Open source is going to win it all: Harvey proves it | E2328

Harvey’s first in-house model is an early example of the open-weight strategy Jason believes could reshape the AI race. The company says it has spent six months pursuing frontier-level legal intelligence with open-weight models and systems that let law firms develop specialized models. Harvey Tenant is based on Kimi K3 and was post-trained with Fireworks Research for long-horizon legal work; Harvey reports promising early performance and cost efficiency. Jason argues that as models approached parity quickly in 2023, models could become commodities and value shift toward data. In Harvey’s approach, firms can train separate models on their own legal data without sharing that intelligence with competing firms.

Jason sees a broader incentive for application companies with valuable expert data to build specialized models instead of handing their data to OpenAI or Anthropic. He argues that frontier providers can learn from startups’ token usage and successful applications, while free-token incentives may also reveal usage patterns and business ideas. Harvey’s history illustrates the tension: OpenAI led its $5 million seed round and gave its founders early access to GPT-4, but the discussion’s estimate that Harvey may have spent $10 million a month with OpenAI was explicitly speculative. Another speaker said 99% of Harvey’s spend still comes from frontier models. The speakers also questioned whether businesses will trust a provider that might compete to build the “final company,” particularly when customers’ sensitive information is involved.

OpenAI’s potential public listing adds pressure to explain its business economics. It was reportedly told to expect a listing by 2027 or sooner; CFO Sarah Friar reportedly said it could happen next year, perhaps even this year, while describing an IPO as a milestone and another fundraise rather than an endpoint. OpenAI confidentially filed its prospectus in June and pledged not to retain data from businesses using its models. Aliyah Howe said retention matters especially when companies hold sensitive information entrusted to them by their own customers. The speakers said they lack churn data to judge who is using AI less and why, but see cost to serve and profitability as unresolved questions that will draw scrutiny.

Jason compared AI growth claims with ride-hailing subsidies: one quarter included one billion rides and $2 billion in losses, or about $2 per ride, alongside a $500 incentive for drivers completing 100 rides in a week. He argued that AI users can switch more easily among models, cloud-provisioned assistants, and open-source tools than ride-hailing customers could switch after subsidies helped establish habits. That flexibility could make it harder for AI companies to preserve margins. One speaker predicted that enterprises will eventually run the overwhelming majority of corporate tokens in-house, moving 99% of spend away from frontier models; remaining frontier use might include distillation, potentially costing $1 million a month or $5 million a year. The speaker attributed the view that Anthropic might see itself as the last private company to Gavin Baker, while making clear the idea was not confirmed.

The discussion also focused on what makes workplace automation useful: systems must meet goals without breaking or demanding constant intervention, with people involved where systems cannot safely handle a task. One speaker described using Grockbot, a cloud-running tool that can access X after login, to gather news, competitor intelligence, potential guests, and event-staff leads for recurring Slack or Google Sheets reports. They estimated an automated research system could save five to ten hours weekly. The same speaker described an SDR agent that finds potential podcast and event sponsors and checks them against the CRM, and planned to have sales staff review its work despite resistance. They said leaders need to help people with strong taste build systems thinking, and people with systems expertise develop taste, while acknowledging they had no solution yet for the latter.

Alan Guo said Willow makes voice dictation free and earns money from Scribe, which drafts replies using a user’s style, personal context, and enterprise knowledge base. It can draw on teams’ source-of-truth documents and learn user-specific facts over time. Guo predicted Apple dictation could reach 70–80% of specialist tools’ quality within two to three years, potentially making dictation increasingly commoditized.

Stanford computer science undergraduate Ethan Goodheart’s WIND project retrofits a stock golf cart with a steering-column harness and six cameras, without LiDAR or changes to its internal electronics. An NVIDIA Jetson Thor runs the cart’s software, and Starlink supports cloud-inference experiments and sending code while driving. The first prototype worked within two to three weeks, but the team built its own driving software because available open-source models were too large or poorly suited to the golf-cart setting; Goodheart said some leading models could not run even on a $4,000 GPU. WIND’s business future is uncertain, though golf and transportation companies, including Disney, have shown interest and a retrofit kit might cost under a couple thousand dollars. Goodheart also said AI lets students move quickly through less-interesting coursework and devote more effort to subjects they want to learn.