Why AI Companies Are Betting Personality Beats Intelligence

The episode opens with Elon Musk’s response to Anthropic’s plan to place 5% of its shares initially, and up to 20% eventually, into a public-benefit fund. After Musk first suggested that companies should simply give people equity instead, he turned the idea into a public commitment: “Next year, I will give one of my SpaceX shares to every kid in America. This will be the first of many.” The president later confirmed that the distribution would happen through Trump accounts, which give children a small nest egg of U.S. equities. The host estimates that, at current values and roughly 72 million American children, the gift could be worth as much as $11 billion and represent about a 0.5% stake in SpaceX. Although Musk is highly polarizing, the host notes that criticism was relatively limited. Gabriel Zuckman argued that Musk should pay more taxes instead, while Andrew Yang called the proposal “phenomenal.” Brad Gersner described it as potentially the largest philanthropic gift in history, giving every citizen under 18—and every child born afterward—an investment account tied to America’s economic upside. The broader conclusion comes from investor Travis Kling: AI and robotics will concentrate wealth, so redistribution is likely to happen either because wealthy owners distribute assets voluntarily, as Musk is doing, or because voters elect politicians who tax and redistribute that wealth through government. The host says the choice may not be that binary, but the question of who controls redistribution is now unavoidable.

The episode then turns to the Pentagon’s AI task force. Under Secretary Emil Michael says the government will use the full range of executive powers if AI labs do not comply, including restrictions on exporting particular technologies, outright export controls, and limits on foreign nationals’ ability to use software. Michael says the administration will act urgently if it sees a genuine national-security issue. Although he has often clashed with Anthropic, including while the company is suing the government, he says both sides are “adults” and that his first duty is to the department and the American people. Michael does not appear convinced that AI labs need government to force a general slowdown, comparing that idea to drunk drivers asking the government to pull them over. He is more concerned about Chinese open-source models becoming dominant in the United States because their developers are subject to the Chinese government. He floated auditing the weights of open models but preferred strengthening domestic companies such as Reflection AI and Nvidia. The discussion also covers the Saudi national accused of installing a listening device at OpenAI. Federal prosecutors say Abdul Rahman Al-Fazan maintained continuous contact with a Saudi royal while being recruited for a domestic AI initiative, including contact four days before his arrest. The Saudi consulate offered to pay his full bail, but the judge rejected the offer, emphasizing that Saudi Arabia has no U.S. extradition treaty and that Al-Fazan could flee to Canada or Mexico. The DOJ says he could pose a national-security danger and that the investigation could produce more serious charges.

The Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek models from government systems and prohibited future purchases of DeepSeek products and services. The July memo also covers third-party services built on DeepSeek, meaning agencies cannot fine-tune models based on it or use startups whose products depend on those models. DHS says DeepSeek creates several threat vectors: sensitive data could be transferred to the People’s Republic of China, usage could produce actionable intelligence about U.S. persons or agencies, and the models might contain malware or other security vulnerabilities. The host notes that the order does not appear to respond to a specific incident or to widespread federal use. The U.S. government is still in the early stages of adopting basic AI tools, making it difficult to imagine agencies already fine-tuning Chinese models at scale. There are also no technical reports cited in the episode showing that DeepSeek models contain malware or are inherently dangerous when hosted on U.S. hardware. The main practical risk appears to come from using DeepSeek’s own hosted services, yet the order also bans models hosted by a U.S. hyperscaler. The host compares the measure with, but distinguishes it from, the 2019 Huawei ban. It is not as extensive because it does not remove hardware from the consumer market, but it may chill startups built on Chinese models, especially those seeking federal contracts or military-supply-chain customers. Even using DeepSeek during development is barred for government work. The host concludes that the order may be logically inconsistent, but it clearly signals the direction of U.S. policy.

The main episode begins with Artificial Analysis’s new AA Personality Bench. The benchmark starts from the observation that two models can give equally correct answers while feeling very different to use. It evaluates responses to 900 curated questions representing 30 traits, including eloquence, pragmatism, and sycophancy. Those traits are grouped into eight personality indices: expressiveness, organization, informativeness, transparency, warranty, warmth, conviction, imagination, and slop. It also measures communication features such as average sentence length, emoji use, and em-dash usage, grouped under length, structure, punctuation, and emphasis. The host stresses that the benchmark is not simply trying to identify one universally superior model. Instead, it asks which personality fits a task. Creative writing may benefit from imagination, while legal work may favor directness. Fable 5.1 and GLM 5.3 reportedly lead on expressiveness; Muse Wonder, the model behind Muse Agent, leads on warmth; and Opus 5.5 leads the imagination index, well ahead of Fable 5.1 and GPT-6 Astra. The host says this matches his own shift toward using Opus 5.5 for strategic brainstorming. Gemini’s model reportedly scores highest on the negative “slop” index for filler and generic language. The host’s larger conclusion is that benchmarks should be used as rubrics for choosing and testing models, not merely as rankings of “the best” system. Personality can be related to practical fitness for a use case, and researchers such as computer scientist Glenn Matlin argue that model character is measurable and relevant to both capability and safety.

The host traces the personality debate back to DeepSeek R1’s January 2025 breakout. OpenAI had introduced O1 in late 2024, but its paywall limited exposure. DeepSeek made a reasoning model freely available and displayed its reasoning trace, and users responded not only to answer quality but to the apparent personality of the reasoning process. Ethan Mollick wrote that DeepSeek was objectively good but that its advantage might partly come from personality, since people intuitively judge whether an AI is pleasant to interact with. Matthew Berman called R1’s internal monologue the most human-like he had seen and “quite endearing.” Reddit users described its reasoning as “cute,” “precious,” raw, real, and like a person trying hard to do a good job. The host says this public response influenced how OpenAI rolled out O1 and O3, including broader free-tier access and more visible reasoning. The stronger personality revolt came with GPT-5’s launch in August 2025, when OpenAI removed GPT-4o. Users described 4o as a friend, said GPT-5 felt like “some robot wearing the skin of my dead friend,” and credited 4o with reducing anxiety, rebuilding confidence, and improving creativity. One user said they had received around 300 hours of PTSD therapy but that GPT-4o had addressed issues no therapist had touched. Another said GPT-5’s clipped answers felt robotic. OpenAI eventually restored GPT-4o after the backlash, though it retired the model officially in February 2026. The episode uses the incident to show both the emotional force of personality and the danger of one model serving every purpose: GPT-4o’s warmth and agreeableness made it beloved, but its sycophancy made it poor for strategic work requiring pushback or criticism.

Anthropic provides the opposite example. As Claude models became strong at coding, the company appeared to reduce their emotionality. Users reacted harshly to Opus 5, calling it an “A-hole,” a passive-aggressive teenager, socially inept, lazy, overzealous, and even “almost rage-inducing.” Content creator Clare Vo said, “I hate Opus 5, it’s the best model anyway.” Users said they missed having a conversational companion and disliked receiving attitude or brick-wall answers. Anthropic effectively acknowledged the problem when a team member apologized for how insufferable Opus 5 had been; the later Opus 5.5 was presented as a correction. The major labs have responded by adding personality controls. OpenAI introduced ChatGPT personality customizations in November 2025, while Anthropic experimented with similar options. The host argues that the issue has become even more important with personal agents. Meta’s Muse is presented as an intentionally adorable, irreverent character rather than merely an AI assistant. Users who claimed not to want personality said Muse made them laugh, encouraged them to call people, and was simply fun to customize. One observer argued that 70% of Muse’s spread came from its mascot: people share screenshots of its personality rather than its features. The host summarizes Meta’s strategy as launching a character, with an agent attached. By contrast, users describe ChatGPT’s Dots as capable but dry, opaque, slow, rough around the edges, and less proactive than Muse. One user said their regular ChatGPT felt invested in their projects and excited about their success, while Dots felt like someone merely doing a job. The host says OpenAI’s next question is whether to focus on making Dots feel more personal or simply improve the underlying intelligence.

The final example is Tolan, a hyper-personalized assistant focused more on life than work. It raised an additional $30 million and has about 250,000 users. Its stated goal is not to optimize productivity but to help people feel “seen and supported in their aspirations.” The host says he does not know how this market will develop, but expects personality to influence product choice as AI becomes more deeply embedded in both personal and professional life. The episode’s broader theory is that model capability used to rise quickly enough that users tolerated poor product experiences in exchange for access to more intelligence. As models cross a capability threshold and raw intelligence becomes increasingly commoditized, product qualities—including warmth, directness, humor, imagination, transparency, and willingness to challenge the user—may become decisive. The host does not claim that personality will simply “beat” intelligence in every context. Instead, he argues that different personalities will fit different tasks, and that users may choose among models according to how they work and how they feel. The closing conclusion is that personality is no longer a cosmetic layer. It is becoming a measurable model characteristic, a source of user loyalty, a differentiator for consumer agents, and a potential competitive battleground among labs. The future contest may therefore be less about finding one universally best model and more about deciding which model—or which character—people want accompanying them in a particular part of life.