Everyday AI Podcast: EP 429

AI News That Matters – December 30th, 2024

Episode Overview

This episode is the year-end edition of the “AI News That Matters,” where host Jordan Wilson breaks down the most pressing stories in AI as 2024 comes to a close. Jordan brings his signature candid, jargon-free style to spotlight breakthroughs, corporate maneuvers, paradigm shifts, and the ethical debates shaping AI right now, all aimed at helping everyday listeners stay ahead in their careers and businesses.

Key Discussion Points & Insights

1. DeepSeek V3: A New Open Source Model Leader

[05:29]

• DeepSeek (China) released the open source V3 model, directly benchmarking with powerhouses like OpenAI’s GPT-4o and Anthropic’s Claude Sonnet 3.5.
• The model is available on GitHub, designed for accessibility and innovation. Its MOE (mixture of experts) architecture optimizes both performance and resource use for developers without “the resources of larger corporations.”
• V3’s licensing allows reproduction, modification, and commercial use. It’s seen as:
  • “A cost-effective alternative and strong contender for organizations focused on developing specialized applications in coding and mathematics.”
    — Jordan [07:11]
• This comes amid increasing demand for flexibility and custom AI solutions.
• Notable Quote:
  “It is a fraction of the cost of others and it's open source y’ all.”
  — Jordan [08:53]
• Jordan predicts 2025 will bring even more LLM innovation out of China.

2. Meta’s Plan: AI Characters on Social Media

[10:24]

• Meta will soon launch AI-generated characters as user accounts on Facebook—complete with bios and profile pics.
• The AI Studio features “hundreds of thousands” of personas and will expand globally.
• Concerns include misinformation, appropriateness, and verification of AI vs. humans.
• Host’s Honest Take:
  “I don’t personally see the value of this. News that is not accurate spreads pretty quickly on Facebook. Now when you add this 'human' element … Not sure I’m a huge fan of it.”
  — Jordan [14:26]
• Hopes Meta will clearly label AI users, but information is still scarce.

3. OpenAI Restructures (Again): A Move to Public Benefit Corp (PBC)

[16:41]

• OpenAI announced plans to restructure as a PBC, better positioned to raise the capital needed for its ambitious research.
• The non-profit arm remains but will hold interest in the new, for-profit PBC.
• Controversy: Elon Musk and Meta object; critics fear profit overtakes ‘public benefit.’
• Jordan’s Take:
  “People are griping and complaining about this … What do you expect? … You need to have money to do that—many billions of dollars.”
  — [19:37]
• Lawfare with Musk continues, which Jordan calls more “for show,” lacking real legal merit.

4. Google Contractors Using Anthropic’s Claude to Evaluate Gemini

[25:10]

• Leaked info via TechCrunch: Google contractors score Gemini responses against Claude 3, comparing truthfulness and verbosity.
• Claude often “prioritizes safety more than Gemini”—an important distinction given AI’s risks.
• Anthropic’s policies forbid competitor training using Claude outputs; Google insists they only evaluate Gemini, not train it.
• Industry Scoop:
  “All models use outputs from other models to evaluate. Period.”
  — Jordan [30:18]
• Jordan shares behind-the-scenes info: Top tech companies often use other models’ outputs in evaluation and sometimes, by inference, in training.
• Contractors often lack expertise, especially on sensitive matters like healthcare.

5. Google Doubles Down: Gemini is the Top Priority for 2025

[34:21]

• Company aim: Gemini to reach half a billion monthly users—becoming Google’s 16th product at this scale.
• CEO Sundar Pichai stresses urgency—he wants “AI” to mean more than “ChatGPT” in public discourse.
• Pledges integrated, multi-device, multi-modal Gemini assistant to rival OpenAI.
• Employees wary of potential high costs; management says no $200/month plans are coming.
• Jordan’s Assessment:
  “Google … dropped the bag. They completely failed in 2023 and early 2024.”
  — [36:32]
• Google’s AI Studio and backend have been strong, but they botched their public, “front end” offering, ceding mindshare to ChatGPT.

6. OpenAI (Re)enters Robotics

[42:32]

• Reports indicate OpenAI has quietly revived its robotics group, closed since 2021.
• OpenAI aims to develop humanoid robots, both through in-house teams and investments (e.g., Figure AI, 1X).
• This trend ties into broader industry moves, with Nvidia and Tesla also deepening their robotics focus.
• Jordan’s Candid Take:
  “Get used to talking about humanoids and robotics a ton … It is both exciting and extremely scary, I’m not gonna lie.”
  — [43:22]

7. AI Manipulation: The 'Intention Economy' Study from Cambridge

[44:07]

• A new paper from Cambridge’s Leverhulme Centre warns that AI can predict and influence human intentions—potentially steering consumer and voter behavior.
• The “intention economy:” AI models learn from behavioral/psychological data to microtarget and manipulate choices.
• Already, big companies use synthetic data generated by LLMs to simulate markets and conduct user testing.
• Key Insight:
  “AI can manipulate humans. Large language models have already been used for a very long time in creating synthetic data.”
  — [45:23]

8. Geoffrey Hinton Doubles Extinction Odds: 10–20% from AI

[46:05]

• “Godfather of AI” Geoffrey Hinton updated his estimate on existential risk from AI to a 10–20% chance of human extinction within 30 years (previously 10%).
• Emphasized: No history of less-intelligent entities controlling more-intelligent ones.
• Calls for regulation and notes profit motives alone won’t ensure AI safety.

9. Redefining AGI: Microsoft and OpenAI Set a $100B Benchmark

[47:20]

• Latest deal between OpenAI and Microsoft defines AGI as “a system capable of generating $100 billion in profits.”
• Previously, Microsoft would lose licensing rights when OpenAI achieved AGI—now it’s pinned to this financial target.
• OpenAI’s latest models (like 01 Pro) are rapidly approaching the capability benchmark, but definitions keep shifting.
• Jordan’s View:
  “At that point, it’s like, yeah, that’s AGI—at least according to OpenAI’s definition. But that’s not the definition anymore.”
  — [49:22]
• CEO Sam Altman thinks AGI will come sooner than expected but may be less world-shattering than feared.

Notable Quotes & Memorable Moments

• “A lot of AI news and large language model news coming out of China... It is a fraction of the cost to use DeepSeek.”
  — Jordan, on the global AI landscape [09:18]
• “Are we living in a simulation? Every, you know... I don’t know, are AIs just going to be running their own social media?”
  — Jordan, on AI personas posting on Meta [11:24]
• “People are griping and complaining... but what do you expect?”
  — Jordan, on OpenAI’s for-profit pivot [19:37]
• “All models use outputs from other models to evaluate. Period.”
  — [30:18]
• “Google absolutely made a... trillion dollar mistake by doing that.”
  — on Google’s failure to prioritize user-facing AI [39:05]
• “It is both exciting and extremely scary—I’m not gonna lie to y’all.”
  — on the rise of humanoid robots [43:22]
• “AI can manipulate humans.”
  — summarizing Cambridge’s new research [45:23]
• “The likelihood of AI leading to human extinction within the next 30 years has doubled.”
  — on Hinton’s latest warning [46:15]
• “Now, according to reports, they’ve changed that agreement... Now it is essentially when an AI system has the track record or the capability to reach $100 billion in revenue.”
  — on the Microsoft/OpenAI AGI clause [49:36]

Timestamps for Key Segments

• Open Source AI Model War: DeepSeek V3 — [05:29]
• Meta’s AI Personas on Social Media — [10:24]
• OpenAI’s Public Benefit Corporation Shift — [16:41]
• Google Contractors Use Claude for Evaluation — [25:10]
• Google Makes Gemini Top Priority — [34:21]
• OpenAI’s Humanoid Robotics Push — [42:32]
• AI Manipulation / Cambridge Study — [44:07]
• Geoffrey Hinton on Human Extinction Risk — [46:05]
• Microsoft/OpenAI Define AGI by Profits — [47:20]

Final Thoughts

Jordan wraps the episode by urging AI professionals and enthusiasts to share the knowledge, join the free newsletter at youreverydayai.com, and explore the trove of resources—over 430 episodes to date.

“If you gained benefit from Everyday AI this year, please tell someone about it ... there’s like thousands of hours of free AI info, sorted by category, no matter what you care about, by the world’s leading experts.”
— Jordan [51:08]

(Summary faithfully represents the original episode’s tone: practical, direct, and occasionally blunt, with a skeptical but optimistic lens on AI’s social impact.)