Edition 16 · 28 September 2026
what stays valuable when AI can do the work?
Back from Bali, Claude gets simpler, all roads lead to Zuck and what to do in a world of AGI.
📌 TLDR:
Back from Bali
Claude Opus 5.5 is crazy good (at least for design) + Claude becomes one convo
Aus govt got hacked?
All roads lead to zuck pt 2
And my thoughts on what stays valuable when AI can do the work
💼 1. What’s new in the business
I'm back from Bali. Not a huge amount to report on the business front this week, but I have been playing around with Claude Opus 5.5, and it is nuts!
It's been absolutely cracked for design. I put together custom slide decks in less than five minutes (more on that below).
Also been mucking around with video editing in Astra and Opus 5.5 and I feel like we’re less than 6 months away (probably closer to 3) from showing a reference video and asking one of the models to edit it in the same style.
Anyway, it’s been a crazy week in AI (like usual), let’s get into it.
🤖 2. AI news
Claude is becoming one conversation
Anthropic is bringing Cowork into ordinary Claude conversations. The rollout lets people ask a question or hand over a larger task without choosing a separate mode first. It's rolling out to Pro and Max users across web, desktop and mobile.
This was well overdue.
View @Baxti_7’s post on X
It never made much sense to me that someone should need to understand where they should ask each question. Most people don't care which mode creates a spreadsheet. They care whether the spreadsheet is useful.
Plus the cutoff memory between Chat and Cowork was weird too. I ended up just advising anyone who was confused to just use Cowork everytime. It did everything Chat did but had the power of a computer behind it.
Expect to see ChatGPT work do the same thing.
Stripe's data-centre chart shows America is building on a different scale
Stripe revealed the data-centre chart of capacity by country, as below.
It separates capacity that's operational today from additions in the pipeline, measured in gigawatts. Reading the bars approximately, the US has a little over 30 GW operational and roughly another 70 GW planned. Its combined bar runs past 100 GW, far ahead of the other countries shown.
What stood out to me is how fkn far the American bar is ahead. It really is a microcosm of how much emphasis the Americans are putting on AI.
They really are viewing it as the industrial revolution.
I think there are two useful implications.
First, the AI race is partly an industrial race. The ability to finance and build physical infrastructure matters alongside the ability to train a good model.
Second, I wouldn't confuse a country having less domestic capacity with its businesses having no AI opportunity. An Australian business can use models running overseas. The questions become cost, reliability, where sensitive work can run, and how dependent you want to be on someone else's infrastructure.
Having said that, I feel this image serves as a snapshot of how much importance countries place on economic growth. With the Americans placing economic growth above all.
The education system: reimagined
a16z announced the Horowitz Andreessen Academy, a full-time, in-person school in San Francisco for young people coming out of high school. The model combines personal projects, courses and work experience.
Over the last few months I’ve beens peaking with a lot of people about this - I don't think our education system does enough to encourage agency. We spend a lot of time learning how to answer a question someone else has set, and much less time deciding which questions are worth pursuing.
View @gaganbiyani’s post on X
AI makes that distinction more interesting. If a student can get help writing code, researching an idea or making a prototype, they have more room to actually try something.
Alpha School is another experiment in this direction. It describes a model with two hours of academics and more of the day available for other skills and activities. My friend said he met them in SF and they were about 6-8 years old, rolling around on roller skates asking about gross margins (crazy).
The best from the founder of Alpha School is that if the students over the duration of their education, cannot generate at least $1mil in net profit their school fees could be returned for free (insane offer I know).
I think in our lifetime we will see major disruption of the education system.
The Aus govt got hacked by AI?
The government disclosed an incident involving an OpenAI research agent and a Medicare statistics portal. It happened on 18 June, and was publicly discussed on 24 September.
According to the government's account, the agent worked around access blocks, reached public and non-public information, and wrote files to an internal server. At the time of the announcement, no personal information was believed to have been accessed, and the investigation was ongoing. The Prime Minister's statement
I think it's worth separating what happened from the more dramatic interpretations of it. This wasn't evidence that everyone's Medicare records had been stolen.
But the actual breach happened in July and Aus govt was only notified in Sept.
I think this is just the first in a litany of crazy AI breaches and cybersecurity flaws that will be coming our way.
Xi touches down to the US for the first time in a decade
Not so much directly AI related (although probably a good thing nonetheless).
But Xi Jinping touched down in the US for the first time in a while.
View @AlexAllone1’s post on X
The US-China summit produced movement on trade, which is a welcome change from another round of escalation.
The White House says the two countries reached consensus on recommendations for more favourable tariff treatment on $30 billion of non-sensitive goods in each direction, alongside work on trade and investment mechanisms.
Both sides have an interest in keeping some commercial channels functioning, even while they compete fiercely elsewhere.
For AI specifically, I wouldn't read a deal on non-sensitive goods as the end of competition over chips, compute or strategic technology. It does show that cooperation and competition can happen at the same time - which is a good precedence for any AI related talks.
👷 3. Builder’s notes
Claude Opus 5.5: the design improvement is crazyy
This one has been absolutely cracked for design. it has been going around like wildfire on Twitter.
See this one tweet here where a person got Opus 5.5. to make a timeline of civilisation.
View @IterIntellectus’s post on X
Personally, I put together custom PowerPoints in literally less than five minutes (won’t show for the sake of confidentiality).
Prompt
Using the brand system design, make a slide deck that has more animations, designs and is more interactive. Use the uploaded deck as reference (I then uploaded a deck I had used previously).
What’s so cool is that the time to value keeps decreasing.
Before you needed massive complicated prompts to get a reasonable output.
Now you can just show a reference and get it to generate a similar output.
Anthropic says Opus 5.5 performs around the level of Fable 5.1 on most work, generates output more than 30% faster than Opus 5, and costs about 40% less on its typical-workload tests. The 40% figure combines pricing and token efficiency.
For me, the useful combination is quality, speed and the cost of trying again.
If the first version takes five minutes, asking for a different structure or a cleaner visual explanation becomes much less painful. You can explore two or three directions before deciding which one deserves attention.
So there’s more autonomy given to judgement, to ask questions like,
What is the point of this deck? Which numbers support it? What should the reader understand after each slide? A model can produce something attractive and still miss the argument.
If you haven’t already, give it a good red hot go.
Higgsfield is giving away money?
Higgsfield (a video gen platform) now lets developers generate images and video through its API, with a prepaid dollar balance and separate billing from its website subscriptions. It also offers a way to estimate a request's cost before running it.
I tried it out for a specific image generation. But tbh it’s still really expensive.
There’s the new jengetsu model which is all the rage and generated outputs like the below.
View @BasedDaedalus’s post on X
But what is cool is the possibilities this opens up.
Imagine a workflow that takes a product brief, produces a set of creative variations, and puts them somewhere a person can review. The generation step becomes part of a repeatable process instead of another website someone needs to open and operate.
I wouldn't assume it's automatically cheaper than a subscription. That depends on the model, the number of attempts and how much you actually use (the baseline sub is $78 AUD though), so it is cool that they opened it up. They also let you get a free video gen on the Jengetsu model when you connect MCP access (i.e. connect higgsfield to chatgpt).
Having said that it did work really well as a marketing stunt…
Which leads me to another talking point.
Higgsfield has just passed a $1 billion annualised revenue run-rate, 18 months after launch.
View @alexmashrabov’s post on X
Which is absoluletly insane growth.
From my understanding they have a team of 150 people working on just content. I still don’t think it’s fully been priced in how much the algorithms and the social media platforms dictate our lives and will continue to do so.
Which leads me to my next point, the almighty social media overlord… Zuck
Meta Muse: all roads lead to Zuck, part two
I wrote previously that all roads lead to zuck…
And after this week it may be even more true.
My mum mentioned Muse the other day.
That probably tells you more about why I'm bullish on Meta than another benchmark comparison would.
The reality is, most people don't follow AI releases. They don't care which company is leading on a particular test. They want something helpful that fits into their life without requiring them to learn a whole new system.
View @finkd’s post on X
Meta has a very strong starting position there.
Muse is its personal agent, accessible through its own app and WhatsApp. Meta launched it in the US and says most everyday use is free, with paid plans for people who want more.
I think there are three core advantages here.
Distribution. Meta already has places where people spend time and talk to each other. Getting someone to try an assistant inside a familiar app is a different problem from persuading them to adopt a new one. They also are the one messaging app that doesn’t allow external access. E.g. if I’m trying to create a personal agent - Meta won’t let me access the data on their sites (feel like there will be an anti-trust thing on this in the future - they literally have a natural monopoly).
Familiarity. People who have never heard of an AI benchmark know WhatsApp, Instagram and Facebook. That doesn't automatically mean they trust Meta with everything, but it removes some of the unfamiliarity around trying the product.
The wider business. Meta actually doesn’t really need to make money off of this. They could literally give it away for free and funnel all your data back into the big algorithm and make more money off of ads. They’re also making a play at being the next platform here (see below)
There are obvious caveats. Having distribution doesn't mean people will hand over meaningful parts of their lives. Reliability matters. Privacy matters. So does whether the assistant actually saves you time once you've checked its work.
And the other companies aren't standing still.
But…
The developer side could deepen this advantage.
Meta is expanding connectors and tools around its products, while its glasses developer updates open more ways to build experiences. Useful third-party services give people more reasons to stay.
When you have a personal agent it has to be able to do everything.
There are so many tool calls it has to have.
You can’t really build all the tools by yourself.
So opening up this platform to developers is a play to make this the new platform, which is a huge play, and I think they’re in a better position than most to execute on this.
Meta releases new glasses (lowkey sick)
The hardware is worth its own mention.
At Connect, Meta introduced VR glasses weighing about 100 grams, with a separate tethered puck handling compute, battery and storage.
I don’t think this is getting enough attention.
They are providing a cinema, a console and a workspace in 100g.
They went for the famous Steve Jobs, an iPod, a phone and an internet communicator (but Zuck just doesn’t have the same drip tbf)
But, it is still really interesting to see them be so bold.
I can see a world where you do some of your work through glasses, talk to your Muse agent to handle the admin bullshit, and connect with your friends and build your business through Meta's apps.
Separately, Meta announced Muse for its AI glasses, Ray-Ban Meta Audio, and updates to its wider glasses range. The VR device and the everyday AI glasses are different products; both point towards more ways to interact without opening a phone.
I think glasses could become a much more natural interface for an assistant. If you're already wearing them, asking for help can become part of what you're doing rather than a separate task.
Whether people want to wear them for hours is another question. Comfort, privacy, battery life and the usefulness of the experience will decide that.
But the combination interests me: the agent, the devices, the developers and the distribution all reinforce them as the ecosystem and the new platform.
And we still haven’t even talked about the creator economy set to hit $3t by 2030s (from $250b now).
Nor the fact that most businesses need social media to function.
But I really do think all roads might actually just lead to zuck.
4. Differentiated take: what stays valuable when AI can do all the work?
Recently, I’ve had a variety of people ask me about business in a post-AGI world.
Now I’m not going to pretend to have all the answers but rather just take informed positions.
Generally speaking, I've been finding it harder to think confidently about what the world may look like in three to five years (not in a doomerism type of way by the way).
My useful planning horizon has been moving closer to six months, maybe twelve. Beyond that, I can have a view, but I wouldn't want to pretend it's a reliable forecast.
You can also get stuck arguing about whether something counts as AGI or not AGI. I generally think AGI is here - it just hasn’t permeated the economy in the same way or in a normalised sense.
But imagine saying to someone 4 years ago - yeah my agent just booked me a flight.
Your what? Just did what?
We can literally fire off a text and have something done for us within seconds - which would have been reserved for the ultra rich literally no less than 5 years ago.
Even terms like agents are taken for granted. We have a multitude of these little AIs that are running around doing whatever our heart desires all for the price of a monthly subscription.
Having said all of this, the question I keep coming back to is more practical:
If AI becomes normalised, where is a good place to build?
I think that's a useful question whether you're building software, running a service business or making things yourself.
Take a service business. The deliverable might be a report, a website or a set of designs. But the customer may also be paying for someone who understands their situation, knows what good looks like, makes decisions and takes responsibility when something goes wrong.
Those are different parts of the value. If you only sell the production step, cheaper production creates an obvious problem. If you own the outcome and the relationship, you have more to work with.
It's why I keep coming back to things that only become more valuable as technology gets better.
Community, art, entertainment, travel, creating, sport, wellness and services built around real relationships.
These are at the core of what it means to be human.
Yes, one may enjoy their job, but is there anything better than driving down the highway yelling your heart out to your favourite song on a sunny Sunday arvo.
Or watching the nail-biting finish of an AFL game?
Or watching the sunset fall over a mediterranean sea while you’re sitting on a balcony in Santorini, sunkist from a day at the beach?
Not to get too romantic, but in a world where AI can build anything I think a good place to build is anywhere that deepens the ‘je-ne-sais-quoi’ of the human spirit
On a more tangible note, I think there’s a version of the future where more people get to become craftsmen, artists and business owners because the administrative and technical work becomes easier. Someone with taste and a clear point of view can bring an idea into the world without needing a whole team to get started.
I think everyone will likely be a creator. Everyone will be growing their business online, building a community and a story as people make more things. The imperfect, rather than the perfect, will be sought after.
Of course, everyone else gets those tools as well. Being able to make something is only part of it. You still need a reason for people to care, a way for them to find you, and enough trust that they come back.
So that's where my thinking is at: use the tools aggressively, but keep asking which part of the business becomes more valuable as the tools improve.
If the cost of producing your main deliverable fell by 90%, what would your customers still come to you for?
I'd be interested to hear how other people are thinking about this.
🎵 5. Carve outs
I’ve had a number of people ask me about what to read about AI predictions etc. Please see four readings I prescribe (hope these are helpful).
Shoutout to Wruck for these readings from way back when.
Also good are Zuck and Bill Gates’ manifestos (always a bit to know if these people have misaligned incentives when they write these articles but they’re at least positive)
Sports
I’ve never been so toey watching the Knights get up for a game (carn Kalyn - if you haven’t watched these highlights, they’ll get you so fired up for the grand final
I’d hate to be a freo fan 😕
Until next time folks,
Alex