{"id":98193,"date":"2026-07-07T20:50:25","date_gmt":"2026-07-07T20:50:25","guid":{"rendered":"https:\/\/www.europesays.com\/ai\/98193\/"},"modified":"2026-07-07T20:50:25","modified_gmt":"2026-07-07T20:50:25","slug":"why-large-language-models-are-an-economic-dead-end","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ai\/98193\/","title":{"rendered":"Why large language models are an economic dead end"},"content":{"rendered":"

Comment: AI critiques and sceptics like myself are consistently chided for not seeing the revolutionary potential of large language models. Rutger Bregman is a prominent Dutch historian and champion of the \u201cabundance\u201d agenda who claims that left-wing \u2018AI denialism\u2019 is a danger on par with climate denialism, or worse. It\u2019s an odd argument, the climate denialists have effectively won and the hydro-carbon intensive tech sector is playing a big part of their victory lap.<\/p>\n

The substance of Bregman\u2019s argument is that left-wing AI abstainers will ensure AI is controlled by the oligarchs and authoritarians. It is not clear to me why championing the technology being used to monopolise the sum total of human intellectual work for venture capitalists is the answer. In fact the high priest of abundance, Ezra Klein, was recently named<\/a> as a member Peter Thiel\u2019s elite secret society Dialog.<\/p>\n

Being a left-wing critic of AI is a bit of bizarro world experience. I find myself constantly insisting upon capitalist arguments about poor unit economics<\/a>, the lack of return on investment, revenue and use cases that fall far short of revolutionary claims.<\/p>\n

Doomerism and abundance<\/p>\n

To avoid stubborn questions about the actual utility and economics of LLMs, proponents such as Bregman appeal to post-scarcity abundance fantasies and teen sci-fi scenarios of apocalyptic risk. They arm themselves with theories of exponential growth and lines going up from the stock market to Moore\u2019s Law or the Kardashev scale. I am partial to the theory of the exponential baby<\/a>.<\/p>\n

The field of AI risk and alignment has been driven by philosophers theorising super-intelligence as much as engineers thinking about applied tasks. Prominent among them is Nick Bostrom, the philosopher and cryonics enthusiast whose super-intelligence thesis informs theories of Artificial General Intelligence (AGI). Bostrom is very important to the proliferation of \u201cdoomer\u201d narratives that evade political economy. Simultaneously fantasist venture capitalists<\/a> are simply declaring<\/a> that AGI is here. For them abundance is real, ie a spectacular accumulation of political and economic power that allows them to seemingly dictate the future.<\/p>\n

\"\"<\/p>\n

From Nick Bostrom\u2019s Super Intelligence (2014) Oxford University Press<\/p>\n

The key finding that excites Bregman is the METR (Model Evaluation and Threat Research) time horizon graph. He thinks it\u2019s an \u201ceven scarier\u201d graph than Al Gore\u2019s cherry picker climate chart.<\/p>\n

\"\"<\/p>\n

\"\"<\/p>\n

Want to stay across stories like this? Start your day with Newsroom’s top stories and analysis in your inbox.<\/p>\n

\"\"<\/p>\n

All the lines go up!<\/p>\n

METR is a non-profit staffed by researchers and insiders from the AI industry who assess performance benchmarks. As a non-specialist there is plenty for me to glean. My favourite METR study<\/a> is that software developers believe they have become superusers through AI but in fact have become 19 percent slower.<\/p>\n

METR\u2019s Time Horizon graph has been described by as \u201cthe most misunderstood graph in AI\u201d<\/a> and a \u201cload-bearing institution\u201d for tech-stock valuations. Much of the excitement around the METR Time Horizon measure comes from its inclusion in AI 2027<\/a> , a popular piece of AI doomerism. The graph measures the ability of different models to complete increasingly difficult software tasks measured by the time it would take a software expert to complete. METR to their credit are circumspect about whether this is the right metric for model evaluation let alone super intelligence.<\/p>\n

In June, Anthropic\u2019s Mythos made headlines with a time horizon of 16 hours<\/a> compared with GPT-2\u2019s nine seconds in 2019. Is runaway intelligence here? Does this finally change everything?<\/p>\n

Put me down as \u201cNo\u201d. This headline number represents a 50 percent task completion rate. AI adoption is currently held back by massive costs and reliability. 50 percent task completion is not a revolution in software engineering. With \u201cmessy\u201d tasks that reflect human complexity, success rates come down to about 30 percent. Gary Marcus regards Mythos\u2019 benchmark as an impressive feat in the old field of \u2018Symbolic AI\u2019 but not an advance towards AGI. He views LLMs as a dead end.<\/p>\n

Software and code need to respond to human systems as they exist. Even if \u00a0we got to a high task automation rate there is still the massive problem of security. The code AI makes is highly insecure and you\u2019ve potentially taken experts out of the process of understanding where the risks are.\u00a0Businesses and government would be reckless to switch their software and workflows with these systems.<\/p>\n

The multi-trillion dollar industry<\/p>\n

Among Bregman\u2019s collection of lines are projections of data centre build-outs and nominal increases in AI revenue. To return to his concerns about oligarchy and democratic deficits these are great indices of over-accumulation and financial engineering that threaten to be the breaking point for the coming financial crisis.<\/p>\n

\"\"<\/p>\n

\"\"<\/p>\n

Data Centre Projection and Anthropic\u2019s \u201cARR\u201d<\/p>\n

There are headline numbers offered by the likes of McKinsey of $7 trillion<\/a> in pledged builds by 2030. Everyone from Nvidia, Meta, Oracle, and SpaceX are putting these pledged build-outs through highly leveraged special purpose vehicle shell corps via a highly volatile network of circular deals.<\/p>\n

One example is Oracle, which has taken on $340 billion in debt to build \u201cStargate\u201d for OpenAI. OpenAI needs to provide $60b<\/a> a year to Oracle but the reality is it lost $38b last year with just $13b in revenue. AI differs greatly from previous software business models as there is no network effect, the CapEx costs are high and ongoing, and at present they lose more money on their paying customers than free-riders.<\/p>\n

But what of Anthropic\u2019s rising fortunes?<\/p>\n

The AI industry does some creative accounting called \u201cAnnual Realised Revenue\u201d which basically means take your best month or quarter and average it out over a year. As the\u00a0Wall Street Journa<\/a>l reported Anthropic\u2019s numbers: \u201cIt is unclear what accounting methods Anthropic has used to book revenue and costs, as the company isn\u2019t yet required to follow the financial-reporting requirements of a public company.\u201d<\/p>\n

What\u00a0Ed Zitron has found\u00a0<\/a>is that this can all be explained by horse-trading between Anthropic and SpaceX. Anthropic\u2019s has received two months of heavily discounted compute from SpaceX which Anthropic used to leak cooked numbers to the press. SpaceX was then able to use Anthropic\u2019s pledge of $45b of future spending as proof of revenue for its recent IPO.<\/p>\n

More recently when Anthropic and OpenAI moved to token-based billing for AI, which starts to reflect real costs, everybody blew through their AI budgets. Uber burnt through its yearly allocation in just four months. Comedically the chief technology officer spent\u00a0US$1200 in tokens\u00a0for a two-hour session meant to impress other C-suite executives.\u00a0<\/p>\n

Taking AI seriously<\/p>\n

So how should the left take gen-AI and LLMs seriously? This is a Silicon Valley power grab, to \u201ceat the world\u201d as US venture capitalist Marc Andreessen would put it. The AI industry has produced confidence machines that pander to the fantasies of bosses, middle managers and corporate consultants. In Aotearoa New Zealand AI is allowing the Government to do DOGE-style cuts and wage war on \u2018woke\u2019 bureaucrats. The assumption being that Silicon Valley helps them fill in the blanks when it\u2019s time to do real work, just don\u2019t ask how much tokens cost. \u00a0<\/p>\n

These are indeed dangerous times but in the way that is grounded in material realities not teen sci-fi.<\/p>\n","protected":false},"excerpt":{"rendered":"Comment: AI critiques and sceptics like myself are consistently chided for not seeing the revolutionary potential of large…\n","protected":false},"author":2,"featured_media":98194,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[6744,3013,309,13786,66,40657,1651,51218,134,51217],"class_list":["post-98193","post","type-post","status-publish","format-standard","has-post-thumbnail","category-agi","tag-agi","tag-artificial-general-intelligence","tag-business","tag-comment","tag-news","tag-newsroom-pro","tag-opinion","tag-otago-university","tag-technology","tag-university-of-otago"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/98193","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=98193"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/98193\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/98194"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=98193"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=98193"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=98193"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}