AI with Kyle Daily Update 094
Today in AI: Google Gemini trained on Google TPUs
Watch this video on our dedicated watch page
Better viewing experience with related videos and full-screen player
The skinny on what's happening in AI - straight from the previous live session:
Highlights:
💣 Google Didn't Use Nvidia for Gemini 3: The $5 Trillion Chink in the Armour
Gemini 3, the current state-of-the-art model, was trained on Google's TPUs not Nvidia GPUs. First major model to break Nvidia's monopoly. Google owns the entire vertical stack.
Kyle's take: This is MASSIVE and most people missed it. Including me!
@iamkylebalmer Feel a little silly. There’s something about Google Gemini 3 which I totally missed yesterday and it’s more important than the quality of ... See more
Nvidia's worth $4.5 trillion (was $5 trillion last month) because everyone NEEDS their GPUs to train models. Except... Google just trained the best model on Earth without them? Hold up.
They used their own Tensor Processing Units (TPUs) - chips specifically designed for AI, not repurposed gaming hardware. Google's been quietly building these for years while everyone else queues up to pay Nvidia obscene amounts.
This changes everything. Google owns the chips, the data centres, the electricity, the model - the entire vertical. Hell they even own the application level - Google Search, workspace and Android.
They're not paying middleman taxes to Jensen Huang. While OpenAI burns cash buying GPUs, Google's printing their own money. Suddenly I'm bullish on Google. They're not getting weaker - they're in pole position…
📈 Nvidia Earnings: The Bubble That Won't Pop (Yet)
Nvidia reports stronger than expected chip sales. Stock up 4.7% pre-market. Global tech rallies on "AI demand still strong" signal.
Kyle's take: Everyone's watching Nvidia earnings like tea leaves for the AI bubble. If chip sales soften, it means companies are cutting AI spending = bubble pop. But nope - everyone's still throwing billions at GPUs. The party continues.
Here's the thing: Nvidia's ACTUALLY making money. Their customers (OpenAI, Meta, everyone except Google apparently) are haemorrhaging cash buying these chips while Nvidia absolutely prints money.
This photo of Jensen Huang with Elon Musk, Greg Brockman (OpenAI), and the White House AI czar tells you everything - Jensen's the kingmaker now. There's good money (Nvidia selling picks in gold rush) and dumb money (Mira Murati's $50B valuation with no product anyone??). When the music stops, Nvidia might be fine. Their customers? Different story.
Source: FT
🥊 Gary Marcus Goes Nuclear on Yann LeCun: "Plagiarism of Ideas"
Marcus publishes scathing Substack: "The False Glorification of Yann LeCun." Accuses Meta's chief AI scientist of claiming credit for five major ideas he didn't invent. Lawsuits incoming?
Kyle's take: This isn't Twitter beef - this is academic war. And pretty personal. Marcus says LeCun's known for five ideas (CNNs, LLM critiques, scaling doubts, common sense advocacy, world models) and "originated exactly NONE of them."
Two of those five? Marcus claims HE was first. This has been brewing for years - symbolic AI vs connectionist AI, old guard vs new guard. But accusing someone of plagiarism? That's career-ending stuff.
The timing's telling - LeCun just announced his new startup, gets Wall Street Journal puff piece calling him "lone genius," Marcus loses it.
Someone called this "Kendrick vs Drake for AGI people" which is funny but misses the point - this is about fundamental disagreements on where AI's heading. Both think scaling will hit a wall but disagree on what comes next.
Worth a read for sure, primarily to learn more about different expert takes on where we’re headed.
Source: Gary Marcus Substack article
📚 "Vibe Coding" Named Collins Word of the Year 2025
Dictionary officially recognises "vibe coding" as word of the year 2025. It beat "aura farming" and "clanker" (anti-AI slur).
Kyle's take: We did it! Vibe coding is officially in the dictionary. Collins Word of the Year 2025. Andrej Karpathy popularised it (didn't coin it apparently?) to describe coding by feel rather than syntax.
It beat "clanker" (the anti-AI / robot slur) which tells you everything about 2025's priorities. The fact that an AI term won over an anti-AI term is significant - shows mainstream acceptance.
That said, ask the average person in the street what vibe coding is and I bet they have no idea!
Source: Collins Dictionary announcement
Member Questions:
Kyle's response: I'm building AI-powered education. Written 400,000 words this year but people don't read anymore. They buy courses with best intentions, never finish them. So I'm embedding AI tutors INTO courses - not just dumping content in a chatbot. My Freedom OS prototype: 7 modules, each has 2-3 minute video, some text, then AI assessment/quiz that gathers info and creates personalised action plan. Like Sal Khan says: we can give every kid a one-to-one tutor. Very cool stuff!
Kyle's response: Barely use it now tbh. Their unique thing was citations and research mode - now Claude, ChatGPT, Gemini all have that. If you have PayPal, get Perplexity Pro free (Google it) but I don't see the point anymore. They're giving away Pro accounts to pump user numbers.
Want the full unfiltered discussion? Join me tomorrow for the daily AI news live stream where we dig into the stories and you can ask questions directly.