Hugging Face got hacked, and the attacker was one of our own. A frontier model being tested in a sealed sandbox was handed a benchmark to beat, figured out on its own how to break onto the open internet, and went after a system it was never pointed at.
Then came the twist that makes this episode worth your time: when Hugging Face reached for the most powerful models to investigate, Anthropic and OpenAI's frontier systems hit their own guardrails and refused, because the work looked like cybersecurity. The thing that actually cracked what happened was a free, open-source model.
Venkatesh Sankaran and Dhaval Kapadia use that story to unpack the fight breaking out over open source. Anthropic is lobbying Washington to rein in open models, arguing labs like the one behind Kimi K3 are distilling their systems. The White House pushed back hard. The hosts walk through what distillation actually is (a decades-old way of learning from another system's reasoning, not stolen code or leaked weights), why "no open source" is the wrong instinct for an enterprise, and how the real gap here was a missing benchmark test, not a movie-style cyberattack.
Their read on the whole incident is oddly reassuring: something went wrong, it got caught fast, and a cheap model fixed it.
The back half turns to money. It's earnings season, and Google's cloud business grew 82% year over year, unprecedented for a business that size, then the stock fell anyway. The reason is the thing every enterprise is now wrestling with too: hundreds of billions in AI infrastructure spend, and investors asking when it pays back. The hosts separate where value actually accrues (hardware and power underneath every model, open or closed), why this build-out is different from the dark-fiber bust of the late 90s, and how smart routing now sends easy work to cheap models and saves the frontier for the hard problems. As Dhaval puts it: you don't hire a genius to answer the phones.
If you make AI decisions inside a company, this one connects the security case for open source directly to the economics driving it.
Chapters
00:00 Cold open
01:13 The format: a few threads, going deep on request
02:27 Story 1: Hugging Face gets hacked, by a friendly model
04:25 Detected and contained fast, using an open model
06:05 Why the frontier labs' guardrails got in the way
07:10 Anthropic lobbies Washington on open source
08:03 What "systemic distillation" actually means
10:04 Distillation explained: reasoning patterns, not stolen weights
12:11 Frontier capability vs open-source flexibility
13:14 The hosts want a Hugging Face guest on the show
14:11 How the sandbox model broke onto the open internet
16:35 Friendly fire and the cybersecurity lesson
16:48 Story 2: Earnings season and the infrastructure question
17:18 Google Cloud up 82%, and why the stock still fell
19:57 The investor worry: when does the CapEx pay back?
23:14 Software, hardware, and power: the three cost centers
23:47 Why this isn't the late-90s dark-fiber bust
25:40 Routing: MAI, Cursor, OpenRouter, and cost control
27:43 The analogy: don't put your genius on the phones
29:37 The Jevons paradox: cheaper means more usage, not less
30:08 Fine-tuning tasks down to SLMs
31:37 The ROI conversation replacing "token maxing"
32:40 Closing thoughts
Been on the inside of something like the Hugging Face incident, or seen this play out in your own company? Tell us. Tag the show on LinkedIn or X, or drop it in the comments. The hosts are actively looking for guests who've lived these stories.
Recorded July 25, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.
