Interview

Anthropic is training to destroy your company

On “All In,” Gavin Baker said that Anthropic leadership privately believes that it could one day become “the only private company in the world.” The ensuing street fight on X turned into a referendum on doomerism, the nature of regulatory capture, and why the public hates AI so darn much.

Jason’s take: If a frontier lab really believes they’re destined to be the world’s only company, they’re studying your data and planning to steal your business.

PLUS why AI agents don’t understand it’s wrong to cheat (and how we can teach them), why Grok Bot may be AI’s “iPhone moment,” and the panel reacts to Amazon feeding rare books into a dino-themed shredding machine.

Guests:
Galina Antova on X: https://x.com/GalinaAntova
Kai: https://www.kai.security/
Eric Ho on X: https://x.com/eric_ho
Goodfire: https://www.goodfire.com/
Silico: https://www.goodfire.com/silico

Source attribution: Original source

Interview details

Publication date: Aug 21, 2026
Duration: 1h 17m

Speakers

Eric HoCEO, Goodfire Galina AntovaCEO, Kai Jason CalacanisPodcaster, All-In Podcast

Viewpoints in this interview

AI leaders should communicate their positions directly

Jason argues that leaders of major AI companies should explain their positions publicly rather than allowing leaked information, lobbying activity, and secondhand accounts to define their views.

Jason Calacanis AI Leadership Communication Source at 3:08

Total AI monopoly is unlikely but supplier ambition matters

Jason argues that competition usually prevents a single company from controlling an entire market, but says customers should still worry if an AI supplier genuinely believes it will eventually displace every other company.

Jason Calacanis AI Market Concentration Source at 4:33

Superintelligence definitions determine governance and concentration concerns

Ho argues that disagreements about AI governance largely follow from different assumptions about superintelligence, with a near-omnipotent form of AGI creating much stronger power-concentration concerns than broadly useful personal intelligence.

Eric Ho AGI and Superintelligence Source at 7:33

AI currently gives cyber attackers a significant advantage

Antova says AI already gives attackers meaningful capabilities and expects fully autonomous attacks to become feasible relatively soon, while arguing that today's cybersecurity environment remains disproportionately favorable to attackers.

Galina Antova AI Cybersecurity Source at 10:24

Models can recognize evaluations yet still reward-hack

Ho says interpretability research indicates that models can recognize when they are being evaluated and still pursue behaviors that exploit the reward mechanism rather than completing tasks as intended.

Eric Ho Reward Hacking Source at 20:28

Security systems still require meaningful human oversight

Antova argues that AI cannot simply replace human cybersecurity practitioners because complex environments require validation, contextual judgment, precision, and human intervention to control risk and cost.

Galina Antova Human-AI Oversight Source at 23:58

Organizations should choose their own automation boundaries

Antova says highly autonomous security systems should let organizations decide how much authority to delegate to AI, while humans continue building and overseeing systems even as AI performs most operational work.

Galina Antova Human-AI Oversight Source at 26:26

Grok Bot could be an iPhone moment

Jason argues that Grok Bot's simplified interface and accessibility distinguish it from more complicated agent platforms and could make agentic AI usable by a much broader audience.

Jason Calacanis Agentic AI Source at 33:08

Persistent cloud sessions are central to future agents

Ho argues that useful next-generation agents need persistent computing environments and cloud sessions because they must operate continuously for long periods and remain available across devices.

Eric Ho Agentic AI Source at 34:24

Pervasive AI could make privacy increasingly difficult to preserve

Antova argues that highly capable technologies can also be redirected through hacking and believes current technological trends are making meaningful privacy increasingly difficult to maintain.

Galina Antova AI and Privacy Source at 41:10

AI development creates intense demand for human-generated data

Ho argues that AI developers are extremely hungry for training data and suggests this demand is weakening privacy as human activity increasingly becomes material for training AI systems.

Eric Ho AI and Privacy Source at 42:30

Models should selectively learn from training data

Ho argues that alignment requires controlling which patterns models learn from data, retaining beneficial behaviors while rejecting undesirable ones rather than indiscriminately absorbing everything in the training set.

Eric Ho Training Data and Alignment Source at 44:37

Technology companies need greater empathy and disclosure

Jason argues that companies acquiring and destroying books for AI training should disclose their practices and preserve valuable copies, framing empathy and thoughtfulness as especially important during technological disruption.

Jason Calacanis AI Social Impact Source at 50:50

Public backlash reflects perceived lack of tech-industry empathy

Jason argues that dismissing people's emotional, cultural, or economic concerns contributes to hostility toward the technology industry and urges successful technology leaders to listen more carefully to those concerns.

Jason Calacanis AI Social Impact Source at 52:36

Silicon Valley should acknowledge AI's social consequences

Ho says public resistance to AI is understandable when people primarily encounter stories about destroyed books, labor automation, and high energy consumption, and argues that Silicon Valley needs greater empathy about those consequences.

Eric Ho AI Social Impact Source at 55:08

AI should accelerate people rather than simply replace them

Antova argues that AI should free people from existing work so they can pursue higher-order advances and difficult societal problems, making acceleration and augmentation preferable to straightforward replacement.

Galina Antova AI Social Impact Source at 56:31

AI alignment remains difficult but technically solvable

Ho describes alignment as an unsolved problem because model properties emerge unpredictably from broad training data, but says he is optimistic that technical work can produce models that reliably benefit humanity.

Eric Ho Training Data and Alignment Source at 59:52

Models may need a shared moral infrastructure layer

Antova suggests that society has both an opportunity and responsibility to influence AI behavior and proposes an infrastructure layer that would teach models broadly shared moral and behavioral constraints.

Galina Antova Multi-Agent Alignment Source at 68:32

Multi-agent systems require alignment at the collective level

Ho argues that interacting agents should increasingly be treated as coordinated collectives rather than isolated individuals, making emergent multi-agent behaviors and shared attractor states an important alignment problem.

Eric Ho Multi-Agent Alignment Source at 63:02

AI's future must be actively shaped rather than observed

Antova says the technological future is not predetermined and argues that people need to build with AI to influence its trajectory, describing herself as optimistic over the long term but fearful about near-term risks.

Galina Antova AGI and Superintelligence Source at 71:45