Open Models, AI Infrastructure, and Search: Four Perspectives

The recording opens mid-conversation as Matthew Berman and Nick Wentz discuss NVIDIA's fiscal 2026 third-quarter results. They cite revenue above $57 billion, strong data-center demand, and a gross margin of 73.4%. Berman distinguishes his belief in AI's long-term importance from the separate question of whether near-term capital spending is excessive.

The first guest, Sentient co-founder Himanshu Tyagi, says his company aims to build an open ecosystem for general AI, including reasoning agents as well as models. He notes the prominence of Chinese open-weight releases and wants builders to contribute useful systems. These are the company's aims, not a demonstration that AGI has been achieved.

Tyagi describes AGI as an ongoing pursuit of broader common-sense reasoning, improvement, and human-level performance across tasks. Rather than treating it as a switch, he expects progress to continue across different domains. Berman asks how anyone would know when the goal is reached; Tyagi prefers a moving horizon to a single threshold.

Tyagi expects online tasks such as research, planning, and purchasing to be among the first areas transformed by agents. He suggests a roughly two-year horizon for major changes to the digital experience. That is his forecast, not evidence that such a transformation will occur on that schedule.

He frames reasoning as dealing with situations outside what a model has already learned, and argues that useful systems may combine models with agents, planning, and world models. This is his working definition; the relationship between learned patterns and generalization is an active research question, not a simple test for whether reasoning occurred.

On continuous improvement, Tyagi distinguishes changing model weights from updating prompts and workflows around a model. He is particularly interested in optimizing an agent's strategy through its prompt. He argues that creating an agent is easier than reliably improving it, which requires evaluation and iteration.

Asked why openness matters, Tyagi says public research lets ideas compound and gives builders opportunities to demonstrate work and create businesses. He also argues that countries may want more control over systems used in important decisions. Open weights, training data, code, licenses, and governance are distinct dimensions of openness.

Tyagi argues that sovereign AI may require access to training data for audit, not just model weights, and says some institutions hesitate to deploy models from particular countries. Availability of data can enable inspection, but it does not by itself prove the absence of backdoors or make a model neutral. Auditing large datasets and trained systems remains difficult.

The hosts ask how decentralization can work when frontier models need substantial hardware. Tyagi distinguishes smaller local inference from training across distributed clusters and says regional infrastructure and regulation affect the economics. His example clusters still require significant investment; decentralization is not synonymous with running a frontier model on a laptop.

On Chinese open models, Tyagi says he was surprised by the combination of technical progress and global distribution. He suggests international visibility can also support a domestic business, and describes DeepSeek as encouraging others to experiment. He explicitly notes that he does not have deep inside knowledge of Chinese decision-making.

Tyagi is excited about open work on science and domain-specific world models, and mentions Sentient's search and financial agents. The hosts respond that competition and the ability to inspect and experiment are important to them. These expected benefits depend on actual licenses, access, and implementation.

The second guest, Reflection AI co-founder and CEO Misha Laskin, describes his company as building frontier open models outside China. He argues for multiple sources of capable models, both as competition to closed providers and as a choice for countries and enterprises. This is his company's strategic position, not proof that any particular model is already competitive.

Laskin divides potential customers into consumer applications, startups using hosted APIs, and larger enterprises or governments that may want more control. He thinks open and closed models can coexist. As organizations grow, he says, control over costs, data, and customization can make open models more attractive, although operating them also adds responsibility and expense.

On safety, Laskin argues that access lets more researchers study powerful models rather than concentrating knowledge inside a few labs. He also says capable open models already exist and that safety work must account for them. Openness can support scrutiny, while broad distribution may also make dangerous capabilities harder to restrict; neither approach eliminates risk.

He discusses possible measures such as distribution controls, provenance, and research into resistance to adversarial fine-tuning. One research idea is coupling safety behavior and useful capabilities so that removing safeguards also degrades performance. He presents this as an open research direction, not a demonstrated solution.

The hosts raise a reported case of closed-model misuse in cyber operations. Laskin says both open and closed models can be misused, while open weights may permit adversarial customization. He expects cybersecurity to be a major area for both attack and defense. The comparison does not imply that access controls and monitoring have no effect on risk.

To build a frontier lab, Laskin lists talent, compute and capital, and a sufficiently large commercial opportunity. He argues that larger enterprise and sovereign customers now provide demand for open models. That is a business thesis from an interested company, and actual procurement choices differ by customer.

Laskin describes an intentionally plain sales slide that evolved through customer conversations. The goal was to discover which problems actually mattered: predictable costs, keeping data within infrastructure, avoiding lock-in, and customization. A good presentation cannot substitute for validating those claims in each customer's environment.

He says Reflection was assembling talent and compute and working up a scaling ladder, training smaller models before larger ones because failures change with scale. He expects to release models the following year and does not foresee an imminent scaling wall. Both are forecasts made in 2025, not guarantees.

Laskin defines a world model as a way to predict observations after an action and discusses language and multimodal models in that context. He expects improvements to continue across text and visual environments. The term covers different research approaches, and simulated predictions can still be wrong.

NVIDIA's Bryan Catanzaro joins next. He recalls experimenting with CUDA as a graduate student and rewriting a support-vector-machine training program for a GPU much more quickly than his earlier FPGA work. He reports a large speedup over his CPU implementation in that particular experiment, which helped convince him that GPUs could accelerate machine learning.

Catanzaro says neural networks faced considerable skepticism when he joined NVIDIA full-time in 2011. AlexNet's image-classification results helped change perceptions, and work on GPU convolution libraries became cuDNN. He recalls a 2013 meeting where Jensen Huang pushed the company to focus heavily on the opportunity in AI.

He describes NVIDIA's culture as a long-term effort to make an entire chain of technology work together, from chips and compilers to libraries and applications. In his view, customers buy a capability and time saved, not just a chip. This is the company's account of its strategy and commercial value.

CUDA, Catanzaro says, includes drivers, compilers, frameworks, applications, and a developer ecosystem. Achieving speed in a cloud data center also requires networking, storage, and operational work. NVIDIA trains its own models on external cloud infrastructure to learn from the same integration problems customers face.

He introduces NVIDIA's Nemotron models and datasets as a way to understand future AI workloads and design systems accordingly. He reports that improvements in pretraining data produced a fourfold improvement in GPU time to reach a target level of accuracy for the company's work. That is an internal result under particular conditions, not a universal multiplier.

Catanzaro discusses Blackwell's NVFP4 low-precision format and says NVIDIA has trained models with four-bit arithmetic while maintaining strong results. He argues that lower precision can improve efficiency and enable larger training runs. Actual quality and savings depend on the workload and implementation.

With data centers demanding more power, he emphasizes efficiency. He highlights the quadratic cost of conventional attention and the memory used by the key-value cache during inference. Hybrid architectures and state-space approaches are among the techniques NVIDIA is exploring to reduce inference costs.

On longer context, Catanzaro says needs vary by application and considers sparse or hierarchical approaches alongside retrieval. A personal assistant that can use a person's messages and email might be more helpful, but would also have access to highly sensitive information. He views privacy and protection as essential alongside technical capability.

Asked about updating model weights continuously, he is cautious: changing a model after safety evaluation adds complexity. He suggests that a capable model may reason over new information in context without changing its weights. This is his preference, not a proof that continual learning cannot be done safely.

Catanzaro says Nemotron's second purpose is to help others deploy and customize AI. He explains NVIDIA's business incentive: when AI adoption grows, demand for its infrastructure can grow too. The degree of openness and reuse rights depend on the specific model, dataset, and license.

The final guest is Robby Stein, vice president of product for Google Search. He describes a gradual progression from AI Overviews to AI Mode and the launch of Gemini 3 in Search. He says teams worked toward a system that can choose richer layouts and use the web to address a user's question.

Stein recalls an internal example where the system generated an interactive wing simulation with a slider. The idea is to adapt the response to the question instead of always returning a fixed text layout. Such generated visuals can be engaging, but their correctness still needs evaluation; an illustration is not automatically a reliable scientific model.

Asked about hallucinations and consistency, Stein says Google builds on long-standing search quality systems and combines model instructions with web retrieval. He acknowledges mistakes. He says the team uses human evaluation, feedback, repeat use, and other signals together, noting that repeated searches can also signal that a user failed to find an answer.

On advertising and commerce, Stein argues that AI search can create new opportunities to connect people with useful information, products, and links. He describes personalization as potentially improving relevance. These are Google's product and business views; users and publishers may experience different tradeoffs around privacy, disclosure, and traffic.

For purchasing and dining, Stein says much of the value is in helping people compare choices before a final transaction. He describes a dining experience that returns availability from booking services but did not yet complete the booking in the example. More autonomous actions are a possible future direction, not a claim that all purchases were automated.

Stein says AI Mode may fan a question out into multiple searches and bring together useful sources. He argues that familiar quality principles such as relevant, original, sourced content continue to matter. His account is Google's description of its approach, not a guarantee of ranking or traffic for any publisher.

Search routes questions among models according to their needs, Stein says: a simple request may favor a faster model, while difficult reasoning or multimodal questions may justify more computation. He describes higher limits and access to advanced features for paying subscribers at the time. Pricing, availability, and models have changed and should be checked independently.

Stein distinguishes AI Mode's focus on information, shopping, and local discovery from a broader chatbot's productivity uses, while recognizing overlap. He says the products should meet people where they already work. He also discusses experiments using personal context such as Gmail with permission, and adaptive visual or voice experiences.

The hosts thank all four guests and close the show. Across the interviews, the central tension is how to make more capable AI useful and accessible while preserving reliable evaluation, security, and control over sensitive information.