GPT-6 Astra Is Here + This Week's Best Stories

Hi!

It’s been a while! I’ve been busy plastering my face all over TikTok, Instagram, YouTube and LinkedIn, helping people make sense of AI and marketing. The last two years have been quite a ride, amassing 400,000 social followers and over 175 million views, but I really missed writing this newsletter, and so it's back!

Hope you enjoy each week's most interesting stories in AI, tech and marketing, all in one place, clearly explained.

This week's best stories:

OpenAI has launched GPT-6 Astra, with improvements in browser control, research, analysis and producing documents, spreadsheets and presentations that follow your existing templates.

My take: this is the first time I've really let go... trusted AI with a whole package of tasks, including building my new (forthcoming) website and sorting out technical jobs across my email, CRM system and hosting. Those connected jobs are where a lot of the time and effort goes.

I just tell Astra what I want to achieve, and it just cracks on. Sometimes for a very long time.

It's very hard to say if it's better than Claude Fable 5.1, but I've been spending more time in ChatGPT and Codex during this early access alpha testing period. It's just so bloody helpful to have this virtual colleague that can genuinely take jobs and run with them.

In my testing, Astra asks a lot fewer questions than Claude models and just gets on with it. I like that. And I can send it a new instruction while it's working to change direction (so-called steering prompts), midway through work to clarify what I want. It's very steerable, so handing over more of the work doesn't mean losing the ability to shape it.

There's some useful evidence beyond the demos. In a customer test published by OpenAI, legal AI company Legora says Astra checked 41 documents in minutes and found all four deliberately planted financial errors.

I've done similar things by going through my entire existing website, scanning for broken links, SEO issues, internal linking opportunities. It doesn't seem to struggle with large bodies of documents, data, or code, which is really useful.

Another thing that's really impressed me is the improvement in browser control and computer use, meaning that if there's not a connector or MCP server for your app, it just goes and takes control of it. Whereas previous iterations of that technology have been slow, clunky, and easily derailed, Astra always seems to get the job done, which again feels like a real colleague. Love that.

Things to try: give it a complete, useful assignment. And use /goal, which basically tells it to keep working until that condition has been met.

It could create an entire ad campaign with all the images, copy, a research-based plan on how to spend the budget and do the targeting etc. It could even build you a dashboard to track it all. I found that, in those kinds of tasks, it was actually quite prone to creating throwaway websites to showcase all the stuff that it had built for you, which blew my mind.

Getting access: the rollout starts with selected organisations, with ChatGPT Plus, Pro, Business and Enterprise access following over the coming days. OpenAI says usage is included in existing subscription allowances, with credits available for more. I would expect you to start seeing it early to mid next week, but I'll keep you posted!

Semafor's investigation uncovered a network of these “ghost creators” whose videos attracted more than 45 million YouTube views. The business is simple: find a divisive topic, write rage-baiting scripts with AI, pay cheap actors to film them and then collect the YouTube ad revenue.

You come to this story thinking it's going to be about AI-generated fake people. It turns out it's just as cheap and easy to pay real humans to read out a script to their phone camera and send it, to be quickly edited and then published. If the angle and script are geared to generate outrage, it can amass big viewing figures.

This kind of thing's been on my mind a lot because I feel like so many video creators are reading off AI-generated scripts. How many times do you hear the phrase "and the best part" or "and the craziest part" halfway through a video, to pivot to some sort of concluding tidbit? Or “Google just quietly dropped a new AI model…”. No one talks like that.

But I console myself with the thought that this just gives those of us creating authentic, valuable content, all the more reason to show up and win the audience's trust and loyalty.

Anthropic's latest model promises better reasoning, coding and sustained work across complicated tasks. Fable 5.1 is the broadly available version; its twin, Mythos 5.1, has different safeguards and restricted access.

What it means for you: from my usage, this model seems a bit faster, a bit easier to chat through things with, and a bit better at long-horizon multi-step tasks in Cowork and Code. A small but significant update.

Try it on a competitor review, a report assembled from several sources or a small internal tool. Those are the kinds of multi-step jobs Anthropic says it handles better. It will be interesting to see how it compares with OpenAI's new model, GPT-6 Astra, once more people have had time with it.

Instagram is tightening its AI-profile labels and limiting the reach of undisclosed synthetic personas. This targets accounts presenting an AI-generated person, not every creator who uses AI to help make content.

It's part of a wider pushback. LinkedIn already lets people flag posts that seem like AI slop, and there has been a similar push towards authenticity on YouTube.

My take: this is great news for people making something genuine. Whether you're an individual or a business, it's another reason to take creator-style content seriously. Show your face, your process and your actual experience. Give people something a content factory can't convincingly copy (and as mentioned in the headline story, write your own scripts!).

OpenClaw 2.0 makes installation easier and rebuilds its browser app, bringing always-on personal AI agents closer to non-technical users.

For me, Claude and ChatGPT's desktop work and coding tools have come a long way since OpenClaw first landed. My Claude Cowork, Claude Code and Codex setups already run agents autonomously and have bags of context about my work. The gains of having an always-on personal agent just don't seem worth the setup anymore.

My take: if you're starting out, I'd get comfortable delegating real tasks to the tools you already use (e.g., Claude, ChatGPT) before taking on another system to maintain. OpenClaw's openness and control remain reasons to explore it, but they're a different proposition that still require some technical skills.

Google has rolled out controls and AI-search reporting worldwide. Websites can opt out of AI Overviews and other generative search features without that choice becoming a ranking signal for ordinary results. Search Console now shows AI impressions and which pages appear, though not a complete picture of resulting traffic.

My take: for most businesses, I wouldn't opt out. Yes, it is annoying that your content is helping power Google's AI answers, often without getting anything back. On the flip side, you do gain visibility and some traffic, so that's a deal I would personally leave as it is.

Higgsfield, Kling and Seedream et al get loads of attention due to massive influencer marketing campaigns, which full disclosure: I myself have been part of as a creator.

But Google (who don't invest in creator marketing around this) already has an impressively capable AI video suite in Flow. You can reuse characters and visual references, control the opening and closing frames, and upscale finished clips for 4K export. The latest update adds more control and lower-resolution drafting before you spend credits on the final version. It's honestly very good, and you can get started for free.

My take: it's simple to use and capable of producing genuinely useful video assets. Before adding another AI-video subscription, give Flow a proper look. It deserves way more attention than it gets.

"Enshittification" was named the Macquarie Dictionary’s Word of the Year for 2024.

Coined by Cory Doctorow, the basic idea is that a product, particularly tech products and online platforms, starts out being great for users to attract as many of them as possible. When monetization pressure from investors starts to pinch, it shifts to looking after businesses to bring money in, and then further shifts to squeezing both businesses and users to reward shareholders. Ultimately, it becomes a shitty experience for everyone.

Famous examples include HP printers, Adobe software, Uber taxis, TikTok.

But a16z, (who invest in tech platforms!) have published an essay arguing that “enshittification” isn't real. Guest writer Ruby Justice Thelot points to growing usage and reported satisfaction in social media apps, suggesting that they've actually gotten better.

I think we were all bought into the idea of enshittification. But I appreciate the counterargument in this piece. If I think about my personal experience of social media, my feeds are rich with incredible content from incredible people, and the much-maligned algorithms are what makes that possible. Even the advertising feels more relevant and easier to control.

What it means for you: I think there's a good argument for saying we’re living through a golden age of social media, where anyone, regardless of audience size or resources, can build an audience either around themselves or their business.

That was the original promise of the internet that people like me got excited about 25 years ago. It's just that back then there weren't the discovery mechanisms for the best ideas and the best content to rise to the top and reach the right people.

Now that system is in place, and open to anyone with a smartphone. It might not be this way forever, so if your life or work would benefit from building an audience, it's time to start, now!

MapQuest reached No. 1 on the US App Store after refusing to rename Lake Ontario “Lake America”. An easily overlooked brand suddenly became the alternative people wanted.

I myself was surprised to see Google adopt these political renamings, first the Gulf of Mexico and now Lake Ontario, for US users. Google cites its policy of following official government naming data. But as far as I can tell, they weren't legally obliged to do this. From a brand point of view, it was such a daft own goal. People are even boycotting Google because of this.

My take: it's good to see an underdog with a good product rewarded for a basic bit of standing by its principles. A clear stance can sometimes do more for a brand than an expensive ad campaign.

Elon Musk now predicts AI will outperform humans at virtually every digital task by the end of 2027.

Sound familiar? In April 2024, he was already predicting AI smarter than the smartest human probably in 2025, or within two years. The wording isn't identical, but the recurring promise is. Anyway, still hasn't happened and probably won't again.

My take: he's not stupid, but I don't think he's very truthful either. It's just another example of an AI leader hyping the product they're selling, happily buying into the myth that AGI/ASI (Artificial Super Intelligence) is around the corner.