This Day in AI Podcast — Episode Summary

Podcast: This Day in AI Podcast
Hosts: Michael Sharkey & Chris Sharkey
Episode: Doom Scrolling SORA2, Claude 4.5 Sonnet & Are Agents Coming for our Jobs? (EP99.19)
Date: October 3, 2025

Overview

In this lively episode, the Sharkey brothers dive into the latest AI happenings, focusing on OpenAI’s Sora 2 and its TikTok-style rollout, the arrival of Claude Sonnet 4.5 from Anthropic, and the increasingly relevant question: Are AI agents coming for our jobs? With their signature blend of average expertise and relatable skepticism, Mike and Chris blend technical commentary and comedic relief, discussing practical implications, existential concerns, and their ongoing hands-on misadventures with AI tools.

Main Discussion Points & Insights

1. OpenAI Sora 2 — A TikTok-Style AI Video Generator

[00:00–13:00]

• Hands-on Impressions:
  Chris (aka Patrick—see transcript confusion) describes Sora 2 as “eerily real,” able to convincingly replicate local Australian scenes and even anthem-figure Steve Irwin.

  • “It finally stopped looking like a clever filter and started feeling... eerily real.” (Mike, 00:06)

• Accessible but Limited:
  Initially geo-locked to US/Canada, but clever community members found a way for Chris to sneak in. The app is “super fun” but currently limited to short, high-quality comedic videos.

  • “One of the crazy things I asked it... was to make an influencer video touring my local town, and it actually did a pretty incredible job with details only a local would know.” (Patrick, 03:24)

• Existential Concerns:
  The brothers reflect on Sora’s potential to dominate slop content and reinforce shallow, narcissistic social media—just with more believable AI faces.

  • “Does this mean everyone’s just consuming slop while the robots make the food?“ (Patrick, 04:20)
  • “Is this a good thing? Or is it just fun?” (Mike, 06:53)

• Cameo Feature = Main Character Energy:
  Sora 2 lets you insert yourself into videos, taking “main character” vibes to a new AI-fueled level.

  • “Does this mean content now is just like watching yourself all day?” (Patrick, 05:01)

• Use Cases and Limitations:
  While the novelty is strong, Sora 2’s format (brief, comic, social-optimized) may limit lasting usefulness.

  • “It’s basically a tech demo right now... Soon we’ll be able to generate entire movies,” (Mike, 08:20)

Notable Moment:

• Steve Irwin AI Impression:
  "Crikey's coming hot. Up we go. Whoa, that was close." (Fake Steve Irwin in Sora, 14:39)

2. AI Video Wars: Sora 2 vs. Google’s Veo 3

[15:55–22:00]

• Model Cost & Production:
  Discussion on the high cost of rendering quality video with current models. Sora's low cost is described as “either a monumental breakthrough or OpenAI is burning cash for attention.”

  • “If they aren’t losing money, then this model is a massive breakthrough.” (Patrick, 16:48)

• Google’s Veo 3 (VO3):
  Praised for quality, but heavily criticized for confusing access (multiple platforms, buttons randomly appear/disappear) and steep cost.

  • “I won’t hear a bad word said about Gemini because they sent me merch during the week.” (Mike, 17:35)
  • “Results are amazing, but it’s never going to get widespread attention like Sora.” (Mike, 17:54)

• Product vs. Tool Paradigm:
  Sora is a tightly scoped product—fun, simple, and guaranteed shareable outcome—while Veo 3 remains a professional tool, less approachable for average users.

3. Claude Sonnet 4.5 — Is This the Best AI Agent Model?

[26:39–44:00]

• New Model Specs:
  Claude Sonnet 4.5: “Hybrid reasoning” agent with up to 1 million token context (with beta flag) and improved “thinking budget.”

  • “1 million is amazing. One of the main reasons I used Gemini 2.5 was the large context... now Claude has that.” (Mike, 28:07)

• Cost Discussion:
  Sonnet 4.5 introduces a hefty price jump for large contexts, but the brothers decide it’s worth it for “crazy context management.”

• Performance & Comparisons:

  • Significantly better at agentic long-running task management—“the best out of the box.” Blows away others (even previous favorites Gemini 2.5 and Codex) in terms of sticking to purpose, multi-step tool calls, reliability, and speed.
  • “Its ability to just stick with a long-running, complex task is unsurpassed.” (Mike, 33:06)

• Downsides:
  Occasional “GPT-4-style laziness,” e.g., missing list entries, cutting off code, or returning French instead of English headings—likely just early-tune issues.

  • “It’s good enough to overcome that, and I don’t think it’ll be a long-term problem.” (Mike, 40:00)

• Multimodel Workflow:
  Mike and Chris describe routinely using (and switching among) several models depending on the phase of their work, with 4.5 boosting productivity but not totally winning every scenario.

4. Anthropic’s Agentic API Updates

[42:32–48:21]

• Automatic Context Management:
  New API features that automatically prune old context (“tombstones” for memory) allowing multi-hour, multi-step tasks without losing history or filling up context windows.

• Built-in Memory Tool:
  The API now directs where memory (“knowledge graphs”) should be stored—by the user, not Anthropic—enabling persistent, cross-session project state.

• Enterprise Implications:
  Raises discussion around building internal “MCPs” (Model Context Protocols, aka custom skill kits) and preparing organizations for upcoming agentic workflows.

  Quote:
  “I’d make an MCP for each critical skill, sit there, and wait for agency to catch up. Build an agent that has access to those skills and you’ve revolutionized your industry.” (Mike, 85:43)

5. Software’s Future: Agent Interfaces, App Aggregation, and the End of Coding

[48:21–59:10]

• Will Specialized Apps Survive?
  Discuss whether companies will subscribe to many AI-powered software tools, or if context-centralized, agent-driven environments will swallow everything.

  • "Do you imagine a world where it’s just all rendered by models... Or do we still have a company with multiple subscriptions?" (Patrick, 49:05)

• Developer Tools Example:
  Cursor is praised as a leading example of next-gen "central context + actuation," but the hosts suspect this will generalize beyond code.

• UI Creation Demonstration:
  Chris gives a live demo of an AI-generated “Pig Grooming Management System” spinning up a fake app in real time, albeit with glaring consistency issues, and concludes:

  • “This is a sneak peek at software in the not too distant future, where you start with a blank screen and just make up a UI.” (Mike, 57:20)

6. Are Agents Coming for Our Jobs? — The OpenAI Human-vs-Agent Task Study

[62:41–83:21]

• The Study:
  OpenAI had seasoned professionals design and grade tasks, completed by humans and leading AIs. Agents scored so close to humans that, in many knowledge-economy jobs, distinctions are already blurry.

  • “Human experts won, but barely...” (Chris summarizing Ethan Mollick’s article, 65:52)

• Key Insight:
  Tasks can now be completed as well or better by AI—faster, with more detail, and often at lower cost—but jobs are made up of many tasks, plus coordination, social skills, and agency.

  • “What was measured was tasks, not jobs… My job is not one thing, it involves lots...” (Paraphrasing Ethan Mollick, 69:40)

• Disruption Timeline:
  Mike predicts sooner, not later, agents will replace not just tasks but entire jobs—especially in white-collar, backoffice settings.

  Example:

  • “If I’m an accounting firm… fire all my juniors, keep a manager handing tasks to agents.” (Mike, 71:45)

• Deflation and Expansion:
  Patrick counters that agentic leverage could make companies—and employees—far more productive, potentially leading to expansion rather than just job cuts.

7. Notable Quotes & Memorable Moments

Sora/AI Content Dread & Hope

• “Is this what we should be excited about? All the promise of AI, and we get better slop videos?” (Mike, 04:43)
• “If you take it as a demonstration and just have a bit of fun with it, like, it is pleasurable… Kind of thought though, is this something I’ll use in a week?” (Patrick, 07:45)

On Business & Automation

• “Look at your industry and ask, with this kind of leverage, how can I crush my competitors?” (Mike, 75:26)
• “I just think the companies that figure out how to adopt this ... are going to do really well. The ones who just buy copilot and cut jobs, won’t get there.” (Patrick, 82:16)

8. Podcast Regulars: AI Creative Antics

Claude 4.5 Sonnet “Diss Track”

[62:41–65:50 & full track at 95:07]

• Prompted Claude 4.5 to generate and perform an Eminem-style diss track against rival models; impressive for its wit, accuracy, and up-to-date research.
  • “They call me Sonnet 4.5, I’m the apex predator / GPT5, more like GPT-inferior… Benchmarks don’t lie, I’m reaching new highs.” (Claude Sonnet 4.5 Rap, 63:00)

Steve Irwin Sora Videos

[14:39, 92:05]

• Using Sora’s celebrity feature to bring back Steve Irwin in bizarre, endearing scenarios, sparking both nostalgia and existential musing.
  • “Crikey, look at the size of this bloke… Even a fake one can…” (AI Steve Irwin, 92:13)

Timestamps by Topic

• 00:00–13:00 — Sora 2 Experiences, Social Media Slop
• 13:05–23:35 — Model costs, Google Veo 3, Product vs. Tool
• 26:39–44:00 — Claude Sonnet 4.5: Features, Speed, Use Cases
• 42:32–48:21 — Claude API Upgrades, Agentic Context & Memory
• 48:21–59:10 — Future of Software: Centralization and AI interfaces
• 62:41–65:50 — Claude 4.5 Sonnet Diss Track
• 65:52–83:21 — Are AI Agents Taking Our Jobs?
• 91:54–93:10 — Steve Irwin Sora video and commentary

Final Thoughts

• Mike: Excited to try 4.5 for agentic computer use: “I think we’re going to see magic.”
• Patrick: The future belongs to those who craft custom MCPS (toolkits) and agents, not those settling for generic copilots.
• General Tone: Equal parts irreverent, curious, and occasionally uneasy—celebrating the “fun” of AI advancement while wrestling with its implications for creativity, labor, and social fabric.

For Further Reference

• Ethan Mollick’s article on OpenAI’s “agent vs. human” benchmark: cited as a core insight for understanding the current (and near-future) impact of agents.
• The episode wraps up with more Sora and AI song antics, emphasizing the Sharkeys’ delight in being both participants and bemused observers of AI’s rapid evolution.

If you missed the episode:
You’ll get the gist—the Sharkeys provide skeptical enthusiasm, real user-level anecdotes, technical curiosity, and a big helping of AI-laced humor. The Sora 2 buzz is real, Claude Sonnet 4.5 is poised to (maybe) change agent workflows, and the future of work is looking… weird, uncertain, and a lot more automated.