Hollywood Strike Update and AI Roundup with Trey Kollmer

Episode Summary

Executive Summary: The episode centers on Hollywood’s writers’ and actors’ strikes and how generative AI is reshaping bargaining over credit, compensation, training data, and digital likeness rights. Trey Colmer argues studios are partly accommodating near-term loopholes, but the deeper fight is over whether AI can be trained on creators’ work and whether new laws are needed. The conversation broadens into model reasoning, long-context limits, creative generation, and where value will accrue across the AI stack.

Main Topics: WGA/SAG-AFTRA strikes and AI clauses (Priority: 5/5): The discussion updates the status of the writers’ strike and the newly joined actors’ strike, focusing on AI-related demands around credit, residuals, likeness control, and protections against studio loopholes. Copyright, training data, and fair use (Priority: 5/5): The speakers unpack legal theories around copyrighted training data, derivative outputs, copyright management information, and whether current law can handle model training at scale. AI creativity and writing workflows (Priority: 4/5): They debate how well models can write stories, jokes, intros, and commercials, with fine-tuning, few-shot prompting, hierarchical scaffolding, and generate-and-mine workflows as key techniques. Reasoning, robustness, and model interpretation (Priority: 5/5): A large portion of the conversation explores whether models really reason, why they fail on slightly modified tasks, and what mechanistic interpretability and benchmark design can reveal. Long-context models and architecture limits (Priority: 3/5): They discuss Claude 100K and similar long-context systems, noting that more context helps with document processing but does not reliably improve reasoning and still suffers degradation at long ranges. Where value accrues in the AI stack (Priority: 4/5): Nathan and Trey debate why NVIDIA has captured far more market value than TSMC or ASML, and they speculate about durable value capture in chips, cloud compute, frontier models, and new AI-native apps. Dramatic irony and near-future media disruption (Priority: 3/5): The closing 'Reality Writer’s Room' frames current events as a story about warning signs being ignored, especially Microsoft’s Sydney incident, deepfakes, digital actors, and the changing economics of entertainment.

Key Arguments: Generative AI already affects bargaining because studios can use it to shift writing credit and compensation, such as paying for rewrites instead of first drafts or replacing paid story work. The AMPTP’s latest offer appears to block some loopholes by stating AI-generated material cannot count as assigned material or a first draft for rewrite purposes. The unresolved core issue is training data: unions want limits on using writers’ scripts to train or fine-tune models that could replace them, even when studios own the copyrights under work-for-hire. Copyright law may treat model training differently from outputs: courts may reject derivative-work claims for outputs while still scrutinizing copying into training sets or removal of copyright management information. AI tools can already help with structure, outline generation, and some commercial writing, but creative work still tends to be too noisy and subjective for reliable end-to-end automation. Few-shot prompting and fine-tuning are on a continuum; model behavior shifts with examples, and fine-tuning can materially improve structured writing tasks. Long-context windows increase working memory but do not necessarily improve reasoning; they often still fail on longer transcripts or complex summaries. Model 'reasoning' likely exists in some form, but it is brittle; small prompt changes, irrelevant details, or adversarial strings can route the model into different modes. Future architectures may need recurrence, selective memory, backtracking, or hierarchical decomposition to support better reasoning, creativity, and robustness. Value across the AI stack may accrue to hardware, cloud operators, frontier model providers, and app-layer products that create new experiences rather than simply automate existing workflows.

Data Points: Episode length: Over 2 hours - The conversation is described as a wide-ranging discussion spanning multiple topics. White House voluntary commitment signatories: 7 companies - OpenAI, Anthropic, Google, Microsoft, Meta, Amazon, and Inflection agreed to the commitments. Frontier Model Forum official members: 4 companies - OpenAI, Anthropic, Google, and Microsoft officially joined the forum. Copyright Office inquiry questions: 34 questions - Trey notes the Copyright Office issued a detailed notice of inquiry on AI and copyright. Claude context window: 100K tokens - Nathan references Claude 100K as a major improvement for long-document processing. Annual context period: 1 year and 5 days - They note the speed of progress since Stable Diffusion’s release. NVIDIA share price increase: From $143 to $485 - Nathan compares NVIDIA’s year-to-date stock move with other AI-related stocks. NVIDIA market cap: $1.2 trillion - Used to illustrate how much value NVIDIA has captured relative to suppliers. ASML market cap: $270 billion - Compared against NVIDIA and TSMC in the hardware value-capture discussion. TSMC market cap: $445 billion - Used as another comparison point in the AI semiconductor stack. Samsung market cap: $472 billion - Mentioned as a comparable company alongside TSMC. NASDAQ year-to-date performance: Up 35% - Used as a benchmark against AI semiconductor stocks. ASML year-to-date performance: Up 40% - Nathan notes ASML is up, but far less than NVIDIA. OpenAI revenue run rate: $1 billion - Referenced as a reported/credible scale estimate for frontier-model commercialization. Training scale claim: 100,000x GPT-4 compute - Trey cites Mustafa from Inflection saying the next models may be trained on vastly more compute. Anthropic example project: 150 applications - Nathan says Claude 2 summarized roughly 150 charity-evaluation applications as a first step.

Pivotal Quotes: "You're copying these works and using it to create a system which can just replace a lot of the underlying creators." — Trey Colmer: On why training-data use remains the central labor and copyright concern for writers. "It does seem like both things are kind of going on somehow." — Nathan LeBenz: On the idea that models can be both genuinely reasoning and still brittle or pattern-driven. "We have this kind of mad science project in search of a capability, which then kind of went in search of a problem." — Nathan LeBenz: On the broader arc of LLM development and how use cases are being discovered after capability breakthroughs.

Implications: Listeners should expect more conflict over training rights, likeness rights, and AI compensation while model capabilities keep improving. The episode suggests the industry is moving toward new legal regimes, new creative workflows, and a sharper split between automation that helps and automation that replaces.