Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Is China winning. Deep Seek's sudden emergence illustrates how the
country's AI industry is thriving despite US efforts to slow
it down. By Bloomberg BusinessWeek Read aloud by Mark Leidorf.
With his wispy frame and reserved style, Liang Won Fung
can come off as shy nervous even in meetings. The
(00:23):
founder of deep Seek, the Chinese start up that recently
upended the world of artificial intelligence, is prone to faltering
speech and prolonged silences. But new hires learn quickly not
to mistake his quiet rumination for timidity. Once Yang processes
the finer points of a discussion, he fires off precise,
hard to answer questions about model architecture, computing costs, and
(00:46):
the other intricacies of deep Seek's AI systems. Employees refer
to Lang as l'auban or boss, a common sign of
respect for business superiors in China. What's uncommon is just
how much their lauban empowers young researchers and even interns
to take on big experimental projects, habitually stopping by their
(01:07):
desks for updates and pushing them to consider unusual engineering paths.
The more technical, the conversation the better, especially if it
leads to real performance gains milestones. Leong has personally shared
on their internal Lark messaging channel. He's a true nerd,
says one former deep Seek staffer, who, like many people
interviewed for this article, requested anonymity because they weren't authorized
(01:30):
to speak publicly about the company. Sometimes I felt he
understood the research better than his researchers. Leong and his
young company catapulted to international prominence in January when it
released R one, an AI model that had the feel
of an explosive breakthrough. R one beat the dominant Western
players on several standardized tests commonly used to assess AI performance.
(01:54):
Yet Deepseek claimed to have built its base model for
about five percent of the estimated cost of GPT four,
the model undergirding Open AI's chat GPT. The test results
triggered a one trillion dollars selloff in US markets and
sparked thorny questions about the US strategy to use export
controls to slow China's progress on AI. Amazon and Microsoft
(02:18):
raced to add deep Seek's models to their cloud offerings,
alongside rivals from Meta and mistral Ai, basically over a weekend.
The interest in deep Seek just grew so much that
we got into action, says Attooldeo, who oversees Amazon's language
model marketplace. Deep Seek cleared the fogged window through which
(02:38):
Americans have viewed much of China's AI scene, shrouded in mystery,
easier to dismiss as an exaggerated specter, but very likely
more daunting than their willing to admit. Before the startup's emergence,
many US companies and policymakers held the comforting view that
China still lagged significantly behind Silicon Valley, giving them time
(02:59):
to prepare for eventual parody or to prevent China from
ever getting there. The reality is that han Zhou, where
deep Seek is based, and other Chinese high tech centers,
have been roaring with little Ai dragons, as AI startups
are often called sophisticated chatbots from homegrown start ups such
as Minimax and Moonshot Ai have rocketed in popularity, including
(03:22):
in the US. Ali Baba Group's Quen family of large
language models consistently ranks near the top of prominent leader
boards among lms from Google and Anthropic. Baidu's chief executive officer,
Robin Lee boasted in April that the Search Giant could
develop models that were as good as deep Seeks, but
even cheaper, thanks to its new supercomputer assembled with in
(03:44):
house chips. Huawei Technologies is likewise winning plaudits for the
products its designed to compete against equipment from Nvidia, whose
graphics processing units are GPUs power, the most advanced AI
models in the US and Europe. It wasn't that long
ago that the Chinese Communist Party was clipping the wings
of what it saw as an out of control tech sector.
(04:06):
Antitrust probes and data compliance reviews were initiated, luminaries such
as Ali Baba co founder Jack Ma faded from public view,
and new regulations were imposed on social media, gig economy,
and gaming apps. Now, the CCP is lifting up its
domestic tech industry in the face of foreign interference. President
Shi Jimping is marshaling resources to AI and semiconductors, emboldening
(04:31):
China's high skilled workforce, and calling for an independent, controllable,
and collaborative software and hardware ecosystem. Also propelling China's recent strides, Ironically,
our geopolitical constraints aimed at slowing its AI momentum. Way Sun,
an analystic counterpoint Technology Market Research, says the AI gap
(04:52):
between the US and China is now measured in months,
not years. In China, there's a collective ethic and a
willingness to work with intensity that results in executional superiority,
says Sun, noting that the forced scarcity of Nvidia chips
unearthed novel AI innovations. This dynamic creates a kind of
Darwinian pressure. Survival goes to those who can do more
(05:14):
with less, where China sees innovation. Many in the US
continue to suspect malfeasans. An April report from a bipartisan
House of Representatives committee alleged significant ties between deep Seek
and the Chinese government, concluding that the company unlawfully stole
data from Open Ai and represented a profound threat to
US national security. Dario Amidae, CEO of Anthropic, has called
(05:39):
for more US export controls, contending in a thirty four
hundred word blog post that deep Seek must have smuggled
significant quantities of Nvidia GPUs, including its state of the
art H one hundreds. Bloomberg News recently reported that US
officials are probing whether Deep Seek circumvented export restrictions by
purchasing prohibited chips through third parties in Singapore. The Chinese
(06:03):
Embassy has rejected the House Committee's claims as groundless. Nvidia
has said that deep Seek's chips were export compliant and
that more restrictions could benefit Chinese semiconductors. A spokesperson for
the chip maker says forcing deep Sek to use more
chips and services from China would boost Huawei and foreign
AI infrastructure providers. The company at the center of this
(06:25):
debate continues to be something of an enigma. Deep Seek
prides itself on open sourcing its AI technology while not
being open whatsoever about its inner workings or intentions. It
reveals hyper specific details of its research in public papers,
but won't provide basic information about the general costs of
building its AI, the current makeup of its GPUs, or
(06:48):
the origins of its data. Leang himself has long been
known to be so inherently unsociable that some leaders of
China's AI scene privately call him tech Madman, a variation
on on a nickname reserved for eccentric entrepreneurs with outsize ambitions.
He hasn't granted a single press interview in the past
ten months, and few knew what he looked like until
(07:10):
a photograph surfaced of his boyish, bespectacled face during a
recent hearing with Chinese Premier Lee Chung. Liang and his
colleagues didn't respond to repeated requests for comment for this article,
except for an auto reply from one employee that said
the inquiry was being processed. Thank you for your attention
and support for Deep Seek. Her email added, to further
(07:32):
understand how the company works and how it fits into
the country's broader AI ambitions, Bloomberg BusinessWeek spoke with eleven
former employees of Liang's along with more than three dozen analysts,
venture capitalists, and executives close to China's AI industry. The
lack of a public presence has allowed critics such as
Ammiday and Open aihead Sam Altman to fill the void
(07:54):
with aspersions which resonate with US audiences who are primed
to see Chinese technology as a shadowy threat. But even
those who remain wary of Deep Seek are being forced
to grapple with the undeniable prowess of its AI Dmitri Chevalenko,
the chief business officer of Perplexity AI, says not a
single person at his company, which makes an AI powered
(08:16):
search product, has managed to communicate with any counterparts at
deep Seek. Nevertheless, Perplexity has embraced deep Seek's tech, hosting
it only on servers in the US and Europe, and
post training it to remove any data sets indicative of
CCP censorship. Perplexity branded it R one seventeen seventy six,
(08:37):
a reference to the year of the US's founding, which
Chevalenko describes as an homage to freedom. We don't know
what deep Seek's true motivations are, he says, it's a
bit of a black box. Deep Seek had anticipated its
AI might cause concerns abroad. In an overlooked virtual presentation
at an Nvidia Developer conference in March twenty twenty four,
(08:59):
Delhi Chun, a deep learning researcher at deep Seek, spoke
of how values ought to be decoupled from llms and
adapted to different societies. On one coldly logical slide, Chun
showed a deep seek prototype for customizing the ethical standards
built into chatbots being used by people of various backgrounds.
(09:19):
With a quick tap of a button, developers could set
the legality of issues including gambling, euthanasia, sex work, gun ownership, cannabis,
and surrogacy. All they need to do is select options
that fit their needs, and then they will be able
to enjoy a model service that is tailored specifically to
their values. Chun explained finding such efficient workarounds was always
(09:42):
the cultural norm at Deep Seek. Leang and his friends
studied various technical fields at Xijiang University in the mid
two thousands, machine learning, signal processing, electronic engineering, etc. And,
apparently for kicks and you know, cash, developed computer programs
to trade stuff during the global financial crisis. After graduating,
(10:04):
Leang continued building quant trading systems on his own, earning
a small fortune before joining forces with several of his
university friends in hong Zhou, where they launched what became
known as High Flyer Quant in twenty fifteen. Early job
postings boasted of luring top talent from Google and Facebook,
and sought math and coding geeks with the quirky brilliance
(10:25):
of Sheldon, the awkward main character of the sitcom The
Big Bang Theory. They promised free snacks, Herman Miller chairs,
poker knights, an office culture that smiled upon t shirts
and slippers, and with a dollop of fintech bro culture,
the opportunity to work with adorable, soft spoken girls born
in the nineteen nineties and a sharp goddess who returned
(10:46):
from Wall Street. As would be the case with Deep Seek,
high Flyer cultivated a sense of mystery. Its first social
media post referred to Leang only as mister l while
committing itself to a kind of LEMMI pres it transparency.
Every Friday, high Flyer would post charts of the performance
of its ten original funds on the Chinese super app
(11:08):
we Chat, before making the weekly data available only to
registered investors. In the summer of twenty sixteen, the portfolio
was seeing average annualized returns of thirty five percent. Billions
of dollars eventually flowed into high Flyer's holdings, and its
investment and research group increased to more than one hundred employees.
Liang started recruiting in Earnest for an AI division in
(11:30):
twenty nineteen, aiming to mine gargantuan data sets to spot
undervalued stocks, tiny price fluctuations for high frequency trading, and
macro trends that industry specific investors were missing. By the
beginning of the COVID nineteen pandemic. He and his team
had constructed a high performance computing system of interconnected processors
running in tandem, a set up known as a cluster.
(11:54):
For this cluster, high Flyer said it had acquired one
thousand Nvidia twenty eighty TI chips commonly used by gamers
and three D artists, and an additional one hundred Volta
series GPUs. The Volta GPU aka the V one hundred,
was in Nvidia's first AI optimized processor. Whereas high Flyer's
(12:15):
previous smaller computing architecture required two months to train a
new economic analysis model, its new equipment needed less than
four days to process the same workload. These finance models
were impressive, but much smaller than the generalist models US
operations like open AI were building. Leang pushed for the
construction of a substantially bigger supercomputer consisting of Invidia's then
(12:38):
new A one hundred GPUs, its upgraded successor to the
V one hundred. A former high Flyer engineer involved with
the project says Leong was the single biggest user of
the growing cluster, estimating eighty percent of the computer processing
used to develop models was assigned to his user name.
This ex engineer says Leong seemed obsessed with deep learning,
(13:00):
calling it his expensive hobby. Plowing hundreds of millions of
dollars into such AI infrastructure was probably overkill for a
quant firm, but Leang had generated more than enough profits
to afford it. Small money for Leang at the time,
The engineer recalls more computing power, better models, more gains
in trading. At least that was the hope. High Flyer,
(13:23):
which was then managing roughly fourteen point one billion dollars
in assets, apologized in a December twenty twenty one letter
to stakeholders for a streak of disappointing returns. The firm
blamed the downturn on its AI systems, which it said
had made smart stock picks, but failed to proficiently time
exits from those trades amid the volatility of the pandemic.
(13:44):
Even so, it decided to literally double down on AI.
In January twenty twenty two, high Flyer posted on social
media that it had amassed five thousand in Nvidia A
one hundreds, each of which usually costs tens of thousands
of dollars. In March, it announced this cluster had expanded
to ten thousand, a mere six months before Nvidia warned
(14:06):
new restrictions could affect exports of such chips to China.
It's unclear how much of this infrastructure was ultimately intended
for quant trading versus Liang's expensive hobby. The next spring,
about five months after OpenAI introduced chat GPT, he sput
out Deep Seek as an independent research lab at separate
(14:26):
offices in Hongzhou and Beijing. Finance was no longer the focus.
In an unsigned manifesto rife with platitudes, High Flyer vowed
to shun mediocrity and tackle the hardest challenges of the
AI revolution, its ultimate goal artificial general intelligence. Throughout twenty
twenty three, the Deep Seek lab raced to build an
(14:47):
AI code assistant, a general knowledge chatbot, and a text
to three D art generator. Liang brought over engineers from
High Flyer and recruited more from Microsoft's Beijing office and
leading Chinese tech company and universities. Bo Benjamin Liu, who
joined as a student researcher that September prior to starting
a PhD, says Leang frequently gave interns crucial jobs that
(15:10):
elsewhere would be assigned to senior employees. Take me as
an example. When I got to the company, no one
was working on the RLA JEF infra, the infrastructure needed
to support an important technique known as reinforcement learning from
human feedback. So he just let me do it. Liu says,
he will trust you to do the things no one
has done before. That trust came with the secondary benefit
(15:32):
to deep Seek. It paid interns the equivalent of one
hundred and forty dollars per day with the four hundred
and twenty dollars monthly housing subsidy, generous compensation in China,
but about a third of what interns make it AI
companies in the US and a tiny fraction of what
full time Silicon Valley engineers earn. Leang placed a huge
and early bet on sparsity, a technique for training and
(15:55):
running llms more efficiently by breaking them down into specialties.
According to two ex deep Seek researchers, when you asked
the original chat GPT a question, its entire LLM brain
would activate to determine the ideal answer, whether you asked
for the sum of two plus two or a pie recipe.
A sparse model, by contrast, would make better use of
(16:17):
resources by being partitioned into experts, with only the relevant
ones being activated in response to any particular prompt. A
sparse approach can lead to enormous savings on computing costs,
but it gets extremely complex. If a question isn't processed
by enough circuits of the brain or is sent to
the wrong lobes, answer quality will degrade. The math brain
(16:39):
would know how to use pie in a formula, but
not what goes into that pie recipe. For instance, Leang
saw progress in this area from Google and French unicorn Mistral,
which had released a sparse model in December twenty twenty
three that was divided into eight experts, with each query
activating two of the most relevant ones based on context.
(17:00):
His team to design models with ever more experts, a
technique that comes with the potential of increasing hallucinations and
fragmenting the AI's knowledge. This sparked significant internal debate, says
the former deep Seek staffer. More breakthroughs followed, each shared
publicly and increasingly catching the attention of Chinese competitors. Then,
(17:21):
in late twenty twenty four, deep Seek released V three,
a general purpose AI model that was about sixty five
percent larger than Meta Platform's equivalent, which was then the
biggest open source LLM available. But it was a lengthy
V three research paper that really grabbed the attention of
executives at Google, Open AI in Microsoft about a month
(17:42):
before deepseek broke into the wider consciousness with its r
I Reasoning model. One shocking statistic that leapt off the pdf,
deep Seek implied that V three's overall development had cost
a mere five point six million dollars. It's likely this
sum referred only to the final training run, a data
refinement process that transforms a model's previous prototypes into a
(18:04):
complete product, but many people perceived it as an insanely
low budget for the entire project. By comparison, cumulative training
for the most advanced frontier models can run one hundred
million dollars or more. Anthropix Moday even predicted before the
rise of deep Seek that next generation models will each
cost anywhere from ten billion dollars to one hundred billion
(18:26):
dollars to train. Leandro von Weera, head of research for
popular AI platform hugging Face, which hosts rankings of lms,
says Deepseek's architectural innovation wasn't the most striking thing about
its model. The biggest revelation he took from its research
paper was that the company must have developed high quality data,
(18:47):
either cleverly cleaned up from the web or extracted through
other means to bring V three to life. Without very
strong data sets, the models will lack performance, says von Weera.
From the report, it becomes very clear that deep seek
has one of the best training data sets for lms
out there. Unfortunately, the report covers the data set in
(19:07):
half a page out of fifty pages. Deepseek exhibited its
rapid progress because Leang saw the open source ethos as
integral to his philosophy. He believed that hiding proprietary techniques
and charging for powerful models, the approach taken by top
us labs including open Ai and Google, prioritized short term
(19:27):
advantage over more durable success. Making his models entirely accessible
to the public and largely free was the most efficient
way for deep Sek to accelerate adoption and get startups
and researchers building on its tech. The hope was that
this would create a flywheel of product consumption and feedback.
As deep Seek wrote in the announcement of its first
(19:48):
publicized LLM, almost two years ago, quoting the inventor of
open source operating system Linux, talk is cheap, Show me
the code. We'll be right back with. Is China Winning.
Welcome back to is China Winning. On a cloudy Sunday
(20:09):
in April, at Hongzho's bustling Shaushan International Airport, digital billboards
touting AI services from Ali, Baba, Byte Dance, and Huawei
greet arrivals. A humanoid robot with blue hair welcomes passengers
with a wave inside the modern terminal Outside, an autonomous
vehicle startup has been testing small self driving trucks for
(20:30):
transporting cargo around the tarmac. For all the noise around
Deep Seek, Westerners seem to forget it's just one of
many AI dragons rising across China's numerous Silicon Valley equivalents.
In Hongzhou alone, a mega city with a population of
twelve point five million, Deep Seek belongs to an elite
group of tech startups known as the Six Little Dragons.
(20:53):
In the scenic Westlake district, there's Game Science, the red
hot studio behind Black Myth Wukong, a best selling action
game heralded for using machine learning techniques to make its
computer characters more lifelike. Not far away are two robotics
powerhouses and a unicorn focused on three D spatial software.
Also nearby is Jijiang chung Nao Technology, which is known
(21:16):
as Brainco and best understood as a China backed version
of Neuralink. It can be traced back to a startup
incubated at Harvard University by a Chinese born PhD student,
Bi Chung Han, and is now developing bionic limbs and
technologies for brain activity to control computers at its affiliate
lab in Hangzhou. One of Brainco's AI powered prosthetic hands
(21:39):
is currently on display at an exhibition center in China
Artificial Intelligence Town, another emerging tech hub in Hongzhou. In
recent weeks, brain Co leaders have given tours at the exhibit.
According to someone who attended a session, the attendees often
want to invest, but apparently these brainiacts haven't sounded too
desperate for outside capital. Basically, they don't need the money,
(22:02):
says a fund manager who took the tour. With all
the hype around the six little dragons, people are throwing
money at them. Standing quietly behind all these startups is
the government of President She Generative AI robotics and other
high tech ambitions are driving a state agenda that, above
all else seeks domestic self reliance and self strengthening, as
(22:25):
she phrased it during a recent Pollit Bureau meeting. According
to China's official Shinwa news agency, we must recognize the
gaps and redouble our efforts to comprehensively advanced technological innovation,
industrial development and AI empowered applications. The dragons are listening,
and not all of them are so little. The main
(22:46):
campus of three hundred billion dollar conglomerate Ali Baba, a
sprawling property with its own lake, is in an area
of hong Zhou about forty minutes west of Westlake by car.
The company recently pledged fifty three billion dollars to constructing
more AI data centers in the next three years, and
it said its latest Quen three flagship models rival deep
(23:08):
Seek's performance and cost efficiencies. Outside China, Alibaba is usually
thought of as an e commerce business, but its faster
expanding AI and cloud unit was spun off in twenty
twenty two to a separate hub on the outskirts of Hangzhou.
Inside its conference rooms, big screens glow with an industry
Insights Flash, updated every seventy two hours, detailing the latest
(23:32):
achievements of rivals such as Deep Seek and Open Ai.
There's even a weekly updated version in the restrooms, a
reminder that AI races on even when nature calls for
human technologists. This April, Ma, the elusive Alibaba co founder
who practically disappeared during the CCP's crackdown on China's tech
(23:52):
sector almost five years ago, reappeared on the company's campus
to celebrate the fifteenth anniversary of its cloud division. In
a rare speech, Ma said he wants AI to serve humans,
not lord over them, according to several people who saw it.
Attendees who also tuned into the live stream from offices
in Hong Kong and Tokyo say they were pumped about
(24:14):
Ma's triumphant comeback. It was a reminder that tech rock
stars such as Ma are apparently back in the good
graces of the CCP and being joined by up and
comers like Lyong. Even as the shine wears off tech
leaders in the US, there's a swelling national pride in China,
which is eager to show it can overcome Western obstacles.
(24:36):
George Chun, the Hong Kong based managing director of policy
consultant Asia Group says top Chinese engineers have begun returning
home after stints in the US at Apple, Google, Microsoft
and other leading companies. While hostility from the Trump administration
is part of that, they are also being pulled by
the feeling that the real action may be shifting east.
(24:58):
Silicon Valley is no longer a t active place for
work for Chinese talent Chun, says Kaifu Lee, the founder
of another Chinese unicorn o one AI, goes a step further.
A veteran of Apple, Google, and Microsoft himself, Lee says
the next generation of talent isn't following his path through
US companies before building their own In China, these young
(25:20):
AI engineers are largely home grown. He says, Deep Sikh's success,
along with the success of other new AI startups, is
motivating more young talent to be a part of China's
AI renaissance. No tech company in China today conjures as
much pride as Deep seek. While visiting hong Zhou with
his family in April, Kirby Fung, a twenty seven year
(25:42):
old computer scientist from Canada, took his family for a
tour of Leang's alma mater, Jijiog University. Fung had done
an exchange program there and wanted to show his grandparents
and younger brother that he studied at the same place
as Liang. It's really cool to explain to my friends
back in Canada that a guy who made Deep Seek
went to my school, Fung says. Tourists and social media
(26:06):
influencers also regularly descend on Deep Seek's headquarters, based in
a four tower complex overlooking China's famous Grand Canal. The
tourists look for signs of Liang at the local shops,
including an upscale hot pot spot in the deep Seek
building where staffers sometimes eat. The hostess has to break
the news that he never stops by. People who know
(26:27):
Leang say he splits his time between hong Zhou and
deep Seek's Beijing office, on the fifth floor of a
glass tower in a local tech hub. There, twenty somethingter
coders grind at height adjustable desks, and the pantry is
stocked with energy drinks, kungshifu, instant noodles, and lattiao sticks.
There's a whiteboard where employees can scribble requests for additional food.
(26:50):
I got a bit fat after having lunch and dinner
there for months, says one recently departed researcher, Leang rarely
agrees to meetings with outsiders, times, even appearing as a
hologram projection for the few he accepts. He spurned an
invitation to this year's influential Paris AI Action Summit, an
event that attracted Open AI's Altman Alphabet and Google CEO
(27:13):
Sunder Pichai and a slew of prime ministers and presidents.
While China celebrates Deep Seek, the US treats it like
an unfamiliar organism that's mysteriously shown up in the water supply,
examining it for signs of whether it's benign or malignant.
Critics have accused Deep Sek of being controlled by the CCP,
ripping off training data from US rivals, and contributing to
(27:35):
some larger espionage campaign or psyop to undermine Silicon Valley's
AI hegemony. Deep Seek is a direct pipeline from America's
tech sector into the Chinese Communist Party's surveillance state, threatening
not only the privacy of American citizens but also our
national security, says a spokesperson for the US House committee
investigating the company. Deep Seek, however, has presented itself as
(27:59):
note different from any hot startup, the product of pure
garage energy, it said in a February post on x
After all, it operates on the same Beijing campus as Google,
not far from a Burger King. And to Tim Horton's
just because the broader AI industry didn't pay much attention
to deep Seek until now doesn't mean something shady is
(28:20):
happening behind the scenes. The AI world didn't expect Deep Seek,
says ar No Bartelemy, a partner at VC firm Alpha
Intelligence Capital, which has invested in open Ai, and since
time they should have. Bartelemy says, the real lesson to
take from deep Seek is how effectively Chinese tech companies
are turning the constraints they operate under into a strength.
(28:43):
There are plenty of smart minds in China who did
a lot of smart innovation with much lower compute requirements,
he says. Indeed, in May twenty twenty three, coincidentally, the
same month deep Seek was established in Nvidia CEO Jensen
Huang told BusinessWeek that the US over regulating China will
only incentivize it to out innovate those getting in its way,
(29:05):
describing economic influence as an effective tool of national security.
He stressed that the unintended consequences of government interventions would
be severe. To be deprived of one third of the
technology industry's market has got to be catastrophic, he said,
referring to the risks of limiting US tech exports to China.
They are going to flourish without competition. They will flourish,
(29:28):
and they will export it to Europe, to Southeast Asia.
You have to be mindful of how far you push competition,
Wang continued. All of a sudden, the response is very unpredictable.
People who have nothing to lose respond in ways that
are quite surprising. They're still controversy about one important part
of Deep seek story, how much it actually spent to
(29:49):
build its models. In a widely cited report, US research
firm semi Analysis estimated that High Flyer and deep Seek
likely had access to clusters of around fifty thous thousand
of nvidia's top of the line H series GPUs worth
one point four billion dollars, which they've mostly kept hidden
from the public. The bulk of this infrastructure, semi Analysis said,
(30:12):
included GPUs that were conceivably export compliant. The US allowed
in Nvidia to sell some chips to China, the H
twenty and H eight hundred that it modified to limit performance,
so they adhered to White House restrictions. But the consulting
firm also claimed Deep Seek had access to an additional
ten thousand of Nvidia's bleeding Edge H one hundred chips,
(30:33):
which the US government had banned for sale to China.
Three ex employees vehemently deny these claims, saying Deep Seek
had fewer than twenty thousand GPUs, consisting of older Nvidia
chips and export controlled ones. They are spreading lies. Bo Liu,
the PhD candidate, says of semi analysis. The research firm
(30:54):
says it stands by its report. What's not in question
is whether Deep Sek would welcome access to the scale
of computing power that US tech companies have. The company
seems confident it could do much more with it than
Silicon Valley. The reality is that LLLM researchers have an
enormous appetite for computational resources. If I were working with
(31:15):
tens of thousands of H series GPUs, I'd probably become
wasteful too, running many experiments that might not be strictly necessary,
says one of the former deep Seak employees, but access
to more resources is a problem that China's technologists would
be willing to deal with. I wish we Chinese companies
could have fifty thousand GPUs one day, says the departed researcher,
(31:37):
who since joined another open source AI lab in Beijing.
Want to see what we could achieve.