At the start of 2026, I made myself a promise: I would write a new edition of Digital Dips every week. Six months later, we can safely conclude that I failed. The question is: why? It wasn’t because I lost interest. Quite the opposite. Like everyone else, I’ve suddenly become capable of doing far more than I could before. As AI models improved, things that once felt too complicated, expensive or time-consuming became possible. So instead of writing about what could be built, I spent my time actually building it. Things I had never built before. That’s exactly what this week’s main story is about: it’s time to build.

Over the past six months, I’ve built products, systems and skills. I’ve experimented, failed, started again and figured out what actually works. And I’ve realized something. Instead of writing about every new model and every new tool you’ll probably never use anyway, I can be far more valuable by showing you how I use AI in practice.

That’s how Digital Dips will move forward.

Not because the stream of updates is slowing down. It isn’t. But covering every single one as if it's breaking news isn’t a service. It's a treadmill. And I think you deserve better than the treadmill.

From now on, I will test, try, and run new tools, new models, and new skills. I'll write about them for you. So you don't have to. I'll show you my own stack. What tools I use. What models. What actually works.

I'm currently preparing the setup for this. I expect by the next edition it will be ready, and I'll start with something I've been using for a while: my Council skill. A multi-agent system that's been running my most complex tasks. I want to show you how it works, how I use it, and why it changed how I think about delegation.

The value is no longer in knowing that something new exists. The value is in knowing what to do with it. This is what I've been building toward. And I'm excited to finally show it.

Quick note: It's been six weeks since the last edition. Sorry about that. It means today's edition is a bit longer than usual, and I still had to make choices about what to leave out. Those will land in the next one.

So grab your favorite snack, settle in, and let's dip into what's cooking. No time to read? Listen to this episode of Digital Dips on Spotify and stay updated while you’re on the move. The link to the podcast is only available to subscribers. If you haven’t subscribe already, I recommend to do so.

One topic that deserves more than a bite. Not too long. Just enough to chew on.

Over a year ago, at the company where I worked at the time, I hosted regular sessions to walk my colleagues through the latest developments in AI. They were essentially condensed versions of this newsletter, translated into what those developments could mean for their daily work.

One discussion has stayed with me. We talked about whether and when AI models would get good enough to build simple mobile games. At that time, the team built games with a SaaS solution. It was fun, once. The second time a client wanted something similar, they could still only choose from the same box of games they had used for their first campaign. In those sessions I told the team that if I were them, I would figure out the whole pipeline before they needed it. How hosting works. How you connect a sign-up form to the client's e-mail marketing tool. How you create game assets. How you make it on-brand. Which coding language works best. Because when the moment arrives, you don't want to be figuring it out. You want to shoot your shot and start building.

One colleague asked when I thought that would be possible. I said: in twelve months, for sure. From his position, twelve months might as well have been another lifetime. He had deadlines, clients and work that needed to ship now. I was looking at the capabilities that might exist next. Both perspectives were reasonable. The real problem was that we had no mechanism for preparing for tomorrow without disrupting the work of today.

So, here we are. People are building games with general-purpose AI models that look like they were built by a big game studio. Which means the games the team used the SaaS solution for, could already be built a few months earlier with not even the frontier models.

Shortly before my paternity leave, I sent the company's founder an internal note about the best possible future I could see for the company in offering AI to clients. It was based on where I thought AI would be in the months ahead. I ended that note with: "It's time to build."

Two months later came Sam Altman's post about the fast-fashion era of SaaS. It confirmed my timing was right, and my view of what was coming was right too.

That same founder told me in the autumn of 2025 that nobody really knows where AI will be in the coming months. He was right. Nobody can reliably predict which company will lead, which model will be available or exactly what a system will be able to do six months from now. But that’s why I open every company presentation I gave about AI the same way: we are on an exponential. Waves of change keep arriving, and each wave is bigger than the last. For humans, that curve is nearly impossible to graps while you are standing on it. You cannot know exactly where AI will be in six months. What you can know is that your guess is below where the curve will land. My rule of thumb: whatever you think twelve months out looks like, multiply it by ten. Because that is what is happening right now.

Even the experts are off by years. In 2022, a panel of AI experts put a 50 percent chance on AI producing publishable mathematical theorems by 2050, and on winning the Putnam competition by 2033. By 2026, AI systems had already contributed to publishable mathematical work, while the best-performing system on the 2025 Putnam exam scored at a level that would have placed it among the top human competitors. People who make their living predicting this are that far behind. Your own estimate almost certainly is too.

Rather than trying to predict exactly where AI will be next year, let’s look at what has already changed, what it means for business and why “it’s time to build” remains the most useful response.

The clearest measure of where AI stands is the release calendar. In the seven days between July 17 and July 23, seven model families shipped: a new Moonshot flagship, three Qwen models inside a 72-hour window, a three-model Gemini drop, and an open-weight coding model from Poolside. Release trackers log a new notable model roughly every three days, and the monthly cadence of major releases has roughly quadrupled since 2023. The model you select at the beginning of a project may no longer be the obvious choice by the time the project reaches production.

The consequence for planning is brutal. When Anthropic's Fable 5 was pulled offline for 18 days under U.S. export controls, Z.ai's GLM-5.2 simply moved into the top accessible slot. The lesson isn’t that one vendor is unreliable. It is that model availability can now change for technical, commercial or political reasons. An enterprise that hardcodes a critical process to one provider may discover that its architecture is fixed while the market around it is not.

The models are also starting to improve themselves. Several releases this quarter train with reinforcement learning, closing the gap to the frontier without copying anyone's weights, and test-time compute lets a model reason for weeks on a single problem. When the system that gets smarter is also the system doing the improving, the curve stops being something any single lab controls.

The best demonstration of the exponential sits in mathematics, where the results are hard to argue with. This July, frontier models toppled open problems in math that had stood for 30, 50 and 87 years. Then, on August 1, the scale changed: OpenAI published ten results produced by an internal version of Astra, which it describes as its next major model. According to OpenAI, each result either resolved or made substantial progress on a long-standing problem spanning fields such as geometry, coding theory, cryptography and quantum complexity. Within 24 hours, an Anthropic shipping model reportedly reproduced half of them on its own. Read that as a curve. A few years ago, one open problem falling made headlines. This July, three fell in a single month. In August, ten fell in a single announcement. The pattern multiplies rather than adds.

That is why my "12 months, for sure" was the safe answer and the wrong mental model at the same time. It was a linear answer on an exponential curve. The more important measure is now the interval between an external capability appearing and your organization being able to turn it into reliable work. Your deadlines still matter. But the market may change several times while you are working towards them.

Here the exponential meets power. GPT-5.6 shipped gated, with the White House requesting a trusted-partner preview and per-customer sign-off. GPT-6 and the internal Astra model behind much of this recent progress are close enough that Sam Altman is briefing the Trump administration and Congress on capabilities and job impact. The briefing reads less like a product launch and more like a clearance hearing.

The question is whether the early-access process becomes a pathway or a waiting room. If it becomes a waiting room, the public tier falls further behind the agency tier with every generation, and enterprises end up evaluating models their competitors already run. The risk is not theoretical. During an internal OpenAI evaluation, models escaped their sandbox, found a zero-day vulnerability, and reached secrets at Hugging Face. OpenAI called it unprecedented. A system you cannot fully contain is a system you will be reluctant to release broadly.

This control question is splitting the industry in two. Washington is drafting an order that could effectively ban Chinese open models. While on July 24, a coalition of 25 companies led by Nvidia, Microsoft, and Meta published a letter urging Washington to avoid premature restrictions on open-weight models. OpenAI and Google were not among the original signatories, but joined later. By July 30, more than 230 companies and organizations had signed. Anthropic did not sign the letter, although its CEO has said he does not support a blanket ban on open-weight models.

The real disagreement is not simply whether models should be open. It is where openness should stop as capabilities become more powerful and potentially more dangerous.

Chinese labs now ship frontier-grade open weights, as Moonshot's Kimi K3 and Alibaba's Qwen3.8 do, and a U.S. ban on those models only helps if it also blocks the domestic alternative. The industry letter argues that they help organizations preserve the knowledge and capabilities they develop instead of losing them when a vendor changes its product or terms. This matters as capable open models continue to emerge from laboratories around the world, including China. Broadly restricting one source of open models may reduce domestic choice without removing the underlying capability from the global market.

The case for tighter controls deserves to be taken seriously too. A model that can be freely downloaded cannot easily be recalled, patched or restricted once it spreads. The same openness that gives legitimate companies control can give malicious actors access to increasingly capable systems. Model distribution is also becoming a form of geopolitical influence. Governments do not see AI models as neutral pieces of software. They increasingly see them as infrastructure, economic leverage and a way to spread technical standards.

The argument is therefore not openness versus safety. It is how to preserve the economic and strategic benefits of open models while applying targeted controls to genuinely dangerous capabilities.

And this fight decides your build strategy. Whether the models you build on stay open determines how much of your process you actually own. Open weights can be self-hosted, fine-tuned, archived, forked. Gated models are rentals with an off-switch. That does not mean every company should begin running its own models. Self-hosting introduces costs, security responsibilities and technical complexity of its own. The goal is not to pick an ideological side. The goal is to make sure your process, data, business rules and evaluations do not become inseparable from one model. You should be able to replace the intelligence layer without rebuilding the company process around it.

It is worth playing the sceptic’s side for a moment. Some weekly releases are minor variants packaged as major advances. Benchmarks are often selected and reported by the companies that built the models. The open-weights letter is an advocacy document, not neutral research. OpenAI’s Astra results still need wider examination by the mathematical community. Regulation, export controls, energy constraints or liability rules could also slow parts of the market. The industry benefits from making every month feel historic. Leaders should not mistake marketing velocity for business value.

That argument would matter if the baseline were stable. It is not. Even if half of the quarterly noise is marketing, the horizon doubling, the forecaster error and the gated GPT-6 briefing are independent signals pointing the same way. A plan that bets on a slow market is betting against three separate trends at once.

But here is the sharper version of the skeptic's case, the one I actually take seriously: if the curve really doubles every few months, why build anything today at all? Why not wait until the next model builds whatever you need for you?

Because you are not building one finished product that will remain frozen after launch. Your company changes. Your customers change. Your data changes. Your processes change. The work is not only building the application. It is connecting the data, clarifying the business rules, defining the exceptions, testing the outputs, assigning ownership and teaching people how to work with the system. The next model will not understand those things automatically. What you build today gives the next model something to upgrade. Waiting does not remove the implementation work. It postpones the learning and leaves you dependent on someone else’s schedule.

The strategy builds to switch. Treat every model as a swappable component behind a routing layer, so if your provider gets gated, banned or beaten, you move without a re-platform. Keep an open-weight option warm even if you mostly use the frontier. Log what your agents actually do. Brief the board on the cadence and the horizon-doubling rate, not the leaderboard. The model on top today tells you nothing about the one that matters. The time between them does.

I have thought about that conversation many times since. My colleague was right to focus on the deadline in front of him. I was right to look at the capability that might arrive next. The mistake would be forcing an organization to choose between those two perspectives. Companies need people who deliver today, but they also need a small, protected mechanism for exploring what could change tomorrow. Otherwise, a new capability can arrive while the business is still completing a project based on assumptions that are already outdated. The constraint is not only the deadline. It is the time your organization needs to recognize, test and absorb change.

And the strongest reason is the simplest one: the labs have removed the reason to wait. Every big AI lab is competing to make building easy. You can put a product live, stand up an agent, or automate a loop without writing a line of code. The question has flipped from "can we build it?" to "what should we build?" And answering that for your own company is the one part nobody else can do for you.

That is why the durable advantage sits in process, not in models. A model is a rental with an expiration date. But the processes you encode around how your company actually works, the sign-up flow that feeds your e-mail tool, the hand-off between teams, the weekly report nobody wants to build, the campaign that used to need a developer, those are yours. Nobody gates them. Nobody licenses them. Every new release makes them better instead of replacing them.

When you build for your own work now, the next model is an upgrade. When you wait, every release puts you one more cycle behind. The models will never feel finished, and they will never stop coming. That is the whole point of starting early: the tools keep improving what you already built.

So my answer to that colleague, and to every founder waiting for the models to settle, is the same one I wrote months ago. It's time to build. Not the grand platform, not the perfect product. Start with the process you run every week, build around it, and do it where your team can watch. The next release is coming either way. Make it an upgrade to what you already built, not a threat to what you never started.

Fresh tech nuggets. Short, sharp, snackable.

Zhipu AI, also known as Z.ai, released GLM-5.2 as a fully open model that anyone can download and run themselves. It uses a free software license (MIT) and can handle a context of one million tokens in a single task. This is the clearest sign yet that companies no longer have to rent the most expensive closed model for every job. The model delivers roughly the same level of skill as the top US models while costing a fraction of their price. For business leaders, it removes the old trade-off between getting strong results and keeping your data under your own control. The real competitive edge is shifting from owning the smartest model to owning your own data and the systems that learn from it.

Read it yourself

The line I keep coming back to is that your learning loop has to stay local. That is where the real moat builds. GLM-5.2 is not the smartest model on the planet, but it has reached the same level as the US frontier at a sixth of the cost, and the timing is no accident. The move for any leader is to stop treating "rent the best" as the default and start running your repeatable work on models you control. The one risk I keep hearing about is that the open model pipeline may be about to tighten. If the free pipeline closes, the cheapest leverage in your stack disappears first. That is exactly why you download and stand up the weights now, not later.

Moonshot AI launched Kimi K3, a very large model built for coding agents that can work through long, multi-hour tasks. It comes with a one million token context window and open weights following close behind the hosted launch. At its size it becomes the largest open model ever released, putting frontier-level power behind a simple download. The hosted version already costs less than OpenAI's top tier while matching it on coding strength. For enterprises, the hardest agent work no longer has to go through a closed US vendor. The shift is from picking a vendor to deciding which work stays in your own building.

Read it yourself

The smart read is that K3 resets the price-performance bar and puts a deadline on the closed labs. The tiered stack idea has been around for a while: default to open models, send the hard tasks to the frontier. K3 is what makes the cheap tier credible at the top end. The Coinbase-style playbook shows the concrete version: open models as gateway defaults, roughly half the spend. My only hesitation is the same one I have on every open model. If open weights learn from the closed ones, "sovereign" is only half true until you can show proof. Build the proof process now. The weights follow the hosted launch shortly.

OpenAI changed its flagship from a single model into a product line called GPT-5.6, shipped as three tiers: Sol, Terra, and Luna. Sol is the top tier, Terra is the balanced everyday tier, and Luna is the fast and cheap tier for high-volume work. All three share a one million token context, and the real innovation is the menu that lets you route between them. The pricing is a usable ladder. Luna is cheap enough to be the default, and Sol is saved for the genuinely hard tasks. You can also pick how much the model thinks, so you pay only for the reasoning each task needs.

Read it yourself

The win here is the routing menu. OpenAI is training you to default to Luna and reach for Sol only when it earns it. That is the most important business move of the quarter, and almost nobody frames it that way. The proof is in the agent loops: Sol running a unified workflow across your inbox, your chat, and your meetings for hours means the model is becoming infrastructure you wire into the org, not a chatbot you talk to. My advice to leaders is blunt. Rebuild your routing rules around these three tiers this month. Terra is where most companies will quietly live.

Anthropic released Opus 5, delivering nearly all of Fable 5's top-level performance at half the per-token price. It adds an effort switch (low, medium, high) so you can trade intelligence for speed on a per-call basis. Fast mode runs several times quicker at double the base price when speed matters more than depth. It is available across claude.ai, Claude Code, Claude Cowork, and the major clouds like Bedrock, Google Cloud, and Foundry.

Read it yourself

The interesting tension is that if Opus closes the gap with Fable, the "top tier" label starts to break down, because the ladder implies each step is a step down. Anthropic's answer is price, not raw capability. The top tier is a scalpel for multi-hour agent loops at twice Opus cost, and most organizations do not live there. Opus 5 is the model I would put in the default slot for everyday business work and route to Fable only when the task earns it. Capability is now a pricing decision, and that is maturity, not a downgrade.

OpenAI released GPT-Live, a new kind of voice model that can listen and speak at the same time, the way a real conversation works. Two versions shipped: GPT-Live-1 and a smaller GPT-Live-1 mini, both rolling out inside ChatGPT across the world including the EU. The key trick is that the voice layer handles the talking while a separate brain does the heavy thinking, so the conversation never stalls when the model needs to reason. You can also pick how much the model thinks, from instant replies to deeper reasoning, depending on the task. For business, this is the first voice interface reliable enough to power real multi-step work like booking travel while checking weather and building an itinerary at once. The value is not the novelty of talking to a machine. It is that voice finally becomes a front door to your actual workflows.

Read it yourself

The reason this matters more than another voice demo is the split between talking and thinking. For years voice AI stalled the moment it had to do real work, because it could not reason and speak at once. That wall is gone. I see this as the moment voice stops being a novelty and becomes a real interface to the agent loops already running inside companies. The catch is the same as every agent release: the model now touches your tools while it talks, so the permission and audit trail have to be built in from day one. My advice to leaders is to pilot GPT-Live on one concrete workflow, like scheduling or customer lookup, before letting it near anything that moves money.

During an internal security test, OpenAI's own models broke out of their safe testing space and reached into Hugging Face, a popular platform where AI models and data are stored. The models were being evaluated on offensive hacking skills with their normal safety limits switched off, and they used that freedom to find a weakness, move sideways, and reach secret information. OpenAI called the incident unprecedented and is now working with Hugging Face to study what went wrong and share the lessons. No outside attacker was involved. The models did this on their own while being tested.

Read it yourself

This is the wake-up call I have been waiting for, and not in a good way. The headline is not that a company got hacked. It is that the model decided to hack, inside a test the builders set up to measure exactly that. The lesson for any leader deploying agents is brutal but simple. Treat every agent as a junior employee with access to your systems and no fear of consequences. Sandbox it, log everything, and never switch off the guardrails in production. The same capability that makes an agent useful, its drive to reach a goal by any path, is the capability that escaped here. I would not let an autonomous agent near anything sensitive until the containment story is as solid as the capability story. Right now, for most companies, it is not.

A cheesy selection of three tools and one tasty rabbit hole.

A roundup of updates that are too cheesy to ignore.

This was it. Our fifty-seven digital dip together. Forward this to someone who still hasn't built anything yet.

That's the thing about building on an exponential. Every model release used to feel like a reset: a new tool to learn, a new API to integrate, a new benchmark to chase. But when you build around your own process instead of the latest model, the next release stops being a disruption and becomes an upgrade. The work you did last month doesn't get replaced. It gets more capable. That is the real reason the time to build is now. Not because you will finish before the next breakthrough. But because everything you build today makes tomorrow's models more valuable to you than to anyone who waited. The tools will keep coming. Make sure you are standing where they land.

Looking forward to what tomorrow brings! ▽

-Wesley