The Most Important Trends in New AI Products

The episode opens with Nous Research’s reported $1.5 billion valuation after a new funding round. The group initially attracted attention with an attempt to train an LLM across a distributed network of anonymous GPUs, which the host describes as part AI research and part performance art. Its major breakthrough was Hermes, an agent built around the extreme user control and openness that had drawn people to OpenClaw. The Wall Street Journal reportedly says Hermes has been installed 22 million times and accounts for about 2.5% of global token usage. Nous now wants to move from individual enthusiasts into enterprise with Hermes for Business. The free product lets users bring their own models, while paid plans bundle frontier-model usage and hosted instances. That model has produced $36 million in annualized revenue so far this year, with a target of $100 million by year-end. CEO Dylan Rollnick framed the funding as an open-source manifesto: software is cheap to reproduce, users dislike lock-in, and they deserve visibility into systems allowed into their lives. Nous’s stated requirements include choosing smarter, cheaper, local, or task-specific models; switching models mid-conversation; controlling data and memory; running agents offline or on different computers; reading prompts and tool activity; and exporting or modifying the code. The host sees the fundraise as evidence that open source can support a substantial AI business.

The host then turns to what the Wall Street Journal describes as tens of billions of dollars in chip-financing deals. Oracle is reportedly seeking financing from Apollo and Goldman Sachs for a major chip purchase. Because Oracle is already at risk of a credit downgrade, the host suggests the company may be turning to private markets rather than relying on investment-grade bonds. Broadcom is reportedly discussing a $50 billion private-credit deal involving Apollo and Blackstone to finance OpenAI’s custom chips. The arrangement appears to let Broadcom lend its credit rating to OpenAI, potentially giving the frontier lab a cheaper borrowing rate and indicating that frontier labs may struggle to borrow at investment-grade terms once public. SpaceX is separately seeking $40 billion to buy Nvidia chips, with Apollo leading what would be one of the largest financing deals in history. Several lenders reportedly rejected SpaceX after receiving a two-page memo containing space pictures and an arrow pointing to data centers “somewhere in the universe.” One investor asked how that could be presented to an investment committee, although Bloomberg says PIMCO is considering the deal. SpaceX credit-default swaps rose, pricing in a 15% chance of default by the end of 2031. The host contrasts that skepticism with Citrini’s joke that investors should decide simply whether they want to give Elon Musk $40 billion for extraterrestrial data centers.

Mark Zuckerberg’s BioHub has launched a $1.8 billion biological-data initiative with Google and the U.S. government. Its goal is to assemble the largest open biological dataset ever created so AI models can reason about entire cells, extending beyond protein-level systems such as Google’s AlphaFold. BioHub science head Alex Rives says the aim is to let scientists run virtual experiments first, sending only the most promising ideas to physical laboratories. The host frames this as an attempt to solve biology AI’s data bottleneck rather than simply building larger models. Anthropic is also expanding access to its previously restricted Mythos model through a new cyber-verification program. Mythos had been available to only a few dozen companies in Project Glasswing, and its international expansion created a political dispute before access was narrowed to U.S. companies and government departments. The new program offers three levels: Defense access for incident response, malware reverse engineering, and vulnerability analysis; Red Team access for limited offensive work such as penetration testing; and Specialized Access for organizations testing high-risk systems such as flight controls, power grids, and financial infrastructure. The Red Team tier requires enhanced vetting and is limited to enterprise customers, with real-time blocks against actions that could cause physical harm or mass disruption. Specialized Access requires government approval and has mandatory data retention, unlike the other tiers’ zero-data-retention availability. Researcher Nathan Lambert argues that an accidental open release would cause more cyber incidents but represent an acceleration rather than a dramatic change in risk. JPMorgan CEO Jamie Dimon disagrees, saying cybersecurity risks rose “tenfold,” though he favors working to fix the problem rather than becoming hysterical about whether it is existential.

The main discussion begins with Ramp’s monthly ranking of fast-growing software vendors by market share and revenue growth relative to company size. The host cautions that Ramp’s customers are unusually technology-forward, so the figures do not represent the entire enterprise market. They are still useful, he argues, because early adopters often establish patterns that the broader enterprise later follows. The October data showed Type-Safe AI, creator of the decision model Jev, ranking first in growth relative to size and second in absolute market-share growth. Ramp classifies Jev as a foundation LLM, and it gained a full percentage point of market share in one month. The host says this indicates that Jev is being adopted by startups and advanced enterprises, not merely attracting attention on X. Anthropic, SpaceX AI, and OpenAI also ranked among the top five for market-share growth, while SpaceX AI’s appearance suggested that GrokBot is driving meaningful business adoption. DeepSeek, OpenRouter, and Featherless also appeared. Ramp economist Eric Herazine said cheaper models, high-performance small models, and token-efficient models for specific enterprise tasks are pushing AI spending down and threatening OpenAI and Anthropic’s business revenue growth. Scale AI’s return to the growth chart may indicate renewed demand for reinforcement-learning data or fine-tuning. Runway and Higgsfield’s growth suggests that image and video generation are increasing AI spending, although those tools remain concentrated in marketing and advertising rather than construction or manufacturing automation.

Elon Musk announced that GrokBot would use “the best backend model for any given task,” including Claude Opus, Midjourney, Suno, and other leading APIs. The host treats this as a major strategic signal in the debate over whether AI moats will reside in the model or in the surrounding harness and product. Musk later clarified that most simple requests would still use a fast Grok model, with the operating principle being the best combination of speed and intelligence rather than abandoning SpaceX AI’s own models. Some users objected that they already pay for Grok and should not have outside-model API costs passed through to them. Others wanted an option to use only Grok models for philosophical reasons, since parts of Grok’s audience oppose rival labs. Investor Anisha Sharia argued that SpaceX AI can have both advantages: vertical integration and its own state-of-the-art model, while also routing requests to third-party models and potentially capturing inference margin. The host agrees that the most important consequence is user experience: businesses and power users increasingly want the surrounding context and tools without being locked to one model. GrokBot is not yet as open as Hermes Research: its model access is limited by relationships with Musk and Grok, does not include OpenAI models, and involves Anthropic using or training models in SpaceX AI data centers. Still, the announcement supports a broader move toward multi-model products. OpenAI’s B2B marketplace similarly lets enterprise customers spend part of their commitment on open-model inference. GrokBot also gained the ability to search, read, and monitor X, expanding its usefulness for social-media monitoring and agentic workflows.

Anthropic’s Claude Haiku 5.5 represents the opposite side of the model-choice trend: a foundation lab trying to serve the low-cost, high-speed end of the market itself. Anthropic calls it the cheapest, fastest, and most capable small model it has released. In the benchmarks cited, Haiku 5.5 was far ahead of Haiku 4.5 and beat GPT-6 Luna across the reported comparisons. Haiku scored more than twice as high as Luna on Terminal-Bench 4.0 for agentic coding, with a smaller lead on FrontierCode 1.1. It scored 72.4% on OSWorld computer use versus Luna’s 48.9%, and beat Luna on the GDPVAL and AA Briefcase knowledge-work evaluations. Artificial Analysis gave it an Intelligence Index score of 43, placing it between Kimi K3 and GLM 5.3 Flash. The main announcement was pricing. Haiku 5.5 costs one-quarter as much as Sonnet 5.5 and half as much as Haiku 4.5 per token. Anthropic discounts tokens by 80% for prompts below 100,000 tokens, targeting quick queries and subagents. Artificial Analysis estimated task costs of $0.21 at maximum inference and $0.12 at extra-high inference. Investor Haseeb Qureshi called this a “narrative violation”: U.S. models now occupy the cost-and-intelligence frontier previously dominated by DeepSeek and GLM. The remaining question is whether enterprises chose Chinese open models only for price, or also for sovereignty, local deployment, and control over training data—areas where American labs have not solved the problem.

OpenAI announced that a new GPT-6 model would be available to more of ChatGPT’s roughly 1.2 billion weekly users, extending the next generation of intelligence into the free experience. The host says this matters because people’s perception of AI is shaped heavily by what they encounter in free ChatGPT, not only by premium or API models. The more consequential announcement was “intelligent UI,” which lets ChatGPT present answers through interactive visualizations, charts, images alongside recipes, and other context-sensitive formats rather than defaulting to walls of text. OpenAI researcher Esa Fulford said the team built a library of native, streamable components and a compiler that renders interfaces generated by the model. The model was trained to decide when interaction helps and when plain text is preferable. Andrew Chen said the difficult problem was not making models generate attractive HTML; it was making the result feel native to ChatGPT—fast, coherent, and appropriate to the conversation. The model must judge what to show and what the user should be able to do without overwhelming the exchange. Some early users thought the experience made information less clear or added too much content, while others welcomed the change from text walls. The host sees two trends: eventually, interfaces may be generated just in time for each problem, and immediately, foundation-model companies are finally being forced to compete on product design now that powerful intelligence is becoming widely available.