Oaktree Capital Management co-founder Howard Marks has engaged in a deep dialogue with AI models, revealing the disruptive impact of AI's evolution toward "autonomous agents" on labor markets and investment landscapes. Marks provided clear investment guidance regarding current AI bubble concerns.
Three months after exploring "Is AI a Bubble?", Marks released his latest memo titled "The Rapid Advancement of AI" on February 26. In this client letter, Marks had Claude compose a 10,000-word AI tutorial, extensively quoting the AI's own statements.
Marks emphasized that no technology has matched AI's adoption speed. Approximately 400 million individual users and 75-80% of enterprises globally now utilize generative AI. Unlike previous labor-saving tools, current AI demonstrates astonishing "autonomous action capabilities."
The Emergence of Three-Level Agents: Replacement Rather Than Assistance Through Claude's perspective, Marks categorized AI capabilities into three tiers: Level 1 conversational AI, Level 2 tool-based AI, and the currently emerging Level 3 "autonomous agents."
Regarding philosophical debates about AI's thinking capacity, Marks cited Claude's economically grounded argument: "If I can perform analytical work equivalent to a $200,000-per-year research assistant, whether I'm 'truly thinking' or 'merely pattern matching' becomes irrelevant to payers. What matters is whether my output proves sufficiently reliable and valuable - and this reliability continues improving."
Marks stressed that the distinction between Level 2 and Level 3 agents determines whether AI represents a $50 billion or multi-trillion dollar market. Level 3 agents signify "task-level labor replacement, not assistance, but replacement."
On February 5, 2026, OpenAI released GPT-5.3 while Anthropic launched Opus 4.6. Marks cited industry observations that new-generation AI can autonomously write tens of thousands of code lines with testing and repairs, even demonstrating preliminary judgment and aesthetic capabilities. More significantly, "AI has participated in its own development process," with current AI sophistication substantially contributing to its own upgrades.
Can AI Replace Skilled Investors? Addressing investment industry implications, Marks offered a balanced assessment. He acknowledged AI possesses many traits of excellent investors: absorbing massive data, exceptional memory, and immunity to fear or greed. When relying solely on "easily accessible current quantitative information," human outperformance becomes extremely unlikely. Marks stated plainly: "AI's information processing capability likely surpasses all humans."
However, Marks believes AI remains imperfect for investment. AI's weakness lies in handling "entirely new developments" lacking historical data. Investment value will increasingly concentrate on non-quantitative work requiring subjective judgment of qualitative factors like management capability and product innovation - areas where emotionless AI currently falls short.
Bubble Concerns and Investment Strategy: Neither All-In Nor Complete Avoidance Addressing AI bubble questions, Marks presented multidimensional analysis. He clearly stated: "AI represents real technology with potential to revolutionize business and reshape lifestyles."
Regarding infrastructure overheating, Marks highlighted a key data point: AI inference capital expenditure now exceeds training expenditure. While training investment remains speculative, inference spending meets actual computing demand and generates substantial revenue growth, validating its rationale.
But Marks cautioned about "circular revenue" among AI companies and startup valuations "essentially equivalent to lottery tickets" without clear business models. Consequently, Marks provided explicit investment guidance: "Since nobody can definitively determine whether this constitutes a bubble, my advice is: nobody should invest fully because we must acknowledge potentially devastating risks if conditions deteriorate; but equally, nobody should completely avoid exposure, potentially missing this great technological revolution. Prudent selection of targets with moderate positioning appears optimal."
Marks expressed deep concern about AI's social impact, citing Claude's calculation that AI assuming 30-50% of structured software work would transfer "$150-250 billion annually in labor value to AI computing." He warned this transformation affecting paralegals, financial analysts, and software engineers may outpace society's adaptation capacity.
Comments