{"id":1643,"date":"2023-04-18T19:54:43","date_gmt":"2023-04-19T00:54:43","guid":{"rendered":"http:\/\/haleyai.com\/wordpress\/?p=1643"},"modified":"2023-04-18T20:07:44","modified_gmt":"2023-04-19T01:07:44","slug":"will-oligarchs-own-the-future-of-ai","status":"publish","type":"post","link":"http:\/\/haleyai.com\/wordpress\/2023\/04\/18\/will-oligarchs-own-the-future-of-ai\/","title":{"rendered":"Will Oligarchs Own the Future of AI?"},"content":{"rendered":"\n

Here we go again. \u00a0We are set back a bit this morning (two weeks ago now) with a recent Tech Crunch article about Anthropic<\/a>, perhaps the most inspiring company touting safe AI.\u00a0 They seek to raise billions to compete with not so Open AI.\u00a0 Before commenting on the article, how about a little context?<\/p>\n\n\n\n

Open AI, the company, was founded in 2015 precisely to, as stated on Wikipedia today<\/a>, “freely\ncollaborate” […and make] its patents and research open to the public.  Things began to change in 2019 as Open AI\ntransitioned to a for-profit corporation. \nUltimately, this year, Open AI has written that they will no longer\n\u201cshare [any] details for commercial and other reasons\u201d.<\/p>\n\n\n\n

Last year, in a\npodcast with Sam Altman and Reid Hoffman<\/a>, the CEO of Open AI suggested that\nonly a handful of companies could provide the foundational AI models on which\neveryone else will build \u201cthe middle layer\u201d. \nThis suggests that foundational AI will simply become part of \u201cBig\nTech\u201d, raising familiar questions of who\nowns the future<\/a>.<\/p>\n\n\n\n

Open AI has been stodgy with being open since it initially refused to share the GPT-2 model in 2019, eventually doing so, arguably due to pressure from the AI community, in particular.\u00a0 Open AI has not shared any subsequent model.\u00a0<\/p>\n\n\n\n

Open AI (little \u2018o\u2019; not the company) does not mean\ncommercial AI becomes impractical.  The\nintent of open AI is to keep fundamentals of nature, like math, electricity,\nand fire (including nuclear power), from becoming private property at the\nexpense of society.  Making a living\nharnessing them and applying them innovatively should remain fair game.<\/p>\n\n\n\n

Our entire tech industry is built on shared intellectual\nproperty, most notably the open-source software movement.  Without open-source, much of modern life\nwould be stuck decades in the past.  All\nthe progress in machine learning over the last few decades would have been\nimpractical without open-source operating systems and programming languages,\nsuch as Linux and Python, in particular.<\/p>\n\n\n\n

AI models are a little different.  They have two critical parts.  One is the source code that implements\nthem.  Typically, this is Python code\nwhich runs on tens to thousands of GPUs, which are massively parallel matrix\nmanipulating machines.  Essentially,\ngiven data, the algorithms written in Python adjust the matrices until the\nerror in predicting things about the training data is minimized (or nearly so).<\/p>\n\n\n\n

This second part, the resulting contents of the matrices\nafter training (a.k.a., the model weights), is where the controversy of open AI\nstarted with GPT-2.  Through GPT-3, Open\nAI was quite good about publishing details of the algorithms used in its\nmodels.  The AI community easily\nreplicated such models, with various modifications.  Fair enough, that\u2019s one degree of open AI.  Many believe it\u2019s not enough.<\/p>\n\n\n\n

Better is the general, open-source attitude among AI\nresearchers, including many with commercial affiliation, and especially the\nHugging Face community sponsored by Meta. \nBut having the source code of a model is not \u201cdemocratic\u201d enough.  Wherein democratic, means practically\navailable to anyone and everyone. \nPractically available to everyone requires both the source code of a\nmodel and the weights resulting from its training to be available.<\/p>\n\n\n\n

Just a few details on the models and their weights.  The transformer architecture has been refined\nsignificantly but remains basically unchanged over the last 5 years.  The source code for producing the state of\nthe art is widely available and gradually evolving as techniques improve.  In order to produce a state-of-the-art model,\nmassive amounts of data are needed. \nWhether training a language model from text or a multi-modal model with\ntext, images, etc., we have enough readily available data to approach the\nstate-of-the-art results democratically. \nWhere it becomes less democratic is the cost of computing the model\nweights given the training data. <\/p>\n\n\n\n

The amount of computation required to train is model is\n(naively) proportional to the number of training iterations times the size of\nthe model.  For the most part, the amount\nof computation is proportional to the number of parameters in the model.  The weights are simply the values of those\nparameters after optimizing the model by training it with the data.<\/p>\n\n\n\n

Table stakes for a good language model, which generally\nrequire over 10 billion parameters, is 1023<\/sup> floating point\noperations.  For example, Google\u2019s\nChinchilla proved a 70B parameter model superior in many regards to models many\ntimes its size.  Meta\u2019s more recent LLaMA\nbenefits from additional improvements.  A\nChinchilla-scale model was trained on 2,048 A100 GPUs on over 1 trillion tokens\nof text in 21 days.  At arms-length,\non-demand processing last year, this would cost roughly $1 million.<\/p>\n\n\n\n

Commercial Open AI would have us believe that this is just\nthe tip of an iceberg.  That $1 million\ntoday will be $1 billion tomorrow.  Open\nAI would have us believe that we can\u2019t afford to keep up as they build models\n10 to 100 times larger.  Well, the jury\nis out.  There have already been models\nwith 3 to 10 times as many parameters as GPT-3 which have fizzled quickly.  But the intent is clear.<\/p>\n\n\n\n

Unfortunately, according to the article, inspiring Anthropic now aspires to be one of the oligarchs of AI.\u00a0 The article states that Anthropic\u2019s investor pitch deck claims, \u201cWe believe that companies that train the best 2025\/26 models will be too far ahead for anyone to catch up in subsequent cycles.\u201d\u00a0 It goes on to assert that AI will automate large swaths of the economy in very few years.\u00a0 Such hyperbole may further inflame unfortunate calls for a moratorium on AI.<\/p>\n\n\n\n

We like Anthropic\u2019s approach to Constitutional AI and are big\nfans of continuous, self-supervised learning, as well as reinforcement learning\ngiven human feedback.  These aspects have\nmaterially advanced the safety and the instruction-following and conversational\nabilities of language models recently. \nBut in doing so, they require an order of magnitude less compute than\nthe brute force pre-training discussed above.<\/p>\n\n\n\n

Anthropic thinks building a model with \u201ctens of thousands\u201d of\nGPUs will produce magic.  We\u2019ll see.  There are a few stubborn facts in the\nway.  One problem is that increasing the size\nof a model and the amount of training data must be balanced<\/a>.  Loosely speaking, one cannot just double the\nnumber of parameters without doubling the amount of training data.  One problem with doubling the amount of\ntraining data is that we\nare running out of data<\/a>.  Another is\nthat we are already approaching the asymptotes that we can get from numbers of\nparameters and amounts of training data. \n<\/p>\n\n\n\n

The basics of the learning curves for models of more than a\nfew billion parameters is that the inflection point of diminishing returns is\npassed quickly, somewhere between 10 and 100 billion tokens of training\ndata.  After \u201cjust\u201d a few 100 billion\ntraining tokens, a model with 30 to 120 billion parameters begins to look\nasymptotically close to \u201cfitting\u201d the training data.  And the 30 billion parameter model fits the\ndata over 95% as well as the model 4 times its size.<\/p>\n\n\n\n

Whether or not size matters, other innovations are coming\ninto focus now that we have sufficient scale. \nHopefully, we can avoid Big Tech, including Open AI and Anthropic,\nowning our future by more openly sharing models, including their weights.  If not, we can expect the innovations and\nadvances in AI to slow as proprietary interests slow the exchange and\nexperimentation that has produced staggering advances in the last decade.  Either way, if the limits of scale alone are\nindeed near, it\u2019s not the end of the world.<\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

Here we go again. \u00a0We are set back a bit this morning (two weeks ago now) with a recent Tech Crunch article about Anthropic, perhaps the most inspiring company touting safe AI.\u00a0 They seek to raise billions to compete with not so Open AI.\u00a0 Before commenting on the article, how about a little context? Open … <\/p>\n

Continue reading “Will Oligarchs Own the Future of AI?”<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[955],"tags":[959,960,922,961,956,970,957],"class_list":["post-1643","post","type-post","status-publish","format-standard","hentry","category-artificial-intelligence","tag-antropic","tag-constitutional-ai","tag-language-model","tag-llama","tag-open-ai","tag-openai","tag-pyton"],"_links":{"self":[{"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/posts\/1643","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/comments?post=1643"}],"version-history":[{"count":5,"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/posts\/1643\/revisions"}],"predecessor-version":[{"id":1659,"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/posts\/1643\/revisions\/1659"}],"wp:attachment":[{"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/media?parent=1643"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/categories?post=1643"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/haleyai.com\/wordpress\/wp-json\/wp\/v2\/tags?post=1643"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}