🧾 Weekly Wrap Sheet (05/30/2025): Lag, Leverage, Loops & the Lab from China
DeepSeek distills intelligence, NVIDIA electrifies it - and the context layer remains missing in action.
🎬 TL;DR
Most people are still speculating about what AI might do one day. The problem? That 'one day' happened two weeks ago.
AI adoption isn’t a software install - it’s a cultural reset. Klarna and Duolingo showed us what happens when tone fails, even if the tech doesn’t.
AI isn’t advancing linearly. It’s compounding. We’re critiquing prototypes while they’re rewriting the playbook mid-flight.
NVIDIA’s earnings weren’t just strong - they were a dispatch from the frontlines of a global re-platforming.
We’ve built intelligence. What we’re missing is memory. Until we solve for that, AI will keep forgetting who it's working for.
DeepSeek released upgraded models - with better reasoning, open weights, and dramatically lower compute.
⏳ The Lag Is the Real Threat
The most surreal part of this AI moment isn’t the tech, it’s the lag. The lag between what most people think is possible and what’s already happening.
I talk to friends in medicine, law, media - smart, thoughtful people - who speculate about what AI might do one day. But the "one day" came and went two demo days ago. We’re moving at a blink-and-you’ll-miss-it pace, and most people haven’t looked up yet.
This isn’t a rant about hype. It’s a warning about asymmetry. The tech is accelerating, but awareness isn’t. Policy isn’t. Public conversation isn’t. Even our imagination is lagging behind.
And that’s what makes this moment so destabilizing. There’s a growing gap between those building the future and those about to be blindsided by it.
The future doesn’t need your permission to arrive. But it desperately needs your attention.
🧠 AI Isn’t Just a Product Decision. It’s a Cultural One.
There’s been plenty of schadenfreude over Klarna walking back its AI rollout and Duolingo bungling its messaging. But behind the stumbles are strategic lessons. They didn’t have the wrong tech. They had the wrong tone.
So, what does good AI leadership look like?
AI = Leverage, not Layoffs. If the headline is "we cut 100 jobs," you’ve already lost the room. Show how AI empowers teams and elevates experience.
Adoption is a Workflow Rewire. Don’t start with reorgs. Start with use cases. Observe. Learn. Then rebuild.
Trust is Your Throttle. Change creates fear. Narrate the shift like a human, not a shareholder update.
Split the Org. A stable core + an experimental edge. Run parallel operating systems. Isolate risk. Feed insights back.
Lead with Empathy. If it touches people, design with dignity. Especially in high-emotion functions like support, HR, or legal.
AI multiplies what you already are. Trust or fear. Clarity or chaos.
🚀 Acceleration Is the Feature
Most AI critiques miss the plot. They judge today’s quirks like they’re final versions - as if the tech isn’t still learning, compounding, updating while you’re complaining. Anyway, here are a few datapoints that might ruin the next “AI isn’t that good” take:
🧮 In 2021, GPT-3 scored 5% on the MATH benchmark. That’s not a typo. Five. Percent. The paper said it would take new algorithms to do better. In 2024, GPT-4 and Claude 3.5 scored over 90%. Apparently the new algorithm was “try again.”
🎥 In 2020, making a two-minute AI video cost $15,000 and 50 hours of HD training footage. In 2025, you can type “a cat doing yoga in a thunderstorm, cinematic” and get it in 12 seconds for pocket change.
💼 GPT-3 barely understood legal text. GPT-4 scored in the 90th percentile on the Bar Exam. Which, frankly, is higher than some actual lawyers.
🧠 In 2022, an “AI assistant” helped autocomplete your sentence. In 2025, GitHub Copilot writes, tests, and debugs full stacks of code. Microsoft has agents that run entire enterprise workflows. Somewhere a middle manager is being replaced by XML.
🗣️ In 2023, AI voices sounded like your dentist trying improv. In 2025, GPT-4o can gasp, flirt, interrupt, and joke in real time. It's not just talking. It's doing a table read.
Critiquing today’s AI for what it can’t do is like criticizing the Wright brothers for not offering lie-flat seats. The thing is off the ground. That’s the part to focus on. As Demis Hassabis put it: “Whether AGI arrives in 2, 5, or 10 years - they’re all short timelines.”
💰 NVIDIA’s Earnings = The State of AI
NVIDIA’s Q1 2026 results weren’t just earnings. They were a geopolitical artifact - evidence of how power, policy, and silicon are reshaping the world.
They lost access to a $50B China market overnight. Wrote off $4.5B in unsellable chips. Still grew revenue 69% YoY to $44B.
🔌 Sovereign AI Is the New Arms Race. Nearly 100 AI factories are underway as nations race to localize their stack.
🌐 Full-stack control is the moat. Spectrum-X and NVLink now make up billions in revenue - and cement NVIDIA’s grip on everything from networking to orchestration. Once you buy the chip, you buy the software, the switches, the roadmap, and the worldview.
🧠 Inference demand is exploding. As models go from autocomplete to agents, token throughput has gone parabolic. Microsoft processed 100 trillion tokens last quarter, 5x YoY. That’s a generational shift in compute demand.
🛠️ Export policy is reshaping strategy. Jensen didn’t mince words: “The question is not whether China will have AI. It already does. The question is whether one of the world’s largest AI markets will run on American platforms.”
You can call them a chipmaker. But really, they sell the hardware that turns electricity into intelligence. And the world is buying.
🗂️ The Layer AI Forgot: Memory
I'm a daily (hourly?) AI user. But these tools still treat me like a stranger. I repeat myself across apps, models, and chats - stuck in a Groundhog Day loop of contextless first dates with memoryless machines.
The problem isn’t intelligence. It’s memory.
We need a Personal Memory Vault.
🔐 Encrypted, private by default
🔄 Portable across models
🧩 Modular, permissioned, composable
📜 Versioned and auditable
Think: Plaid, but for personal context.
Let developers request memory bundles with user consent. Let users track what’s being used. Let the vault plug into any assistant, agent, or app.
It’s not just about usefulness. It’s about control. Big Tech will offer memory next. But it’ll be a lock-in tactic. We need this layer to be open, sovereign, and user-owned.
I don’t want a hundred context engines. I don’t want to be platform-loyal out of sunk-cost guilt. I just want to stop going on first dates with my own data.
🔍 Seek Deep and You Will Find: Scale, Speed, and a Growing Threat
This week, DeepSeek dropped two new models on Hugging Face - and both are telling.
First up: an updated version of its flagship R1 reasoning model, now scaled to 685B parameters and released under a permissive MIT license. The upgrade brings sharper reasoning, better handling of complex tasks, and benchmark gains that nudge it closer to OpenAI’s o3 and Google’s Gemini 2.5 Pro, but with weights you can actually download.
The real strategic move, though, might be its distilled counterpart: DeepSeek-R1-0528-Qwen3-8B. Trained by fine-tuning Alibaba’s Qwen3-8B on outputs from R1, it clocks elite performance - outscoring Gemini 2.5 Flash on AIME 2025, and nearly matching Microsoft’s Phi-4 on HMMT. All that in a model that runs on a single 80GB GPU.
Why it matters: DeepSeek isn’t just scaling up - it’s scaling down with purpose. This two-track strategy lets them play in both cutting-edge research and commercial deployment. If others keep chasing parameter counts, DeepSeek might win by distillation.
But there’s a catch. As TechCrunch noted, the new model appears to be even more heavily censored than its predecessor, particularly around political criticism. That’s the tension at the heart of DeepSeek’s rise - open-weight, but not necessarily open expression.