Nvidia CEO Jensen Huang just said it out loud: “I think we’ve achieved AGI.” Last year, saying this would have sounded like science fiction. Today, the CEO of NVIDIA says we’re basically there. At the same time, Anthropic has been quietly testing a new model called ‘Mythos’ with select customers. Insiders call it a “step change.” Dramatically higher scores in coding. Academic reasoning. Cybersecurity. Some even say it is far ahead of any other AI model in cyber capabilities. It’s part of a new Capybara series, larger and smarter than their current frontier model Opus. More expensive too. Not ready for the public, yet. And reportedly so powerful that rollout needs to be slow, because of security risks. And they’re not alone. OpenAI is in the same direction. Their upcoming model, code-named ‘Spud’, is described as a massive leap with serious economic impact. They even renamed their department to “AGI Deployment.”… We’ve mentioned it in the previous edition of Digital Dips: Morgan Stanley warned that a massive leap in AI capability could hit in the first half of 2026. They we’re not lying. 2026 looks like the year we hit the steep part of the curve.

On today’s menu: Europe building AI on its own terms, models that improve themselves, Claude stepping onto your desktop as a digital co-worker, OpenAI making hard cuts to double down on what it believes will move the global economy, and memory getting 6x lighter while the race to AGI gets louder.

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.

Fresh tech nuggets. Short, sharp, snackable.

Mistral AI has launched Forge, a new platform for enterprises to build AI models using their proprietary data. This shift caters to the need for AI solutions that internalize company-specific knowledge, moving beyond generic models. Key partners like ASML and the European Space Agency are already on board. As AI becomes a cornerstone of enterprise tech, the focus shifts to strategic autonomy and customization.

Read it yourself

It's refreshing to see Europe investing in AI infrastructure rather than just policy tweaks. Mistral's Forge stands out by anchoring AI in the unique knowledge of each enterprise rather than offering a one-size-fits-all solution. The real battleground isn't over who has the best model but over who can provide genuine ownership of context without compromising on security or control. This approach could redefine AI utility in Europe where regulatory landscapes are challenging and ownership matters. More tailored solutions like this are likely what we'll see shaping the future of AI in tightly regulated sectors.

MiniMax M2.7 isn't just another AI model; it's actively shaping its own development. By running over 100 iterations, it improved its performance by 30%. The model also monitors research workflows, covering 30-50% of tasks, and excelled in machine learning competitions, achieving a 66.6% medal rate. This marks a shift towards AI models capable of self-evolution, indicating a future where AI systems might operate with minimal human intervention.

Read it yourself

That's really fascinating: MiniMax M2.7 participated in its own development. Watching it autonomously run over 100 loops, analyzing failures, tweaking code, and improving performance is impressive. And it doesn't stop there: handling up to half the research workflow and excelling in competitions shows its capability. MiniMax is clearly moving towards fully autonomous AI processes. This isn't a theory anymore; it's a glimpse into a future where AI models manage themselves.

Google AI Studio has launched a revamped development experience that bundles everything needed to create production-ready apps directly in the browser. Key features include the new Antigravity coding agent and built-in Firebase integration for databases and authentication. This streamlined approach could reshape the market by making scalable app deployment as easy as writing a prompt. The first-party infrastructure Google provides could be a game-changer for developers.

Read it yourself

This is where it gets interesting. And a little painful for some parties. While startups like Bolt, Replit and Lovable focused on making creation feel smooth and fun, Google quietly bundled the part that actually matters in production. Backend, hosting, auth, database. Live in one tab. And for free. The others charge monthly fees and still ask you to glue together five tools before you can ship something that doesn’t break. And when things do break? Token bills climb. Context disappears. The AI starts rewriting parts of your app you didn’t touch. The real gap wasn’t building nicer prototypes. It was owning the layer underneath. And Google owns that layer. With millions of developers already inside their ecosystem, fixing the “front door” is a much smaller problem than building world-class infrastructure from scratch.

Google just unveiled TurboQuant, a new algorithm that reduces large language models' working memory by a stunning 6x. This shift doesn't compromise accuracy and speeds up processing by up to 8x. Historically, AI's memory demands have taxed both technology infrastructure and budgets. With TurboQuant, running AI becomes vastly more efficient, cheaper, and scalable. The ripple effect could dramatically alter existing memory and GPU market dynamics. Ultimately, it suggests a potential fading into the background of expensive AI hardware costs.

Read it yourself

Google may have just taken a needle to the AI memory hype, and the market reacted fast. In simple terms: AI systems need a lot of temporary memory to remember your conversation, and that memory has become one of the most expensive parts of running AI. Google now claims it can shrink that memory by 6x without making the model worse, and even make it up to 8x faster. If that holds, companies won’t need as much high-end memory hardware to serve the same number of users. That matters because AI has pushed memory prices up sharply, based on the belief that more AI always means more chips. But if smart software reduces the need for that hardware, the economics change. The same machines could handle more users, longer conversations and lower costs. Nothing breaks overnight, contracts are signed and demand is still strong, but strategically this is big: when software removes a bottleneck, entire markets quietly reset.

Claude has taken a big step forward, now capable of autonomously controlling your entire computer. This includes opening apps, navigating browsers, and filling spreadsheets. Running initially on macOS, it enhances productivity by allowing tasks to be assigned from a phone and completed on a desktop. The strategic move positions Claude as a "digital employee" capable of handling complex computer tasks, adding another layer to what AI can do right now.

Read it yourself

Anthropic just delivered the biggest AI product launch of the year so far. In 30 days they rebuilt OpenClaw, the open-source agent that pulled 250K GitHub stars, but this time wrapped in enterprise guardrails. Where OpenClaw let you text an agent from WhatsApp to control your desktop, Anthropic shipped Dispatch: persistent threads from phone to desktop inside their own app. Where OpenClaw used Discord and Telegram as control panels, Anthropic added Code Channels with controlled bridges. Where OpenClaw offered full OS access and 100+ unreviewed skills (and the malware that came with it), Anthropic introduced curated plugins, admin controls, permission prompts and sandboxed execution. Even the always-on “heartbeat” daemon has an opposite: Claude only runs when your desktop is open, friction by design to prevent runaway automation. The strategy is obvious. Let open source test the edges. Then ship the safe version for the enterprise before anyone else does. Yes, there are gaps. No WhatsApp or iMessage integration yet. No headless mode. Your Mac must stay awake. But make no mistake: this is a digital employee. Claude can use your screen like a human, any app, any browser, any spreadsheet, and you can text it from your phone and return to finished work. Meanwhile OpenAI hired the builder of OpenClaw, announced internal reorganizations, and is still running agents in cloud sandboxes without local file control or true async desktop handoff. The demand signal is clear: executives want an AI that works on their machine, not in a demo browser. Anthropic shipped it. Others are still aligning slide decks.

OpenAI is shutting down its Sora app and discontinuing all video generation capabilities inside ChatGPT. Despite a $1 billion deal with Disney only four months ago, the company is pivoting away from video to focus on its newest language model, codenamed Spud. This decision follows the significant revenue growth displayed by competitors like Anthropic, which have thrived by concentrating on text and code solutions. Expect future announcements on timelines and data preservation for current Sora users.

Read it yourself

OpenAI didn’t tweak the video strategy. They killed it. App gone. API gone. No video inside ChatGPT. And Disney’s $1 billion deal? Dead in four months. This is not a product clean-up, it’s a full reset driven by economics. While Sora burned GPUs and delivered $1.4 million in consumer revenue, Anthropic quietly built a $19 billion run rate with one surface: chat, code, enterprise. No video circus. OpenAI looked at the numbers and followed the money. So now ChatGPT, Codex and the browser become one focused product, Instant Checkout is cut, and the Sora team moves to robotics. Even the product org gets renamed to “AGI Deployment,” which tells you how leadership wants this story framed. Either they are closer to something big, or this is the sharpest strategic pivot we’ve seen in years.

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-three digital dip together. Forward this to someone who sees AI as a feature, while others use it to change the game.

If this edition made one thing clear, it’s this: The models are not the bottleneck anymore. Anthropic ships digital employees. OpenAI restructures around deployment. Mistral AI builds Forge so enterprises can train on their own data. If you’re a leadership team thinking about autonomy, control, and real operational impact, this is where I can support you. Not by adding another layer of tooling. But by helping you translate these shifts into structure. Mapping how your organization actually runs. Identifying where knowledge should live inside models. Designing workflows where agents don’t just assist, but own outcomes. The labs have made their move. If you want to make yours, I’m here to help you. Reply and let’s design how this fits inside your business.

Looking forward to what tomorrow brings! ▽

-Wesley