Forward Future Live: Industrial AI, Satellites, Deepfakes and World Models

Matthew Berman and Nick Wentz open the February 20, 2026 edition of Forward Future Live. Their guests are IFS CEO Mark Moffat, Vantor CEO Dan Smoot, Pindrop Security CEO Vijay Balasubramaniyan and Runway co-founder and CTO Anastasis Germanidis. The episode begins with the hosts’ reactions to recent AI news.

The hosts discuss the Gemini 3.1 Pro announcement and an announced 77.1% ARC-AGI-2 score. They compare it with the earlier model’s reported result and watch visual examples including an SVG and a city-planning demonstration. These are results and examples discussed at the time, not an independent assessment of general capability.

Neither host has extensively tested the model yet. Berman says he plans to try it in OpenClaw both as a conversational assistant and behind the scenes in automated workflows, and report back on the experience.

Asked about his recent OpenClaw uses, Berman describes a custom CRM and a searchable knowledge base collecting articles, videos and posts. What excites him most is when the system connects records across those workflows—for example, linking a contact with an article about that person’s company.

He also describes a meeting workflow: Fathom transcripts are processed for action items, which are added to a task list and assigned in HubSpot. He says the assistant can check subsequent email and mark a task complete when it sees that the promised work has been done.

Wentz asks how someone without much technical experience should get started. Berman says that, at the time, he would first learn basic AI-assisted coding: running OpenClaw involves security concerns and attack surfaces that require attention. He expects hosted assistants to become more accessible, while potentially offering less flexibility.

The hosts discuss changing Anthropic documentation and subscription restrictions as they understood them that week. Berman reports that his own authentication stopped working and later resumed; this is a historical account of his experience, not a statement of current terms.

Berman says he is testing Sonnet 4.6 for tool calling and invokes Cursor through OpenClaw for coding. He discloses that Cursor had provided him with generous access to tokens, a relevant detail when considering his assessment.

The hosts introduce Mark Moffat, CEO of IFS, whose work focuses on industrial software. They ask how the company’s history and recent growth relate to its emphasis on industrial AI.

Moffat says IFS has spent decades focused on industrial verticals and the lifecycle of assets and services, from capital allocation and construction through operation and maintenance. In his view, this accumulated domain knowledge and operational data make the current AI shift an important opportunity for the company.

Asked what IFS actually does, Moffat describes enterprise software for capital-intensive businesses, including utilities, energy, construction and telecommunications. A field technician might use it to organize planned maintenance or an emergency repair.

He distinguishes language-model assistance with voice or written reports from other uses of AI beneath the surface. A system could compare a technician’s findings with previous repairs to help predict faults, schedule work by urgency, and suggest which parts to carry.

Moffat argues that applying AI effectively requires understanding the full workflow and the people doing the work. He describes spending time with engineers, planners and manufacturing staff to locate specific operational problems before building a feature.

As an example, he describes work with Boston Dynamics and Eversource Energy on inspection of underground utility ducts. A Spot robot can collect video, sound and sensor readings in a hazardous environment; IFS software coordinates the robot and the resulting work.

In his account, the system can relate a detected issue to previous faults or manufacturer information, use models including Anthropic’s, and initiate a workflow: order a part, create a work order and send a human technician to repair it. This is the company’s description of the proposed closed loop, not an independently verified performance result.

Asked about world models, Moffat says IFS builds smaller specialized models where appropriate but does not expect to be the sole source of innovation. He describes early partnership conversations around simulating physical engineering environments, without naming a partner or claiming a completed deployment.

Wentz asks how a company can trust models in critical workflows. Moffat acknowledges that mistakes can threaten lives and says IFS uses simulations and extensive testing instead of accepting a model’s output at face value. His numerical reliability language is not supported by a disclosed measurement in the interview.

On computing capacity, Moffat says IFS had not experienced a shortage for its own needs and credits relationships with providers including Microsoft. He says the company continues to watch the situation.

Moffat distinguishes digital twins and engineering design systems from world models that might represent the physical instance. He says the company sees them as complementary and is pursuing partnerships, while acknowledging that this work is still early.

He views changes in trade, industrial investment and energy infrastructure as business opportunities for IFS, alongside customers’ need to manage supply-chain changes. This is the perspective of a software supplier serving those industries.

On robotics, Moffat says software will need to coordinate both human and robotic workforces. He would emphasize hazardous or unsafe environments, while acknowledging that the extent to which robots substitute for human labor is unknown.

Moffat is optimistic about jobs and cites a forecast of both job creation and displacement. He calls for companies and governments to support reskilling. His optimism is a forecast, not a guarantee that every worker who learns AI tools will keep a job.

The next guest, Dan Smoot, leads Vantor, formerly known as Maxar. He describes a satellite-imagery and geospatial-intelligence business that is trying to turn its image archives into applications, rather than only supplying raw data.

Smoot says the company’s satellites collect high-resolution Earth images and that its maps support two- and three-dimensional views and change detection. His claims about coverage and revisit frequency describe company capabilities; they should not be read as continuous observation of every location.

Vantor also images objects in orbit. Smoot says satellites can be turned to photograph other satellites, giving operators another way to inspect and document activity in space.

One proposed application is navigation when GPS is unavailable. Smoot says a drone can compare camera imagery with a previously built three-dimensional landscape map, while ground systems handle some of the processing. He mentions military and delivery uses, but the accuracy figures he cites are not independently tested in the interview.

He expects autonomous navigation to combine multiple sensors rather than depend on GPS alone. The hosts ask whether the same approach could eventually be used in a car; Smoot says it is possible, not that the interview establishes such a deployment.

The hosts ask about Vantor’s Google partnership and other relationships. Smoot describes applying Google Earth AI models to historical imagery to find objects and changes, then producing text summaries for analysts. He says a version can run in isolated customer environments.

Smoot gives maritime monitoring as an example: imagery analysis could help distinguish vessel types, follow movements and direct analysts toward areas worth checking. He presents this as a defense and intelligence application of the data.

He says automated analysis can shorten the time between acquiring images and preparing a report. That speed claim comes from Vantor’s CEO; the conversation does not include a controlled comparison or error rates.

The company’s competitive position, Smoot argues, comes from both owning the imaging hardware and accumulating an archive of data. The same infrastructure also requires security against jamming, spoofing and threats to satellites and ground links.

Space is large, but operators still track many satellites and pieces of debris. Smoot describes automated warnings that can give an operator time to plan an avoidance maneuver. He distinguishes active, maneuverable spacecraft from debris that cannot easily move out of the way.

Asked about space-based data centers, Smoot sees a possible role for orbital power and communications but stresses that operating in space introduces difficult engineering and environmental questions. He does not provide a deployment timeline.

He distinguishes geospatial digital twins built from measurements of actual locations from generative world models. In his view, the two could complement each other, but navigation and some training scenarios depend on geographic accuracy.

Smoot also sees civil uses beyond defense: urban planning, disaster preparation, flood mapping and monitoring activities such as illegal fishing or deforestation. He describes these as potential and ongoing applications, while the interview does not independently verify each outcome.

Asked about the delay between an event and an image, Smoot says some government use cases have taken as little as 15 minutes. He recounts a company demonstration timed at about twelve and a half minutes from capture to delivery. Those examples are not a promise of that latency everywhere.

On privacy, Smoot emphasizes securely delivering requested imagery into customers’ private environments. His answer focuses on customer data protection and access; the conversation does not establish a complete policy for the privacy of people pictured by satellites.

The hosts introduce Vijay Balasubramaniyan, co-founder and CEO of Pindrop Security, to discuss synthetic voice and video. He says deepfake-detection technology exists, but argues that adoption by the platforms and organizations receiving the content has lagged.

Asked about education versus technology, Balasubramaniyan says familiar visual and audio cues can become unreliable as generation improves. He cites a company study on people’s detection accuracy; its result depends on the study’s design and should not be generalized to every listener or every type of fake.

He describes fraudsters using cloned voices to impersonate a relative and create a false emergency. The practical difficulty is that a convincing voice can exploit urgency before the recipient has time to verify the story.

Pindrop’s approach, he says, looks for anomalies in audio and video as well as identity signals. He distinguishes two questions: whether an interaction involves a real person at all, and whether that person is the individual they claim to be.

The company primarily serves organizations such as banks, insurers, healthcare providers and retailers. Balasubramaniyan describes attackers combining breached or stolen information with synthetic voice and video to impersonate customers; he says Pindrop examines enterprise interactions for signs of fraud.

He claims a steep increase in AI-generated fraud calls observed among customers and gives percentages for some organizations. Those are his account of Pindrop’s observations, not a representative estimate for all calls or institutions.

He contrasts early voice-cloning systems that needed extensive recordings with newer tools that can work from much shorter samples. Berman notes that his own public videos supply extensive audio and images; Balasubramaniyan says the concern extends to both consumers and hiring.

Balasubramaniyan cites figures about fraudulent job candidates and suspected state-linked applicants. The hosts do not establish the sample, definitions or verification methods, so the figures cannot be treated as prevalence across the general job market.

Despite the threat, he is optimistic that detection can improve. He explains that a synthetic voice must reproduce many aspects of a person’s delivery, while a detector searches for inconsistencies. His assertion that detection is much cheaper than generation is a company perspective, not a demonstrated universal result.

He compares broader deployment of detection to the adoption of earlier online security measures and argues that carriers, social platforms and other intermediaries should participate. Detection is probabilistic: the discussion itself returns to the importance of measuring accuracy and earning trust.

Berman asks whether synthetic avatars that do not impersonate a real person should also be identified. Balasubramaniyan favors labeling AI-generated content and distinguishes malicious bots from agents authorized to perform useful tasks for a person or institution.

For agents, he says identity and authorization matter alongside detection: who authorized the system, what actions can it take, and what economic authority has been delegated? He argues that benign and malicious automation occupy a spectrum.

The conversation turns to political misinformation. Balasubramaniyan recounts cases in which synthetic material was treated as real and others in which genuine material was falsely called synthetic. His examples are presented as his account of investigations, not independent verification of each event.

Asked whether a detection provider could become an arbiter of truth, he says companies must demonstrate accuracy repeatedly and be held to a high standard. He does not suggest that one provider’s conclusion should be immune from scrutiny.

For families, he recommends agreeing on a verification protocol before an emergency: be wary of urgent requests for money, use a shared secret word, and hang up and call the person back through a known number when in doubt. His emphasis is to pause before acting under pressure.

The final guest is Anastasis Germanidis, co-founder and CTO of Runway. The hosts ask how video generation has progressed and how it relates to the company’s work on world models.

Germanidis traces Runway’s progress from early, low-resolution experiments to a production video model in 2023 and more recent multi-shot video with audio. He describes Runway as an applied research company that trains multimodal models and deploys them in creative tools.

Users can start with a text prompt, an image or existing video. Germanidis says some outputs are now used directly in distributed work, including film and television, rather than only for rough prototypes. This is his description of customer use, not a measure of how common such use is.

He defines a world model as a system that simulates an environment and the consequences of actions within it. Runway’s GWM1, he says, developed directly from its Gen-4.5 video model. He reports improvement on internal physics-related benchmarks while acknowledging substantial room to improve.

Germanidis groups applications into experiences for people and tools for autonomous systems. He imagines interactive environments, games and conversational characters, alongside simulated data generation and evaluation for robotics or autonomous vehicles.

He says Runway’s navigable world model is a potential foundation for games, but keeping a generated world consistent over long periods is still a bottleneck. His expectation of future progress is a projection, not a guarantee that today’s models can support a complete game.

Asked whether world models are necessary for general intelligence, Germanidis says he considers them important. His reasoning is that many hard problems involve interacting with the physical world; a simulator could offer faster feedback for experiments and reinforcement learning.

The host asks how the team measures progress toward such ambitions. Germanidis points to practical creative applications as well as benchmarks for gravity, fluid dynamics and other physical behavior. He says visual inspection can show qualitative improvement, but quantitative measures also matter.

Berman asks about Hollywood and the concerns of people working in visual effects. Germanidis says Runway works with studios and expects generative tools to be combined with traditional production, for example by translating an actor’s performance into another character.

He predicts that cheaper production will broaden who can make films and increase the number of works, while taste and talent will remain important. His specific forecasts about scale are speculative and come from a company building the tools.

On performers and celebrities, he expects familiar people and reputations to retain value, perhaps with new ways to license a likeness. He also notes the possibility of AI-generated influencers, without claiming to know how the market will settle.

The hosts ask about fan fiction and intellectual property. Germanidis cites Runway’s Lionsgate partnership, which he says included adapting models using the studio’s library. He anticipates negotiated arrangements and guardrails for using characters, while leaving the details unresolved.

Asked when an ordinary person might create a feature-length film with these tools, he predicts the technical capability is close. He distinguishes generating a long film from making a compelling one, which still requires vision, taste and skill. The timeline is his prediction at the time of the interview.

For film students, he says narrative, cinematography, shot selection and character work remain useful. Faster iteration may let a filmmaker test more ideas, even when the final production is shot conventionally.

Could an AI system write and direct an entire high-quality film without a person? Germanidis says contemporary language models still produce generic narratives, though they can help brainstorm possibilities. He expects that limitation to improve, but does not claim it has already been solved.

The hosts close by reflecting on the range of topics: industrial maintenance, satellite imagery, synthetic-media fraud and generative video. They thank the four guests and the production team.