Hard Fork

‘A.I.-Washing’ Layoffs? + Why L.L.M.s Can’t Write Well + Tokenmaxxing

March 20, 2026

Key Takeaways Copied to clipboard!

  • Recent tech layoffs at companies like Atlassian, Block, and potentially Meta are being publicly linked to AI investment, raising concerns about 'AI-washing' where AI is used as a convenient justification for cost-cutting driven by other financial pressures. 
  • Current large language models (LLMs) struggle with high-quality literary or creative writing due to post-training alignment (RLHF) that prioritizes helpfulness and safety over stylistic variability, a limitation even AI leaders acknowledge. 
  • The intense internal competition among tech employees to maximize AI token usage, tracked via internal leaderboards, incentivizes high consumption, which some executives view as a proxy for productivity, despite the risk of creating perverse incentives (Goodhart's Law) and massive infrastructure costs. 
  • Measuring productivity solely by token consumption ("tokenmaxxing") is viewed as a flawed metric, drawing parallels to outdated programming metrics like lines of code or aircraft weight. 
  • The practice of tracking and incentivizing high token usage is already spreading beyond technical roles into fields like marketing, potentially harming employee morale and focusing on activity over actual value. 
  • Unlimited free access to powerful AI models creates a unique employment lock-in, where employees might be unable to afford the cost of their necessary token usage if they switch jobs. 

Segments

AI Layoffs and AI-Washing
Copied to clipboard!
(00:00:30)
  • Key Takeaway: Tech layoffs at Atlassian and Block, with Meta expected to follow, are being framed by CEOs as necessary adaptations due to AI, prompting the concept of ‘AI-washing’ as a convenient excuse.
  • Summary: Atlassian cited rising bars for growth and changing skill mixes due to AI for its 10% staff reduction. Block’s CEO, Jack Dorsey, claimed a fundamental shift necessitated immediate cuts after a period of significant over-hiring, including a $68 million event with Jay-Z shortly before. The narrative linking layoffs to AI adoption is seen as powerful, capable of positively influencing stock prices, similar to past crypto name changes.
Meta’s Reported Layoffs Context
Copied to clipboard!
(00:10:16)
  • Key Takeaway: Meta’s reported 20% workforce cut is occurring alongside massive AI infrastructure investment, signaling to the market a cost-control measure while aggressively pursuing AI.
  • Summary: Mark Zuckerberg stated that projects previously needing large teams can now be handled by single talented individuals. Meta is investing $135 billion in capital expenditures this year for AI infrastructure. The layoffs signal to investors that the company is managing expenses while making its largest bet in company history on AI.
Worker Anxiety and Unionization Potential
Copied to clipboard!
(00:14:50)
  • Key Takeaway: Workers face anxiety over whether using AI tools proves their utility or confirms their work is automatable, potentially leading to increased interest in unionization as a protective measure.
  • Summary: Executives may view the fear generated by layoffs as a positive byproduct, reducing employee restiveness seen during the 2020 era. Unlike manufacturing workers who could negotiate automation impacts through unions, unorganized tech workers lack this leverage. The prospect of unionization at major tech firms like Meta is suggested as a potential future development.
LLMs’ Weakness in Creative Writing
Copied to clipboard!
(00:19:46)
  • Key Takeaway: Despite advancements in coding and math, LLMs struggle with compelling literary writing because their post-training alignment prioritizes verifiable, helpful responses over ungrounded, surprising creativity.
  • Summary: The guest, Jasmine Sun, argues that earlier models like GPT-2 and GPT-3 were sometimes more stylistically compelling because they lacked the restrictive post-training layers (like RLHF) that enforce a ‘helpful assistant’ persona. The process of grading AI output often relies on flawed rubrics, such as counting exclamation marks or checking fan fiction for ‘factuality,’ which fails to capture the essence of art. True literary voice and style stem from lived experience, which LLMs fundamentally lack, making their metaphors ungrounded.
AI Editing and Writer Self-Improvement
Copied to clipboard!
(00:38:35)
  • Key Takeaway: Writers can leverage LLMs effectively as collaborative editors by fine-tuning the AI with their personal archive and retrospective notes to develop a custom rubric based on their unique voice and aspirations.
  • Summary: The guest co-developed a qualitative rubric with Claude based on her past work, focusing on elements like leveraging her ‘insider anthropologist position’ and register switching. This custom editor prompts the writer to generate specific scenes or memories rather than inventing content, pushing for self-improvement rather than simple text generation. This collaborative ‘centaur model’ approach is seen as crucial where personal perspective matters.
Token Usage Leaderboards
Copied to clipboard!
(00:47:20)
  • Key Takeaway: Tech companies are tracking employee AI token consumption via leaderboards, using it as a new, albeit imperfect, proxy for productivity and incentivizing adoption of new agentic workflows.
  • Summary: Tokens are the atomic unit of AI labor, and advanced agentic tools can consume millions, leading to extreme monthly costs, with one top Claude Code user reportedly spending over $150,000 in a single month. Leaderboards are intended to motivate engineers to embrace the new way of working, but they risk triggering Goodhart’s Law by incentivizing wasteful token usage rather than valuable output. This mirrors historical attempts to measure programmer productivity using flawed metrics like lines of code.
Measuring Progress by Tokens
Copied to clipboard!
(00:58:23)
  • Key Takeaway: Measuring programming progress by lines of code is analogous to measuring aircraft progress by weight, suggesting token usage is an inadequate metric for AI productivity.
  • Summary: Historical context shows that measuring progress by simple inputs (like lines of code) has been historically flawed. While high token usage might correlate with productivity for some, it is not the definitive way to measure output. The industry is expected to realize this quickly due to escalating AI model costs.
Tokenmaxxing’s Economic Spread
Copied to clipboard!
(00:59:22)
  • Key Takeaway: The incentive structure of token-based leaderboards is spreading from tech into non-technical fields like marketing, often negatively impacting morale.
  • Summary: Managers across various sectors are attempting to incentivize and track AI usage, leading to non-technical employees being evaluated on AI adoption rather than pure creativity. This mirrors past issues, like traffic leaderboards at Gawker, which fostered competition over quality output. One marketing professional noted her bonus is now tied to AI usage, even when she felt her previous evaluation method was working fine.
Token Budget Consequences
Copied to clipboard!
(01:01:50)
  • Key Takeaway: Token budgets are becoming a tangible factor in job retention and negotiation, with some employees effectively trapped by the cost of their necessary AI consumption.
  • Summary: Some individuals have faced repercussions for spending too much on AI tools, while others are now negotiating token budgets during job interviews. Employees at major AI labs with unlimited free access face a unique situation where quitting might become unaffordable due to the high external cost of their required token usage.