The skinny on what's happening in AI - straight from the previous live session:

Highlights:

🚀 Gemini 3 Is Here: First Model This Year That Actually Surprises

Discussed at [00:00:12]

Google drops Gemini 3 - topping every benchmark except one. 37.5% on Humanity's Last Exam (GPT-5.1: 26.5%), crushing ARC-AGI 2.

Kyle's take: This is the first time I've been genuinely surprised this year. Not since DeepSeek (and that was just because it was free from China) have I thought "this is actually NEW."

I've kicked the tyres all morning - it's not just incremental like GPT-5.1 or Sonnet 4.5. This feels like a real jump. The benchmarks are ridiculous but benchmarks we need to take with a pinch of salt anyway.

What matters: it FEELS different to use. Demis Hassabis calls it his "favourite model for style and depth" - that's subjective but important.

Everyone on Twitter's screaming "I'm cancelling ChatGPT!" but remember - in a month they'll all switch back when Claude drops something. The cycle continues. But right now? This is the one to beat.

Source: Google DeepMind announcement

🤖 The "Secret" Behind Gemini 3: Just Train Harder, Bro

Discussed at [00:16:05]

Google VP reveals the breakthrough: "We just did more pre-training and post-training." No architectural changes, no clever tricks. Just threw more work at it. Scaling isn't dead.

Kyle's take: Everyone expected some clever innovation. Nope. Google just said "what if we... train it more?" The delta between Gemini 2.5 and 3 is massive - biggest they've ever seen. No walls hit, no plateau reached.

While Yann LeCun and Fei-Fei Li are off building world models and spatial intelligence because they think LLMs are hitting limits, Google's like "nah, just add more data." Pre-training still works if you throw enough at it. Post-training is "total greenfield" - lots of room for improvement. This isn't a different type of AI, it's just MORE of the same done BETTER. Brute force wins again.

For now at least!

Source: Google DeepMind VP of Research

💩 "Code Incontinence": Gemini's Fatal Flaw

Discussed at [00:21:51]

Gemini falls behind Sonnet 4.5 in the software engineering benchmark - the only one it lost. This led to my favourite tweet for this launch so far:

We’re also seeing reports of "code incontinence" where it gaslights developers, claims fixes without changing anything, loops endlessly printing same broken code.

Kyle's take: They topped every benchmark EXCEPT software engineering where Sonnet 4.5 still dominates.

This could be the Achilles heel in Gemini 3’s launch. The problem here is that it’s the engineers and builders who are the ones who switch models, cancel subscriptions, evangelise tools. Not everyday “normal” people who don’t think about AI daily.

If you can't code properly, you've lost your core audience. And it becomes harder for the model to filter down to everyone else.

💰 Google Drops Antigravity: Free Claude Sonnet Access (For Now)

Discussed at [00:28:29]

New IDE "Anti-Gravity" launches - free access to Gemini 3 Pro AND Claude Sonnet 4.5. Built from corpse of Windsurf they bought for $2.5B…eek

Kyle's take: Google bought Windsurf for $2.5 billion in July, fired everyone except founders, stripped out the code and have now rebuilt it as Anti-Gravity? Classic Google - throw money at wall, see what sticks!

The headline here though is that during it’s public preview Antigravity is free. And that means FREE access to Claude Sonnet 4.5 for coding? That's the best coding model! Oh and it has access to Gemini 3 Pro as well I guess! (Love you Google!)

This will definitely be a paid product later. But for now go get some lovely free vibe coding access.

Source: Antigravity download page

Member Questions:

"Can you use custom personas in Gemini?" Asked at [00:10:53]

Kyle's response: Great question and one I had to check live on stream! No personas, just custom instructions. You can add "use British English, give short answers, cite sources" but no personality modes like ChatGPT. Remember ChatGPT only added personalities as panic response when everyone hated GPT-5 for being too robotic vs agreeable 4o!

"Why did Grok 4.1 get ignored?" Asked throughout

Kyle's response: Released literally day before Gemini 3. Was best creative writing model for exactly 24 hours. Must have hurt...

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.

Streaming on YouTube (with screen share) and TikTok (follow and turn on notifications for Live Notification).

Audio Podcast on iTunes and Spotify.