[HN Gopher] Developing AI models or giant GPU clusters? Uncle Sa...
       ___________________________________________________________________
        
       Developing AI models or giant GPU clusters? Uncle Sam would like a
       word
        
       Author : rntn
       Score  : 57 points
       Date   : 2023-11-05 19:45 UTC (3 hours ago)
        
 (HTM) web link (www.theregister.com)
 (TXT) w3m dump (www.theregister.com)
        
       | wslh wrote:
       | Is this some kind of new Clipper chip [1] I understand the
       | differences but I see this as a type of allegory?
       | 
       | [1] https://en.wikipedia.org/wiki/Clipper_chip
        
         | mpalmer wrote:
         | I'm not sure what connection you're seeing here. How does a set
         | of reporting requirements resemble the attempt to encourage
         | telecom businesses to add a backdoor in communications
         | infrastructure?
        
           | wslh wrote:
           | I mentioned explicitly the word allegory.
           | 
           | First, the term backdoor applies loosely to the Clipper chip
           | since it would have been public that this chip could be used
           | for accessing private information. I think lock is a better
           | security term. Backdoors generally are secret. My memories
           | from that time are that the term backdoor was also an irony.
           | 
           | Second, the link is about informing private information
           | because there is a security concern. It is not because the
           | government want to have stats to inform the public about how
           | to run AI nodes.
        
             | mpalmer wrote:
             | No, I think backdoor is the appropriate term, not least of
             | all because the first sentence of the Wikipedia entry you
             | linked identifies the Clipper chip as such.
             | 
             | As long as we're nitpicking word choice, the word allegory
             | does not apply here. An allegory is an intentional
             | narrative device employed by an author or artist. Two
             | things you find to be similar does not constitute an
             | allegory.
             | 
             | The link you identify is pretty tenuous. With the Clipper
             | chip, the government wanted access to _live communications_
             | , not information about the infrastructure the
             | communications were passing through. That is not the case
             | with the recent EO.
        
               | wslh wrote:
               | Sorry, if I can say it differently it reminds me of the
               | Clipper chip because of the security concerns for the use
               | of technology in a specific way. AI now, crypto before.
               | Please take it as a personal connection then.
        
       | amzn12333 wrote:
       | Just do it in China or Russia then.
        
         | threeseed wrote:
         | There are export controls for the highest end GPUs.
         | 
         | And history has shown that companies don't just move to
         | China/Russia because the US market is so lucrative.
        
           | cscurmudgeon wrote:
           | Where did Apple manufacture its phones?
        
             | threeseed wrote:
             | Apple never moved to China.
             | 
             | And their phones are manufactured in China, India and
             | Vietnam.
        
           | Dalewyn wrote:
           | 1. Highest-end silicon is not strictly necessary. Most people
           | here can surely appreciate just how fast microprocessors in
           | general are today?
           | 
           | 2. The export controls have exceptions, loopholes, and people
           | with strong wills.[1][2]
           | 
           | 3. China is slowly but steadily freeing itself from import
           | dependence.[3]
           | 
           | [1]: https://tech.slashdot.org/story/23/08/21/203241/china-
           | keeps-...
           | 
           | [2]:
           | https://mobile.slashdot.org/story/23/09/09/1849221/huawei-
           | sh...
           | 
           | [3]: https://mobile.slashdot.org/comments.pl?sid=23071327&cid
           | =638...
        
           | roenxi wrote:
           | For most of history, China wasn't the wealthy superpower it
           | has been since around 2020.
           | 
           | There is an argument that China has the world's largest
           | economy and it is nestled in the middle of the region that is
           | best known for high-tech manufacturing. It is still possible
           | that they'll fumble this somehow, but the fundamentals are
           | solidly on the side of China becoming the place to do AI
           | training.
        
         | cscurmudgeon wrote:
         | That's what will happen and we will wondering how China got to
         | be number one in AI (like we do now for China and
         | manufacturing).
        
         | Tyr42 wrote:
         | China has similar rules, but tighter u less it's for national
         | security.
        
       | akira2501 wrote:
       | The order itself is ridiculous election year pandering. It
       | imagines we are suddenly in an "age of AI" and that it's clueless
       | meddling is required to establish an industry that already exists
       | and to prevent algorithmic dark patterns that are already
       | entrenched from forming.
       | 
       | It then sets every federal agency out on a quest to identify and
       | then create a plan to ameliorate all the "scary AI boogey men"
       | sci-fi fever dreams that have been associated with the deployment
       | of a technology that doesn't even actually exist yet.
       | 
       | And, of course, more H1-B visas, because.. you know.. we wouldn't
       | want to be "left behind."
       | 
       | [0]: https://www.whitehouse.gov/briefing-room/presidential-
       | action...
        
         | ketzo wrote:
         | Edit: I read H1-B as a generalization for "skilled immigration"
         | - I now see that people have issues with the H1-B program,
         | specifically.
         | 
         | The H1-B thing is kind of a non-sequitur. I think anyone
         | involved in tech (software, hardware, anything) should be
         | _wildly_ in favor of massive H1-B increases, AI-related or not.
         | 
         | The U.S. has the opportunity to cement itself as the center of
         | the world as far as technological progress from now into
         | eternity. Why wouldn't we take every brilliant immigrant we can
         | get?
         | 
         | If you're a software developer and you are afraid for your
         | job/wages, I get that intuitively; but how many times do we
         | have to learn the lesson that "more people making software
         | leads to _more_ software jobs"
        
           | jgalt212 wrote:
           | > I think anyone involved in tech (software, hardware,
           | anything) should be wildly in favor of massive H1-B
           | increases, AI-related or not.
           | 
           | Not if you like high wages, and you want them to stay that
           | way.
        
             | mschuster91 wrote:
             | There's way too much work to do in the AI space and - even
             | worldwide - not enough skilled people by far.
             | 
             | The danger, IMHO, isn't wage dilution anyway (that one can
             | be counteracted by politics) - it is that sensitive
             | knowledge will make its way back to China and other current
             | or potentially hostile nations, and it would not be the
             | first time either that this happens.
        
             | ketzo wrote:
             | My final paragraph addresses that, but to expand:
             | 
             | Devs have been worried about offshoring/immigrants
             | replacing them/lowering their wages for decades.
             | 
             | The only outcome we have _ever_ observed from having more
             | people building software is that dramatically more software
             | jobs have become possible and in-demand.
             | 
             | Also, and I understand why you might not make this
             | argument, but let's be honest: software engineering wages
             | are _extraordinarily_ high. Slowing or even reversing that
             | growth by small amounts _on average_ would still leave
             | millions of extremely well-paid jobs.
        
               | srackey wrote:
               | Except we have no control to compare to.
               | 
               | Software is eating the world after all. Seems likely that
               | demand would be high anyway. Perhaps if there were no
               | H1Bs, entry level grads would make $250k, with the
               | average senior dev making $750k+.
        
           | akira2501 wrote:
           | > The U.S. has the opportunity to cement itself as the center
           | of the world as far as technological progress from now into
           | eternity.
           | 
           | I don't accept this premise and I don't think serves as a
           | reasonable excuse for government interference in labor
           | markets. Even if you do accept this, then the solution
           | sacrifices long-term labor stability for short-term labor
           | monopolization.
           | 
           | Either way, I don't see this as a positive outcome, and I
           | regret every administrations attempt to expand the program
           | using any excuse that happens across their desks.
           | 
           | > but how many times do we have to learn the lesson that
           | "more people making software leads to more software jobs"
           | 
           | The connection between this outcome and increased H1-B visas
           | for mostly _corporate sponsors_ is sketchy, at best.
        
             | threeseed wrote:
             | a) There has been no change to H1B visa intake since 1990.
             | 
             | b) There is no evidence that H1B visas have caused labor
             | instability in the IT market.
        
             | Al-Khwarizmi wrote:
             | Why do you see giving more visas as a government
             | interference in labor markets?
             | 
             | In my view it's the opposite, a perfectly free labor market
             | would be one where anyone can apply to a job. Restricting
             | immigration by denying visas _is_ a government interference
             | in the market. So more visas means less interference.
             | 
             | (Note: I'm not claiming anything about whether it's a good
             | idea or not).
        
             | xtreme wrote:
             | The government is already interfering with the labor market
             | by denying work visas to many US-educated students,
             | limiting number of employment based green cards, handing
             | out diversity visas through random lottery, and in general
             | dictating who should or should not be eligible for work
             | permits.
        
               | arthurcolle wrote:
               | Not trying to speak for GP but I think they would
               | probably tell you that what you're talking about is
               | _also_ an interference and should go away as well :)
        
           | dukeyukey wrote:
           | H1-B increases are a band-aid at best. They are a shitty,
           | substandard substitute for an actual, proper, modern working
           | visa system.
        
           | oceanplexian wrote:
           | Oh hell no. H1B isn't there to bring in the best and
           | brightest; those workers would qualify for skilled worker
           | visas. It's there for low-skill workers, who are set up to be
           | abused by corporations under the threat of deportation, while
           | replacing American workers.
           | 
           | The US doesn't need to "cement itself" as anything. Thanks,
           | but no thanks. Ironically, your kind of attitude is the same
           | one that gave rise to the insane, populist politics of the
           | last few years and has done more harm to immigrants than any
           | other single policy.
        
             | fnordpiglet wrote:
             | There are only 30,000 skilled worker visas issued per year.
             | There are additionally 65,000 H1B visas issued, with an
             | additional 20,000 issued to folks with a masters or higher
             | from a US university.
             | 
             | H1-B is specifically for specialized occupations and
             | generally requires a minimum of a bachelor's degree and
             | specialized skills in demand. It is, in fact, called the
             | H1-B Specialty Occupations Visa.
             | 
             | I suspect you're thinking of the Diversity Visa program,
             | which offers a lottery of 55,000 visas annually to anyone
             | (except for some eye brow raising exceptions).
             | 
             | There are also other programs like migrant worker programs
             | that allow unskilled labor into the country for a limited
             | time to fill seasonal work gaps.
        
           | idrios wrote:
           | The US should be giving more freedom to existing H1-B holders
           | so they aren't indentured to the companies that employ them
           | for 10+ years, at H1-B minimum salary, while waiting for them
           | to generously sponsor their green card. Make H1-B not tied to
           | employment so companies stop abusing them.
        
         | threeseed wrote:
         | a) The election is in 2024 and this isn't a top 10 issue for
         | voters.
         | 
         | b) It is inarguable that we are in an age of AI.
         | 
         | c) There are legitimate concerns with AI that shouldn't be
         | blindly waved away as "boogey man". Especially with what we
         | have been seeing in Ukraine with their use of autonomous and
         | remotely controlled drones. Adding an AI layer into the mix
         | needs to be regulated.
         | 
         | d) Number of H1B visas are set by Congress not the President.
        
           | akira2501 wrote:
           | > The election is in 2024 and this isn't a top 10 issue for
           | voters.
           | 
           | How many days from today until that election day?
           | 
           | > It is inarguable that we are in an age of AI.
           | 
           | We have companies hawking large language models. It's
           | entirely arguable that we are in an "age of AI." You're about
           | to be in an age of "no more easy CMOS gains." The
           | intersection between these two points is going to be
           | interesting.
           | 
           | > There are legitimate concerns with AI that shouldn't be
           | blindly waved away as "boogey man"
           | 
           | This is why we have courts and a legislature with committee
           | powers. I do not believe that an eager top down federal
           | agency approach is going to solve real problems without
           | creating more encumbrances than it's worth.
           | 
           | This is also why it's pandering. Look at the list of issues
           | they bring up, those most definitely rank with voters.
           | 
           | > Number of H1B visas are set by Congress not the President.
           | 
           | Well then it's potentially even a bigger problem. They're
           | going to change prioritization for those applicants against a
           | limited pool.
        
       | downvotetruth wrote:
       | Wonder how much of a write off Nvidia will claim for Mellanox due
       | to the 100 GB/s limit.
        
       | Ukv wrote:
       | > One requires reporting of any model trained using more than
       | 10^26 integer or floating point operations total
       | 
       |  _On an unrelated note, OpenAI announces GPT5 will be trained
       | with fixed-point arithmetic._
        
       | zitterbewegung wrote:
       | Why are they using parameters or ops using trained when an
       | objective tests that exist or compare to other state of the art
       | systems which makes more sense. This executive order is a joke
       | when you could just almost train a bit less and get the same
       | output...
        
       | WalterBright wrote:
       | By hobbling American AI development, it just hands AI leadership
       | to other countries.
        
         | mdale wrote:
         | The regulations as presently defined are pretty lax IMHO; it
         | seems people jump to the all regulations are a unbearable
         | constraint perspective without some measured look at what is
         | being proposed?
         | 
         | I get the don't give an inch or they will take a mile, but also
         | some regulation structures are probably warranted by such a
         | fundamental shift on par with the advent of the Internet.
         | 
         | Back then we both got it right with DMCA in that big Internet
         | companies could thrive in the US; but also problematic
         | concentration of Monopolistic power in big Internet companies.
         | 
         | Will have to do some iteration to see what makes sense with the
         | advent with this new computing model / capabilities at this
         | scale.
        
       | arthurcolle wrote:
       | I know that its more the capabilities angle that the US
       | Government is mainly interested in here, rather than copyright
       | infringement... But just curious, could packages like google's
       | fully-homomorphic-encryption or the difficult to build Pyfhel be
       | used to mitigate government intrusion into model training /
       | datasets?
       | 
       | I know there is some work being done to 'extract' training data
       | from models (although given the compression, not entirely sure
       | you could really extract _everything_ but curious if anyone is
       | seriously working on training models on purely encrypted data
        
         | PeterisP wrote:
         | FHE does still add a huge overhead (measured in orders of
         | magnitude, not percentages), so it would not be applicable for
         | large models that are comparable with state of art models. You
         | can apply FHE to neural networks in general (e.g.
         | https://developers.googleblog.com/2023/08/expanding-our-full...
         | lists it as one of potential use cases), but the unsaid
         | assumption is that they are relatively small NNs, and that's in
         | an era where we know that simply using a larger model brings
         | meaningful improvements to the outcome, and the main limiting
         | factor for people training their own models is the cost of
         | compute.
        
       | thomastjeffery wrote:
       | Dear Uncle Sam,
       | 
       | You have been sorely misinformed about AI. Even the name itself
       | has been used to mislead you! Artificial Intelligence _does not
       | exist today_. While new systems may be  "intelligent" in their
       | designs, none of them is "an intelligence".
       | 
       | Artificial Intelligence has been the pursuit of Computer Science
       | since the earliest days of software design, back when the AI
       | department at Bell Labs developed Programming Languages. What is
       | called "AI" today is no more than the newest efforts in that
       | _pursuit_. Despite the excitement at these newer efforts, _the
       | goal that is AI_ is as mysterious as it ever was.
       | 
       | So what's new? Inference Models. These allow computers to
       | navigate _ambiguous data_. This was entirely impossible before,
       | and is only somewhat possible today. While computers do not get
       | completely stuck on ambiguity the way they used to, they are
       | still unable to _conclusively resolve_ that ambiguity. They can
       | only be trained to guess. With careful training on very large
       | datasets, some impressive results have been obtained.
       | Unfortunately, those impressive results are always closely tied
       | to embarrassing mistakes. This is a _feature_ of contemporary
       | inference models.
       | 
       | The goal in mind is to perform this process well enough to have a
       | novel and useful system. Many believe that once a model is big
       | enough and trained well enough that it will become _reliably_
       | more accurate. There is no conclusive evidence that is the case.
       | While they are often introduced as a  "limitation", behaviors
       | like "overconfidence" and "hallucinations" are _features_ of
       | these systems. In order to remove a feature from a system, a
       | _new_ system must be invented. In the mean time, let 's consider
       | what _does_ exist, rather than get lost in our own dreams.
       | 
       | So what _should_ you be worried about? Contemporary inference
       | models are powerful enough to create _convincing results_. What
       | does that mean for the people of the United States? We need to
       | recognize the reality in front of us: people can tell lies. Data
       | alone is never a reliable _source_ of information. This has
       | always been true, but the inherent difficulty in storytelling has
       | made some lies _impractical_ to tell. As technology improves, so
       | does each person 's ability to tell a story.
       | 
       | We have all watched the journalistic integrity of our world
       | suffer at the hands of Social Media, and at the failure of large
       | corporations to moderate content. At best, people have grown to
       | distrust scientific discovery and leadership; and at worst,
       | untamed hate speech has lead to genocide.
       | 
       | So what can we do about it? Readers need to be able to
       | differentiate content, not by its substance, but by its _source_.
       | The good news is that this has been a solved problem for 50
       | years. All that an author needs to attach their identity to their
       | writing, is to provide readers their unique public key and a
       | signature of their work. Unfortunately, the best tools to do
       | this, including GPG, are very technical and difficult for the
       | layman to use. It should be the priority of the United States,
       | for the sake of national security, to improve this landscape by
       | creating (or motivating the creation of) easy-to-use public-key
       | encryption software.
        
       | Spivak wrote:
       | Y'all just need to read the executive order (or at least the
       | white house fact sheet) itself it's got way more stuff than this
       | article covers and to my eyes is overwhelmingly positive. It's
       | like the white house read all of the complaints people on HN have
       | had about inappropriate uses of LLms and bundled them into one
       | huge omnibill telling every federal agency to stop people from
       | behaving recklessly with AIs that are are mirror of the average
       | Redditor.
       | 
       | Outside of reporting that you're working on a new huge model and
       | some maybe down the road future NIST guidelines they don't
       | actually restrict the production of models at all. It's all about
       | telling the humans jumping on the move fast break things train
       | that no you can't use these tools for for <obviously dystopian
       | thing> like renter screening, price fixing, and judicial
       | sentencing.
       | 
       | Good on the white house for recognizing that the harm these
       | models produce is almost entirely the humans connecting them to
       | the real world.
        
       ___________________________________________________________________
       (page generated 2023-11-05 23:01 UTC)