Matthew Berman

Matthew Berman is the founder of Forward Future, an AI media company built around his YouTube channel, The Future Today newsletter, and a weekly live show. He previously founded Sonar Technologies, an enterprise messaging business acquired by Marchex. His AI coverage combines hands-on model testing, research explanation, and interviews with industry leaders, with particular attention to the instructions and interventions needed to make a model useful in everyday work.

From a customer’s text messages to Sonar

Matthew moved to San Francisco in 2009 to pursue technology and entrepreneurship, then spent nearly two years at 500 Startups. Sonar Technologies began with a problem described by Taylor Wang at Mayvenn: customers were texting the business, but its team lacked a good way to answer at scale. Matthew built a prototype over six weeks of nights and weekends. Mayvenn used it immediately; after finding more customers, he left 500 Startups to pursue Sonar. Christine Tsai, his boss there, became an investor.

Sonar grew into an enterprise messaging platform serving customers including Zillow, Bird, and DriveTime. It enabled businesses to manage personalized text conversations across industries such as real estate, lending, travel, and automotive sales. Marchex acquired Sonar on December 17, 2019, for $12.5 million in cash and stock, with up to another $1.5 million in stock contingent on sales performance. The acquisition added messaging capabilities to Marchex’s voice analytics, supporting its effort to connect customer intent across both channels.

Matthew also wrote publicly about mistakes in building the team. In his 2016 hiring retrospective, he described ignoring reference-check warnings when the team liked a candidate and relying too heavily on gut feelings during interviews. The team introduced a shared scoring rubric to focus its hiring discussions. He also abandoned the assumption that early employees would accept sharply reduced salaries in exchange for equity and enthusiasm for the company.

Matthew continued leading Sonar after the acquisition. In his May 2023 farewell, he reflected on nine years of founding and running the business, naming customer value and the team as his proudest achievements. He had not yet settled on his next venture, but committed to going deeper into artificial intelligence. Sonar had used AI for years, and the rapid development of large language models had renewed his interest in the field.

Testing AI against his own production work

At Forward Future, Matthew works alongside co-founder Nick Wentz on tutorials, news analysis, interviews, and practical guides. His May 2025 interview with Satya Nadella addressed agents and the changing software stack. Their jointly credited June 2025 Thomas Dohmke feature explored coding agents and software development.

Matthew’s September 2026 Astra review shows how he tests those capabilities himself. Alongside miniature 3D worlds and a city-building game, he assigned tasks from his weekly production routine: researching videos, developing arguments, and preparing presentation decks with speaker notes. An initially repetitive data-center deck needed a request for more varied layouts. Further revisions addressed the requirements of recording a video, including source screenshots, reveal order, and space for his face on screen.

His enthusiastic assessment retained the friction involved. The model often fulfilled a request narrowly when he wanted it to go further; more explicit requirements and longer goals produced more ambitious results. He also described unfinished projects and interventions, including fixes to game controls and stopping a render that was slowing his Mac. These project observations show what he obtained and how he directed the system, rather than establishing a controlled benchmark result.

Setup, specialization, and model behavior

Matthew’s practical positions concern both what models can do and how much work users must do to benefit from them:

• Point-and-click local AI: In the local-AI discussion explicitly directed to Matthew, he argued that setup should approach the simplicity of opening Cursor or installing Codex. Users need to know which tasks their models and hardware can support, but software should perform that matching instead of requiring them to master configuration documentation.
• Specialization has a cost: His stated uncertainty about fine-tuning concerns whether a smaller, specialized model justifies its preparation and maintenance when a strong general-purpose model already performs well with appropriate context. The attraction is more efficient execution; the test is whether it solves the customer’s task within budget and whether the customization can be managed without becoming another daily responsibility.
• Research explanation: Matthew introduced Anthropic’s alignment-faking research in December 2024 and its language-model interpretability research in April 2025. His contribution is public explanation and attention to model behavior; the underlying research belongs to its authors.

Forward Future also publishes Tools & Experiments that let readers try structured agent workflows. Relay routes a coding task through planning, execution, and independent review using the models available in the user’s environment. Loop Library and Loop Skill help users find, adapt, and design repeatable workflows. These are company projects, alongside Matthew’s individually credited model reviews.