Forward Future Live: AI Inference, Personalized Audio and Building Software with Lovable

Matthew Berman and co-host Nick Wentz open their October 31, 2025 livestream with interviews about inference, voice interfaces and software creation. Their guests are Lin Qiao of Fireworks AI, Raiza Martin of Huxe and Anton Osika of Lovable. The show’s news and earnings discussions reflect what was known or being debated at the time.

Qiao describes Fireworks AI as an inference cloud that tailors both serving infrastructure and models to an application. Her vision is for models to adapt to changing use, rather than treating a general-purpose model as a fixed appliance. This is the company’s product direction, not a claim that every deployed model already learns automatically.

Asked whether one large general model or many specialized models will dominate, Qiao points to information inside applications: user preferences, interactions and workflows. She estimates that a substantial share of the world’s data is held within applications, though her percentages are illustrative rather than measured in the interview.

Qiao says businesses have long used product analytics, but turning application data into improvements for generative AI remains difficult. Fireworks aims to supply tools that help developers adapt a model without having to build a large specialist team or manage all the infrastructure themselves.

Qiao says Fireworks raised $250 million at a $4 billion valuation. She credits the team and customers, including AI-native startups building new kinds of experiences. The hosts ask about celebrating the funding, and Qiao jokes that the team plans to upgrade its coffee machine.

Berman asks how continuous adaptation works in practice. Qiao begins with product traces and feedback: the team must be able to tell whether an output is useful before it can use that signal to guide further training. Her comparison with human learning is an analogy for reward and feedback, not evidence that models learn like people.

Qiao distinguishes a released general base model from subsequent specialization for a domain. She says a model can explore possible responses and receive positive or negative feedback, allowing developers to train it toward a specific task without repeating the original pre-training process.

A reward can be implemented as a simple program that evaluates the result, Qiao says. In some tasks it is easier to check whether an answer is good than to produce the best answer directly. More complex evaluations can incorporate tools, databases, business logic and production-like environments.

Qiao describes testing different model versions and their interactions with surrounding systems. Training quality may rise and then degrade, so teams need to monitor the results and select a suitable checkpoint before trying it in production. That process entails evaluation and selection, not a guarantee of uninterrupted improvement.

Wentz asks whether Fireworks will expose these tools or operate them for customers. Qiao says experienced developers want low-level controls, while a broader audience needs a simpler interface. Her proposed design layers accessible abstractions over composable APIs, with customer feedback shaping the product.

Qiao announces that a limited preview of the company’s reinforcement fine-tuning product is planned for the following week. She invites feedback about how customers use it and what works. The timing is a statement made during the 2025 livestream.

The hosts ask about Fireworks’ AMD partnership. Qiao traces the company’s hardware approach to her team’s work at Meta and PyTorch: developers should be able to access the tools and hardware that best fit their applications without doing all of the integration themselves.

She says rapid hardware and model releases create an operational challenge. Fireworks had onboarded Nvidia H200 and B200 systems and was working on newer hardware; she also mentions AMD and specialized chips. Supporting more choices means keeping up with different architectures and quickly changing workloads.

Berman asks whether Fireworks will focus on open models or also serve proprietary frontier models. Qiao says she would like to serve the best models regardless of origin, while acknowledging her own preference for open source from her PyTorch background.

Qiao argues that model leadership changes rapidly and sees the movement of different providers up and down leaderboards as a sign of convergence. She attributes this in part to overlapping public data, labeling resources, researcher mobility and similar architectures. That is her interpretation of the market, not proof that all models are equivalent.

For Qiao, the distinction moves to data that application providers possess but base-model developers may not. She sees an opportunity to adapt models using that data and to optimize quality, speed and cost together, rather than assuming improvement in one must always sacrifice the others.

Asked how Fireworks differentiates itself in a crowded infrastructure market, Qiao points to product and model co-design. She says successful application companies look beyond simply wrapping an API: they study where the general model and their own product diverge, then adapt the model as the product changes.

Qiao says this process is currently complex, slow and resource-intensive. Fireworks wants to make it easier for more developers to control an application-specific model and improve it as usage supplies feedback. The hosts thank her and note that an opposing view favors a single general model.

The hosts turn to Anthropic’s October 2025 research on signs of introspection in language models. They describe the researchers’ finding as limited and discuss whether a model can report aspects of its own internal state. Berman explicitly frames the implications for consciousness as an open question.

Berman wonders how many human-like behaviors would change people’s judgments about AI. Wentz cautions that people tend to anthropomorphize systems and that a human-centered lens can also distort how animal intelligence is assessed. Their discussion is philosophical speculation, not evidence that a model is alive.

Wentz summarizes the experiment as introducing controlled changes in model activations and testing whether the model notices them. He cites roughly 20% success in the discussion, while Berman says he has only scanned the research. The result concerns a particular experimental setup and is not a general measure of consciousness.

The hosts discuss 3I/ATLAS, the third known interstellar object observed passing through the solar system. They reject speculation that it is an alien spacecraft or a threat to Earth. NASA describes it as an interstellar comet and says it posed no danger to Earth; claims about its exact age remain uncertain.

Berman raises Nvidia CEO Jensen Huang’s view that AI is not in a bubble and contrasts current demand for GPUs with unused fiber during the dot-com buildout. Wentz notes Huang has a financial interest in the industry and argues that today’s large incumbents have stronger established businesses than many dot-com companies.

The hosts remain concerned about concentration among a handful of large technology companies. Their observations about GPU utilization and comparisons to past bubbles are interpretations, not a comprehensive valuation analysis or a guarantee that AI investments will pay off.

Berman and Wentz introduce Raiza Martin, co-founder and CEO of Huxe and formerly a product leader on NotebookLM. Martin describes Huxe as a mobile app offering personalized, proactive audio based on information such as a user’s interests, calendar and email.

Martin contrasts that idea with a turn-taking chatbot: instead of waiting for the user to ask, a service might prepare material it expects the person to find useful. She compares the experience to turning on the radio in the morning, except the station is tailored to the listener.

Berman says voice could make otherwise inaccessible time useful, such as while driving or walking. Martin adds a different motivation: people want to stay informed and productive without spending the entire day looking at screens. She hopes audio can also make room for more thoughtful long-form content.

On competition, Martin identifies three elements of her product strategy: audio as a modality, personalization and proactivity. She thinks people often assign each app a particular job, so Huxe should focus on doing one valuable job well rather than responding to every move made by a larger competitor.

The hosts discuss the different roles familiar apps play in daily life, from checking on family to browsing a marketplace. Martin uses these habits to explain why a product can earn repeat use by fitting a specific routine instead of trying to replace every other service.

Asked about NotebookLM, Martin says she joined Google Labs in 2022, before ChatGPT’s launch. Josh Woodward asked her to help create a new applied AI business. She had returned to college and wanted a way to interact with textbooks and other lengthy reading material.

Martin recalls that the idea of talking to a document was not initially obvious to colleagues. The team had to develop the product without an established template. She says ChatGPT helped people understand the category, and that NotebookLM’s audio overviews produced a major growth spike in 2024.

The hosts ask about future hardware for voice interfaces. Martin points to earbuds as an existing, socially familiar device people already wear for music and calls. She is interested in what changes when the audio stream becomes personalized, though the effects on behavior and data remain open questions.

Berman imagines earbuds with additional sensors, while Martin contrasts their social acceptance with the discomfort some people feel around camera-equipped glasses. These are ideas about possible interfaces, not announced capabilities of Huxe.

Martin says she regularly interviews Huxe power users and has heard a recurring pattern: they play a briefing while brushing their teeth in the morning. Some return at lunch or the end of the day. These anecdotes describe users she spoke with, not a representative usage study.

Asked about mixing professional and personal information, Martin says some users connect both types of accounts. Huxe experiments with a productivity-focused daily briefing followed by a feed of personal interests. She describes the design challenge as deciding which information is useful at each moment.

Looking ahead six to twelve months from the interview, Martin hopes to explore passive, generative and personalized content that can lead naturally to follow-up questions. She describes these as product questions to investigate, rather than a fixed roadmap or a proven outcome.

Martin says she sees different patterns among users: some want a customizable news stream, while others follow a learning rabbit hole and prefer listening to reading. Her examples suggest possible user groups but do not establish population-wide demographic behavior.

After Martin leaves, Berman says he has tried Huxe and likes its proactive direction. The hosts compare this idea with ChatGPT Pulse. When asked about hallucinations, Berman reflects that his own experience may have made him complacent; both stress the need to check sources and citations.

The hosts discuss OpenAI’s 2025 restructuring and a reported possible IPO at a $1 trillion valuation. They explicitly identify the IPO as a Reuters report and say it is not confirmed. Their comments about what an eventual investment might return are personal speculation, not established financial outcomes.

A displayed chart estimates ownership after the restructuring, including Microsoft, the nonprofit and an employee vehicle. The hosts react to the reported percentages and speculate about motives. The chart represents approximate figures reported at the time and should not be read as current ownership.

Anton Osika, Lovable’s co-founder and CEO, explains the product as a way to turn an idea into a working app by describing it to AI. Some requests require further iteration and clarification; users may build without directly inspecting code, a practice commonly called vibe coding.

Osika traces Lovable back to GPT Engineer, an open-source tool he built after GPT-4 made more capable intelligence available through an API. A demonstration that generated files for a snake game attracted attention and suggested to him that more people wanted help turning ideas into software.

He says he and co-founder Fabian Hedin developed that idea through a prerelease application before launching Lovable. Osika describes the company’s rapid growth to $100 million in annual recurring revenue; the speed comparison he makes is a company claim, not independently established in the interview.

Osika attributes the launch to three areas of work: improving the agent chain to avoid common errors, supporting backend functions such as authentication and data storage, and making the interface approachable. He says the team kept focusing on product improvements as usage grew.

He recounts meeting entrepreneurs who built businesses using the product, including one customer who made $3 million. When asked for context, Osika clarifies that this was gross revenue over 48 hours, not annual recurring revenue, and that the founder already had a Brazilian education business, audience and distribution channels.

The hosts ask about claims that Lovable was dying. Osika disputes a search-traffic chart and says its paid-customer data showed growth. That response is the founder’s account; neither chart alone independently establishes the company’s overall financial health.

Osika says freelance developers, designers and product managers use Lovable for tasks ranging from prototypes to internal tools. He cites usage at major companies as examples. The users and applications differ substantially, so a prototype should not automatically be assumed to be a production-ready system.

For founders, building software is only part of the work. Osika emphasizes distribution through marketing or sales, talking to customers, identifying friction and improving the product. He also says creating something can be rewarding even when its maker does not turn it into a business.

Wentz describes a build-first culture in which a working prototype can communicate an idea better than a static specification. Osika repeats the phrase “demo, don’t memo”: for a new AI interaction, actually using a demonstration can reveal qualities that are difficult to convey in a document.

Berman asks about building an AI company in Europe. Osika points to European AI firms and argues that strong technical talent and a long-term team culture can be advantages. He says he has not personally felt regulatory friction; that is his experience, not a general finding about all European startups.

Asked whether Lovable uses a 996 work schedule, Osika says it does not. He values people who care deeply about the customer and the product more than attendance for its own sake, and says he sometimes has to tell highly committed employees to go home.

For people using AI to build software, Osika recommends determination and curiosity. When something becomes confusing, he suggests asking questions, learning from mistakes and rebuilding if needed. His optimism about AI-assisted builders’ productivity is a personal view rather than a guarantee for every project.

For a business, he recommends staying open to ideas while remaining determined to find product-market fit. Identify who could pay, test the simplest credible version of the hypothesis, and talk to potential customers about what they would need before buying.

Berman observes that much of this advice applies to any startup: founders need persistence, customers and a way to sell. Osika agrees that AI makes some aspects easier while increasing competition, because other founders can use the same tools.

Asked about rivals such as Replit and Bolt, Osika emphasizes culture, talent and continuous product improvement. He believes customer trust and brand can compound over time. He also mentions a Shopify integration that combines a Lovable frontend with a store backend.

The hosts ask about growth loops. Osika points to community projects, portfolios and ambassadors who help other builders succeed. Sharing projects can attract new users, while helpful community members can improve the experience for existing ones.

For the next twelve to eighteen months, Osika says he cannot predict the pace of AI progress. He discusses integrations with other business software and the possibility of building mobile apps and agents, presenting these as directions rather than firm release commitments.

After Osika leaves, the hosts return to the AI bubble debate. They cite an investor’s comparison between unused fiber during the dot-com era and high current demand for GPUs. Current utilization is a useful data point, but it alone cannot settle whether future investment returns justify valuations.

They discuss Alphabet’s third-quarter 2025 earnings and its first quarter above $100 billion in revenue. The company’s displayed figures include 650 million monthly active users for the Gemini app, 34% year-over-year Google Cloud revenue growth and more than 300 million paid subscriptions across Google One and YouTube.

The hosts consider whether Google can turn its advertising strength and distribution into a durable AI business. Berman compares Gemini’s monthly active users with a reported weekly ChatGPT figure, but those periods differ and should not be treated as a direct like-for-like ranking.

Wentz raises the question of advertising revenue per user and whether AI products can preserve the economics of search. They discuss subscriptions, the cost of paid access and the possibility of more relevant advertising. Their ideas about future monetization are speculation.

They turn to Meta’s earnings, discussing revenue growth, a large tax charge and the market’s reaction. The hosts express an opinion that the accounting charge should be understood separately from recurring operating performance; their prediction about subsequent share movement is not a fact.

Finally, they discuss Amazon’s earnings and layoffs. The hosts cite AWS revenue of roughly $33 billion and 20% year-over-year growth, then consider how cloud infrastructure, Amazon’s own chips and third-party hardware fit into its AI business.

Wentz notes AWS’s contribution to Amazon’s operating income, and the hosts contrast cloud-service economics with the lower margins of physical retail and delivery. They also recognize that future competition could change cloud margins.

The final story is a lighthearted discussion of a meeting between the leaders of Nvidia, Samsung and Hyundai. The hosts joke about the drink in a photograph, thank the three guests and close the show.