The Artificial Intelligence Show

Episode #189: Is Claude AGI? AI Change Management, Nvidia-Groq Deal, Meta Acquires Manus, Yann LeCun Speaks Out & OpenAI Preps AI Device

Date: January 6, 2026
Hosts: Paul Roetzer and Mike Kaput

Main Theme

This episode explores the rapid advancement of AI models—especially Anthropic’s Claude Opus 4.5—and asks if we are now on the threshold of artificial general intelligence (AGI). Paul and Mike break down the extraordinary jump in capabilities, the way this is changing knowledge work, and what it really means for businesses, workers, and the global tech landscape. The discussion weaves in topics like organizational change management, the path of AI adoption, major news from Nvidia, Meta, and OpenAI, and the ongoing social and ethical implications of exponential AI progress.

Key Discussion Points and Insights

1. Is Claude Opus 4.5 AGI?

— Starting [05:41]

• Benchmarking Breakthroughs: Claude Opus 4.5 achieves a “time horizon” of nearly five hours—a measure of the longest continuous expert-level task an AI can complete. This is the highest ever published, exceeding previous benchmarks by a large margin.

• AI Progress Speeding Up: Data from Meter and Epoch AI show that AI capabilities are now doubling every 7 months, with projections that AIs could routinely handle week-long human tasks in 2–4 years.

• Buzz Among AI Insiders: A flurry of insider commentary erupted over the holidays:

  • Igor Babuschkin (xAI co-founder) remarks on December 26, “Opus 4.5 is pretty good.”
  • Andrej Karpathy (OpenAI co-founder) replies, “It’s very good. People who aren’t keeping up even over the last 30 days already have a deprecated worldview on this topic.” [08:07]
  • Jackson Kernion (Anthropic researcher) declares, “Opus 4.5 is as much AGI as I ever hoped for.” [10:40] He later clarifies that Claude code can now write, fix, and refine complex software autonomously, an experience the public hasn’t seen unless they code.

• Anecdotes of Disruption:

  • Google principal engineer J. Dogan openly admits Claude code replicated in one hour what her team had spent a year building:
    “I gave Claude code a description of the problem. It generated what we built last year in an hour.” [15:44]
  • Former Google/DeepMind engineer Ronan Anil (now Anthropic): “If I had agentic coding and particularly Opus, I would have saved myself first six years of work compressed into a few months.” [17:09]

• Broader Implications:

  • Dave Holtz (founder of Midjourney): “It’s crazy. I can sense the limitations, but I know nothing is going to be the same anymore.” [18:58]
  • Elon Musk’s take: “We have entered the singularity.” [19:50]

• Key Quote (Dario Amodei, Anthropic):
  “We are not far from the world…where AI is writing 90% of the code, and then in 12 months, we may be in a world where AI is writing essentially all of the code.” [09:30]

• Hosts’ Reflections:

  • Paul and Mike share personal stories about using large models as “co-CEOs” or strategic partners, noting that 95% of their output on complex documents now comes from AI, with the human in a guiding and verifying role.
  • Paul reflects:
    “I would estimate I spent over break maybe 10 to 15 hours talking to co-CEO in this thread. If I hadn’t had this GPT, we’re talking about 200 plus hours of manual research, note taking, summarization, writing.” [24:30]
  • The takeaway: AI is already transforming high-level business planning and not just coding or research.

2. Organizations and the “People Problem”—Not a Tech Problem

— Starting [31:48]

• Many companies are failing to get significant ROI from AI, not for technical reasons, but due to human factors: fear, resistance, lack of urgency, lack of buy-in, and poor change management.
• Paul’s core point:
  “If your company isn’t generating significant ROI from AI adoption, then you have a people problem.” [31:58]
• Backed up by Jack Soslow’s take:
  “Technology is the easy part. Finding the problem is harder. But the hardest part…the part almost nobody wants to do, is the human work of driving change…” [32:16]
• Mike adds: “Individuals are racing ahead with AI because there are no real barriers in your own life to experimenting with this, except time.” [37:00]
• The hosts warn: The gap between AI-forward professionals and laggards is widening at breakneck speed; organizations that fail to cultivate AI literacy and embrace the tools will become obsolete quickly. [37:30]

3. OpenAI’s “Head of Preparedness”—Are They Worried About Truly Dangerous AI?

— Starting [38:19]

• OpenAI posts a new C-level role to manage risks from frontier AI:
  “You don’t create a role like this at this salary unless you genuinely believe the next models you’re releasing…could cause physical or systemic harm.” — Mike [39:57]
• Mandate: Preparedness for abilities like self-improvement and continual learning, which could make models less controllable and more dangerous.
• Paul: “There was quite a bit of buzz…that Google has made advancements in continual learning that have not been publicly acknowledged yet… tells me the labs are much farther along in both of those areas than we might be aware of.” [41:30]

4. Sal Khan’s “1% Solution” for Worker Retraining

— Starting [42:00]

• Sal Khan (Khan Academy) calls for a “1% of profits” contribution by AI-benefiting companies to fund massive global worker retraining—an annual fund of $10B.
• Paul supports the principle but points out the deeper problem:
  “What jobs are they training for? What are they providing $10 billion to prepare them for when no one seems to know what jobs three, five years from now look like?” [47:10]

5. Jevons Paradox and the Job Debate

— Starting [47:30]

• Aaron Levie (Box CEO) argues that AI's efficiency will create more demand and more work, not less, due to Jevons Paradox.
• Paul disagrees:
  “Demand for knowledge outputs is not infinitely elastic… Just because we can create more doesn’t mean there’s more demand…” [51:20]
• Both hosts warn that, even if a future of abundance is possible, the 3–5 year transition will be “messy” and painful for many.

6. “Vibe Revenue” in AI Startups—Is Product-Market Fit Real?

— Starting [55:20]

• Greg Eisenberg warns of “vibe revenue,” where AI products see short-lived growth due to novelty, leading to high churn and no retention unless true daily value/utility is present.
• Paul reflects: “You can’t build a business and just always keep replacing it with new stuff. You have to build a world class customer experience…” [56:33]

7. Salesforce's Retreat on LLMs

— Starting [57:56]

• Salesforce execs are reportedly emphasizing “deterministic automation” over LLMs for some tasks, due to inconsistencies and customer demand for predictable workflows.
• Paul: “They’ve kind of painted themselves into a corner where they’ve bet everything on AI that isn’t fully ready yet…” [59:47]
• HubSpot is cited as taking a more gradual, additive approach.

8. NVIDIA’s $20B Non-Exclusive Deal with Groq

— Starting [63:25]

• NVIDIA secures Groq’s language processing unit (LPU) IP and key talent to integrate ultra-fast inference chips for real-time AI workloads, without a full acquisition.
• Paul: “Seems like a brilliant move… a prelude to many more acquisitions to come.” [64:54]

9. Meta Acquires Manus for $2B+

— Starting [66:21]

• Manus, a Singapore-based startup known for autonomous AI agents, is acquired to boost Meta’s business AI and catch up to rivals.
• Paul: “They’re going to keep spending billions… but Meta would be at the bottom of my list [for impact in 2026].” [67:43]
• Comes amid internal turmoil and leadership changes.

10. Yann LeCun Speaks Out—Leaves Meta, Criticizes LLMs

— Starting [68:34]

• LeCun leaves to form Advanced Machine Intelligence Labs, focused on “world models” learned from video, not language.
• He describes Meta's LLM push as a “dead end,” reports Llama 4 was a flop, and accuses the company of “fudging” performance benchmarks:
  “My integrity as a scientist cannot allow me to do this.” [73:18]

11. OpenAI’s Upcoming Audio-Centric AI Device

— Starting [74:14]

• OpenAI is unifying its audio model teams ahead of a 2026 device launch. The device may have little/no screen and could be “pen”-shaped.
• Aim: To offer more natural conversation, support for interruptions, and “emotive” responses.
• Paul: The “smart pen” theory could make sense, leveraging founder Sam Altman’s own note-taking preferences. [76:13]

12. Richard Socher’s 2026 Predictions

— Starting [77:41]

• Socher (You.com CEO) sees “reward engineering” jobs for tuning agent incentives, predicts 10-person unicorns, and expects the first hints of superintelligence by year’s end.
• He also notes:
  “Every knowledge worker becomes a manager of AI agents…” [79:15]
• Validates the emerging need for new business models and thought around agent-driven markets.

13. OpenAI’s Industry Prompt Packs

— Starting [80:35]

• OpenAI launches “prompt packs”—carefully curated, plug-and-play prompt templates for different sectors and functions (sales, HR, government, etc.).
• Mike: “Three years after ChatGPT comes out, the vast majority of knowledge workers just simply still don’t know everything that’s possible…” [81:28]
• Paul: These are ideal to help less fluent coworkers “get to value” quickly and encourage adoption. [82:22]

Notable Quotes and Memorable Moments

• Andrej Karpathy [08:07]: “People who aren’t keeping up even over the last 30 days already have a deprecated worldview on this topic.”
• Jackson Kernion (Anthropic) [10:40]: “Opus 4.5 is as much AGI as I ever hoped for.”
• Google engineer J. Dogan [15:44]: “I gave Claude code a description of the problem. It generated what we built last year in an hour.”
• Ronan Anil (Anthropic, ex-DeepMind) [17:09]: “I would have saved myself first six years of work [with Opus].”
• Dave Holtz (Midjourney) [18:58]: “It’s crazy. I can sense the limitations, but I know nothing is going to be the same anymore.”
• Elon Musk [19:50]: “We have entered the singularity.”
• Paul [31:58]: “If your company isn’t generating significant ROI from AI adoption, then you have a people problem.”
• Jack Soslow (Kiraday) [32:16]: “The hardest part…is the human work of driving change, sitting with people, earning trust, refining the product until it fits their hands...”
• Yann LeCun re: Meta [73:18]: “My integrity as a scientist cannot allow me to do this.”

Timestamps for Key Segments

• [05:41] Main topic — Is Claude OPUS 4.5 AGI?
• [10:40] Jackson Kernion’s “AGI as I ever hoped for”
• [15:44] Google engineer: Claude code matches a year’s work in one hour
• [18:58] “Weeks where decades happen” and Elon Musk's "singularity" claim
• [21:30] The tipping point for knowledge work and complex planning
• [31:48] AI ROI is a people, not technology, problem
• [38:19] OpenAI’s “Head of Preparedness” post — Are they bracing for dangerous AI?
• [42:00] Sal Khan’s 1% proposal for worker retraining
• [47:30] Aaron Levie’s Jevons Paradox optimism—Does more AI mean more jobs?
• [55:20] AI “vibe revenue” vs. true product-market fit
• [57:56] Salesforce hedges on LLMs, re-emphasizes deterministic automation
• [63:25] Nvidia’s $20B Groq deal
• [66:21] Meta acquires Manus; is Meta falling behind?
• [68:34] Yann LeCun leaves Meta, blasts LLMs, and details internal dysfunction
• [74:14] OpenAI's audio device — rumors and theories
• [77:41] Richard Socher’s predictions for the age of AI agents
• [80:35] OpenAI prompt packs for business and industry

Tone and Style

The episode is deeply knowledgeable, frank, and personal. Paul and Mike bring an unvarnished, hands-on perspective as operator-learners and industry connectors, while remaining pragmatic and grounded in the human and organizational realities their audience faces. The tone mixes urgency (“the world is changing faster than you realize”), caution (“there will be pain and disruption in the near-term”), and encouragement to experiment and embrace change, coupled with skepticism about optimistic Silicon Valley narratives.

Summary

Episode 189 is a must-listen for anyone navigating AI adoption or anticipating the labor and structural changes ahead. It delivers a compelling and granular view of real-world AGI progress, the new toolkit for business leaders, and the deep “people problem” that may be standing in the way of true transformation. The hosts don’t shy away from hard questions about jobs, company strategy, or industry hype—and their synthesis of leading-edge news with direct organizational experience is invaluable for professionals charting an AI-powered future.