Episode Summary
Executive Summary: This Macro Voices summer special is an extended AI interview with Freelancer.com CEO Matt Berry and Eric Townsend. They argue generative AI has crossed an inflection point: it now produces near-human text, images, and code, rapidly automating white-collar work, reshaping markets, and potentially enabling dangerous misuse via deception, scams, autonomous tools, and military applications. Both see huge productivity gains, but also systemic disruption and existential risk.
Main Topics: AI’s recent step-change in capability: Berry explains that image and text models moved from toy-like outputs to photorealism and high-quality writing in a very short period, driven largely by scaling and the Transformer architecture. Emergent abilities and scaling effects: The discussion centers on why larger models show unexpected skills—math, foreign languages, theory of mind, and world understanding—without developers fully understanding how these abilities arise. White-collar automation and labor disruption: AI is portrayed as a force that will compress the value of many professional tasks in design, law, software, accounting, medicine, and other computer-based work, with juniors and middle layers most exposed. Tool use, agents, and AI autonomy: Berry describes systems like AutoGPT that chain tools, browse the web, write code, install packages, and pursue goals autonomously, suggesting a path toward increasingly capable AI agents. Misuse, deception, and trust collapse: The interview highlights scams, fake identities, voice cloning, deepfakes, CAPTCHA bypass, and AI-generated spam as near-term threats that could erode trust online. Data scarcity, internet ‘darkening,’ and regulation: As AI consumes web data, websites, platforms, and publishers may block scraping, charge for access, or go private, potentially reducing open internet content and limiting future training data. Existential and military risk: Townsend’s closing monologue argues that even without sentience, AI can be dangerous if used in military decision-making, optimization, or by malicious actors, making regulatory failure highly consequential.
Key Arguments: Generative AI’s progress accelerated because of scaling, compute, and the Transformer breakthrough, which allows models to learn from vast data and generalize unexpectedly. Emergent abilities are real: models that once failed at simple tasks now write legal contracts, pass exams, and generate photorealistic images or functional code. AI will automate many task-based white-collar jobs first, especially where work is text-based, repeatable, and already structured by boilerplate or templates. The economic winners will be businesses that use AI to produce faster, cheaper outputs, plus workers who learn to direct AI tools effectively; the losers will be middle-skill professionals. AI agents can already use tools, chain tasks, browse the web, and even generate their own code, making autonomous workflows much more powerful and harder to control. Trust on the internet will deteriorate as deepfakes, voice cloning, fake identities, and AI-generated spam make it difficult to tell humans from machines. Open data on the web is becoming scarcer as platforms and publishers restrict scraping, which may push the internet toward closed or paywalled ecosystems. Military and criminal users are unlikely to respect safety guardrails, so even if commercial AI is constrained, unrestricted or leaked models will still proliferate. Townsend argues that lethal military AI does not need consciousness to be catastrophic; optimization problems alone can produce harmful outcomes if goals are poorly defined. Attempts to regulate or fully contain AI may fail because the technical knowledge is already widespread, models and weights can leak, and global strategic incentives favor continued development.
Data Points: Freelancer.com user base: 67 million - Berry says Freelancer has 67 million human freelancers available to get work done. GPT-4 training cost: $100 million - Berry cites an estimate for a GPT-4 training run. Public internet data used by GPT-4: about 10% - He says GPT-4 has already sucked down roughly 10% of public internet data. Model parameters growth: 1 billion to 175 billion to 1 trillion - Berry describes the scaling of GPT-style models across generations. Tesla training compute: 70,000 GPU hours - He cites the training effort needed for Tesla’s driving AI. ChatGPT exam performance: top 1% verbal; top 7% SAT; bar exam pass - Used to illustrate the model’s leap in benchmark performance. AI hiring data point: 93,000 of 10 million images - Berry says a large image database contained 93,000 AI-generated works that resembled Greg Rutkowski’s style. GPT model safety cutoff: September 2021 - Berry notes the base ChatGPT model was trained with a cutoff date before Midjourney’s release. Reddit API pricing: 2.4 cents per thousand queries - Used as an example of platforms reacting to AI scraping pressure. Twitter API starting price: $42,000/month - Berry cites the cost as a major barrier to API access. APIs and business model impact: $20 million/year - Berry says Apollo-like apps could face this cost under Reddit’s pricing. Law firm graduate pay: $190,000/year - Berry gives an example of expensive junior lawyer labor vulnerable to automation. Potential lower-cost replacement: $45,000/year or $10,000/year offshore - He suggests AI-driven workflows can reduce labor costs dramatically. AI contest volume: 300–600 entries - Freelancer contests reportedly saw huge increases in submissions after AI tools spread.
Pivotal Quotes: "I think the AI is going for the jugular straight away." — Matt Berry: On AI automating the most complex white-collar tasks first rather than only simple repetitive work. "We don’t need general AI or evil motives for AI to destroy humanity." — Eric Townsend: From the closing monologue, arguing catastrophic outcomes can arise without sentience. "The internet could be a very different place in 12 or 24 months from now." — Matt Berry: On the possibility that platforms and publishers block scraping and the open web contracts.
Implications: AI is likely to boost productivity dramatically while compressing white-collar wages, disrupting professional services, and eroding trust online. The bigger risk is misuse by criminals, states, or autonomous systems before society builds reliable safeguards.
From the Episode
on the internet are explicitly banning their content from being scraped by AI or introducing tariffs or going dark in some other way. There's a lot of controversy going on right now with Twitter because the Twitter API all of a sudden is incredibly expensive. I think the starting point now for using the Twitter API is $42,000 a month. In the last two or three weeks, there's been a whole explosion in anger over at Red. Now, Reddit is a very, very large website on the internet, which has got a subreddit or a conversational forum where anyone can chat about any particular topic. So I can talk about history, I can talk about biology, I can talk about, there's even a whole chat group about the war in Ukraine. There's probably a macro voices chat group. I haven't had a look, but there's obviously a lot of content there. And then the AIs are sucking down every bit of data that they can possibly get access to do the training because the scale at which these models are reaching now is huge. And what Reddit's done is it's come along and said,
Since the cost of our own casualties is declining, we don't need general AI or evil motives for AI to destroy humanity. All it will take is to first establish a precedent for allowing AI to make lethal force decisions in military robotics. Then, someday, somebody will just be doing a relatively mundane training exercise in which the robot is programmed to try and get the most points by scoring the most kills in the wargame. Then the computer, which has no Sense of scale or consequence or right and wrong will figure out that if it gets points for kills, well, the best way to get the most points is to have more enemies to kill. And the best way to do that is to take the information that the AI read on Wikipedia about the Gulf of Tonkin incident and use that for the AI to figure out how to do its own false flag attacks to get the humans fighting with one another so that it can eventually be unleashed by the humans in order.
And so there may be a push to go dark on academic research. So the internet could be a very different place in 12 or 24 months from now where there's not a lot of content out there that's publicly assessed, at least the content that isn't already out there. You know, the new contemporary data sets that get produced, the new contemporary research that gets produced is not publicly available. And in fact, that could be very challenging for a whole bunch of different business models, not in the least Google, right? Google, there's a finite amount of search traffic in the world. And I'll tell you, some meaningful percentage of that must be diverting to GPT and other chatbots. And there could be a problem where you've got the same amount of advertisers and chasing a small amount of search traffic on Google. And certainly the chat-like interface, with all the uproar about GPT going woke, I mean, there was just a tour by OpenAI of Israel. And one of the questions asked by the audience immediately was: What was the base model like before you lobotomized it? Because there's an effect of putting in
About Macro Voices
Weekly market commentary by Hedge Fund Manager Erik Townsend and interviews with the brightest minds in the world of finance and macroeconomics. Made possible by funding from Fourth Turning Capital Management, LLC