chat is cooking (again)

Yo yo folks,

Nice to have you join me for today.

📌 TLDR:

• Content grind continues

• ChatGPT has their (dev) day

• There’s like 20 new model releases (so bear with me)

• Skills I’ve been working on

• And my unfiltered thoughts on applied AI in the workplace

💼 1. What's new in the biz

Been doing a lot more on the content front lately.

Swanny and I set each other a challenge for posting. Mine was to get to 30 posts over 30 days (in September) and get to 500 followers.

Currently sitting at about ~885 and hit 30 posts (including trial reels), so was pretty cool to achieve that.

Probably most thanks to this video that went semi-viral: around 100,000 views.

Weirdly, thought this was one of the worst videos I posted tbh.

Instagram post

Consequently, I've been playing around with trial reels too. Honestly, it's been a really interesting case study in how to capture attention and how to script things.

I was kind of winging it before. Now it's becoming a bit more structured.

One thing I'd really love is a tool that makes it easy to split-test trial reels. Different hooks, different scripts, different edits. Then see what actually gets people to stop and watch.

But also don’t want it to be really AI-sloppy. Will report what I can cook up.

It feels like there's still a lot of room for tools that help creators with that process. I’ve gone hunting for a few and couldn’t find any.

I'm working on systemising it and thinking about what could be productised. A lot of the work is the same stuff repeated: come up with an angle, script it, make a version, test it, learn something, and go again. Like I spent way too long thinking about different variations and then editing those for trial reels.

Also, as a result of those posts, this is the first newsletter for a few of you, so welcome.

If we haven't met, I'm Alex. I help businesses implement AI, and this is where I share what I'm building, what I'm finding useful, and a few thoughts along the way.

Feel free to hit reply with any questions. Always happy to chat :)

2. What's new in AI

ChatGPT gets an always-on agent

OpenAI had its (Dev)Day this week. Two announcements in particular really stood out to me.

The first was dots, its always-on agents. They have their own cloud computer, can work across connected apps, and keep making progress between conversations. They've started rolling out to eligible Pro and Business Premium users, with Enterprise access in beta. How dots work

View @OpenAI’s post on X

We saw this direction when OpenClaw came out.

For ourselves we got inundated with over 330 enquiries for setting up OpenClaw in less than 50 days.

It was just a matter of time until this was done by OpenAI.

(Which should come as no surprise since they bought the founder of OpenClaw).

We’ve seen Instinct and Meta's Muse. Now OpenAI has its version. I imagine Anthropic will come out with something similar.

It's a pretty digestible format for someone who isn't technical. Give your agent a goal, let it work, and come back when it needs your judgement.

I gave it a crack today and it’s pretty cool. I can actually call my own agent too + it has all the existing context that Chat has access to.

So gave him a buzz and he was rather helpful.

It’s a little bit laggy but still very impressive - can imagine this being insanely cool in the next 12 months.

For someone already deep into Codex, Claude Code or their own agent setup, this doesn’t change a whole lot. The appeal is that someone else handles the setup and gives you a familiar place to use it.

TBPN posted an interview with Sam Altman from DevDay covering how he uses dots, the open ecosystem, and model-picker exhaustion.

I think they’re moving towards more general usage.

That last point feels particularly relevant. Most people don't want another decision about which model or mode to use.

View @tbpn’s post on X

For me personally, what I'd love is something closer to "Hey Siri", except it can actually do the work and has the context to understand what I mean.

I am also convinced that a screen-free paradigm is the next one. And the future of agents lives seamlessly with you (not reserved to just your phone) - e.g. a wearable or something.

Accordingly, I’m surprised OpenAI didn’t release something about hardware. Feel like they were due for this, particularly since buying Jony Ive’s startup for $6.4b.

My guess in the shorter term is that more of this work moves onto cloud computers, so the agent can keep going when your laptop is closed. Kind of stupid that for the average person you have to keep your laptop open.

In the longer term we will see wearables and a whole ecosystem of AI powered home appliances, it’s just a race to get there.

Your ChatGPT subscription can follow you into other tools

The second announcement was Sign in with ChatGPT.

Eligible Plus and Pro users can now use their plan allowance in participating tools. OpenAI announced 16 partners, including Devin, Notion, Vercel and OpenClaw. Usage counts towards your existing limits. DevDay announcement

This is a massive dub in my opinion.

Normally, if I build an AI app, someone has to pay for the model usage through my app or configure their own API account. Both create friction.

Letting someone bring the subscription they already pay for makes trying a new tool much easier.

There are limits, though. This doesn't mean you can plug a subscription into any commercial app and get unlimited subsidised usage. OpenAI's developer guide sets out eligibility for open-source projects, locally running personal projects and selected private apps. Integration guide

But the direction is interesting.

As models get more interchangeable, the account people already have, the tools connected to it, and the context stored there become increasingly valuable.

Additionally if I’m a developer and I want people to be able to use my app without them having to use a lot of tokens, I’m going to encourage people to get a more advanced ChatGPT subscription (or if they don’t have one - to get one).

We were already doing this with some people for OpenClaw setups, so super smart for them to come out and actually encourage this.

I assume Claude may come out with something similar - or at least they should.

Honestly just a game of “whatever you can do, I can do better.”

You can learn more below.

Meta's Muse moves into small business

I talked about Muse last week.

Meta has now added Muse for Small Business, with connectors including QuickBooks, Shopify, Canva, Stripe and HighLevel. It can work with business information to draft campaigns, review financial performance and help plan growth. Nothing publishes, sends or spends without approval. It's currently available in the US and Canada. Meta's announcement

This is a pretty natural next step.

The person most willing to pay for an agent is someone who can see a return from it. Save them time, help them follow up with customers, or take a recurring job off their plate.

I.e. a business owner.

Connecting the tools is a big part of making that possible.

An assistant that knows your business, can reach the relevant information, and leaves you something ready to approve is a much more useful proposition than another empty chat window.

Another dub for Meta. I'm keen to see what the Australian rollout looks like.

20 new models released this week (so bear with me as I recap them)

A quick summary of the new models

To condense it down:

Sonnet 5.5 and Sol are the cheapest everyday options here. Opus costs twice as much per token, while Astra (the most capable OpenAI model) costs five times as much, so I’d save those for tasks where the extra capability actually matters. Argon matches Sol’s pricing at launch, then moves up to Opus’s. The real comparison is what it costs to finish the job: a cheaper model isn’t always cheaper if it needs more attempts.

Google's Argon announcement however is particularly interesting. Google reports leading results on several evaluations, including Zapier's AutomationBench, which tests business workflows. Initial access is through its trusted cyber-defender programme, with broader access to follow.

Pretty cool for Google considering they haven’t won one just yet.

But whenever you use these models, there is still a lot of vibe checking involved. A benchmark can give you a reason to try something. It can't tell you whether it understands your brief, follows your preferences, or makes a spreadsheet you can actually use.

So easiest way to test out the changes?

Just use them on your everyday task.

My quick vibe check:

• Opus 5.5: elite at design in my testing. The everyday workhorse along with the Sol models.

• Sonnet 5.5: the practical everyday option. Half Opus’s API price and generally easier on the usage allowance.

• GPT-6.1 Sol: elite at writing, especially when I give it a good reference. Also gets used regularly as the everyday workhorse.

• GPT-6 Astra: really good at difficult coding, but it burns through my allowance quickly. Its standard API rates are five times Sol’s.

• Fable: great for solving Nobel math prizes - wouldn’t really touch it otherwise (pretty similar to Astra)

It can all be a bit overwhelming.

The easiest test is still your own work: give each model the same brief, compare the result, and check how much usage it took to get there.

That's the really the only comparison I (and you should) care about.

The dilemma of agents and marketplaces

So two parts to this (really good article details this in further detail)

1/ agents and the platforms that power them are becoming marketplaces where you can buy and sell.

2/ marketplaces should be very wary of the agents that they let access their data (I’ll explain below).

On the first point, I think the bigger shift is that these agents are competing to become the place where people make decisions and buy things.

OpenAI and Anthropic already have marketplaces for partner software. OpenAI's lets eligible enterprise customers put part of their existing spending commitment towards approved products.

But consumer commerce is where this gets really interesting.

Say I text my agent:

"Book me a flight to Melbourne next Friday."

How does it choose?

Does it browse airline websites? Compare prices through Skyscanner? Use a direct connection to Qantas? Which options does it show me, and which never make the shortlist?

That selection process becomes incredibly valuable.

Already Instinct founder Noah Shinn discussed this on Invest Like the Best this week, talking about wanting to keep Instinct free but monetise by making a transaction-fee business model on the roughly US$1 billion a year in transaction volume flowing through the platform.

View @patrick_oshag’s post on X

So the value chain is now moving away from the previous marketplaces/search engines and up one to the agents that are helping users buy more.

Accordingly, I’d be very cautious if you’re an established marketplace.

Amazon, for example, blocked Meta's Muse from shopping on its site. Reporting

They do 50% of all ecommerce. So really Muse and other agents need Amazon more than Amazon needs them.

If my agent chooses a product for me, I might never browse the marketplace or see its sponsored listings. The customer relationship starts moving towards the agent.

Which means a lot less money for the marketplaces that host this supply.

That doesn't mean advertising disappears. It raises a new question: will an agent recommend the best option for me, or the option that pays it most?

There's an opportunity for new businesses here, which every agent platform is trying to capture.

👷 3. Builder's tools

A couple things I’ve been building this week

/client-portal

One thing that we’ve systemised is generating a portal to track all the usage, skills and workshops that we conduct for companies.

It’s one thing to build something, it’s another to make sure that the work we do for people is all viewed in a trackable manner.

Now when I want to spin this up for another person I can just say /client-portal and in one click get a customised dashboard spun up.

See what you can produce in just one click now. It will even pull the relevant recordings and dynamically include them into the portal.

/reel-audio-split

One cool thing I played with this week was separating the spoken track from the background music in a reel.

Sometimes I see a reel with music that I think would be good but there’s audio over the top.

How to fix this?

Threw it into Claude and it was able to get me the two tracks separated. Then played the sound out loud for shazam and shabang - it got the right song

I know this probably sounds super minor but I promise you, LLMS literally couldn’t do this a few months ago.

I've turned the workflow into a skill, so I can ask for it again without explaining the whole process.

Give it a reel, ask it to separate the audio, and get two tracks back: the vocals and the backing track. Claude or ChatGPT can run the audio tools when they're connected to an environment that supports them.

🤔 4. Differentiated take: My unfiltered thoughts on implementing AI into businesses.

I’ve been thinking a lot about how AI gets diffused into the economy. I think there’s a narrative from the tech-bro world that this UBI (universal basic income) is imminent, and no one will have a job.

I have a lot of friends that are in the startup world raising money, growing quickly, hiring aggressively. Where talk of AGI is evident everyday.

How the ‘singularity’ has occurred.

Software is dead.

The internet is dead.

There’s 5 years left to escape the permanent underclass.

There’s 10 years left to build a business.

I also speak to service-based businesses everyday.

Where software still runs on systems from 1990s.

Or everything runs on paper.

People might spend 3-4 hours a day on menial, mundane drudgery.

And so the dichotomy is rather striking. 

It reminds me of when I was in Barcelona last year (stay with me now)

I remember trying to get a seat at a cafe in the middle of winter.

1:30pm.

Wednesday afternoon.

Absolutely fkn packed.

Around me, people were drinking with their friends, having chats, and debriefing on how shit work had been that day.

Meanwhile, my feed was full of people talking about AI changing everything.

One of the biggest technological changes of our lives is happening.

And yet, a lot of people are still mostly interested in enjoying theirs.

And I kind of see this everyday with ordinary businesses.

The capabilities are unbelievable. But getting people to use them consistently is a completely different job.

Change is hard.

People have existing habits.

They use AI for a handful of things they understand, while updates arrive faster than they can keep track of them.

And most people generally don't care about AI itself.

They care whether work gets easier.

This seems particularly evident in Australia.

We have a fairly good life here. The weather is good, people are nice, and plenty of business owners would rather enjoy what they've built than spend every evening learning a new tool.

Someone running a business has customers to serve, staff to manage and invoices to chase. A new model release is competing with all of that for attention.

And we’re not just talking about an upgrade in an existing technology - we’re asking for people to change the entire way that they operate.

I mean changing CRMs can be difficult and time-consuming. But people already understand what a CRM does.

Now ask them to work in a completely different way.

In previous engagements, we assumed too much prior knowledge. Implementation sometimes took 1.5 to 2 times longer than expected.

You have to map out the process, explain what's changing and make sure everyone understands it. You can't move at the speed of the most technical person in the room.

You kind of have to move at the speed of the lowest common denominator.

And so when I hear of all this doom and gloom around AI, how there will be the permanent underclass, I think two things.

1/ Maybe this was what the hype was like in the dotcom era. And,

2/ I think this will take a lot longer than all the tech bros think it will.

Altman even speaks of something similar here:

View @tbpn’s post on X

And even with the changes that are being implemented the best starting point is often the mundane drudgery they already want help with.

Emails. Reports. The information they keep searching for. The follow-up that gets forgotten.

It’s not like it’s solving all business problems.

The two biggest ways we’ve tried to increase adoption is to

1/ outline the outcome

2/ empower people to see it for themselves.

If you can teach someone to fish, it’s much more powerful than fishing for them.

And if you can tell them they’re going to eat the best fish ever (as opposed to worrying about the functionality of how to cast a fishing rod) then they’re going to be much more incentivised to catch the damn thing.

Let people see a useful result, then help them take responsibility for finding the next use case.

I think this change could be as significant as the industrial revolution. I also think the implementation lag is longer than people spending all day with frontier models tend to assume.

I don't know whether that takes five years, twenty years or a generation. But a model being able to do something doesn't mean an organisation is ready to hand it the job.

There is a lot of work between those two points.

And I think that's where a lot of the opportunity is.

🎵 5. Carve-outs

• Sports:

  • Devastating to see the knights lose. I’m not even a knights fan but thought they might be able to get it done. Have to listen to the Joey Johns and Matty Johns podcast if you haven’t

    • https://www.youtube.com/watch?v=HMyCHhHBrPE

  • Serious question: what does one do without sports? Feeling lost.

• Bring on Aussie summer!

Peace,

Alex