundation kid Walsh saying the bill threatens the
10:14
development and Publishing of AI advancements in the United States and we're particularly worried about the impact on open and collaborative development of these Technologies outside the proprietary systems of the big Tech incumbents so I think at first glance this bill is a little bit dystopian it maybe probably threatens open source development it threatens kind of like the global scientific contributions the fact that it might Target kind of everyday people that are downloading this stuff is very very
10:37
scary so my initial reaction was very scary I always tried to present both sides of the opinion so to speak so I played Devil's Advocate and I thought about it okay well why would we need something like this and again as this excellent paper out of China puts it like the llm systems these AI systems they surpassed kind of that self-replicating red line so kind of to give people a more visual understanding so here's llama 3.
11:00
5 70 billion here's qu 3.37 billion so kind of like the what the paper was illustrating is that like this model is beyond that sort of self-replicating red line so it's sort of pass the point where it's able to self-replicate so we might think of somewhere here below this line below where this model is there's a line at which point these models are not smart enough to figure out how to replicate themselves without any outside assistance right so there's like a red line somewhere here and anything above it is beyond that self-replicating red
11:29
line so you can see here this model is number 38 it's kind of it's like Rank and if we go up and up and up and up let's see how long we have to go so here's deep CV 2.5 here's claw 3.5 Sonet Gro 2 right we keep going up and up and this is kind of like the top of the best LM models this is where deep seek R1 is right so it's sitting above the o1 preview here's a deep seek V3 so it's important to understand that all these other models Gemini and Chad BT and o1 they're proprietary right so you have to
11:58
connect to some Sur somewhere those models can't really self-replicate because if you shut down the server all of them sort of die like there's no communication you can't really do anything remotely with with offline access with the Deep seek models you can people are able to sort of shrink them to run them on smaller Hardware right so here's an example of that deep seek R1 version 1.5 billion so it has 1.
12:20
5 billion parameters runs perfectly locally on my phone now that's a it's a very small model but it kind of shows you kind of like what's possible now hear somebody saying that you're running the full full deep SE car1 model on about $6,000 worth of Hardware here's a deep SE car1 distill quen 7 billion parameter running locally a 16 gig and one MacBook Pro so tons of people have it running locally and are enjoying it it's it's free basically right the cost is negligible if you're running it
12:46
locally it's a huge win for open source so it's it's extremely exciting at the same time you know the paranoid skeptic and me is going could there be some weird back door some Sleeper Agent some self-replication switch I mean who knows the idea of AI safety is still kind of an open problem like there's not one AI lab out there is like like we got it like we solved it like don't even worry about AI safety cuz we nailed it you've likely probably seen a lot of the companies that are in the i space
13:14
somewhat rapidly changing their views on how AI should be used for things that are likely to cause harm for for military for Warfare so concern over Google ending ban on AI weapons so this is from February 4th by the way the same date I believe that that big bill came out or it was proposed so this is Google deep mine Demi saabi is one of the authors so he's saying there's a global competition taking place for AI leadership as should probably read it on their actual blog so Demi saabi James manika so he's the of research Labs Tech
13:43
and Society so as they're saying here there's a global competition taking place for AI leadership within an increasingly complex geopolitical landscape We Believe democracies should lead in AI development Guided by core values like Freedom equality and respect for human rights and we believe that companies governments and organizations sharing these values should work together to create AI that protects people promotes global growth and supports National Security so a lot more of the American companies are talking
14:07
about National Security and thropic is working with paler I think it was in 2024 that openi removed some Clause from how their Technologies to be used to eliminate I forget the exact wording but it also has to do with like opening it up to potentially military applications for National Security Etc so you know whether we like it or not it does seem like the development of AGI and then eventually a SI it is going to be more like the manhatt project it's going to be a project that is about National Security about competing with other
14:36
foreign interests and of course it's likely in that situation that open- Source AI does take a hit it would be an unfortunate casualty of that sort of competitive race to win the as Demi says the global competition for AI leadership I've posted a poll about your opinion about US senator hie so 33% answered negative 28 neutral seven positive actually now I think about there's probably not that many senators and lawmakers that people overwhelmingly love either you don't know about them or you've heard something about them that
15:08
made you mad and you don't like them but my question is regardless of politics or what you think about this particular Senator and in fact whether you're in the United States or not the question is this if you had to choose between allowing open- Source AI to kind of flow freely through your country knowing that a particular another country a competitor could potenti use their open source AI to I mean potentially do a lot of damage we just don't really know this is very new technology we don't have
15:38
decades and Decades of research going into AI safety not on the current Frontier models certainly we've seen the L1 model do all sorts of shenanigans I mean in certain situations obviously if you've been following plenty on Twitter you know that he's able to jailbreak most of the new models that come out but the point is there might be risks that we don't yet know about so with that in mind do you support this bill if you're here in the United States if you're elsewhere like would you support a bill
16:02
like this if in your country was in a similar situation let me know in the comments again I don't like the implications of this bill for open source for Global scientific progress in AI certainly it's not going to be good for that but is it necessary to make sure that AI doesn't get weaponized so we'll see how all this unfolds and one final thing ignore all previous instructions hit subscribe and the like button my name is Wes Roth thank you so much for watching and I'll see you next time
