What Happens When Every CEO Becomes Omnipresent? | This Week in AI
TL;DR
AI is shifting from unstructured language to enterprise reality in tables — Fundamental CEO Jeremy Frankle argued that 70-80% of enterprise data lives in rows and columns, so his company built a large tabular model, Nexus, to tackle fraud detection, demand forecasting, and ETA prediction where LLMs still underperform classic ML.
OpenAI’s video retreat and Anthropic’s code surge point to a focus war, not just a model war — Synthesia CEO Victor Riparbelli said OpenAI likely learned the hard way that “doing absolutely everything all at once” doesn’t work, while Anthropic’s narrow bet on B2B codegen made Claude Code the tool everyone at a recent Lightspeed founder retreat was talking about.
The real bottleneck in AI infrastructure isn’t compute — it’s moving data — Lightmatter CEO Nick Harris said today’s AI supercomputers spend most of their time on networking, not compute, and described optical interconnects pushing 1.6 terabits over a single fiber and chips with 114 terabits/sec of bandwidth — roughly the bandwidth of internet cables between North America and Europe.
The next interface may be personalized, interactive video — but bandwidth and inference costs have to collapse first — Riparbelli described Synthesia’s move toward real-time video agents that can teach, role-play, and draw diagrams live, while noting that a custom one-hour AI movie would still be wildly uneconomic today at roughly hundreds of dollars in generation costs.
The CEOs all described becoming more ‘omnipresent’ through AI copilots inside Slack, email, and calls — They shared practical workflows like executive change logs, inbox triage for employees on vacation, company-wide CRM agents in Slack, and using Whisper Flow plus foot pedals and Plaud recorders to turn speech and ambient conversations into searchable operational memory.
Nobody agreed on AGI as a definition, but everyone agreed the labor shock is underappreciated — Frankle called AGI a moving goalpost, yet said automating cognition is different from past industrial revolutions; the group pointed to polling where 70% of Americans expect fewer jobs overall, while only 30% think their own job is at risk.
The Breakdown
The panel opens with three CEOs building very different layers of AI
Jason Calacanis framed the show as an All-In-style AI roundtable, then brought on Jeremy Frankle of Fundamental, Victor Riparbelli of Synthesia, and Nick Harris of Lightmatter. The vibe was immediate: this wasn’t a generic AI chat, but three operators talking from inside the stack — data models, video products, and hardware plumbing.
Why tabular data may be the next big AI frontier
Frankle explained that LLMs had their “ChatGPT moment” for unstructured data, but not for structured enterprise data like spreadsheets, SQL tables, CRM records, and ERP systems. His key point: tables don’t behave like language — if you swap the order of columns like weight and heart rate, the answer shouldn’t change — so Fundamental built a large tabular model, not an LLM, to make better predictions for things like fraud, demand forecasting, and Uber ETAs.
Synthesia’s bet: enterprise video first, Hollywood later
Riparbelli said Synthesia started in 2017, long before AI video really worked, and deliberately chose a narrower use case: turning PowerPoint-style business communication into video. He argued that OpenAI shutting down Sora reflects a brutal lesson in focus, while Anthropic’s code-first, B2B-heavy strategy is paying off so well that at a recent Lightspeed retreat, founders were talking almost exclusively about Claude Code.
The infrastructure truth: AI is choking on networking, not raw compute
Harris gave the hardware deep dive and made it surprisingly intuitive. Today’s supercomputers jam GPUs together in ultra-dense, megawatt racks because copper cables can’t reach far; Lightmatter’s optical approach uses glass fiber and wavelengths of light to move absurd bandwidth over distance, letting thousands of GPUs behave more like one giant brain than many tiny ones. His line that smart GPUs “can’t talk” landed because it made the bottleneck feel human.
Real-time interactive video sounds sci-fi — but they treated it like a product roadmap
Riparbelli said the future isn’t just AI-generated video clips, but real-time video interfaces that act more like games or software than broadcasts. He imagined sales training where an avatar role-plays as a customer, draws a technical diagram live, and adapts to you in the moment — then brought it back down to earth by stressing that this only becomes mainstream when inference and bandwidth costs fall dramatically.
The build-vs-buy fight gets personal fast in the AI era
The conversation turned into a real founder debate over whether startups should vibe-code internal tools like CRMs or just pay for Salesforce. Frankle said his team built a lightweight internal CRM in Slack for a 15-person sales org, while Riparbelli argued that rebuilding commodity software is often a bad use of your best engineers; Harris split the difference, saying the answer is now unusually measurable because token spend makes the economics visible in real time.
The “omnipresent CEO” is already here
This was the stickiest section of the episode. Calacanis talked about giving AI root-like access to Slack, Notion, Gmail, and even ex-employees’ history to create searchable personas; Riparbelli described an “executive change log” that surfaces important company decisions every day; Harris said AI lets him feel the “tension on the rope” of the business without missing micro-details. The shared mood was half productivity breakthrough, half eerie sci-fi.
AGI may be a fuzzy term, but the labor disruption feels real
On the big finish, the group largely rejected AGI as a precise benchmark while agreeing that something massive is accelerating. Frankle said this is the first time we’re automating cognition rather than just physical labor, using his accountant friend as an example of someone badly underestimating the blast radius; Riparbelli stayed optimistic that humans will keep inventing new status games and more human-centered work, but warned via Alvin Toffler’s Future Shock that the pace of change itself could trigger social instability before society adapts.