Tuesday, September 23, 2025

Benedict's Newsletter: No. 610

NO. 610   FREE EDITION   SUNDAY 21 SEPT 2025
SPONSORED BY ATTIO
AI-native CRM

Attio is the CRM for the AI era. Connect your email and Attio instantly builds your CRM - with every company, every contact, and every interaction you've ever had, enriched and organized.

You can also build AI-powered automations and use its research agent to tackle some of your most complex business processes, freeing you to focus on what matters the most: building your company.

Join industry leaders like Flatfile, Replicate, Modal and more.

Try the CRM that's built for the future. Start a free Attio trial today

News

Meta glasses

Meta announced a new version of its smart glasses. See this week's column. LINK

Gemini in Chrome

Google is adding a 'Gemini' LLM assistant into Chrome, barely a week after a US antitrust judge decided Google doesn't have to sell Chrome. Gemini is also at the top of the US App Store, partly aided by the 'Nano banana' image generator going viral. With the models today largely commodities (no, not if you're an enthusiast, but for any casual consumer), distribution is this month's battlefield for consumer mindshare. LINK, TRENDS

TikTok resolution?

After a year of muddle, it now appears that the USA and China have a TikTok deal, with a consortium of US investors and tech companies reportedly including Oracle, Silverlake and Andreessen Horowitz (where I worked from 2014 to 2019) buying an 80%+ stake in the US operation from Bytedance, moving it to a new non-Chinese platform and running it locally. The US government would have a board seat. It's not clear what any of that means, though, and we don't have a price ($35-40bn has been suggested). WSJ, BLOOMBERG

Intel's party round

After the US government took a 10% stake in Intel last month, this week Nvidia bought a 4-5% stake for $5bn, together with a deal for Intel to make custom CPUs for Nvidia. I wonder if they expect a quid pro quofrom Trump on Chinese export restrictions (though China has now supposedly decided to stop Chinese companies from buying, anyway). LINK

China's Nvidia dance

Meanwhile, the FT reported that China's government has told Chinese tech companies not to buy Nvidia's new RTX Pro 6000D, the product launched in the last few months specifically for the Chinese market, to comply with US export restrictions. There are too many moving parts here to speculate with any confidence, even if you were an expert on Chinese semis and PRC industrial policy, which I am not: is this brinkmanship with Washington, pressure for a better chip, support for the domestic industry, or all of the above? LINK, BACKGROUND, NVIDIA COMMENT

ChatGPT usage data

I've observed a few times that we have few tangible metrics for actual LLM usage by consumers, besides a range of rather fuzzy ('does your company use AI?' is a particularly meaningless one asked by the US BLS) and quickly outdated studies by economists. This week, though, OpenAI co-published a 100-page study on sampled real usage data. Most notable to me is a breakdown of queries by type (analysed using, well, ChatGPT, obviously). While Anthropic reports 36% use for programming (see below), for ChatGPT this is apparently only 4.2% (though this doesn't include API use by other tools). And the headline: 700m WAU in July were averaging 18bn weekly messages, an average of 26 each. LINK

Agents and democratisation

Advertising and e-commerce platforms have always struggled with how to help the long tail of smaller businesses use all of the sophisticated and complex tools that they create. This week, Amazon launched an 'agent' product to walk businesses through the selling platform. LINK

And reflecting a much broader trend, it's also expanding its creative ad tools, with a new agentic model that it claims analyses your business and best practices to generate both display ads and video ads. I don't think many people realise that there are going to be a LOT more video ads - both on the web and on 'TV', as every legacy broadcaster tries to open up streaming inventory to SMEs, and they all want to use AI to help create the ads for people who could never have afforded to shoot or run a traditional TV campaign. LINK

Meanwhile, YouTube is trying much the same for 'Shorts' creators. LINK

The week in AI

Reddit wants to be paid a lot more by AI model labs, since it has a window on a lot of current collective human opinions. LINK

Not just India - Google is launching its discounted Gemini package in Indonesia. LINK

Google launched a protocol for payment between AI agents. Of course, launching is a lot easier than traction. LINK

MCP ('help agents talk to each other') is further along the curve, and GitHub is trying to make it a feature by adding its own directory. LINK

Following Meta last week, OpenAI released new guidelines on teen safety. LINK

Musk chaos

Elon Musk followed the assassination of a right-wing American polemicist called Charlie Kirk by posting a bunch of inflammatory tweets calling for violence. I think most people have long since settled their opinion about Musk, but it does occur to me that there are twoMusk Delusion Syndromes - there are the people who ignore and deny everything he says, and the people who ignore and deny everything he's achieved. The challenge is that you have to see both - yes, he really did create that, and yes, he really did say that. We might remember that Henry Ford had two sides to his personality as well. LINK, FORD

H1B chaos

Each year the USA runs an annual lottery for 85k new H1B 'skilled worker visas' - about 400k people are in the US on this visa, of which about 250k are in tech (and about 70% from India and another 12% from China). On Friday Donald Trump announced that each visa would now have a $100k fee, beginning on Sunday.

The initial release of this rule was that this 1: applied to current visas and 2: would be an annual fee (the Commerce Secretary said this clearly and repeatedly). Google's immigration lawyers advised that it also appeared that the $100k fee was chargeable on entry. meaning that if you had an H1B and flew into the US on Monday, you'd need to pay $100k at passport control, and so Google told H1B employees not to leave the country. Microsoft said the same. An Emirates plane leaving SFO held at the gate while the captain let Indian tech employees get off. After 24 hours, the government then said that this is a one-off fee that only applies to new grants.

While the announcement itself showed the level of competence we're now accustomed to from this administration, the underlying issue has room for debate - there is a view that the H1B visa is abused by contracting shops to bring in cheap labour that could be filled in the USA. Adding a $100k fee addresses that - but also makes it a lot harder for startups without lots of ready cash to bring in a few key workers. In a normal world, this would be a detailed policy discussion, not something to bodge out on a Friday night, but part of Trumpism is a legitimate frustration with the very real decline in US state capacity - the ability to get anything done or any law passed - and that tempts it to do something, even if badly.

Another part of Trumpism is here too, though - the president can exempt anyone he feels like from this rule, which is a perfect opportunity for 'favours'. LINK, GOOGLE, ANNUAL FEE, CLARIFICATION

About
What matters in tech? What's going on, what might it mean, and what will happen next?

I've spent 25 years analysing mobile, media and technology, and worked in equity research, strategy, consulting and venture capital. I'm now an independent analyst, and I speak and consult on strategy and technology for companies around the world.

Ideas

Search referral traffic to the FT is down 25-30%. For the Daily Mail, it's more like 90%. LINK

This is a good list of the issues big companies have run into trying to deploy LLMs in the last 24 months. I've heard a lot of these as well. A lot of this boils down to 'imagine you'd given everyone on your company 'the web' in 1997 and then looked for ROI" - that would not have gone well. Start with tangible use-cases and problems to solve and work out, not cool-but-unscalable demos and 'AI-washing'. Also, sort out your data first. LINK

Scam centres in Myanmar might hold as many as 100k people against their will, forced to try to trick Westerners out of their savings. I don't remember this in vintage sci-fi. LINK

Swarms are the new thing in Ukraine drone warfare. Sci-fi was all over this. LINK

Marc Benioff thought that Salesforce's 'Agentforce' AI product line would be a lot quicker and easier to set up than his own employees expected. They were right. LINK

Amazon is extending its logistics platform to competitors Walmart and Shein. LINK

A decade ago there were still a bunch of regional messaging apps, and Kik had 40m MAUs in India. Today it shut down: WhatsApp won pretty much everywhere except the USA and East Asia. The CEO had a post-mortem. LINK

Outside interests

Iran's most famous translator was… not really writing translations. Fascinating. LINK

Data

Anthropic released an updated version of its activity report. LINK

The Daily Mail complained to the UK's competition agency (the CMA) that Apple isn't letting it into Apple News UK. You might or might not care about the complaint itself, but there are a lot of interesting internal operating metrics - Apple News is big. COVERAGECOMPLAINT

How many web searches does ChatGPT have so far? Maybe a low single-digit percentage of Google? LINK

Dentsu's 2025 CMO survey. LINK

Elon Musk's xAI has 64m monthly consumer users (for context, OpenAI is reporting 700m weekly actives). LINK

Preview from the Premium edition

Glasses

There's not very much we can say about 'smart glasses' today that we couldn't have said at any point in the last decade. Imagine something that looks like a fairly normal pair of glasses, where the glass is a 3D display with depth of field, that works outdoors in bright light and has a field of view as wide as normal glasses. Give it sensors that can map the world around you, so that you can put a game onto a table, or images on a wall, or ask an AI assistant about anything you see, or ask it to remind you who this person is. And give it some form of control (hand gestures? Voice? Eye tracking? Brain waves?) that makes it useful. Now you have 'augmented reality!'

That's all easy to say, but try building it. There's a lot of engineering and miniaturisation (anything like this would probably have to be driven by a phone to begin with), but that seems within reach. You'd need a lot of AI to light it up, but it's very hard to bet against continued rapid improvement there. And then you need the optics, and this is a lot harder. Magic Leap and Microsoft's HoloLens built bulky optics with a narrow field of view that needed a dim room. They also tried to be complete computers. Neither worked. 

Meta is trying a different way to build towards the objective. Last year's

Upgrade to Premium
You're getting the Free edition. Subscribers to the Premium edition got this two days ago on Sunday evening, together with an exclusive column, complete access to the archive of over 600 issues, and more.
 

No comments:

Post a Comment

January 15, 2026

You know what Americans aren’t talking about very much today after Trump’s threat to detonate the North Atlantic Treaty Organization (NATO) ...