This Day in AI, EP99.07-06-05 (June 6, 2025)

Hosts: Michael Sharkey & Chris Sharkey
Episode Title: AGI Reality Check, Gemini 2.5 Update, Are Your AI Chats Safe & Fun with Veo3

Brief Overview

In Episode 99, Michael and Chris come together in-person for a wide-ranging, irreverent, and self-deprecating discussion of recent AI headlines and hands-on experiments. The brothers review Google Veo3’s video generation, break down the Gemini 2.5 update, debate the realities and risks of AGI, analyze model lock-in and access drama, and voice concerns about AI chat privacy in the wake of new legal demands. Listeners are treated to their signature blend of average hot takes, practical anecdotes, and comedic asides as they navigate an ever-shifting AI landscape where everyone is just trying to keep up.

Key Discussion Points & Insights

1. Experimenting with Google Veo3: Costs, Quality, and Use Cases

• Immediate Take: Veo3’s new video model impresses in some ways but is prohibitively expensive for hobbyists.

  • $3.75 for 5 seconds via Foul, 78 cents/second on Replicate. Quick humor over accidentally prompting “Tech Cage” instead of “Tech Cafe,” leading to a caged man with cats ranting about the podcast. (01:10–02:16)
  • “It sort of does sum up the person who probably wrote that.” – Chris (03:12)
  • Quality Assessment:
    • Highly adherent to prompts, even adding small requested details.
    • Some oddities: teeth, uncanny eyes, audio leveling issues needing post-editing.
    • “It’s 4K as well… I think, using them as cutaways for $3.75, if you just want some sort of establishing shot… I think it would totally work for that.” – Danny (03:53)
    • Caution: Early social media demos were “cherry picked.” Random outputs still trip up.
    • “It’s not something you want to just play around with unless you’ve got some real commercial use…” – Danny (04:26)

• Potential for Marketers: Marketing materials and stock video generation seem realistic uses even today.

2. DeepSeek (Deep Seq) R1 Model Allegations & Model “Train-Stealing”

• Rumors swirl about DeepSeek’s latest model being trained on Gemini outputs, following earlier speculation around OpenAI and Anthropic’s Claude. (05:28)
  • “I think this is the new benchmark, right? Deep See—whichever model they’re training on is clearly the…” – Danny (06:09)
• DeepSeek’s current performance highlighted: creates impressive, functional code and websites, not to be underestimated. (06:29–06:48)
• Benchmarking through “emotional intelligence” and repetition analysis suggests possible synthetic data use.

3. Google Gemini 2.5 Pro Update & “Thinking Budgets”

• New Tune: “A new tune of Gemini 2.5 Pro dropped into general availability… not a lot of data yet apart from them claiming it’s now number one on a bunch of benchmarks.” – Danny (07:10)
• Key Features:
  • More creative, better-formatted responses.
  • Improved instruction-following, e.g., for comments in docs/code.
  • Ability to specify up to 32,000 tokens for pre-answer “thinking budget.”
    • “This one can use 32,000 (tokens) just for the thinking, let alone… final content…” – Chris (08:20)
• Tune Versions: Users can now toggle between March and new June tunes (March version proving popular for some).
• Frequent Updates: The slow-drip release of “tunes” is smart marketing even if it confuses users. (09:34)

4. Personal Model Rankings & Comparison

• Model Variety: Much debate about Gemini, Claude, GPT-4/4.1/4.5, and open source options.

  • “I feel like even if we were cut off for everything except one modern AI model, you could still get a lot out of it…” – Chris (33:09)

• Model Performance:

  • Gemini 2.5 is a go-to for Danny (“gun to my head… it would still be 2.5.” – 31:31).
  • Claude Sonnet 4 is Chris’s daily driver for its parallel tool calling, decent speed, and reliability.
  • GPT-4.1 remains suprisingly good for conversational and vision tasks, but “just one of the most middling, disappointing, mediocre models out there.” – Chris (29:04)

• Agentic Capabilities: Claude’s ability to handle tool calls in parallel is noted as a “profound” development, possibly game-changing. (14:22–15:48)

  • “It did 10 at once… and synthesized those results back… astonishing.” – Chris (14:22)

• Model Switching Fatigue: Frequent retuning and updating causes both excitement and skepticism among users.

5. Infinite Context & AI System Layer

• Context Size Wars:

  • Gemini’s large context (theoretical “infinite” context) stands out in practice:
    • “You can build up the benefit of all these tool calls and… take advantage of that full context…” – Chris (17:04–18:19)
    • Danny: “[The] large context right now is what keeps me going back… I just feel really reassured that it’s taking into account larger chunks of data…” (18:19)

• Challenges Remain:

  • Even with large context, “context drift” still happens without reminders to refocus.
  • The need for goal lists and mechanisms to reunite context from branch-tangent conversations is growing.

6. AI Model Access, Lock-In, and the Windsurf/Anthropic Split

• The Drama:
  • Windsurf (a Claude-powered coding client, alleged OpenAI acquisition) loses Claude API access. Anthropic responds: “It would be odd for us to sell Claude to OpenAI… just trying to enable our customers who are… working with us in the future.” (24:03–25:17)
  • Anthropic’s move seen as the “first rug pull,” establishing a new precedent for strategic model denial.
    • “We predicted… our first episode ever of the podcast that what if they take the models away… that was our fear right at the start…” – Chris (26:31)
  • Attempts to workaround by procuring Claude via AWS or Google Cloud; likely to be recurring industry issue.
  • Theme of “what happens when the model you love is suddenly taken away?”

7. AGI Reality Check: Scaling LLMs & The Limits of Today’s Tech

• Yann LeCun’s View (Meta/Facebook):
  • Clip played at [37:08]: “We are not going to get to human-level AI by just scaling up LLMs. This is just not going to happen. Okay?”
  • “It’s not a PhD you have next to you, it’s a system with gigantic memory and retrieval ability, not a system that can invent solutions to new problems…”
• Hosts’ Reaction:
  • Respect for practical (if unsexy) take: these are amazing general tools but missing true “intelligence.”
    • “My point is this: so what? It can still do heaps of really useful things across many industries…” – Chris (39:15)
  • “We can now build AI systems that will change the world… change everything and are changing everything already.” – Danny (40:13)
  • Concern that AGI hype overshadows real-world deployment and utility.
  • Noted trend: less public “AI doomsday” fear, more talk about practical adoption.

8. AI Safety Rhetoric vs. Product Reality

• Anthropic’s Alarmism:
  • Seen as coinciding with need for fundraising: “every time I’ve noticed they need to fundraise, they just come out and spread doom porn like there’s no tomorrow…” – Danny (46:07)
• Sundar Pichai (Google) Approach:
  • More measured: acknowledges possibility of progress plateau, focuses on pragmatic challenges (context, memory).
  • “He’s talking very practical… like, these are the next problems we need to solve: infinite contexts and better memory…” – Danny (45:14)
• OpenAI’s World Tour Parallels:
  • Safety panic as a recurring marketing technique, not necessarily an honest reflection of risk. (46:26)

9. Court Orders & AI Chat Privacy

• Court Order: OpenAI directed to preserve all chat logs, including deleted and API chats, in ongoing media lawsuits (e.g., NYT).

  • “Like deleted chats, private chats, their logs, like all this stuff…” – Danny (48:12)

• Privacy Implications:

  • Brings up unsettling scenarios: would-be confidential chats now potentially retrievable for legal scrutiny.
  • “It’s an area where I think the GDPR laws should be enforced… The right to know what data they have stored about you…” – Chris (50:19)
  • Especially important as people connect more personal data and organizational data to AI tools.
  • “It’s like, everything I ever see, I’m going to keep…it’s an incredibly violating thing…” – Chris (52:13)

• Reality: Free AI access often comes at the cost of data privacy and consent; proactive legislation and user education are vital for long-term control.

10. The Future: AI Workspaces, Job Disruption, and What’s Next

• Job Automation:
  • Sober reflection on job impact—especially procedural roles. Example: AI taking drive-thru orders isn’t “intelligence,” just automation. (57:33)
• Adoption Curve:
  • “People...underestimate how long it takes society to adapt to these technologies.” – Danny (59:43)
• Long-term Vision:
  • Shift from “software stacks” to “model/MCP stacks”—curated setups of MCPS, agentic tools, models.
  • Belief that those who adopt and learn AI now will simply be more productive and fulfilled, not replaced.

Notable Quotes & Timestamps

• “The show neither impresses me nor disappoints me.” – AI-generated review via Veo3 (02:51)
• “We are not going to get to human-level AI by just scaling up LLMs. This is just not going to happen. Okay?” – Yann LeCun, guest clip (37:08)
• “If you had a gun to my head and said, you can live the rest of your life at this moment with one model, it would still be 2.5.” – Danny (31:31)
• “It did 10 at once… and synthesized those results back. The amount of thinking… was astonishing.” – Chris re: Claude Sonnet’s parallel tool calling (14:22)
• “We predicted...what if they take the models away… that was our fear right at the start…” – Chris on Windsurf/Anthropic (26:31)
• “My point is this: so what? It can still do heaps of really useful things across many industries…” – Chris (39:15)
• “If you’re going to talk about any form of legislation or anything, it’s just consumer protections around the right to not have your data trained on, even if the products [are] provided for free. That should be a human right with AI, really.” – Danny (51:05)
• “This is the first rug pull really that we’ve seen...” – Danny on locking out Windsurf/Coding clients (26:51)
• “It’s not going to be your software stack anymore. It’s going to be your MCP stack or your model stack.” – Chris (59:28)
• “Yeah, it’s not going to be like robots with lasers on their head walking down the street killing people for thought crimes.” – Chris (60:41)
• “When I get that [AI dishwashing/kitchen bot], I’m just done. That’s AGI for me.” – Danny (60:48)

Segment Timestamps

| Time | Segment/Topic | |------------|--------------------------------------------------| | 00:00–03:19| In-person banter, Veo3 first impressions | | 03:19–05:18| Veo3 prompt results, pros and cons | | 05:28–08:20| DeepSeek R1 training rumors & demo experiences | | 08:20–11:54| Gemini 2.5 Pro new tune & thinking budget | | 11:54–14:22| Benchmarks, model variety, Claude tool calling | | 14:22–19:41| Parallel tool calls, agentic workflows | | 19:41–24:03| Context size, infinite context, workflow issues | | 24:03–29:52| Windsurf/Claude rift, model lock-in discussion | | 29:52–34:48| Model rankings, resilience, and alternatives | | 34:48–37:08| Open/closed model ecosystems, open source | | 37:08–41:38| AGI, Yann LeCun interview/clip, hosts’ reactions | | 41:38–47:33| Safety rhetoric, OpenAI/Anthropic/Sundar Pichai | | 47:33–54:42| OpenAI lawsuit, chat privacy, legal demands | | 54:42–59:43| AI workspace future, job disruption, adaptation | | 59:43–64:16| Model stacks, societal change, kitchen AI dreams | | 64:16–end | Closing banter, daily driver choices, humor |

Memorable Moments

• Accidentally generating a “tech cage” video in Veo3.
• Griping over repeated “tune”/model versioning—“None of these labs have a healthy relationship with naming things.” – Danny (12:58)
• Lively digression about 90s CPUs as a parallel to context size progress (23:11–23:40)
• The “first ever rug pull” by Anthropic when they cut Windsurf’s Claude access and its portents.
• Yann LeCun’s no-AGI-with-LLMs quote, met with both agreement and a ‘so what?’ attitude.
• Proposed AI time-keeping/clock metaphors for tool-calling self-regulation (63:02).
• Strong feelings on privacy: right to have chat data hard-deleted (“The basic person is not going to assume, oh, it’s a soft delete…” – Chris, 50:37)
• Mundane but relatable AGI wish: “When I get that, I’m just done. That’s AGI for me.” – Danny, (60:48)

Tone and Style

Conversational, skeptical yet enthusiastic, with frequent self-deprecating humor. The Sharkeys routinely poke fun at their perceived mediocrity, lean into average-guy takes, and approach AI advancements with a critical but practical eye. The episode is balanced, mixing hands-on technical discussion with irreverence and accessible analogies—a podcast for people figuring out AI as they go.

For Listeners/Newcomers

• Curious about AI’s actual impact on work and privacy? This episode covers firsthand experiments and lively, down-to-earth debate.
• Worried about AGI hype or model access rug-pulls? The Sharkeys ground their discussion in the world as it is, not as doomsayers or marketers pitch it.
• Thinking of using AI tools in depth? Insights here about prompt quality, context management, and privacy risks are practical and immediate.
• Like your tech with a side of self-aware humor and skepticism? This episode delivers that in spades.

Closing

The in-person format adds new comedic energy, and, despite poking fun at their own knowledge, the Sharkeys give listeners a timely and nuanced look at the realities of AI development, deployment, and user risk in 2025.

“We don’t often touch. This is the most affection…”

“If you like the show, please do leave a comment, like, and all things. What other. Oh, wait, this is how you be a TikTok influencer…” (64:27)