📩 flying cars, spies at openai and the personal brand paradox

Hola amigos,

There’s a lot to get to this week.

Apologies for the delay in getting this out - I’m really wary of releasing stuff I’m actually proud of, so spent a bit of extra time on this one today

📌 TLDR:

This week I’ve been building something to help me spend less time making content, while questioning how much content we actually need.

Healthy relationship with social media, clearly.

Also, Tesla made a car leave the ground, someone allegedly hid a recording device at OpenAI behind a TV, and you can now pay to put your younger immune cells in storage.

Oh and Meta came out with their newest frontier model and Claude can edit videos like a pro.

Plenty to get through.

Hope you enjoy :)

💼 1. In the business

I’ve been building my own social media creation app.

The idea is to bring together content editing, tracking what other people in the space are doing, and the bits of publishing that currently involve jumping between too many tools.

Mostly, I want to minimise the distance between a thought in my head and something I’m happy putting out into the world, while ensuring the content is engaging and original.

Claude has been surprisingly helpful with the video editing side of this. There’s still work involved, but being able to describe what I want and get closer to it without manually doing every step is pretty cool.

I whipped up the /goal function in claude and just voiced out what I was thinking.

I’m somewhat conflicted though.

A lot of growing on social media seems to come down to understanding the mechanics of the platform. The format, the hook, trial reels, catching a trend at the right time. Having something useful to say is only part of it.

Everyone is so fucking mimetic - it’s actually infuriating.

And it’s often easier to just copy someone else’s proven topic than to have an original thought.

Case in point, I could spend 60 mins developing an idea, scripting, editing and more, and it gets a solid 300 views (this did in fact happen).

Or I can see a trend repurpose it and it can reach over 100k views (which also did happen).

This is despite the fact that I think the video that got 100k+ views is rather shit.

It’s made me think about my social media consumption - particularly on instagram.

Instagram often leaves me feeling like I’ve numbed my brain for half an hour.

X (which I got more recently) I’ve found to be more useful for learning and finding interesting things.

With that said, I’m building something to make creating easier. The core thesis is to allow those that are really good at business to also be really good at getting their name out there.

Just wary because I’d also like to feel good about what I’m helping create.

More on that below.

🤖 2. What’s new in AI

We have fkn flying cars!

I’ve been asking why flying cars aren’t a thing for years.

My cofounder can attest to this.

It seems we were working to solve a variety of different issues and this one didn’t seem to be high on the agenda for anyone.

I’ve always thought in the long run it’s one of those things I’d love to contribute to building.

But alas, Elon appears to have got there first.

Twitter tweet

I think this was quite possibly the coolest product demo I have ever seen.

The demo had it driving up a near-vertical ramp, hanging upside down, then launching off and rotating back onto its wheels. Later, it barrel-rolled over a line of other Teslas.

Twitter tweet

And then there’s the acceleration.

Tesla is claiming 0 to 60 mph in 0.55 seconds for the top-spec Rocket version. That’s roughly 97 km/h in about half a second.

Hahah wtaf.

Hagerty’s interview with Tesla’s engineering lead explains how they’re attempting it. Fans underneath the car create suction to pull it towards the road, helping the tyres grip even at low speeds. Then high-pressure nitrogen thrusters provide an extra shove, so acceleration isn’t dependent on tyre grip alone. (cool video feel free to check it out)

The flight-capable version takes that idea further, using thrusters to lift and control the car in the air.

I’ve seen a few people that have been complaining online that they hadn’t got their vehicle yet after ordering years earlier and this was a largely useless demo.

But seriously get a grip - the car is flying for God’s sake.

Hard not to be inspired by this shit.

The alleged spy at OpenAI

This story gets weirder the more you read.

According to the US Department of Justice, OpenAI engineer Abdulrahman Alfozan was arrested on 9 October, accused of secretly recording conversations at the company.

Staff had found an encrypted recording device plugged into a USB port behind a television in an executive meeting room. It was concealed in toilet paper.

Which is quite a choice of camouflage.

The room reportedly hosted discussions about confidential research, product plans and commercial relationships. So potentially a very valuable place to leave a microphone.

Then there’s the chatbot history. Prosecutors allege that an account associated with his wife, sharing identifiers with his own account, asked about building and concealing an audio recorder. Some prompts included the discovered device’s unique network address.

Security also allegedly found a second, similar recorder in a bag containing his wallet and ID.

Honestly just a bizarre trail to leave behind.

You’d think if you were actually a spy you wouldn’t google the day before how to be one, nor would you wrap up your recording device in toilet paper.

Just goes to show you the stakes that are on the line here.

Meta and Alexandr Wang is everywhere

My feed has been full of Alexandr Wang. The former Scale AI CEO, now leading Meta Superintelligence Labs, has appeared on David Senra’s podcast and Cheeky Pint with Zuckerberg.

Twitter tweet

Feels like there may have been a coordinated PR bot campaign for how much they’ve been coming up on the feed

But alas, there’s reason for the podcast appearances. Meta has released Muse Wonder and Wonder Full, plus Muse Spark Open. Spark Open is the one with downloadable weights.

That means developers can run and customise the model on infrastructure they control, subject to the licence. With Claude, you access a hosted model. You can’t download its underlying weights and run it yourself.

For a company processing thousands of internal documents, that could mean more control over where its data goes and potentially different economics at scale. Open weights don’t mean free computing, though. You still need hardware and people to keep it running.

Is it better than Claude? Probably wouldn’t say that for now. Meta’s published evaluation methodology primarily compares Spark Open with other open-weight models. The useful test as always - is how reliably it completes your actual work, with how much supervision and at what cost.

But for now they have said that it is at the frontier for some tasks - which is very impressive given where they were 12 months ago.

Ultimately though, Meta doesn’t need to beat Claude at everything. If it’s good enough for some jobs, businesses could move those across and keep Claude for the harder stuff.

Putting your younger cells in storage

Not directly AI related but thought it was cool to share nonetheless.

This sent me down a rabbit hole, and I ended up asking my dad about it since he’s a doctor.

Reservoir preserves immune cells collected through a blood draw. The idea is to store them while they’re younger, potentially making them available for future treatments.

The company comes from Cellino, which is backed by Khosla Ventures.

Twitter tweet

The reason younger cells might matter is that our immune system changes as we age. Some immune cells become less effective, and the pool able to respond to unfamiliar threats changes. The bet is that preserving cells earlier could leave you with useful starting material for future treatments.

A few things this could help are cancer treatment (you use your own immune cells for certain therapies, so having younger versions of them would be great), autoimmune conditions and restoring immune function later in life.

So imagine being 65 and needing a treatment made from your own cells. Having a sample from your 30s could potentially give doctors another option to work with.

As medicine becomes more personalised, and AI potentially helps develop new treatments, the cells you preserved decades earlier might become useful.

The big unknown is whether those future treatments will actually need or benefit from that sample. Reservoir is selling preservation today, rather than a guaranteed future therapy.

Regardless, cool to see someone building something that could be truly helpful - rather than automating your emails.

Can Anthropic keep doing 10x?

Anthropic - the makers of Claude have been on a ridiculous run rate (if you didn’t already know)

The question with Anthropic is whether it can keep growing revenue tenfold year on year.

At some point, the numbers get so big that maintaining that pace becomes a ridiculous task.

Reported Q3 revenue was $16.8 billion, up 45% on Q2, after growth of more than 140% the previous quarter.

Still enormous growth. But the rate has slowed, and I wouldn’t assume the next year automatically delivers another 10x. Those quarterly figures alone don’t establish its year-on-year multiple, either.

What really caught my eye was the reported estimate that 5–7% of revenue comes from distillation.

Distillation is where another AI company will use the outputs of a superior model and then copy that into their model.

Very popular because it lets companies jump the gun, not do all the training that the other companies have done and still achieve the same (or similar) output.

In short: some customers are paying Anthropic while building something that could compete with it.

Then you’ve got Meta releasing increasingly capable open-weight alternatives, including Muse Spark Open.

If businesses can run an open model that’s good enough for their particular task, some may decide they don’t need to send as much work through Claude. Running it yourself still has costs, but it gives customers another option and potentially more negotiating power.

Plus Meta was reportedly a large customer of Claude - so if they’ve reduced their spending, they could put some pressure on Anthropic’s growth or pricing.

Heading into a potential IPO, even a slight change in the growth story could matter. Investors aren’t just buying what the business makes today. They’re paying for what they think it can become.

Anthropic can be doing incredibly well and still fall short of some very aggressive expectations.

🛠️ 3. Tools and things worth trying

Be more ambitious with what you ask AI to do

I liked this piece by Felix Rieseberg, who works at Anthropic (makers of Claude)

Twitter tweet

It’s a useful nudge to think beyond the small task you were about to hand over.

If you’re asking for a report, what decision is the report supposed to help you make? Could you get help investigating the options, checking the assumptions and building something you can actually use?

Previously, I’ve found myself asking for a slightly faster version of how I already work.

Very common for a lot of businesses we speak with - they want to put duct tape over a problem that is structurally flawed.

Prompting your way to the answer is probably pretty inefficient - better off fixing the underlying foundations.

Claude inside Word and Google Docs

Claude can now work directly in Word, with document editing, tracked changes and support for working through comments. There’s also a Google Docs, Sheets and Slides integration in beta.

For people who spend their day in documents, this feels useful for a very ordinary reason: less copying things into a chat window and then trying to get the result back where it belongs.

“Work through these comments” is much closer to how you’d ask a colleague for help.

I’m interested in how much of the editing process can happen inside the actual document, with enough visibility to review what changed.

Personally my viewpoint is that most of your work is done inside of a chatbot and all your tools will be connected (e.g. I don’t really ever look at my emails anymore)

And I think this is the end state that we all move to.

But for now this seems like a very useful addition.

Can an agent use your website?

Profound looked at how Muse and Instinct browse and shop.

One interesting finding: agents can use a browser to navigate existing websites through the page’s accessibility structure. Clear labels, understandable content and usable controls matter.

You may already have much of what they need.

The harder question is attribution. In Profound’s observations, these agents didn’t identify themselves by name when browsing, making some visits difficult to distinguish from human traffic.

So your next customer might arrive through an agent, and your analytics might struggle to tell you that.

It’s exploratory research, but I thought an interesting case study into how agents actually operate.

A quick website test

Related to that, try Is Agentic, the site-readability checker referenced in Vercel’s guidance.

Agents are quickly becoming the way that people find and see your brand.

Except most websites weren’t built so an agent could read them. They were built so that humans could interact with them.

So you probably need to check how good agents are at reading them.

Simply use the link below and put your site in.

My initial score was around 73.

I gave my agent the improvement prompt the tool generated, had it make the changes, and got to 100.

That score doesn’t guarantee agents will recommend your business. But it was a very easy way to find things that could make the site harder for them to use, then fix them.

A useful little trick if you’ve got a website.

Editing videos with Claude

I’ve been using Claude to do more of the actual video editing, and it’s getting pretty fkn awesome.

It’s honestly a fantastic experience - the time to move from an idea to finished product is approaching zero, and I’m all for it.

If you’ve ever edited a video, you’ll know how painstaking it is. So a few things I’m finding useful.

WhisperX has been particularly useful. It gives you a transcript with timestamps for individual words. That makes it much easier to line up captions and find the gaps where a cut should go, rather than guessing and chopping off the end of a sentence. (see link for skill)

In one recent rough-cut run, the workflow took 209 seconds of raw footage down to 68 seconds. Which previously would have taken way longer than I would like to admit.

I’ve also built a custom workflow called my reel-edit skill, it uses those timings to remove earlier takes, stumbles and dead air, then adds captions, screenshots, punch-ins and other visuals in my house style. It also lets me make alternate hooks without rebuilding the whole edit. (reply to this if you want me to send it to you!)

FFmpeg: one of the tools underneath the workflow for cutting and preparing footage. Claude can help write the commands, which makes it much more approachable than learning every option yourself. FFmpeg resources.

I still hold the creative direction but these have been so helpful.

Agents haggling with agents

xAI’s hagglebot got me thinking about what happens when negotiating becomes cheap enough to do routinely.

Twitter tweet

Most of us don’t compare every supplier, chase every quote or renegotiate every renewal. The potential saving often isn’t worth the time.

Agents change that calculation.

I.e. your agent can bargain on your behalf until it achieves a good outcome.

Interestingly, this could be one of the easiest ways to put deflationary pressure on the economy.

For example I recently used the new ChatGPT dot feature when I found out the swings had been taken down by my local council and were not going to be replaced.

Instead of just copping that, I took a photo, told my dot to chase this up and it sent the council an email.

And voila, what would have taken me 20 mins before took me all of 30 seconds in Chat.

So the council might be in for a barrage.

🤔 4. Differentiated take of the week: the personal brand paradox

Everyone seems to have a personal brand these days.

I myself am not immune to this (if you’ve been following along appreciate you x).

But I have a slightly uncomfortable relationship with the idea of growing a personal brand.

Obviously there’s enormous upside. Especially when you run a business.

People generally only buy from you if they trust you. And creating a personal brand and putting yourself out there is one of the easiest ways to build trust at scale.

People get to understand how you think before they ever speak to you. They see what you’re building, what you care about and whether you might be useful to them. Over time, that can become an extremely valuable asset.

Gone are the days of requiring permission to start a column, obtaining a television slot or asking for persimission to acquire an audience. You can start sharing what you know today.

That’s an amazing opportunity. And largely a positive thing - I’m not advocating for media to be held behind a select few legacy media conglomerates.

It also means everyone can post anything, all the time. And increasingly, they can get AI to produce it for them.

The result feels like a cesspool of people telling you how shit you are, followed by people telling you how good they are.

I.e. You’re falling behind. You’re using the wrong tools. Your business is broken. Here’s someone who apparently fixed all of that before breakfast lol.

And don’t even get me started on the cesspool of AI content that is out there.

I understand why people make this content. The platforms reward it, and if your business depends on being seen, there’s a fairly obvious incentive to play along.

But I don’t particularly enjoy consuming it.

And I often feel people forgo moral integrity in pursuit of a view.

The great paradox that I’m seeing more and more is that a lot of the content I think resonates with a lower level buyer, i.e. someone who is not as educated, or not as qualified as a buyer for you.

And I’ve increasingly found that modesty has become a commercial disadvantage. Doing good work might be admirable, but it doesn’t give the algorithm much to distribute.

Growing up, I felt like modesty was such a paramount trait. One especially encouraged in Australia.

There’s absolutely nothing wrong with being proud of your achievements. But when it comes packaged as something that is used as a fucking hook, it feels very… distasteful.

Don’t get me wrong - there’s plenty of people who share useful things. And the democratisation of media is largely a good thing in ensuring there is freedom of information.

So I don’t think the answer is to disappear from the internet and congratulate myself for being above it all. That would be a fairly convenient way to avoid the discomfort of publishing anything

I think Hormozi has a good way of thinking of it, he said it was the price he was willing to pay for the impact he wanted to have.

For me, I’m trying to work out how to participate in a way I still respect.

Sharing what I’m actually building feels like a reasonable place to start. What worked, what was harder than expected, what I’m still unsure about.

AI could help with this. I’d love to spend less time fiddling with an edit and more time making the thing worth sharing.

But if I use that saved time to produce five times as much content nobody needed, I’m contributing to the exact experience I dislike.

I’d rather use some of it to have something better to say.

How do you build an audience without slowly becoming the kind of content you avoid?

5. Carve-outs

An AI film example worth a look at

I used to feel pretty confident I could tell when something was AI-generated.

Increasingly, I can’t.

My guess is that within the next 12–24 months, there’ll be plenty of footage where I have no reliable way of telling just by watching it. That’s a prediction, rather than a claim that every part of filmmaking has already been solved.

It does make me wonder how much of our attention will shift from interacting with other people to just scrolling through an AI cesspool.

Until next time,

Alex