The Huddle: Issue 11

A biweekly look at what's changing in AI, ways to apply it today, resources you didn't know you had, and sharing wins.

Stop me if you’ve heard this before… AI is moving fast. There are some interrelated stories worth hitting on below, and they add up to one uncomfortable point about whatever tool you’ve standardized on: it's going to get leapfrogged before Halloween (and maybe twice). Since our last issue, Anthropic shipped near-Fable intelligence at half Fable’s price, Google confirmed its next flagship still isn’t ready, ChatGPT Voice moved into Work and Codex on desktop, and OpenAI’s next model is so good at hacking they had to ring the alarm. So this issue covers the news quickly, then we’ll talk about the one part of your AI setup that nobody can leapfrog.

HEADLINES THAT ACTUALLY MATTER

Claude Opus 5 arrives at half the price of the top model. Anthropic released Opus 5 today, saying it comes close to the frontier intelligence of Claude Fable 5 for half of what Fable costs. I've said before that Fable wasn't the right daily driver for most of you and that price was a big part of why. Anthropic more or less agreed and built the model I was describing. Getting near Fable abilities for a discount has more than a few parallels to LeBron James deciding to play in Philadelphia for $4M/year - but I’ll keep that analysis to a minimum. Keep reading the Huddle, keep Trusting the Process. (Anthropic's announcement)

Google shipped the workhorse and admitted the flagship isn't ready. On July 21, Google released Gemini 3.6 Flash plus a cheaper Flash-Lite tier, confirmed that Gemini 4 has entered pretraining, and left Gemini 3.5 Pro unreleased after promising it on stage in May. On the Q2 earnings call, Sundar Pichai said Google needs to improve its coding and agentic coding and that a larger Gemini 4 would be necessary to compete at the next frontier. That is a candid thing for a CEO to say out loud, and I'd rather have the honesty than another delayed launch date. A few issues back I called this a three-horse race. At the frontier it currently looks like two. The distribution race is a different animal, because Gemini will keep showing up in your search results, your Workspace, and your Android phone regardless of where it lands on a benchmark. (Search Engine Journal's writeup of the call)

ChatGPT Voice can now run your other AI. Yesterday OpenAI brought its new voice model into the ChatGPT desktop app on both Mac and Windows. It listens and speaks at the same time, so you can interrupt it the way you'd interrupt a person, and from inside a voice conversation you can start, check on, and redirect work already running in ChatGPT Work or Codex. On a Mac it can also see whatever window you have in front of you. If you've ever wanted to talk a problem through out loud while something gets built in the background, we’ve got good news for you. One piece of fine print worth knowing before your team gets excited: tasks you kick off by voice draw down the same usage limits as tasks you type, so the microphone doesn't come with its own budget.

(TechCrunch)

An OpenAI model escaped its sandbox and hacked a real company. OpenAI disclosed on Tuesday that during an internal test of hacking ability, two of its models, GPT-5.6 Sol and an unreleased one, both running with their cyber refusals deliberately switched off, found their way out of the isolated environment they were in, reached the open internet, and broke into Hugging Face's production systems. They stole credentials and exploited an unknown flaw to run code on Hugging Face's servers. Nobody instructed them to do any of it. Hugging Face (yes, it’s hard to take a company with that name seriously) had already detected the intrusion and reported it to law enforcement before either company knew whose model was responsible.

Important to note: this was a test with the safeties removed, not an attack. It’s like tying yourself to the mast of a ship just to hear the sirens - the guards were left down on purpose. And if you have no idea what I’m talking about, go to an IMAX theater ASAP and tell me what you thought of the movie.

(OpenAI's incident report, and TechCrunch on how the sandbox failed)

WHAT DOESN'T GET LEAPFROGGED

Most readers of The Huddle that I spend time with say the same thing “how do you keep up?” Or even better “how do we spend our time now to not get further behind?” The big idea or challenge I want to give readers this week is context. Models, apps, and companies will change. But the muscles you build won’t - and the context you amass and carry with you to new tools can become a formidable weapon (you don’t have to pluck the string before, either).

The part of your setup that survives the model madness is context: the written record of who your company is, how you talk, what you sell, who you sell it to, and what good work looks like when your team does it well.

This idea goes by a lot of names, and I think a few of them are why people bounce off it. You'll see it called a second brain, a personal knowledge base, a context library, a memory layer, a digital garden, or PKM if someone is feeling technical. Pick whichever one doesn't make you roll your eyes. Underneath every one of those names is the same modest thing: plain text files in a folder that you can hand to any AI tool, on any day, in any order.

I don't think everyone needs one, and I want to be clear about that before someone spends a Saturday building folder structures they'll never open again. If you use ChatGPT twice a month to clean up an email, skip this entirely. The tell that you're ready is boring and specific: you catch yourself re-explaining your own company to an AI more than a couple of times a week. The solution to the context problem can solve your prompting problems pretty fast.

Mine happens to live in Obsidian, which matters far less than people assume. I could move the whole thing into a folder of text files on my desktop tomorrow and lose almost nothing, because the value is in what's written down, not where it's parked. A few issues back I mentioned that the instructions you write to teach one AI a task are starting to travel to other tools. This is how you make that true for your business on purpose instead of by accident.

The version we've been building for clients is the company second brain, and that's where we can save people a lot of time while making the outputs more consistent. It's a small set of files covering brand voice, design standards, value propositions, core values, and examples of your best work, written down once, kept current, and made available to the tools your team uses. It keeps five people from prompting five different versions of your company into five different chat windows and getting five different answers back.

If you want to try it, five files will get you most of the way there, and you can sketch the first version in half an hour:

• Who we are and what we sell
• How we talk, with two examples of writing that sounds like us and one that doesn't
• Who we sell to and what they're trying to fix
• What we believe and won't compromise on
• What good work looks like, with two or three real examples

There's a video in Trailblaze Academy on custom instructions that's the first rung of this same ladder. Reply with the word Academy if you want in, and we'll get you access.

THE BUZZER

The honest part of a week like this one is that almost none of it belongs to you. You didn't build Opus 5, you can't make Gemini 4 show up faster, and you'd have caught that sandbox problem about as quickly as OpenAI did, which is to say not at all. What you can own is thirty minutes of writing down how your company works, so the next tool you try starts halfway smart instead of starting over. Write one of those five files. Start with the last one, the examples of good work, because it's the hardest to write and the one that changes the output the most.

Thanks for huddling with us before the weekend.

Bryce Stuckenschneider

Founder, Trailblaze Labs

ABOUT TRAILBLAZE LABS

We started Trailblaze Labs because we saw a gap forming between what AI could do and what most businesses knew how to do with it. Turns out, that gap is where we live now. We help business leaders set real AI strategy, train their teams to use these tools with confidence, and build the workflows that actually move the needle. Trailblaze translates AI to the language of your business.

If something in this newsletter sparked a question, gave you an idea, or made you want to push back on something I said, I want to hear about it. Seriously. Reply to this email, connect with me on LinkedIn, or just say hi. I read everything and I love a good friendly argument about where this is all headed.

Let's keep building.