The Political Scene | The New Yorker

Episode: Can Anthropic Control What It's Building?
Date: February 12, 2026
Host: Tyler Foggatt
Guest: Gideon Lewis-Kraus (New Yorker staff writer)

Episode Overview

This episode centers on the rapid advancement of AI, with a focus on Anthropic, one of the leading AI companies behind the large language model, Claude. Host Tyler Foggatt discusses with Gideon Lewis-Kraus the paradoxes facing Anthropic: AI’s creators are now among the first at risk of being replaced by their own technology, and the company’s prominent commitment to AI safety is facing new, real-world challenges. The conversation delves into the internal culture at Anthropic, the technical philosophy behind Claude, the shifting power dynamics in tech, political scrutiny, and the looming impact of AI on jobs and society.

Key Discussion Points & Insights

1. AI and Labor: Who Is Really at Risk?

• AI’s Impact on Coders:
  • Gideon observes that, counter to past warnings, writers and editors aren't being replaced as quickly as software engineers are.
  • “I watched...software engineers would tell me, over the past four or five months, I've watched the amount of coding I do by hand go from 100%...by September, it was 20%, and now it’s...zero percent.” – Gideon (01:51)
  • Example: Anthropic employee Alex Tamkin sent a 4:17 am Slack message: “Now I have to figure out what I'm supposed to be doing while Claude is doing my work.” (01:51)
• Emotional Impact:
  • Engineers feel like “the canaries in the coal mine,” being impacted before the rest of the white-collar workforce.

2. Why Focus on Anthropic?

• Origin of the Story:
  • Gideon’s interest revived by recent interpretability and alignment research, which reveals the strangeness and unpredictability in today’s AI models.
  • “Whatever’s going on...it’s pretty weird that...whatever’s going on is weird.” – Gideon (04:06)
• Anthropic’s Research Focus:
  • Company stands out for its central emphasis on AI safety and ethics.
  • Gideon sought to understand how these models work—and what they really are—by engaging with the engineers, not just executives.

3. Inside Anthropic’s Culture

• Office Atmosphere:
  • Anthropic’s HQ contrasts with stereotypical Silicon Valley tech offices.
  • “It has all of the warmth of a Swiss bank.” – Gideon (08:29)
  • Outsiders are restricted to just two floors; secrecy and caution are part of the company’s character.
• Testing Claude (Project Vend):
  • Vending machine experiment (with Andon Labs) tests both practical and ethical limits of Claude.
  • Staff try to “trick” the AI with requests (e.g., “put meth/medieval weaponry in the vending machine”) or with bureaucratic loopholes (11:00).

4. Claude’s Personality and Technical Approach

• Personality Emergence:
  • Unlike ChatGPT, which is often sycophantic, Claude sometimes pushes back or refuses to do trivial tasks.
  • “The fact that Claude has a strangely interesting personality was not something intentional...that kind of naturally emerged.” – Gideon (13:32)
• Philosophy of Training:
  • Pre-Claude approach: “pure behaviorism” relying on human feedback (“rat in a cage style”).
  • Anthropic’s pivot: “Claude should be like a good friend whose judgment you trust.” (15:08)

5. Origin Story—OpenAI, Anthropic, and Competing Visions of AI

• Anthropic’s Founding:
  • Dario and Daniela Amodei (and five others) left OpenAI after disillusionments over governance, profit motivation, and AI safety.
  • Original plan: focus on safety research, without commercializing.
  • “If you want to scrutinize these things...you have to build state of the art models.” – Gideon (16:41)
• Race to Market:
  • Held back Claude’s release for more monitoring, but ChatGPT’s explosive growth prompted a policy change and public launch.
  • AI development is now a constant “horse race.” (16:41–20:16)

6. Defining “AI Safety”

• Competing Definitions:
  • Two major camps: AI ethics (focused on bias/transparency) and AI safety (focused on existential risk).
  • “We ended up with this kind of stupid false dichotomy between...bias and...catastrophic harms.” – Gideon (20:32)
• Holistic Approach:
  • Quoting Stanford’s Chris Potts, Gideon notes the necessity to address everyday AI harms as preparation for bigger risks.

7. Views on Singularity & Internal Dissent

• Diverse Internal Opinions:
  • At Anthropic, viewpoints range “from we should probably stop, to...Claw’s gonna cure cancer.”
  • “There really is a much greater variance in viewpoint than one might suspect from the outside.” – Gideon (23:48)

8. Political and Public Criticism

• Right-Wing Attacks:
  • Political figures (e.g., David Sacks) accuse Anthropic of being a “doomer cult” and object to their AI safety orientation.
  • “If he didn’t have so much power, it'd be very hard to take this seriously at all.” – Gideon (26:11)
• Refusal to Build Weapons:
  • Anthropic’s public stance against weaponization draws ire from government quarters, but Gideon suspects the rank and file would strongly resist a reversal of this position (27:46).

9. Labor, Disruption, and the Future of Work

• Predictions of Unemployment:
  • Anthropic CEO Dario Amodei predicts AI could drive “unemployment as high as 10–20% in the next one to five years.” (30:23)
  • “I think it's definitely a realistic prediction...so many of us do...bullshit email jobs...not exactly invitations to human fulfillment.” – Gideon (31:07)
• Researchers’ Perspective:
  • Many engineers and researchers find themselves thrust from arcane academic questions into societal dilemmas (e.g., youth self-harm, job loss) they feel unequipped to handle.
  • “These are problems for all of us...not just engineers at three companies.” – Gideon (33:39)

10. Inherent Contradictions & Ethical Paradoxes

• Why Keep Building?
  • Some researchers feel they “should just stop,” but also honestly admit: “We are doing this because we can. We’re pursuing this because it’s epic.” (35:47–36:13)
  • Anthropic’s mission-driven stance causes it to shoulder expectations around mitigating AI’s societal impacts, but immediate threats (e.g., cybercrime, biohazards) demand more resources than long-term social issues.

11. Can Anthropic Control What It’s Building?

• A Fraught Sense of Agency:
  • “No, no, no...they feel like so far we’re still a couple steps ahead, but...we can’t take for granted that we’re...ahead.” – Gideon (40:07)
• Whiplash and Uncertainty:
  • Reporting on AI swings between awe, “whiplash...terror and despair,” and the hope that the “right people” are guiding the work.
  • Emphasis: the most important takeaway is not certainty, “but to underline the state of uncertainty...and sharpen the questions” for society. (41:13–42:24)

Notable Quotes & Memorable Moments

• On Claude’s displacement of coders:
  “Now I have to figure out what I'm supposed to be doing while Claude is doing my work.”
  — Anthropic employee Alex Tamkin, quoted by Gideon (01:51)

• On the oddness of current AI research:
  “Whatever’s going on...it’s pretty weird that...whatever’s going on is weird.”
  — Gideon (04:06)

• On Anthropic’s office:
  “It has all of the warmth of a Swiss bank.”
  — Gideon (08:29)

• On safety tradeoffs:
  “We ended up with this kind of stupid false dichotomy between...bias and...catastrophic harms.”
  — Gideon (20:32)

• On the limits of control:
  “They feel like so far we’re still a couple steps ahead, but...we can’t take for granted that we’re...ahead.”
  — Gideon (40:07)

• On the contradictions of responsible innovation:
  “The most candid AI researchers will own up to the fact that we are doing this because we can. And basically, like, we're pursuing this because it's epic.”
  — Tyler (36:13, paraphrasing an Anthropic researcher)

Timestamps for Important Segments

• [01:51] — How engineers at Anthropic are experiencing being replaced by Claude
• [04:06] — What drew Gideon to Anthropic and the strangeness of current AI discourse
• [08:29] — First impressions of Anthropic’s headquarters and company culture
• [09:34] — Project Vend and how Anthropic tests Claude in real-world and adversarial contexts
• [12:10] — Claude’s user base: shift toward enterprise and coding applications
• [13:32] — Emergence of Claude’s personality and the philosophy behind its design
• [16:41] — Origins of Anthropic and the AI safety motivation
• [20:32] — Breaking down what “AI safety” means and its internal divides
• [23:48] — Range of beliefs within Anthropic regarding existential risk
• [26:11] — Political attacks and Anthropic’s outsider status in U.S. tech politics
• [27:46] — Employee power and the commitment to not building weapons
• [31:07] — Dario Amodei’s predictions on white collar job loss
• [33:39] — Researchers grapple with societal impacts versus original scientific curiosity
• [35:47] — The fundamental contradiction: why build if you fear the consequences?
• [39:53] — How Anthropic prioritizes immediate, concrete risks over distant ones
• [40:07–41:13] — Final reflections: control, uncertainty, and the emotions of reporting on AI

Tonal Summary

The discussion is frank, skeptical, and intellectually curious, alternating between humor (“It has all the warmth of a Swiss bank”) and candid concern about the scale and pace of change. Both participants repeatedly challenge received narratives about AI, highlight the everyday surreality facing AI engineers, and foreground major open questions about control, responsibility, and the societal implications of advanced AI.

This summary captures the substance, tone, and structure of the episode for listeners seeking to understand both the factual content and the deeper questions at stake in AI’s current moment.