The skinny on what's happening in AI - straight from the previous live session:
Highlights
💸 Trump Allows Nvidia to Sell H200 Chips to China - For 25% Revenue Share
Unprecedented revenue share deal opens Chinese AI market. Black market no longer needed, but government gets quarter of all sales!
Kyle's take: This is potentially one of the biggest pieces of news of the year and I haven't seen many people talking about it.
Trump's given Nvidia the green light to sell advanced H200 chips to China - the workhorse GPUs that are used to train frontier models. Previously restricted for national security, now allowed to "approved customers" (who can then sell them on to anyone, let's be real).
This is not a tariff or normal duty - it's a 25% revenue share straight to the US government. That's up from 15% they tried in July. It's basically a bribe to let them sell to China.
Back in July Beijing's previous response was brilliant - "no thanks, we don't want any" - absolute power move. Huawei have been busy getting up to competitive levels with Nvidia but still behind (this is the suspected reason for the DeepSeek delays)
China's been forced to be efficient without these chips - DeepSeek trained on way fewer GPUs than Colossus's 50,000 H200s. Now they get legitimate access. Which could shift things very rapidly.
Taiwan wrapped up in all this makes it a tinderbox. Most Nvidia chips are still made in Taiwan. Which complicates this whole situation a lot.
🎬 Netflix Buying Warner Brothers - For the AI Training Data?
It's not just about streaming rights. Warner's century plus of content becomes AI training goldmine.
Kyle's take: Netflix acquiring Warner Brothers is massive by itself - streaming service buying an old-school studio that they helped destroy…
But Hollywood Reporter points out something crucial: WB’s back catalogue is an AI data asset. Netflix needs massive libraries they actually own to train AI on. Most stuff on Netflix they can't train on - don't have the rights. Only for their own productions which (even though they are busy producing a lot) pales in comparison to a catalogue like WB.
But buy a whole studio? Suddenly you have rights to train on and produce derivative works from Warner's entire library going back decades. WB also owns DC - this is a move like Disney buying Marvel.
Think about it - all those characters with IP, all those episodes and films. Netflix could let consumers create personalised versions of shows, AI-generated episodes.
Ted Sarandos kept saying "innovation" and "world building" (an AI concept) on investor call but didn’t specifically talk about AI training.
🤔 Do AI Personas Actually Work? Ethan Mollick Says No, Everyone Else Says Yes
Ethan Mollick’s new testing shows that telling an AI "you are a physicist" doesn't improve physics accuracy. Giving personas has no effect on accuracy. So why do Anthropic and Google recommend it?
Kyle's take: Ethan Mollick tested one of the most common prompting techniques - giving AI a persona like "you are a great physicist" - and found it doesn't make answers more accurate. Used GPQA and MMLU Pro benchmarks, tested with and without domain expert personas.
People are disagreeing because this contradicts their experience. Even Anthropic's official prompt engineering guide says "give Claude a role." Google says same thing. OpenAI too.
So what's happening? I think personas aren't about necessarily about accuracy - they're about focusing attention. An LLM can roleplay as anything, that's what makes them amazing. But without specificity, they tend towards the generic.
By saying "you are a sales copywriter for a webpage," you're focusing this huge possibility space down to something specific. More context = better performance. Especially when you are working with the AI and having a proper back and forth discussion.
Mistakes come when we don't give enough context and AI fills gaps itself. Anthropic claims roles "significantly boost performance" in complex scenarios. Ethan says "show me the evidence." That’s entirely fair!
My take: just try both ways. Benchmarks are fascinating but what matters is your workflow. If adding a persona works for you, do it.
Kyle's response: The main reason I don't talk about Copilot much is it's not very good! It's built on ChatGPT since Microsoft owns 27% of OpenAI, but in their re-skinning and fine-tuning, they've stripped out capabilities.
Everyone I talk to finds Copilot less capable than ChatGPT. They’ve recently added Claude to some copilots too but it hasn’t made a huge difference.
Main benefit is security and being built into Microsoft 365. I used to do Copilot trainings but stopped because people either thought AI wasn't good (only exposure was Copilot) or were annoyed they had to use it instead of ChatGPT/Claude which they used in their everyday life.
The promise is exciting - AI at the core of organisation, surrounded by all Microsoft apps. But reality? 50 different implementations, none leading in tech. For instance, PowerPoint AI terrible compared to Gamma or NotebookLM.
Microsoft needs time to pull it off. Apparently they are also struggling to sell their AI products too. There’s a lack of demand. They recently halved their sales targets which suggests something is very wrong!
Kyle's response: I don't have a job! Haven't had one since I was 18. Always run my own businesses - first at 11 selling pirate video games (don't tell anyone, especially my mum !). Set up and sold a TV station in Vietnam in my twenties, had digital marketing agency, multiple online businesses. But never a 9 to 5.
Three years ago when ChatGPT came out, I saw it would threaten my marketing business in a year, so started reading and writing constantly about AI. Written over 500,000 words on AI for business in my newsletter. These playbooks are 15,000 words each - five newsletter issues at 3,000 words each. Huge amount of content! Too much!
I'm an interested amateur sitting between AI and entrepreneurship. Talking about how we can USE artificial intelligence. No PhD or computer science degree, although I’m sort of tempted to go back to studying! Instead like a lot of us, I’m working it out as I go. I’m just doing so very publicly! Exciting time because it's new technology and humanity's figuring out what to do with it.
Kyle's response: First up I’m not a developer! I’m from the marketing side.
This is a huge problem now - we can build so much with tools like Lovable but marketing hasn't been worked out.
Market's oversupplied, AI tool directories full of similar projects dying from lack of distribution. Building's the easy part now - anyone can make an MVP or proof of concept.
The difficulty is getting people to use what you built! Gets harder as building gets easier. My advice: build in public. Look at Jack Friks - 67,000 followers coming from him talking openly about money, marketing, process. People follow him, not just his products. When he has ideas, he can validate quickly with audience.
I did this myself - ~250,000 audience right now. Asked if they'd want to learn to give AI workshops for £1,000/hour. Hundreds said yes. That gave me market validation and then I built the product with the first cohort.
You have three basic options: build audience (content), buy audience (ads need capital), or borrow audience (partnerships need connections). Without cash or connections, build your own. Pick up camera, do TikTok/Instagram videos. Learn to explain what you're doing in 30 seconds.
Check out Yoni as another exa mple - 189,000 followers, one video daily for 500 days, started before having product. Build audience around you and journey, not specific product.
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).
