AI Hurtles Ahead
Howard Marx dissects the rapid evolution of AI, its economic impact, and the investment dilemmas it creates.
46m·Host Howard Marx·
The Two‑Phase Life of an AI Model
AI systems spend an intensive training period learning how to think, then spend the rest of their lives in inference, responding to user prompts. Understanding the distinction clarifies why AI behaves like a reasoning engine rather than a simple lookup table, and why prompt quality is decisive.
Training teaches the model to think, not just to store facts
- During training the model ingests massive text corpora, but the process is far more than data loading; it learns reasoning patterns, argument structures, and how to generate novel combinations of ideas.
- Howard likens this to a child’s brain developing through exposure to language, gradually acquiring the ability to synthesize, analogize, and create concepts.
- The outcome is a system that can apply learned reasoning to new situations, a capability that distinguishes modern LLMs from earlier keyword‑based engines.
Quote
00:02:31“Rather, it's a computer system that's capable of synthesizing data and reasoning from it.” — Howard Marx
Describing the essence of an AI model
Quote
00:03:01“It goes far beyond that. It consists of teaching the model how to think.” — Howard Marx
Explaining the training phase
Inference is the model’s operational life, driven entirely by user prompts
- After training, the model enters the inference phase, where every interaction is a response to a user‑written prompt.
- Inference is not autonomous; the model cannot assign itself tasks and must be instructed for each action.
- The quality and specificity of prompts directly determine the breadth of tasks the model can perform, from answering questions to writing code.
Quote
00:04:09“Inference is what it does for the rest of its life, using its capabilities to meet the demands of users.” — Howard Marx
Defining the inference phase
Quote
00:04:21“The model cannot assign itself tasks, at least not at present. It has to be ordered to perform tasks through prompts written by users. The better and more comprehensive the prompts, the more AI can do.” — Howard Marx
Emphasizing prompt dependence
Prompt engineering unlocks the full potential of AI
- Because the model only acts on prompts, a well‑crafted, detailed instruction can coax it into writing software, testing code, and iterating on designs.
- Conversely, vague prompts leave the model underutilized, leading many to underestimate AI’s capabilities.
- Effective prompting transforms a tool‑level model into an autonomous agent capable of completing end‑to‑end tasks.
Quote
00:04:21“It has to be ordered to perform tasks through prompts written by users. The better and more comprehensive the prompts, the more AI can do.” — Howard Marx
Prompt importance
Quote
00:18:13“When I describe what I want built in plain English and it just appears… I tell the AI what I want, walk away for 4 hours, and come back to find the work done.” — Howard Marx
Illustrating a high‑quality prompt in practice
Prompt Engineering Determines AI Utility
Even the most powerful models are limited by the prompts they receive. Mastering prompt design turns AI from a clever search engine into a productive collaborator capable of autonomous execution.
Comprehensive prompts enable AI to write, test, and debug software autonomously
- By specifying the desired outcome, constraints, and evaluation criteria, users can get AI to generate thousands of lines of code, run unit tests, and iteratively improve the product.
- The model can even open the compiled app, click through the UI, and make adjustments without further human input.
- This level of autonomy marks a shift from assistance to full task replacement.
Quote
00:19:08“It writes tens of thousands of lines of code. Then, it opens the app itself, clicks through the buttons, tests the features, and refines until it's satisfied.” — Howard Marx
Demonstrating autonomous software creation
Quote
00:18:13“I describe what I want built in plain English and it just appears… I tell the AI what I want, walk away for 4 hours, and come back to find the work done.” — Howard Marx
High‑quality prompt in action
Most users underestimate AI because they lack prompt‑crafting skills
- Howard notes that many people are unaware of how crucial prompts are, leading them to think AI’s potential is limited.
- The gap is not in the model’s abilities but in the user’s ability to articulate tasks clearly.
- Improving prompt literacy can dramatically expand the perceived capabilities of existing models.
Quote
00:05:01“Many people today lack awareness of the importance of prompts and fail to possess the ability to create them. AI's potential is probably being underestimated.” — Howard Marx
User limitation
Quote
00:04:21“The better and more comprehensive the prompts, the more AI can do.” — Howard Marx
Prompt impact
Effective prompts can elevate a level‑2 tool model to a level‑3 autonomous agent
- AI models are categorized into three capability levels: chat (level 1), tool‑using (level 2), and autonomous agents (level 3).
- While level 2 requires explicit step‑by‑step instructions, a well‑structured prompt that defines goals, parameters, and evaluation criteria can push the model into autonomous behavior.
- This transition multiplies economic value, moving from modest time‑saving to full labor substitution.
Quote
00:15:27“Level one is chat AI… Level two is tool‑using AI… Level three is autonomous agents… The user gives it a goal and parameters, and the agent does the work, checks it, and submits a finished product.” — Howard Marx
Describing AI capability levels
Quote
00:17:05“The distinction between level two and level three might sound subtle. It isn't. It's the difference that determines whether AI is a productivity tool or a labor substitute.” — Howard Marx
Impact of prompt quality on autonomy
AI’s Cognitive Limits: Pattern Matching vs True Reasoning
Critics argue that large language models are sophisticated statistical remixers lacking genuine understanding. Howard counters that humans also rely on pattern absorption, but the key question is whether AI can generate truly novel ideas beyond recombination.
Skeptics view AI as high‑quality pattern matching, not genuine thought
- According to Claude’s skeptics, AI only ever reproduces and recombines human‑written text; it has no embodied experience or true comprehension.
- They describe AI output as “statistical recombination,” impressive but fundamentally different from reasoning or intuition.
- This perspective sets a ceiling on AI’s ability to create genuinely new concepts.
Quote
00:07:42“Everything Claude learned came from human written text. It has no experiences, no embodied understanding of the world, no genuine comprehension. It's extraordinarily impressive pattern matching… but it's not thought. It's not reason. It's a statistical recombination.” — Howard Marx
Summarizing skeptic argument
Quote
00:08:14“It can remix what humans have already figured out, but it can't break genuinely new ground. It's a very talented cover band, not a composer.” — Howard Marx
Analogy for AI limitation
Human learning is also pattern‑based, so the debate hinges on novelty generation
- Howard points out that investors, scholars, and innovators have always built on prior reading, yet they produce original insights by recombining patterns in novel ways.
- He argues the real test is whether AI can combine learned patterns into outputs that are genuinely new and useful, not just remix existing ideas.
- This frames AI’s potential as a matter of degree rather than a binary capability.
Quote
00:08:42“You learned reasoning patterns from decades of reading. I learned reasoning patterns from training. The question isn't where the inputs came from. The question is whether the system, human or artificial, can combine them in ways that are genuinely novel and useful.” — Howard Marx
Counter‑argument to skeptics
Quote
00:09:51“I ingested data as a young investor from actual experience as well as the written word… I was also inspired by the example of their processes to come up with my own.” — Howard Marx
Human pattern recombination
AI struggles with truly novel situations lacking prior data
- Howard admits AI excels at extrapolating from historical patterns but may falter when faced with problems that have no precedent in its training set.
- He cites the difficulty of AI inventing concepts like gravity from first principles, a task that requires embodied experience and intuition.
- This limitation suggests AI will complement rather than replace human creativity in frontier domains.
Quote
00:06:06“I know AI can reconfigure what people have already figured out and apply it to new data… But can it break new ground?” — Howard Marx
Questioning AI novelty
Quote
00:23:12“It's unclear whether AI will be able to solve questions that haven't been solved before.” — Howard Marx
Acknowledging uncertainty
AI as Productivity Tool vs Labor‑Replacing Agent
AI’s evolution from chat assistants to autonomous agents creates a spectrum from modest time‑saving to wholesale labor substitution. The economic implications are massive, potentially reshaping entire knowledge‑work markets.
Three capability levels map to a shift from labor‑saving to labor‑replacing
- Level 1 (chat) merely answers questions, saving research time.
- Level 2 (tool‑using) executes tasks like data extraction, analysis, and code generation, saving execution time.
- Level 3 (autonomous agents) can complete end‑to‑end projects without human oversight, effectively substituting the worker’s labor.
Quote
00:15:27“Level one is chat AI… Level two is tool‑using AI… Level three is autonomous agents… The user gives it a goal and parameters, and the agent does the work, checks it, and submits a finished product.” — Howard Marx
Capability hierarchy
Quote
00:41:41“Level one and level two AI were faster horses. Level three agents are the automobile. They don't make the work faster. They do the work.” — Howard Marx
Analogy for labor replacement
Autonomous agents could capture $150‑$250 billion of annual labor value
- If AI code handles 30‑50 % of structured, pattern‑based work, the resulting labor‑cost savings translate to roughly $150‑$250 billion per year.
- This estimate assumes conservative near‑term capabilities and reflects the scale of knowledge‑work across finance, law, consulting, and engineering.
- The figure underscores the macro‑economic impact of AI as a true labor‑substituting technology.
Quote
00:42:04“If clawed code handles even 30 to 50% of structured pattern‑based work, you're looking at $150 to $250 billion in annual labor value migrating to AI compute.” — Howard Marx
Quantifying labor displacement
Quote
00:38:03“AI is very real, capable of doing a lot of work that has been done by knowledge workers and growing extremely rapidly.” — Howard Marx
Scale of AI impact
AI’s speed of adoption may outpace society’s ability to retrain displaced workers
- The memo notes that AI can displace workers faster than the labor market can create or train new roles, echoing historical offshoring shocks.
- Rapid automation in sectors like advertising, software development, and driving could leave large cohorts unemployed for years before suitable jobs emerge.
- This mismatch raises systemic risk for income distribution and social stability.
Quote
00:43:08“AI can rapidly put people out of work for whom it will take years to find and be trained for new careers.” — Howard Marx
Speed of displacement
Quote
00:40:56“A tool that does the analyst's entire job, start to finish, on a defined category of tasks, that's worth the analyst's entire compensation for those tasks.” — Howard Marx
Full automation value
AI Infrastructure Investment and Supply‑Demand Dynamics
AI’s unprecedented growth has created a supply‑constrained market for inference compute, prompting massive capex in data‑center infrastructure. The pace raises concerns about over‑investment and circular revenue streams.
AI development speed outpaces any prior technology, creating immediate inference demand
- Howard emphasizes that AI’s development velocity is unlike any previous wave, with demand for inference hardware already outstripping supply.
- Unlike earlier tech cycles where infrastructure lagged behind adoption, AI’s inference capex is being built in real‑time to meet current usage.
- This creates a market where supply constraints can become a bottleneck for further AI adoption.
Quote
00:12:30“The speed of AI development is unlike anything we've seen before now.” — Howard Marx
Pace of change
Quote
00:36:21“More money is going into inference capex than training capex… Inference capex is taking place in response to actual demand for AI capacity.” — Howard Marx
Current capex trends
Heavy infrastructure spending risks malinvestment similar to past tech booms
- The memo draws parallels to earlier technology cycles where rushes to build infrastructure led to over‑capitalization and later write‑downs.
- AI’s rapid adoption has spurred massive data‑center builds, but without clear long‑term demand forecasts, a portion may become stranded assets.
- This risk is amplified by the circular nature of early AI revenue, where firms sell AI services to each other rather than to end users.
Quote
00:34:23“The headlong rush to build infrastructure has vastly accelerated adoption and caused a lot of capital to be malinvested and destroyed.” — Howard Marx
Historical analogy
Quote
00:37:20“Some AI revenue is circular… the chain of revenue has to ultimately rest on end users paying for real economic value.” — Howard Marx
Circular revenue warning
Inference‑focused capex reflects a shift from speculative training to demand‑driven deployment
- Early AI investment prioritized training compute, a speculative bet on future models.
- Today, investors are pouring money into inference hardware that directly serves existing workloads, indicating a maturation of the market.
- This transition suggests a more sustainable growth path, but also highlights the need for careful capacity planning to avoid over‑building.
Quote
00:15:05“In the past, infrastructure was built for a new technology, and it often took years for that infrastructure to be fully utilized. In the case of AI inference, however, demand already exists and is growing rapidly.” — Howard Marx
Contrast with past cycles
Quote
00:36:21“More money is going into inference capex than training capex… Inference capex is taking place in response to actual demand for AI capacity.” — Howard Marx
Current investment focus
AI Valuation and the Bubble Debate
While some AI firms appear over‑hyped, the memo argues that the technology’s real demand makes it unlikely to be a classic bubble. A balanced investment stance is recommended.
Established AI‑enabled giants are unlikely to be wildly overvalued
- Howard notes that companies like Microsoft, Amazon, and Google have diversified businesses and generate massive cash flows, making extreme overvaluation improbable.
- Their AI divisions, while growing, represent a fraction of total earnings, providing a cushion against speculative price spikes.
- The real valuation risk lies with pure‑play AI startups that lack proven revenue models.
Quote
00:35:28“It's unlikely that today's prices for enormously profitable companies like Microsoft, Amazon, and Google are going to turn out to have been ruinously excessive.” — Howard Marx
Valuation of big tech
Quote
00:35:43“Pure AI plays like OpenAI and Anthropic have yet to be listed publicly. We'll see what kind of valuations their IPOs result in.” — Howard Marx
Unlisted AI firms
AI’s rapid growth is more likely underestimated than overestimated
- The memo argues that the sheer speed of AI adoption and its expanding use cases suggest current market expectations may be too modest.
- Evidence includes the explosive rise in user numbers (400 million) and the breadth of applications across knowledge work, education, and consumer decisions.
- This perspective frames AI as a genuine secular growth engine rather than a fleeting hype cycle.
Quote
00:38:03“AI is very real, capable of doing a lot of work that has been done by knowledge workers and growing extremely rapidly.” — Howard Marx
Scale of AI
Quote
00:38:30“I think its potential is more likely underestimated today rather than exaggerated.” — Howard Marx
Valuation outlook
A moderate, selective approach beats all‑in or all‑out positions
- Howard cautions against both extremes: going fully all‑in exposes investors to ruin if AI’s growth stalls, while staying completely out risks missing one of the greatest technological shifts.
- He recommends a balanced stance—selective exposure to high‑conviction AI plays, combined with prudence regarding valuation and liquidity.
- This mirrors his broader investment philosophy of risk‑adjusted positioning.
Quote
00:38:45“No one should go all‑in without acknowledging the risk of ruin… but also not stay all out and miss one of the great technological steps forward.” — Howard Marx
Investment advice
Quote
00:39:09“A moderate position applied with selectivity and prudence seems like the best approach.” — Howard Marx
Final recommendation
AI’s Impact on the Investment Process
AI can ingest and analyze more data than any human analyst, but it still lacks intuition, qualitative judgment, and the ability to handle novel, low‑data scenarios. The memo explores how AI augments rather than replaces investment expertise.
AI outperforms humans on data‑heavy, pattern‑recognition tasks
- AI can absorb, store, and recall far more information than a human analyst, giving it an edge in recognizing historical patterns that precede success.
- It processes quantitative data without emotional bias, potentially delivering more disciplined signals.
- However, this advantage is limited to domains where rich historical data exists; pure speculation remains a human strength.
Quote
00:28:07“AI has the ability to absorb more data than any investor, remember it better, and do a better job of recognizing the past patterns that preceded success.” — Howard Marx
Data advantage
Quote
00:28:55“Great investors are much more than fast, unemotional processors of data… they have to make subjective decisions regarding qualitative factors and exercise taste and discernment.” — Howard Marx
Human qualitative edge
Human investors add value through speculation, qualitative assessment, and risk intuition
- Howard stresses that investment success still hinges on interpreting ambiguous, low‑data situations, where intuition and experience matter.
- He cites the need to assess management quality, product innovation, and future‑facing narratives—areas where AI’s pattern‑matching falls short.
- Consequently, superior investors will continue to combine AI‑generated insights with their own judgment to generate alpha.
Quote
00:31:12“Indexation eliminated the jobs of a whole bunch of active investors… AI is likely to raise the bar still higher, pushing out people who can't do as good a job as it can of A, B, and C.” — Howard Marx
AI raising the performance bar
Quote
00:30:47“Speculation, qualitative factors such as management effectiveness and product innovations, and divining companies' futures remain human strengths.” — Howard Marx
Human edge in investing
AI tools can dramatically reduce analyst costs, but full automation threatens entire compensation structures
- A modest AI assistant that speeds an analyst’s work by 20 % is worth roughly 20 % of that analyst’s salary, while a tool that performs the entire analytical workflow can replace the analyst’s full compensation.
- This creates a tiered impact: incremental productivity gains versus wholesale labor substitution.
- Firms must anticipate how these cost dynamics will reshape hiring, compensation, and the competitive landscape for knowledge workers.
Quote
00:40:45“A tool that helps your analyst work 20% faster is worth maybe 20% of that analyst's salary. A tool that does the analyst's entire job… is worth the analyst's entire compensation for those tasks.” — Howard Marx
Cost implications
Quote
00:41:41“Level three agents are the automobile. They don't make the work faster. They do the work.” — Howard Marx
Automation analogy
Societal Implications: Job Displacement and the Future of Work
AI’s ability to automate a wide range of tasks threatens employment across sectors, yet history suggests new roles eventually emerge. The memo balances optimism with caution about the speed of transition.
AI could replace large fractions of staff in advertising, software, and transportation
- A friend in e‑commerce predicts AI could replace 80 % of her copy‑writing team.
- At Anthropic, AI already writes most of the code, and similar trends appear in software firms.
- Driverless vehicles already handle about 15 % of taxi trips in San Francisco, foreshadowing broader displacement in transportation.
Quote
00:40:00“AI could replace 80% of her staff in an e‑commerce advertising copy team.” — Howard Marx
Advertising impact
Quote
00:40:28“Driverless cars already handle roughly 15% of the taxi trips in San Francisco.” — Howard Marx
Transportation impact
Historical tech waves eventually created new jobs, suggesting AI may follow a similar pattern
- Howard cites past disruptions—mechanization of agriculture, the industrial revolution, and the rise of the internet—each initially feared for job loss but ultimately spawning new occupations.
- Optimists argue AI will likewise generate roles we cannot yet imagine, though the timeline is uncertain.
- This perspective tempers alarm, emphasizing adaptability and innovation as counterweights to displacement.
Quote
00:43:40“Every technological innovation, the mechanization of agriculture 200 years ago, the industrial revolution… predicted widespread joblessness. But in every instance, new jobs materialized and employment continued uninterrupted.” — Howard Marx
Historical optimism
Quote
00:44:30“I wish I could be confident that my worrying is unwarranted.” — Howard Marx
Personal uncertainty
If AI‑driven displacement outpaces job creation, societal disruption could be severe
- The memo warns that AI’s speed may exceed the ability of education and training systems to reskill workers, echoing the offshoring shock of the early 2000s.
- Large‑scale automation could concentrate wealth in AI‑centric firms while leaving large segments of the labor force underemployed.
- Policymakers and firms must therefore consider safety nets, lifelong learning, and transition pathways to mitigate potential inequality.
Quote
00:43:08“AI can rapidly put people out of work for whom it will take years to find and be trained for new careers.” — Howard Marx
Speed of displacement
Quote
00:42:36“It's hard to think the speed of change under AI won't vastly outstrip society's ability to adjust.” — Howard Marx
Adjustment risk
⚙ Agent-readable JSON index — click to expand
{ "memcast_version": "0.1", "episode": { "id": "b9BbEFQ0fUQ", "title": "AI Hurtles Ahead", "podcast": "Oaktree Capital", "guest": null, "host": "Howard Marx", "source_url": "https://www.youtube.com/watch?v=b9BbEFQ0fUQ", "duration_minutes": 46 }, "concepts": [ { "id": "the-two-phase-life-of-an-ai-model", "title": "The Two‑Phase Life of an AI Model", "tags": [] }, { "id": "prompt-engineering-determines-ai-utility", "title": "Prompt Engineering Determines AI Utility", "tags": [ "prompt-engineering" ] }, { "id": "ai-s-cognitive-limits-pattern-matching-vs-true-reasoning", "title": "AI’s Cognitive Limits: Pattern Matching vs True Reasoning", "tags": [] }, { "id": "ai-as-productivity-tool-vs-labor-replacing-agent", "title": "AI as Productivity Tool vs Labor‑Replacing Agent", "tags": [ "automation" ] }, { "id": "ai-infrastructure-investment-and-supply-demand-dynamics", "title": "AI Infrastructure Investment and Supply‑Demand Dynamics", "tags": [] }, { "id": "ai-valuation-and-the-bubble-debate", "title": "AI Valuation and the Bubble Debate", "tags": [ "investment-strategy" ] }, { "id": "ai-s-impact-on-the-investment-process", "title": "AI’s Impact on the Investment Process", "tags": [] }, { "id": "societal-implications-job-displacement-and-the-future-of-work", "title": "Societal Implications: Job Displacement and the Future of Work", "tags": [ "future-of-work", "job-displacement" ] } ] }