GPT-5.6 Sol, FIFA AI & Wall Street’s AI nerves
OpenAI's new GPT-5.6 Sol model sparks debate on AI safety and release strategies, while Wall Street expresses growing skepticism over the long-term economics of frontier AI models. The discussion also touches on AI's impact on the FIFA World Cup and a thought-provoking paper comparing LLM anthropomorphism to Age of Empires II "goats."
This week’s “Mixture of Experts” episode delves into OpenAI’s latest release, GPT-5.6 Sol, the growing skepticism from Wall Street regarding AI’s long-term economics, the impact of AI on the FIFA World Cup, and a thought-provoking paper comparing LLM anthropomorphism to Age of Empires II “goats.”
OpenAI’s GPT-5.6 Sol and the AI Race
The panel first compared OpenAI’s GPT-5.6 Sol (including Terra and Luna) with Anthropic’s Fable and Mythos models. While capabilities are seen as a back-and-forth race, a key differentiator noted by Kush Varshney and Lauren McHugh Olende is Sol’s enhanced safety mechanisms. Unlike Fable’s single-layer classifier, Sol employs a “defense-in-depth” approach with extra training, sophisticated guardrails, and a reasoning model to analyze responses. Lauren highlights that despite Sol’s claims of better scientific reasoning, benchmarks for tasks like protein binding still show low accuracy (7%), far from the 30% threshold for concern, indicating a focus on “chasing corner cases” rather than foundational breakthroughs like “chain of thought” reasoning.
Chris Hay criticizes OpenAI’s staged rollout, similar to Mythos, as a “stupid way of releasing models” that creates a “two-class system” and goes against the ethos of open source, which allows for broader problem discovery and fixes. He suggests OpenAI might be using it as a responsible approach to safety. Chris also points out OpenAI’s architectural difference in token efficiency, potentially capping “thinking time” for models (as theorized based on Noam Brown’s work) to prevent them from reaching “bad answers,” thus enhancing cybersecurity. Kush views the actual safety measures as real but the release process as “security theater.”
Lauren discusses the impact on open innovation, noting the 0-12 month gap between open and proprietary models. She questions the efficacy of the 30-day review period for new models, especially when only a select group of 20 organizations gets early access. She argues that without comprehensive, peer-reviewed testing and more technical expertise in government, evaluations can be biased or superficial, as seen with Fable’s jailbreaking incident.
Wall Street’s Skepticism on AI Economics
The discussion then shifted to Wall Street’s growing bearishness on AI, marked by SoftBank’s share drop, Nasdaq losses, and Apple’s decline. Concerns revolve around the escalating price of memory (DRAM, NAND) and the sheer infrastructure costs of frontier AI models.
Chris attributes market jitters to the speculative nature of AI investments, where companies must constantly innovate to stay ahead of open-source alternatives (which are 0-12 months behind). He emphasizes that proprietary AI companies are playing an “infinite game” where they must continuously outpace competitors or risk running out of resources. Kush draws a historical parallel to the British East India Company, which faced financial ruin due to massive investments and external factors, leading to government bailouts and eventually public rejection (the Boston Tea Party). He suggests that the high cost of staying ahead might lead to an unsustainable economic model for current AI giants.
Lauren highlights a disconnect between hardware and software innovation. While the market focuses on “Rammageddon and Rampocalypse” (memory shortages), significant software optimizations (quantization, key cache, flash attention) are actively reducing RAM requirements. She believes the market isn’t fully appreciating these software advancements. The panel acknowledges that the immense costs of training and serving inference need to be offset by revenue, and the current “golden window” for end-users, where subscriptions are likely subsidized, will eventually close as prices adjust to true operational costs.
AI at the FIFA World Cup
The panel discussed a Wired article detailing AI’s expanding role in the FIFA World Cup, from team logistics and strategy (recruiting, penalty analysis, squad selection) to FIFA’s own bespoke AI agent, “football AI Pro,” designed to level the playing field. Kush suggests that investments in AI for sports by smaller nations could yield secondary benefits for other sectors like agriculture.
Chris is skeptical that AI can fundamentally change individual player talent (e.g., make an average player like Lionel Messi) but finds the use of AI to find players with ancestral links for national teams humorous. Lauren points out the current limitations of data collection in soccer, where distant video footage is inadequate for precise analysis. She envisions a future with far more granular data, such as player-perspective video and gaze tracking, which could offer powerful insights but would face significant challenges in implementation and player acceptance due to privacy and the natural flow of the game. The panel also touched upon the debate around technology intrusion in sports, drawing parallels to the divisive VAR system in soccer.
LLMs, Anthropomorphism, and Age of Empires II
Finally, the discussion turned to a paper by Microsoft AI researcher Adrian de Winter, titled “If LLMs have human like attributes, then so does Age of Empires II.” De Winter critiques the widespread tendency to anthropomorphize LLMs, noting that a high percentage of recent AI papers either assume or conclude that LLMs possess human-like traits. His paper uses goats in Age of Empires II, configured to perform basic logic gates, to illustrate how even simple systems can be anthropomorphized, arguing that perceived human-like traits in LLMs are merely a “lens” through which we view the technology, not fundamental properties.
Kush sees de Winter’s “absurdist” approach as a powerful tool to make us aware that LLMs are merely “electricity with NAND gates.” He draws on Vedanta philosophy, distinguishing between observable reality (vyavahāra) and ultimate reality (paramārtha), suggesting that consciousness, even in humans, might be an illusion, and the “goats” help us understand this.
Lauren offers a nuanced perspective, acknowledging that while an “activation vector” for “fear” in a model might not be true emotion, it provides a useful “interface” for understanding and predicting model behavior. However, she notes the randomness and unexplained aspects, similar to our understanding of the human brain. She predicts that the question of “soul” in AI will become a prominent topic in the coming year.
Chris vehemently disagrees with anthropomorphism, calling it a “Groundhog Day” phenomenon akin to past reactions to ELIZA or expert systems. He argues that LLMs mimic human-like behavior because they are trained on vast amounts of human text, just as birds mimic car alarms without possessing deeper understanding. He asserts that, mechanistically, AI models are deterministic circuits, and looking “under the hood” reveals no “soul.” He concludes that current AI models do not possess biological or soulful attributes, though he remains open to future advancements changing this view.