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

Highlights

🐉 China's Kimi K2: Open Source Model Beats GPT-5?

Discussed at [00:00:35]

Moonshot's new Kimi K2 "Thinking" model - 1 trillion parameters, open source, beating GPT-5 on benchmarks. Scored 44.9 on Humanity's Last Exam vs GPT-5's 41.7. Better at coding than GPT-5 (but not Claude)?

Kyle's take: This is another DeepSeek moment - China making state-of-the-art (SOTA) open source models while the West abandons open source. It's not just as good, it's seemingly BETTER than our closed models.

One trillion parameters (GPT-5 estimated at 1.5 trillion. Maybe. We don’t know!), yet outperforming on benchmarks. Yes, take benchmarks with salt - you can train for them. But third-party verifiers like Artificial Analysis confirm it's legit.

This is a massive problem for OpenAI raising for a $1 trillion IPO - why pay when China gives “better” models free? Opensource undercuts the closed source ambitions of OpenAI, Anthropic, Google and xAI.

What about fears of Chinese data security? If you run it locally, there's no connection to China. This is the main draw of open source models - you can deploy them yourself on local hardware (if you have enough GPUs!) or on a cloud server only you have access to. If anything it’s more secure than pinging info back and forth to OpenAI’s servers

Source: Moonshot Kimi K2 release / Artificial Analysis benchmarks

💸 OpenAI Damage Control: "We Don't Want Government Bailouts" (They Do)

Discussed at [00:14:02]

CFO Sarah Friar asked for government "backstop" at WSJ conference. Sam Altman immediately contradicted her. Both now backtracking after massive backlash about potential taxpayer bailouts.

Kyle's take: Is this 2007 all over again - "too big to fail”? That went well before right? Right?

The CFO admitted their data centres have "uncertain lifespan" and that government backing would "drop the cost of financing" so they can borrow at lower rates. Basically having the government secure the investment so that it’s more palatable to investors to get involved.

Translation: if OpenAI fails, taxpayers eat the losses.

Sam went into full damage control. He even wrote in capital letters! (you know he's serious) saying "we don't want government guarantees."

But then he immediately pivots to "governments should build AI infrastructure" - same thing, different words.

They've committed $1 trillion to infrastructure this year alone - $300B with Oracle, $500B Stargate project. Revenue? Still in the tens of billions. That’s great for a startup their age BUT the maths doesn't stack up.

Source: Sam Altman's damage control statement

🤔 Why Gemini 3 Might Already Be Obsolete?

Discussed at [00:11:36]

Google's leaked 1.2 trillion parameter Gemini 3 for Siri might be dead on arrival. China's Kimi K2 at 1 trillion parameters already beating current models.

Kyle's take: I wonder if this is why Gemini 3 hasn't launched - is it not SOTA (state of the art) anymore. Bloomberg leaked it's 1.2 trillion parameters for Apple's Siri (March 2026). But Kimi K2 just dropped at 1 trillion.

Google might have trained a model that's already obsolete. You can't just retrain quickly - that's months and billions of dollars.

The closed-source companies never reveal parameters - this leak was a mistake. And remember that Google's in an impossible position: their cash cow is search advertising (70-80% of revenue) but AI kills search. They need to slaughter their own business model but not too fast or they'll have no cash. Meanwhile China's giving away better or at least competitive models for free. Buckle up.

Source: Entirely speculation!

Member Questions:

"What's the best free vibe coding tool?" Asked at [00:22:29]

Kyle's response: Easy answer now - Google AI Studio! Go to ai.studio.google.com, click "Build" on the left. It's their vibe coding tool, totally free. Has Gemini 2.5 Pro/Flash and access to Nano Banana (best image model). You can even use voice like Lovable. Good for testing ideas, then move to Cursor if needed (hobby version free, then $20/month). But start in AI Studio Build mode, it's the best free option now. Google's UI is confusing as always, but the tool is solid!

"Doesn't China inflate numbers?" Asked at [00:09:55] in reference to Kimi K2 Thinking benchmarks.

Kyle's response: That's why we have third-party verification - Artificial Analysis, Hugging Face, LMArena etc. They independently test all models, Chinese and Western. Every company lies about being "state-of-the-art" - Grok's particularly slippery, claiming best at coding then people use it and it's rubbish.

Remember that the ultimate test is subjective - does it work for YOUR workflows and tasks? Try the major releases yourself. Can't trust any company really, they all make stuff up. That's why independent benchmarks and our own testing matters!

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.