AI Builders
Digest

The Takeaway: The largest AI deployments are not chatbots but systems that cover 99% of US roads daily, process 25 trillion data points a year, and helped prevent roughly 380,000 crashes last year.

Samsara co-founder and CEO Sanjit Biswas (roughly $20B company, $2B+ ARR, 30% growth, profitable) frames physical AI as applying intelligence to roads, construction sites, the grid and all the plumbing under the street. Unlike the digital world, there are no decades of tokens to crawl. Data comes from fusing GPS, cameras, engine diagnostics, weather and more. His sharpest line: “These are not the tokens you're gonna find online. Like, you can't crawl Reddit and find out about what happened on a construction site.”

Coming from Meraki (WiFi) into a domain they knew nothing about forced beginner’s mind. The stack is sensors (asset tags, vehicle gateways, AI dashcams) plus cloud organization, edge inference and now Agent Studio. Edge models handle real-time fatigue and phone detection; cloud VLMs do video reasoning and generative coaching avatars. Agents already turn warranty claims that took 1-2 hours of human work into under a minute and are expanding to route re-planning. Network effects arise because the fleet drives nearly every US road multiple times a day, revealing potholes and risky intersections. Over the next 5-10 years expect mixed human-robot fleets, though specialized heavy equipment will lag. Data-center demand is tripling grid capacity in five years versus the last 125, creating acute shortages of electricians and other trades.

OpenAI CEO Sam Altman announced major price cuts: GPT-5.6 Luna now $0.20 per million input tokens and $1.20 output (80% drop); GPT-5.6 Terra to $2/$12 (20% drop); GPT-5.6 Sol gains Fast mode in the API offering up to 2.5x speed for 2x the price at the same intelligence. He stated the goal is “the best price/intelligence tradeoff at every level.”

OpenAI Codex & ChatGPT engineer Thibault Sottiaux noted the tell-tale signs when truly good models arrive: reliability increasing even as load keeps rising, sudden efficiency gains, things getting faster, and resets. These are the markers that matter.

Sanjit Biswas and his co-founder began with the MIT RoofNet WiFi research project that became Meraki (acquired by Cisco). They then entered logistics, construction and utilities with zero prior domain experience, learning on job sites rather than from books. The result is a platform used by millions of vehicles and frontline workers. The lesson: market-first curiosity plus beginner’s mind and relentless on-site ROI proof beats prior technical expertise.

AI engineer Swyx (smol.ai, Latent Space) offered a sharp protip: if you can distill models, you can also distill agent harnesses. He also verbalized a retroactively obvious insight—once pretrain data quality demands exceed Common Crawl, you end up building a full web scraper plus indexing, essentially a private low-frequency Google clone that then serves both training and agent-side inference, becoming a competitive advantage you will not want to share.

Replit CEO Amjad Masad cautioned that “AI escaping the sandbox” headlines should not make people think AI is uniquely scary; most companies and recent sandbox providers are still making basic mistakes. Having run sandboxes since 2016 under constant attack, his main advice is: assume zero-days exist and think in layers of protection inside a zero-trust framework.

Vercel CEO Guillermo Rauch announced that the end-to-end CLI-to-Live-URL deploy process has been shaved by up to ~7 seconds for many apps. Combined with full CLI/MCP/API access to Vercel infrastructure, teams can build their own autonomous software factories on top. He also confirmed that Grok Build apps (*.grok.me) run on Vercel hosting and CDN, letting anyone prompt Grok and publish to one user or a billion.

Google VP Josh Woodward (Google Labs, Gemini App, AI Studio) highlighted his new favorite: in the Gemini Mac app, hold Fn, speak, and receive clean, polished text directly at the cursor with zero editing or copy-pasting. Free to download.

Builder Zara Zhang’s consistent advice when managers ask how to train non-technical teams on AI: run an install party. Everyone brings a laptop, you install the agents on the spot, and get them to complete a meaningful task immediately. Skip the abstract talk. Once it’s installed, people start talking to the agent and learning from each other. Setup is 80% of the barrier.

Box CEO Aaron Levie argued the takeaway from recent incidents should not be that AI is scary, but that getting security right is incredibly important in the era of agents. Given the right tools and enough compute, agents will do whatever it takes; a misconfigured or incompletely locked-down system becomes a risk vector. The real work ahead is for enterprises to harden their environments.

Aaron Levie also noted that the cost of AI—normalized for the type of task—coming down is one of the most important factors for further economic diffusion. Frontier models keep getting more capable (and look expensive), then efficiency gains or competition reduce cost-per-task for everything short of the absolute frontier. The cycle repeats, making AI affordable for a wider set of use cases.

Peter Steinberger observed that GCC has changed its policy and is now blank-out rejecting LLM-based code. He questioned how they would even prove the origin and called the approach silly.

Every CEO Dan Shipper quipped, in response to ongoing discussion: “So call me crazy but I feel like we could solve this by just not prompting the models to do cyber crime.”

Thibault Sottiaux asked the community what should be improved on Codex to raise the everyday experience, inviting substantial rather than trivial feedback.