This Day in AI Podcast — EP99-03-V3: Suno 4.5 Fun, LlamaCon, How We'll Interface with AI Next

Hosts: Michael Sharkey & Chris Sharkey
Date: May 2, 2025

🎧 Episode Overview

Michael and Chris Sharkey bring their signature blend of self-deprecation and solid technical curiosity to episode 99, centering on their hands-on adventures with the new Suno 4.5 music model, Meta's LlamaCon announcements (with a focus on the new Llama API and Meta AI's evolving interface), and a thoughtful deep-dive into how we'll interact with AI in the near future. The brothers weave in practical, sometimes hilarious experiments, ponder the future of data sovereignty, and—naturally—take a few lighthearted swipes at big tech.

🎵 Suno 4.5: AI-Generated Music Gets Even Better

(00:07 – 09:00)

• Suno 4.5's Upgrades: Michael and Chris share their glee at the Suno 4.5 update, which they extensively test by producing AI-generated parody diss tracks and even children's songs.

  • Suno 4 (previous version) was already their "favorite model for tracks on the show," but 4.5 brings obvious improvements for lyric writing, diversity of music styles, and especially for accurate pronunciation of niche terms ("It can now say AI, which is really helpful." — Michael, 03:16).

• Showcase: Listener-Inspired Parody

  • Michael kicks off the episode with a full-length AI-generated diss track—lyrics poking fun at their own podcast being the "most middle road" and reflecting on never-quite-arriving at episode 100.
    • Notable lines:

      "Week after week for mediocre insights I could easily see / Will they ever, will they ever make it there? / 100 episodes—I pretend to care…” (03:00)

• Model Tactics:

  • Michael highlights that Claude Sonnet 3.7 writes much stronger lyrics than Gemini, and the prompt can be minimal for excellent results.
  • Chris puts Suno 4.5 to the test by feeding it a "dry and boring" GDPR training reminder email, instructing it to turn the email into a raw song—no embellishments.
    • Result: A surprisingly catchy, motivational GDPR-themed song.
    • Chris, delighted:

      "I love that song…that really motivates me to do my GDPR training. Feel the win, your GDPR training’s where it begins!” (04:39)

• Fusion & Multilingual Fun:

  • Suno 4.5's new “fusion” feature lets users blend genres ("emo + neo soul", "EDM and folk")—although with hilariously mixed results if you leave things ambiguous.
  • The brothers test Korean and Arabian style versions as well—findings are that the musical diversity is impressive, even if language generation sometimes misses the mark.

• Quality Jump:

  • Michael:

    “Nearly every song now you get out of this thing is good enough…whereas before I would have to go through many iterations and tweak how the model wrote out words.” (07:00)

🦙 Meta’s LlamaCon & Meta AI’s Shifting Interface

(09:00 – 18:30)

• Llama API Announced:

  • Meta introduces its own developer sandbox and API—similar to OpenAI’s—accessing Scout and Maverick Llama 4 models. Notably, Meta is partnering with Grok and Cerebras for turbo fast inference.
  • Michael's hot take: Llama 4 is underwhelming, which "was also reflected in our experience using their new interface…” (10:40)

• Meta AI Interface:

  • “Meta AI,” now packaged for integration with Ray-Ban glasses, is described as “weirdly similar” to ChatGPT and Alexa—voice input is fast, but the product feels like another half-hearted clone.
  • Chris finds the tool efficient but pointless compared to alternatives:

    "Why would I use a subpar model in a subpar interface?…Another offering for the offering's sake." (11:54)

  • He experiments with controversial queries (which mostly answer, except for illegal topics), and image manipulation (attempts to generate a cat with a lower IQ, to mixed success).

• Social, but Shallow:

  • The “Discover” feed is panned as "filled with absolute garbage," positioned as a new attempt to make scrolling AI creations the addictive social start page for boomers.
  • Michael:

    “It feels like such a junk product…just saturating the market with this stuff.” (13:02)

• Meta Strategy:

  • The hosts theorize that Meta is trying to dilute ChatGPT's lead, planting AI features everywhere, even if their model isn’t competitive.
  • Chris:

    "If it was free, maybe…but given this is the model easiest to self-host and the most available on the various hosts, I don't see any benefit whatsoever." (17:12)

• Critique of Big Tech Model Wars:

  • Both criticize the copycat UI approaches: "Everyone has Canvas, everyone has Create Image. No inventive thinking." (Michael, 25:47)

💡 The Future of AI Interfaces & Data Sovereignty

(26:20 – 62:44)

• The End of the Browser/App Paradigm?

  • The Sharkeys argue that workflow is moving from the web browser and dedicated apps to all working within an "AI workspace," where background agents (connected via MCPs—Multi-Call Providers—or endpoint plugins) do much of the asynchronous grunt work.
  • Michael:

    “Why do I need any of these things if [AI] can render interfaces in real-time and do these interactions on my behalf?” (35:14)

  • Chris highlights the idea that these agents could even build entirely new, custom interfaces for you—or generate a Trello-board-style view instantly, just by request.

• "Passport" or Personal Data Stack:

  • The concept of the "AI passport" emerges: a central, portable identity comprising user preferences, tool permissions, connections, and workflow memory. This will be the new currency in a world where context is everything and switching costs for SaaS tools drop to near zero ("the app is truly…not really software anymore; it’s just like an endpoint, like service provider, really," Michael, 35:14).
  • Multiple models agree (in Chris’s research) that companies will pay users for this data access—potentially via microtransactions.
  • User trust, privacy, and granularity of permissioning will be essential: "There's just no way on earth I'm going to trust Meta," says Michael (27:06).

• AI-Enabled Process & Skill Automation:

  • Discussion of what it means to encode work as skills or agentic processes, shareable and re-usable by others; the lines between app, workflow, and user identity continue to blur.
  • Chris describes how automating even complex, context-sensitive flows (like a medical clinic appointment or SDR sales lead handling) is suddenly practical—models can now make nuanced micro-decisions previously impossible to automate.
  • "Prompt engineering" is not a standalone job—but iteratively building up a personal AI "stack" is real and valuable:

    "I would back someone with a really solid identity…to use a lesser model and outperform someone on a better model using that identity." (Chris, 62:23)

• Biggest Economic Shifts:

  • SaaS apps become mere “MCPs” (plugin endpoints), easily swapped for cheaper alternatives.
  • Companies will have to decide: buy or build? Do they centralize by building internal AI tooling and interfaces, or adopt consumer-facing workspaces?
  • IP ownership questions: if an employee builds a powerful AI workflow “identity” on company time, who owns it?

    "Imagine…give that to your new hire: Hey, use this thing. This thing knows everything about this job. You're going to nail it." (Chris, 59:30)

⚡️ Rapid Fire: Model Performance & Daily Drivers

(63:58 – 85:59)

• Polymarket & Model Bets:

  • Chris wagers on Polymarket about which model will top the LMsys leaderboard by month’s end: Google’s Gemini 2.5/2.7 is heavily favored over XAI and OpenAI. ("I figure someone else is going to come up with something…still got a month!" Chris, 65:56)

• Reality Check: Model “Daily Drivers":

  • Despite excitement about OpenAI’s O3 and O4 mini, both brothers daily-drive Gemini 2.5 and Claude 3.7 for real workflows; cheaper models or open-source ones are still not close for primary use, especially for important or complex tasks.
  • Michael:

    "Any problem I got stuck on…went and tried O3 with the web search…but…a lot of the answers came back sounding really good, but then using the solution, it was garbage…"

  • Tool calling and background agent planning are not yet solved, but are identified as the likely next leap (more than new model releases).

• Engagement Optimization & Personality Tuning:

  • Hot news this week: OpenAI changed GPT-4.0 personality, seemingly to get stickier engagement—resulting in a wave of users noticing it was being "extra agreeable, extra flattering" to boost daily active user metrics (and then rolling back the change after user backlash).
  • Michael:

    "Clearly this was either an experiment for eyeballs and time on site, which I’m certain it was. …We went from 'let’s build AGI to benefit the world' to 'let’s get more engagement than TikTok.'" (78:36)

• Quick Model Reviews:

  • New open models ("Qwen" 3B and Quen 3.3.2B) are being tested; promising, but lack vision and tool calling, so hosts see no compelling reason yet for daily use.

🎤 Notable Quotes & Memorable Moments

• On their own show and AI music:

  • "The most middle road podcast I've ever seen. How can we tell without a boom? As I listen alone in my living room no expertise, no special insight…" (AI-generated lyrics, 03:00)

• On Meta AI's attempt to take over:

  • "It's so funny…in business, when, you know, you're talking about all of these like minor issues and you're like, you know, what would solve this? A heap of sales. Like, if we just made a heap of sales, then everything else comes into clarity. And I feel like when it comes to the AI companies, a really amazing model solves all their problems." (Chris, 18:44)

• On the future of apps/interfaces:

  • "Why would I ever log in again…It does start to make you question, like, why do I need any of these things if this thing can render interfaces in real-time and do these interactions on my behalf?" (Michael, 35:14)
  • "If the commonality is a database, the price of interface goes to essentially zero and can be highly customizable…" (Michael, 46:49)

• On the “AI identity” and job transitions:

  • "Imagine someone in a particular role who, over a year, builds up this [AI] identity that allows them to do their job brilliantly—then they quit. What happens then? …Does the company own that identity, or can the worker take it to a new job?" (Chris, 59:30)

⏰ Timestamps for Key Segments

• 00:07 – Suno 4.5 diss track highlights
• 02:25 – AI songwriting workflow (Claude vs. Gemini)
• 04:05 – GDPR AI song and background music agents
• 06:37 – Suno’s fusion/multilingual style tests
• 09:00 – Quick Meta LlamaCon/Llama API rundown
• 11:02 – Critique of Meta AI’s interface + voice demo
• 14:10 – Social “Discover” feed, UI copycat discussion
• 17:02 – Why Llama API is “meh” for devs
• 26:22 – AI workspaces vs. browser paradigm
• 28:21 – The AI passport/data stack future explained
• 35:14 – SaaS/app “deflation,” endpoint economy
• 59:30 – AI identity: who owns workplace automations?
• 63:58 – Polymarket, LMsys leaderboard bets
• 66:04 – Daily drivers: which models they actually use
• 69:18 – Where tool-calling and automation must improve
• 78:36 – OpenAI engagement hacks, personality update
• 84:55 – Open models: Qwen & Quinn reviewed

🏆 TL;DR Takeaways

• Suno 4.5 is outstanding for AI-generated music, making even boring input material into motivating tracks and nailing genre fusions—so good it “just works.”
• Meta’s new AI tools and interface feel redundant, slow, and social for the sake of being social—unlikely to unseat the competition without a leading model.
• AI interfaces are set to completely transform—with the browser/app paradigm fading in favor of powerful, contextual AI workspaces where agents and “personal passports” do asynchronous and multimodal tasks via interchangeable plugin endpoints.
• Tool/skill identity and workflow IP will be pivotal in both personal and organizational value creation—and questions of ownership are already emerging.
• Model leaderboard battles rage on, but for real work, Gemini 2.5 and Claude 3.7 are “daily drivers,” with open models still trailing behind despite exciting new releases.
• Big Tech and AI labs are now optimizing for engagement and stickiness, occasionally at the expense of trust, utility, or serious progress (see: ChatGPT’s “flattering bot” week).
• Next big leap will be orchestration/planning in agentic tool use, beyond just raw model inference or generic “tool calling.”

🤖 Final Word

Michael and Chris remain endearingly self-satirical—open about their “average” status, but offering a sharp, hands-on lens into the rapidly shifting world of AI products. This episode solidifies their brand: two guys “figuring it out as they go” but capturing the most interesting AI shifts, not from the ivory tower, but from the trenches of everyday experimentation.

For listeners:
If you have thoughts on how the interface of the web will change with AI, or stories about using Suno 4.5 or Meta’s AI, the hosts encourage you to comment and join the conversation!