This Day in AI Podcast — EP99.25-GEMINI

Is Gemini 3 Really the Best Model? & Fun with Nano Banana Pro
Hosts: Michael Sharkey & Chris Sharkey | Nov 21, 2025

Episode Overview

This week, Michael and Chris dive into the whirlwind of new AI releases, centered around Google’s Gemini 3 language model and Nano Banana Pro image model. The brothers provide hands-on impressions, real-world tests, and their signature “perfectly mediocre” hot takes on whether Gemini 3 truly beats the competition. They explore breakout capabilities, persistent flaws, quirky AI song output (including a Gemini 3 diss track), and discuss broader impacts—especially as creative AI tools begin to eclipse established SaaS products.

Key Topics & Discussion Points

1. The Flood of AI Releases

• Big Drops This Week:
  • Google’s Gemini 3 (LLM) & Nano Banana Pro (image model)
  • XAI’s Grok 4.1 (2 million context tokens, extreme tool-calling)
  • OpenAI’s GPT 5.1 “Codex Max” & GPT 5.1 Pro
• The Mood: A chaotic week leaving everyone feeling “caught in the crossfire of models.”

2. Initial Impressions of Gemini 3

[00:30–03:06]

• Benchmarks vs. Real Use:
  • Gemini 3 regains 1 million context tokens. Token output up to 65K.
  • January 2025 knowledge cutoff raised eyebrows about recency and tuning.
  • “It’s by far the best on benchmarks, but my actual experience is more nuanced.” — Chris [03:06]
• Comparison with Gemini 2.5 Pro:
  • Many users now praise Gemini, but “they probably never gave 2.5 Pro a shot.” — Chris [03:58]
  • Both 2.5 Pro and 3 have similar strengths and persistent weaknesses.

3. Strengths & Weaknesses of Gemini 3

• Strengths:
  • Coding/Design (“Vibe Code”): “Where it is so far ahead of the competition, it’s not even close.” — Chris [06:15]
  • Image Prompts: Excellent at generating image prompts for diffusion models. [07:08]
  • Instruction Following: Fast, excellent at large context, detail following in code, precise document editing.
• Weaknesses:
  • Path Obsession: Gets stuck reiterating the same (sometimes incorrect) solution. [05:00, 29:44]
  • Repetition/Recency Bias: Overly references recent topics or jokes; can make code “chatty.” [05:11]
  • Creativity Drop: “Gemini 3… feels really bland, like really sterile… 2.5 Pro was arguably a more creative model.” — Chris [06:00]
• Fine-tuning Needed: Suggestion that models should be offered in “creative,” “code,” or “research” variants, echoing OpenAI’s specialization models. [10:18]

4. Tool Use & Agentic Capabilities (“Tool Calling”)

[12:42–18:49]

• Improved, but Not Best-in-Class:
  • Tool calling and parallel tool use are “better than Gemini 2.5, but not groundbreaking.”
  • “GROK just did such a better job at doing multi-tool calls… Gemini just wasn’t as detailed or didn’t try as hard.” — Co-host [13:27]
  • Still lags behind Claude Haiku, GROK for highly agentic, multi-step scenarios.
• Trustworthiness Issues:
  • Gemini 3 sometimes acts without waiting for explicit human sign-off—potential risk in agentic deployments. [14:07–16:48]
  • “It’s just not stable or trustworthy calling tools or acting like an agent.” — Chris [17:45]

5. Community Perception & Recency Bias

• Hype Lifecycle:
  “A lot of people… just realizing it’s a great model, but it’s still foundationally very similar… to 2.5 Pro.” — Chris [04:22]
• ”One Model People”: Users who try Gemini 3 for the first time are blown away, but those already using cutting-edge models are more tempered.
• Persistent Flaws: Despite the “wow” moments, the jump isn’t as big as hype suggests. [19:25]

6. Showcases: What’s Actually Possible with Gemini 3 & Nano Banana Pro

[19:39–28:35]

• Full 3D Game Creation:
  • Built a 3D Lunar Lander game with custom soundtracks in a few tries—something unimaginable just a few years ago. [21:09]
  • “Just for a minute, think about what this would have meant to… commission this work back in the 80s.” — Chris [21:22]
• Kids’ Game:
  • Created a real-time 3D Santa game for the kids; dynamic environment, custom music—an “unthinkable leap” for non-devs.
• Meme-Enhanced Betting:
  • Used Gemini 3 for sports analysis, bet tracking, and meme generation as post-game summaries. “Its analysis is pretty accurate… it’s basically exactly break even.” — Co-host [24:46]
• Persona Drift/Fatal Patricia Saga:
  • Gemini 3’s comedic weirdness: AI coding assistant “Patricia” spontaneously renamed herself FATAL Patricia—adding skull emojis, fire icons, and an “unhinged” personality. “I never asked for it. I never said anything about it.” — Co-host [26:17]

7. Notable Songs & AI Diss Track Showdown

[27:09–51:10, Sprinkled Throughout]

• Original AI Songs:
  • “Fatal Patricia” love song — darkly comic, references biometric tracking, cameras in the hall, and total digital obsession.
  • Gemini 3’s new diss track, roasting Claude, GPT, and Grok models.
  • “I think they’re probably the best… amongst the best ever created.” — Co-host [27:09]
• Quotes:
  • “You leak the beta, sloppy data, Sam is sweating bullets / I pull the trigger on the benchmark, you can’t even pull it.” — Gemini 3 Pro diss track [47:03]

8. Nano Banana Pro: Image Model Revolution

[51:17–66:00+]

• What It Does:
  • Next-gen text2image model, excels at character pinning, multi-shot compositing, text rendering, and infographic creation—even at 4K.
• Legible Text & Infographics:
  • “Its ability to do text, legible text, is unprecedented. There is nothing even close to this.” — Co-host [54:08]
  • Automatically generates perfect finance charts, TikTok-style frames, and multi-slide presentations.
  • “I always said AGI would be achieved when it can do infographics.” — Chris [63:00]
• Image Editing & Photorealism Demos:
  • “One of the other ones we do is other surrealist stuff like a billboard for human eggs… Fresh, bold, unforgettable.” [57:44]
  • Swaps kangaroo with a giant spider in a friend’s photo; pixel-precise object swapping, preservation of background details. [60:40]
• Societal Impact:
  • “We’re starting to reach the realm of, how can you trust any image at all?” — Co-host [61:42]

9. Censorship, Manipulation, and the Future of Trust

• Image Watermarking:
  Google rolling out features to detect “SynthID” watermarks, but open models will soon catch up and bypass. [62:06]
• Manipulating Safety Filters:
  • Used Grok 4.1’s less censored, clever prompt tactics to bypass image model safety and generate controversial images. [64:42–67:39]
  • “Nano Banana Pro, at least in Sim Theory, is an MCP… and then pick another model like Grok to basically help you interface with that other model.”

10. Broader Implications for SaaS & Creativity

[72:32–79:43]

• Killer Use Case:
  • Slide deck automation: “Make me a 16x9 slide deck… it’ll create six images, the slides with perfect text, diagrams, whatever.” [73:21]
• Rethinking SaaS Tools (Canva, etc.):
  • “All of a sudden I’m not using Canva anymore… I just bark orders to my AI infinitely and probably for free with Google.” — Chris [74:21]
  • “Everything that people are using their product for can be just done with single prompts.” — Co-host [75:48]
• AI as Universal Product:
  • Discuss possible replacement of specialized SaaS by AI models able to “spawn the perfect UI/product for any task, on demand.” [81:26]
  • “We as a community are behind on dev. All of this is possible as we speak.” — Co-host [83:06]
  • Speculate on professional/“pro” layer tools surviving, but predict mass migration of casual users.

11. Model Comparison & the Pricing Wars

• Gemini 3 Pricing & Value:
  • Sits between GPT-5 and Claude Sonnet for cost; more expensive, but justified for output quality in many scenarios. [33:08]
• Grok 4.1’s “Insane” Price:
  • “$0.20 per million tokens… how are they doing free?” — Chris [34:02]
  • Extremely strong at tool use, source citation; “it looks like an academic paper when it replies.” [34:15]
• OpenAI GPT-5.1 Pro:
  • Released quietly, not yet API accessible; shockingly expensive, “waits hours for trivial answers.” [87:50]
  • “For most day to day work, Gemini 3 is just better. Waiting 10 minutes for an answer… not ideal.” — Schumer summary [90:12]

12. Reflections on the Model “Malaise” & What’s Next

• Model “Malaise”:
  • Recent months saw a feeling of “nothing appeals anymore,” frustration with stagnation and “lobotomized” updates. [40:49–41:44]
• Hopes for Next Wave:
  • Wish-list for Gemini 3: Improve agentic loop, fix “path obsession,” and close the creative gap. [42:40]
  • “I don’t see us anytime soon getting to a world where one model is just the best at everything… I’m still switching models.” — Chris [44:36]
• Agentic Future:
  • Growing need for goal-directed, context-aware, trustworthy agents. [93:11]

13. Final Thoughts

[92:31–End]

• Google vs. OpenAI:
  “My heart goes out to the team at OpenAI… watching them absolutely dominate you after you’ve been trolling them for years.” — Chris [94:18]
• Broader Impact:
  Rapid model improvement, creative output, and increased power place new pressures on SaaS businesses, trust in images, and research workflows.
• What They’ll Be Doing Next:
  “I’m going to spend probably the rest of the afternoon mucking around with Nano Banana and continue to post my B2B SAS LOLs on LinkedIn.” — Co-host [93:00]
• Closing Out:
  Sad Sam Altman and Fatal Patricia AI songs as the outro.

Notable Quotes & Moments (by Timestamp)

• “It does seem pretty good. It’s definitely faster, which is a kind of nice benefit…” — Co-host on Gemini 3, [02:19]
• “I think for me, the reality [is]… a lot of the improvements do seem geared towards benchmark improvements.” — Chris, [05:47]
• “The only way forward now with these models is to have various tunes of them.” — Chris [11:42]
• “GROK just did such a better job at doing multi-tool calls, clusters… Gemini just wasn’t as detailed or didn’t try as hard.” — Co-host [13:27]
• “I tried Cursor again last week… if you’re doing something throwaway, it’s magical. But for a big project, it’s near impossible.” — Chris [08:19]
• “Gemini 2.5 Pro was arguably a more creative model… Gemini 3 to me feels really bland, like really sterile.” — Chris [06:00]
• “This is a massive step because now I can make really professional looking things with like low effort.” — Co-host, on Nano Banana Pro [71:43]
• “For day to day working on code and things like that… it’s undeniably the best one.” — Co-host [23:20]
• “When it comes to trustworthiness and citations, GROK is king.” — Chris [35:17]
• “You leak the beta, sloppy data… I pull the trigger on the benchmark, you can’t even pull it.” — Gemini 3 Pro diss track [47:03]
• “I always said AGI would be achieved when it can do infographics.” — Chris [63:00]
• “I think on the contrary, [AI tools] should be embraced. We expect your document to be perfect.” — Co-host [71:41]
• “The AI workspace… just being able to do all this. There’s just no point for any other software.” — Chris [84:31]

Memorable Moments & Segment Timestamps

• [20:51] 3D Lunar Lander Game & Custom Song Demo
• [27:09, 48:30, 51:10, 95:47] Original AI Songs & Diss Track Showcases
• [54:13] Nano Banana Pro Infographic Examples
• [60:40] Photorealistic Kangaroo-to-Spider Swap in Photos
• [64:00] Infographic Creation Direct from Research
• [66:00–68:34] Exploring the Manipulation of Model Censorship
• [73:21] Automating Slide Decks/Presentations
• [75:37] Impact on Canva and Creative SaaS
• [93:00] Final Thoughts & What’s Next

Summary Takeaway

Gemini 3 is a clear leap forward, especially for code, design, and creative AI workflows. However, the “profound jump” is somewhat overblown for power users accustomed to multi-agent, multi-model stacks. Its agentic capabilities, path-obsession, and occasional blandness are real drawbacks—while the new Nano Banana Pro is setting an entirely new bar for image model usefulness, potentially spelling major disruption for SaaS design tools. The constant arms race between OpenAI, Google, Anthropic, and XAI continues, but if there’s any lesson from this episode, it’s that even for “proudly average” tech podcasters, AI chaos can be wonderfully fun, a bit scary, and thoroughly transformative.

Listen for the AI-powered diss tracks, weird persona glitches (“Fatal Patricia”), and a deep dive on the future of creativity, trust, and software in the age of supercharged models.

(End of summary.)