how to keep your context clean

Morning,

📜 TLDR:

• Doing our best to deliver absolute excellence.

• ChatGPT strikes back

• What’s cooking in the world of AI building

Enjoy :)

💼 1. What’s new in the biz

Absolute obsession with giving the best possible customer experience.

So much of business just comes down to this. How can you give the best possible experience to your clients.

A lot of the time I feel life is not so much about learning new things but rather being reminded of the same things - in different contexts

It’s all well and good to make new products and have different ideas, but ultimately these often serve as distractions from the main goal.

Hormozi has a great note about this: the best people in business, do one thing really well, better than everyone else, for a really long time. You then earn the right to do other things once you’ve accomplished that.

Not a whole lot new in the business this week - but being reminded of that, is always a win.

So my big takeaway from this last week: earn the right to experiment.

With that in mind we’ve really been working on just serving the core needs of our typical client.

The main wins we’ve been working on: building out the AI brain, solving people’s mundane tasks with skills, sorting people’s emails (shocking how much time people spend on emails) and empowering people to be able to automate their repetitive tasks themselves.

🤖 2. What’s new in AI

Three things happened in the last fortnight: the price of AI continues to fall off a cliff, the industry split over who's allowed to download a model, and the rise of personal AI agents.

The bottom of the market is in freefall.

Last edition I wrote that models were commoditising. On July 30 OpenAI cut GPT-5.6 Luna, their high-volume workhorse, by 80%. Twenty cents per million tokens in, down from a dollar. The mid-tier came down about 20%. The flagship, Sol, didn't move at all.

Nobody is discounting the top of the range. Everybody is sprinting to the bottom of it, because that's where the volume is, and it came six days after Anthropic shipped Opus 5 pitched as near-frontier intelligence at half the cost. The frontier is a prestige product now. The boring tier is where the money is, and the boring tier is getting cheaper faster than anyone budgeted for.

Here's why that matters if you run a business rather than a lab. It’s all about giving yourself the structure underneath of how to have context that lives outside of a model - so you can switch whenever you want.

We always encourage people to have their own personal context folder that acts as their second brain - you can then point whatever model you want at it and not lose your context.

We do the same thing with any business brain we build. It’s basically a memory layer that sits underneath whatever model you’re working on - so you can switch models whenever you want without losing your context.

On top of that, there’s really no reason why you shouldn’t be trying to use AI as much as possible.

There's a list of things you probably already decided not to automate because the numbers didn't work. Reading and classifying every inbound email. Summarising every sales call. Pulling line items out of every supplier invoice. Tagging every lead by intent. Twelve months ago that maths was genuinely marginal. At this price though, it isn't really maths anymore, it's a rounding error against one hour of someone's time.

So the move this month isn't chasing a new model. It's going back to the list of automations you priced out and killed, and re-running the numbers. A good chunk of that list just became worth building.

DeepSeek sees a massive upgrade

In addition to this, the new deepseek model is (almost) on par with the Opus 4.8 models - which is absolutely nuts.

The column that matters for me is the first one. On Terminal Bench, which tests whether a model can actually operate a computer rather than just talk about it, DeepSeek's cheap model scores 82.7. Anthropic's Opus 4.8 scores 85.0. That's a gap of two and a bit points.

Just a reminder Claude Opus 4.8 was the flagship model no less than 2-3 months ago.

And now, it’s essentially being commoditised away.

As for the way I’m using these different models?

If I want to build something generally the process is: word vomit into claude/chat what I want to build. Use fable and the /grill-me skill to discern what I actually want. Then use a cheaper model to execute as necessary (depends on how many tokens I have).

Fable/Sol does more of the high level thinking. Then use the cheaper, more efficient models to execute.

The Open Source wars spice up.

I mentioned this before, but on July 24 a coalition published a letter called Open Weights and American AI Leadership, arguing Washington shouldn't restrict downloadable models, because openness prevents lock-in and stops capability concentrating in a handful of hands. It launched with 25 names: Nvidia, Microsoft, Meta, Dell, IBM, Palantir, Mistral, Mozilla, the Linux Foundation, Hugging Face, a16z, Y Combinator.

View @JensenHuang’s post on X

Notably, OpenAI, Anthropic and Google had all declined. That framing lasted about forty-eight hours. Jensen Huang posted it, the post did eleven million views, and by the next day the letter had doubled to fifty signatories with OpenAI and Google among them.

The durable story is who's still out. Amazon and Anthropic. Amazon is Anthropic's largest investor and primary hardware partner. Google, also an Anthropic backer, signed anyway. So the two companies whose economics depend most completely on closed weights sold through an API are the last two standing, and one of them appears to have brought its biggest investor along for the ride.

I don't think that makes them wrong. It makes them consistent. But it does tell you the open-versus-closed argument isn't really a philosophical one, it's a P&L one, and you can (largely) predict almost everyone's position from their revenue model.

If you want an awesome breakdown on how the AI wars are right now, please spend 2 minutes and watch this video (it’s hilarious).

View @charliebcurran’s post on X

Personal AI agents went from novelty to category, fast.

In the last few weeks I’ve seen three companies, all building an AI assistant that lives in your messages rather than in a browser tab:

- Poke, from The Interaction Company. Text it on iMessage, SMS, Telegram or WhatsApp. No app, no install, you just give it your number. Backed by Spark Capital and General Catalyst. And in June it became the first AI agent Apple has ever approved on Messages for Business. They were acquired by Cognition Labs for north of $100mil (and to be completely honest I didn’t think it was all that good).

- Orchid, out of Y Combinator. An executive assistant in iMessage that texts you your day each morning, drafts your replies and books your meetings once you approve. Their launch video did 20 million views and a fair amount of backlash (which I think they knew it would cause). Again it really lacks in the proactiveness of the product, but assume they’re just trying to get distribution and then will iterate on the product.

View @orchid_hq’s post on X

- Town, which raised a $55 million Series A in June led by a16z and Forerunner, on top of an $18 million seed. It plugs into email, calendar, Slack, docs and WhatsApp and learns how you work rather than waiting to be told.

There’s also Buzz which is the agent-native Slack.

View @jack’s post on X

The bigger trend I’m seeing is that AI is moving towards more intuitive UI.

Now you can simply tag @AI in your workspace. Or you can just text @AI and have it respond in a place you already live.

The next version of this is the personal AI agent which you can just “speak to” in a really natural way. Ultimately what siri was meant to be, but never was.

And then combining this with the next paradigm of compute - glasses. Story for another day, but that’s where this is all moving I imagine.

Just cool to see AI being incorporated in a more digestible form.

And the author of situational awareness takes a massive hit.

If you've read anything serious about where AI is heading, you've probably read Leopold Aschenbrenner, or read someone quoting him. His essay "Situational Awareness" is arguably the single most influential document on AI's trajectory of the last few years (highly recommend reading it if you haven’t - linked bwlo). He then went and started a hedge fund of the same name to bet on the thesis.

He was right. The fund was up 439% for the year through June. At its peak in July it sat on about $45 billion.

By the end of July it was down 67% for the month and holding around $10 billion. The AI infrastructure names he was concentrated in, SK Hynix, CoreWeave, Nebius, Micron, Bloom Energy, fell between 35% and 47%. At the same time the software stocks he was short moved sharply the other way. Both ends of the book went wrong in the same fortnight, the margin calls came, and he was forced to sell the entire leveraged public portfolio to Ken Griffin's Citadel at a discount. His letter to investors opened with "we let you down this month."

The detail that turns this from a market story into a lesson is the leverage. Reported at up to 400%.

View @beffjezos’s post on X

What’s crazy is he's actually still up around 80% for 2026 even after all this. Add 4x, and a temporary drawdown becomes a forced sale of everything at whatever price the buyer feels like offering.

And this all happened while he was getting married this week (less than ideal)

👷 3. Builder’s notes: keep it clean

This week was about keeping it simple.

Clean out your context. This is the big one. If you're using Claude or anything like it, you've got a CLAUDE.md, an agents file, and a project folder that have all been accumulating dust for months. Every one of those gets loaded into the model's head before it does a single thing for you.

I went through mine properly for the first time in a while and cut a large amount of it: instructions that contradicted each other, rules for a workflow I stopped using in May, half-finished files sitting in the folder doing nothing but taking up room. The output got noticeably better, and it got faster.

So if you’re Claude workspace feels slow it may not be the model, it could just be your context. You're making it read your entire filing cabinet before it answers a one-line question. Same reason a bloated folder full of dead files degrades everything, the tool has to consider all of it. Treat your context files like a fridge. Something in there has gone off, and you'll only find it if you look.

We made a skill for this btw.

Skills are the highest-leverage thing you're not using. A skill is a folder with instructions for a recurring job, which the assistant picks up on its own when the job comes up. You write down how you want something done once, and then you never explain it again.

One that has been hilarious is condensing your output. Sometimes AI can be a bit too verbose. So we have a skill for that. Whenever you want your AI to just cut to the chase just use the /i-have-adhd skill. I don’t have adhd but you get the point (shoutout to Swanny for this one).

If you want this skill or the skill to clean out your context properly - reply to this email and I‘ll send it across (won’t let me upload a zip file)

OpenAI is cooking again: Along with their big discounts, OpenAI shipped a new speech-to-text model on July 28. The headline number: word error rate on the standard multilingual benchmark went from about 40% to about 19%, across 22 languages, at a lower price per minute than the model it replaces.

Halving the error rate is the difference between transcription you have to proofread and transcription you can just use. If you think faster than you type, and most people do, that's not a small change to the tool, it's a change to how you get things out of your head.

Super cool demo here:

Save a lot of different things but never know where they end up?

Sublime is a really cool tool for this. Basically it acts as your digital brain for whenever you see something cool and want to save it.

They’ve been around for a while but was sent this by a friend. Basically just save anything with their in-built clipper and you’ll be able to go to one place to see it all.

Sublime — A second brain with a soul

🎶 4. Carve outs:

• (No differentiated take this week, I really only want to write these when I feel I have something to give)

• If all roads lead to data, are we doing enough to be able to make the best forecasts? Inspo for next week’s differentiated take.

Appreciate y’all. At 539 and counting :)

Until next time,

Alex