📩 is ai going to slow the fk down?

Yo yo folks,

Apologies we’ve been a bit late (been overseas a little bit, plus wanted to see if mid-week was better for posting)

This week, the TLDR is:

• Working from Bali with a bunch of business owners.

• Another AI researcher leaves with a warning (and the rebuttal)

• Google makes AI more useful inside the tools we already have.

• A few tools and skills worth sharing.

• And a thought I keep coming back to: everything is a sales funnel.

💼 1. What’s new in the business

The boys are working overseas this week, in Bali with a bunch of other business owners.

Been awesome hanging out with people doing crazy stuff.

Everyone from venture-backed startups in $50m+ valuations to solopreneurs bootstrapping $1mil+ a year.

We’ve also got some people with big followings, 80k+ on instagram, so it has been a really big learning curve in terms of content.

One thing I’ve done has been creating a core ICP document that documents all the pain points and possible angles to appeal to this ICP.

Every video should cater to this person in some capacity.

My catchphrase is making things that are visual and valuable.

The reality is the algorithm rewards extremism - so producing something that is visually appealing is very important.

And then providing as much value as possible.

If I don’t think there’s something that viewers can takeaway as valuable then it’s hard to be surprised when it doesn’t go off.

In terms of Bali, it’s my first time here but there’s a bit of a weird dichotomy. The villas are gorgeous but the actual external surroundings are a bit grim.

🤖 2. What’s new in AI

Another researcher leaves with a warning

Jacob Coxon resigned from Anthropic last week, warning that Anthropic and OpenAI are prioritising the race to build more powerful AI over safety.

Twitter tweet

At 180 million views and counting. Not to mention the plethora of other interviews and press tours that have been had.

This news story has gotten so big that my dad sent me a message about it.

My initial reaction was that it felt like more fear-mongering. I figured there was a PR angle, and wondered whether the push for regulation could end up protecting the companies already ahead.

With that said, the argument I think is worth taking seriously is about the incentives.

If complying with the rules becomes so expensive that only the biggest labs can afford it, regulation could entrench the incumbents.

This is the regulatory-capture concern I’ve mentioned before: rules presented as protecting the public could also make it much harder for a newcomer to compete.

It feels a bit ‘I want to have my cake and eat it too’.

The tension I see is that these companies want to build the most powerful AI ever, while arguing that access to powerful AI needs tighter controls. There may be good reasons for those controls, but I still want to know who gets to set them, and who benefits.

Twitter tweet

I also worry that poorly designed restrictions could slow American challengers while Chinese competitors keep developing.

Having said that, we have no idea what they're cooking up in these labs. We aren't in there building the models with them.

Coxon says he left before his Anthropic equity vested (which complicates the idea that his departure was simply about boosting the company's value), but who knows how true this is.

When I messaged Dad, I also pointed to the anticipated IPOs (both companies expected to IPO this year)

So calling this all good PR is too simple. Attention can make the technology sound impressive while also making people less willing to trust it.

The question is whether the proposed rules address a demonstrated risk, and whether a smaller competitor could realistically meet them.

I want the safety concerns taken seriously. I also want the commercial incentives examined with the same scrutiny.

Having said all of that, it looks like we are going to see some regulation. And it poses some interesting questions and opportunities.

1/ what does this mean in the race with China?

They definitely are not slowing down.

2/ does this open up scope for a new model company?

Paul Graham certainly thinks so.

Twitter tweet

3/ what does this mean for people who are trying to start up in the same space?

the counter to the above is, if there is a lot of regulation, this could prove very difficult for a newcomer to actually build something new that also complies with the regulation.

4/ what are they actually building in there and how close are we to recursive self-improvement (RSI - expect this term to become much more common place in the future)?

I’m getting the sense we’re closer than not, I think by the end of next year it may be achieved.

Google is making AI wayyy easier to use

Google announced new Gemini capabilities that work across Workspace apps.

You can turn an email thread into a Google Doc, create a spreadsheet from relevant project material, or turn a proposal into a presentation without manually moving everything between applications.

The features are rolling out across eligible plans, including Business Standard and Plus.

This is a practical continuation of what I wrote about previously.

The information is already sitting in your inbox and your files. Being able to ask for a finished piece of work from that information removes another layer of clicking, copying and explaining.

For a business owner, a useful test would be the weekly client update.

Give it the relevant emails and project files. Ask for the update. See how much work remains before you would actually send it.

OpenAI and Cursor are building for longer jobs

OpenAI released its Agents API in public beta on 10 September.

It gives developers access to the infrastructure behind Codex for managing context, using tools and coordinating multiple agents.

Twitter tweet

Cursor launched Projects the same day.

A Project keeps shared context, delegates tasks and can pick up work from a schedule or activity in Slack and GitHub. It’s also rolling out in beta.

Twitter tweet

What’s cool is the ability for work to continue without your input.

You give the system a body of work, and it can carry the relevant information between individual tasks.

My read is that the ability to define the work becomes increasingly valuable as more of the machinery for doing it becomes available off the shelf.

Questions that need to be answered are what should trigger the task? When should you be in the loop?

I’m yet to play around with these properly so will have more to report back on in the next week or two.

Open-Source keeps the pressure on

DeepSeek’s V4.1-Flash release has a pretty ridiculous cost story.

In the OpenDesign benchmark (see below), it scored 81.2 against GPT-6 Astra’s 82.7, reportedly at 1.4% of the cost.

That’s about 98% of Astra’s score for roughly one-seventieth of the price.

There’s a qualification here: this measures a particular set of design tasks. It doesn’t mean DeepSeek is “98% as intelligent”, or that it will perform equally well across everything you throw at it. The screenshot also doesn’t show enough methodology to judge how consistent that advantage is.

But it’s a result worth testing against your own work.

The architecture helps explain the focus on efficiency.

DeepSeek says V4.1-Flash has 552 billion parameters in total, but activates only 8 billion when processing input and 16 billion when generating output. It also reports using one-quarter of the previous generation’s cache memory. DeepSeek’s announcement

In plain English, the model has a large pool of capacity, but uses a smaller portion for each token. Its design allocates different amounts of computation to reading and writing.

That doesn’t necessarily mean it uses fewer tokens. It means the computation behind those tokens can be cheaper.

So how does this apply to you if you run a business?

Test the models on the jobs you actually need done. If a cheaper one produces work you can use, there’s no prize for paying more.

At the same time, a cheap answer that needs twenty minutes of fixing can become quite expensive.

The number that matters is the cost of getting an acceptable result, including your time.

My read is that this creates more room to use cheaper models for routine work and reserve the expensive ones for the tasks where they make a meaningful difference.

It also keeps pressure on the frontier labs. A small improvement becomes harder to sell at a large premium when a cheaper alternative is already good enough for the job.

Scams are getting crazy good (warn your fam pls)

One of the screenshots above describes a man receiving a call that sounded like his wife, asking for his credit-card details to pay for petrol.

Except his wife was at home with him. And she drove a Tesla.

It’s an unverified social-media account. We don’t know what technology was used, or whether the incident happened as described.

But the broader risk is real. The FTC already warns about scammers cloning a family member’s voice to make a fake emergency convincing. FTC guidance

Tools like ElevenLabs and Fish Audio show how accessible voice cloning has become. Both offer voice-cloning products; that doesn’t establish that either was involved in this incident.

I expect impersonation scams to become more common as convincing audio becomes easier to produce.

The uncomfortable part is that recognising someone’s voice can no longer be enough to establish who is calling.

And when you think your partner or child needs help, you’re probably not listening carefully for a slightly strange cadence.

A few precautions are worth agreeing on now (spoke with my mum about these last night):

• Hang up and call them on a number you already know. Don’t use a number supplied by the caller.

• Agree on a private family phrase. Something that isn’t sitting on your social profiles. Use it as an extra check, alongside calling back.

• Treat urgency and secrecy as reasons to pause. Especially requests for money, card details or verification codes.

• Talk this through with your parents and kids. Make it normal to verify an unusual request, even when the voice sounds familiar.

This is going to be a crazy next decade - make sure old people get educated.

👷 3. Builder’s notes

Chat is cooking (again)

Two things to report here

1/ I’ve used Astra this week and 2/ the voice to laptop feature is unreal.

I mentioned Astra last week but this week I’ve actually been able to use it as it’s been made available to the public.

For the everyday person it won’t mean a whole lot of difference. As I mentioned last week it’s just a more expensive model that is also more capable.

More broadly, I think Chat has just been in the lab cooking recently. The app, the transcription, the models, I think I’m seeing a broader sentiment across the board that people are returning to the chatGPT app.

Claude just sucks at writing, and for most people - that’s all they really want help with.

The question becomes - how do I not have my context stuck in one model?

I’ve broken it down before, you can find it here

ChatGPT remote is another one I want to cover. Released a little while ago and I’ve talked about it before but the app is so fkn good.

The transcription is even better.

So i’ve been plugging the laptop into charge and then just walking around in the sun prompting it.

It has all the context because they’ve got this great work view where you can open up specific folders and projects.

Being able to walk around and speak to it changes how the work feels.

For example I walked around and voiced out all the contents of what I wanted to cover this week in the newsletter.

It was able to pull relevant twitter posts I had seen using the browser view, then mould it into specific thoughts and give pushback where necessary.

Then when i went to go write it up - it was all there. (don’t worry I still handwrite this all)

You just need to select the remote option on the mobile app and then pair it from your desktop app.

A skill for slide decks

I also want to share this sick slide-deck skill I made.

So I needed to make a slide for a potential client. I basically gave it a reference deck I’d made previously, said help me turn this into a slide.

I can now actually prompt this inside of Claude code with the /design skill and combine it with the /ai-proposal-deck skill.

This is what it was able to produce. The cream background and the gradient.

Because it is all baked in with our internal WhiteHorse AI brain + my second brain (e.g. the folder structure on my own laptop) it has all the context (I’ve blurred out some of the important stuff). See below.

I literally did it in one-shot. A bit of fine-tuning I have to do.

But I’ll share the skill once I’ve got it really good.

Skills for everyday use

Matt Pocock’s skills are worth a look.

The part that’s super cool is that there is a skill that actually walks you through the implementation of the other skills (even if you’re lost).

He’s definitely a bit more technical than most but the skills are setup so that whatever LLM you use - Claude or ChatGPT can read the /setup skill and help you with the plugins.

He’s most well-known for the /grill-me skill but he has a multitude of other ones that you can find here.

If you’re confused just feed the above url into your LLM and ask how to set it up.

🤔 4. Differentiated take: are friendships conditional?

I was talking with some friends the other day about the idea that everything is a sales funnel.

(some credit to my good friend Amey with his fantastic article here)

Once you spend enough time thinking about business, you start seeing the same patterns elsewhere.

Someone encounters your work. They become interested. Trust develops. Eventually, they buy something.

You can see versions of that in networking and dating too. An introduction becomes a conversation, which becomes a relationship.

I started wondering how far the comparison went.

I had someone object with: “What about going to the gym?”

In an abstract way, you can say it’s part of a funnel.

I go to the gym to become healthier, more productive and a more attractive partner. Perhaps those things improve my chances of finding the relationships and opportunities I want.

But it’s a bit of stretch..

Having a reason to do something doesn’t make it a sales funnel. Maybe you just enjoy lifting weights. Maybe feeling good afterwards is enough.

If the definition expands to include everything, it stops explaining much.

Still, the conversation left me thinking about something more specific:

how much of our interest in other people depends on what they can do for us?

I know people who describe their relationships quite openly in those terms.

They spend time with someone because that person can help them get better at sport. They cultivate a relationship at work because it might help them get promoted.

There’s some honesty to that. We benefit from our relationships, and pretending otherwise would be ridiculous.

I want friends I can learn from.

Shared interests matter.

Enjoying someone’s company is itself a benefit.

But I think there’s a distinction between benefiting from a friendship and making that benefit the condition of the friendship.

If your colleague leaves the company and loses their influence, do you still call?

I studied philosophy at university (if you couldn’t tell by now lol) and spent some time studying Immanuel Kant.

One formulation of his categorical imperative says we should treat people as ends in themselves, never merely as means.

“Merely” is doing a lot of work there though.

We help each other achieve things all the time. That’s part of living together. The question is whether the other person’s interests matter to you beyond their usefulness to your own.

As I’ve gone through life, I feel like I’ve encountered more people who approach relationships through that calculation.

Personally, I find it uncomfortable.

If I don’t get along with someone, I struggle to spend time with them simply because knowing them might be advantageous.

Although that isn’t much of a moral achievement on its own. It might just mean I’m bad at pretending.

The harder test for most is how one behaves when a friendship becomes inconvenient. 

When someone needs support, has little to give back, or is going through a period where their company isn’t especially enjoyable.

That’s where my thesis becomes a little harder to defend.

I think for me it comes back to simply enjoying the company of a person - if I can’t enjoy their company, no matter how successful they are, it’s hard to be friends with them.

A good question to ask yourself:

If knowing this person could never advance my career, improve my status or open a door, would I still want to know them?

A lot of people I think would answer no.

I hope I’m not one of them.

5. Carve-outs

The Odyssey

Finally watched it. Genuinely one of my favourite movies. Probably the best movie experience I’ve ever had (seeing it in iMax was so worth it). The soundtrack stays banging.

The Social Network

The soundtrack is goated and the movie stays undefeated for anyone starting a company (see the boys watching it below).

Until next week amigos,

Alex