Edition 23 - 2025

Quote of the week

“If you want something you’ve never had, you must be willing to do something you’ve never done.”
- Thomas Jefferson (attributed)

The Future of LLMs

In a surprising turn, Anthropic has pulled access to some of its Claude coding models from third-party platforms like Windsurf, reportedly due to reports that OpenAI might acquire the startup. Co‑founder Jared Kaplan explained in a TechCrunch interview that “it would be odd for us to be selling Claude to OpenAI,” and emphasized their focus on sustainable, long‑term partnerships.

This move by OpenAI raises important questions: are the big AI labs moving away from platform-based models toward vertical integration within their own ecosystems? We're starting to see unique partnerships being formed in specific verticals such as coding, social media, phones, and other more niche areas. Is this a sign of greed and control, or simply a recognition that incremental model improvements aren’t delivering the breakthroughs they used to? And what does that mean for startups and developers who’ve been building on top of these APIs?

Meanwhile, Apple has published intriguing research (“The Illusion of Thinking”) showing that current large reasoning models collapse completely when problem complexity crosses a certain threshold. The paper argues that LLMs primarily simulate reasoning rather than genuinely understand and think, and their performance can degrade dramatically when even slight transformations are applied to input.

So, what could spark the next leap forward in AI? We may be looking beyond larger LLMs for the breakthrough. Possible paths include hybrid systems that combine symbolic reasoning with neural networks, modular architectures that mirror human cognitive processes, or grounded models with real-world understanding and memory. At this point, LLMs feel “good enough” for many applications and squeezing out more quality may not be worth the effort. The next visible wave may come not from adding size, but from fundamentally new paradigms.

A Model on Your Phone?

Google has rolled out a new mobile app (Edge Gallery) that allows users to run powerful AI models directly on their smartphones without needing an internet connection. Users can download models ranging from lightweight (around 500MB) to more robust options (up to 4GB) and use them for tasks like chatting, image analysis, and single-prompt responses. These models run entirely on-device, delivering low-latency performance with strong privacy protections—even in areas with limited or no connectivity.

One highlight is Gemma 3n, a compact, multimodal model designed for mobile use that performs competitively with leading models in text and image tasks. While it's limited to knowledge up to June 2024, it’s optimized for fast, local inference. This move marks a broader shift toward making AI faster, more private, and more reliable—especially useful in industries where data security is essential and real-time performance is non-negotiable.

How does this work? To run AI models offline on smartphones, the app uses compact model designs like Google’s Gemma 3n, built to be lightweight yet powerful. Techniques like quantization reduce model size and resource demands by simplifying how data is stored and processed. The app also uses a mobile-optimized engine that runs efficiently on your phone’s hardware without overheating or draining the battery. Once downloaded, the model operates entirely on-device, with no internet or cloud connection, enabling fast, private AI experiences anytime, anywhere.

What are the limitations? Running AI models locally on smartphones has some real limitations. Right now, it mostly works on newer Android devices with strong processors and enough storage. If your phone is older, underpowered, or low on space, the experience may be sluggish or simply not work. iPhones aren’t supported yet, though iOS compatibility is on the roadmap. And because these models run offline, they can’t access real-time information or update their knowledge past their training cutoff. So while it's fast and private, it's not quite plug-and-play for everyone yet.

What could this technology unlock in the future as it improves? As on-device AI becomes more powerful and efficient, it could enable fully private, personalized AI assistants that work anywhere, even in remote or sensitive environments. Think real-time translation, offline customer support bots, secure document summarization, or intelligent camera features that don’t need to send any data to the cloud. Eventually, it could lead to an entirely new class of offline, always-available AI-driven apps that are fast, secure, and highly responsive, making smart computing truly ubiquitous.

Why Winning Matters

Despite sweeping U.S. efforts, including export controls targeting AI chips and leading-edge technology to slow China’s tech rise, the results have fallen short. China has surged ahead in electric vehicles, consumer drones, and solar panels, and is rapidly closing the gap in AI and advanced microchips. Export restrictions, while intended to hinder progress, may have backfired, prompting significant domestic investment and accelerated self-sufficiency on China’s part. As China continues to build momentum across strategic sectors, it raises a deeper, more important question: what’s really at stake if we fall behind in AI?

This isn’t about nationalism, rivalry, or conflict. It’s about economic gravity. Historically, the country that leads in breakthrough technology ends up hosting the companies, infrastructure, and capital that shape the global economy. The U.S. dominated the internet era because it built the internet. The same is true for semiconductors, cloud computing, and social platforms. I talk about China constantly in this newsletter because the AI race is a war. Not in a militaristic sense, but as a battle for where the next generation of opportunity gets headquartered. While it’s not guaranteed that China taking the lead in AI would pull all the wealth, talent, and influence with it (especially given its many internal challenges) it’s certainly possible. And if we underestimate that possibility, we risk handing over the future of tech by default.

Meme of the week

Vibecoding GTM: Build What You Show

Using tools like Lovable, Bolt, Cursor, and Replit, GTM teams are starting to build interactive demos, landing pages, and calculators without waiting on engineering. These tools let marketers and operators get scrappy. You don’t need a ticket in the sprint. You can build what you need in a few hours. You can show value before product actually ships.

Imagine you’re on a customer call. They ask for a feature. You tell them you don’t have it. But by the end of the same call, you’ve built a prototype of it and dropped it in the chat. That changes the entire conversation. It’s not just a sales move, it’s a trust-building one. You’re showing them what’s possible in real time. It’s fast. It’s rough. And it hits.

There’s something powerful about that moment. When someone sees a working version of what they just asked for, they lean in. Curiosity turns into belief. Instead of “we’ll get back to you,” it becomes “let’s keep going.” It short-circuits the usual dance between sales, product, and engineering. You don’t need full buy-in, you need a spark. The psychology of immediacy can close deals.

But the tooling to make this a reality still lags. It’s messy to manage environments, share things securely, or connect them to the broader stack. And honestly, it’s annoying that Figma leaned consumer instead of enterprise. They could’ve owned this entire space. If Figma let teams use existing design systems and componenets to generate new design, it would be the bridge between design and product everyone’s been waiting for. But they didn’t. So now we hack it together.

Good News

Global efforts to combat neglected tropical diseases are gaining serious momentum, with medicine pledges soaring to 28 billion units—up from 19 billion just last year. Novartis has nearly doubled its five-year R&D investment to $490 million, focusing on diseases like malaria, leishmaniasis, and dengue. The boost comes alongside growing support for the Kigali Declaration, now backed by 84 governments and organizations. Since September 2024, an additional $441 million in funding has been pledged, pushing total commitments past $1.8 billion.

Scotland is taking bold steps to crack down on severe environmental harm with a proposed ‘Ecocide Bill’ that would criminalize large-scale ecological damage. The legislation would impose up to 20 years in prison for individuals and unlimited fines for corporations responsible for disasters like oil spills or deforestation. If passed, it would place Scotland at the forefront of global environmental accountability, targeting both companies and the leaders behind their decisions.

CAR-T therapy is showing game-changing results for patients with multiple myeloma, a blood cancer long considered incurable. A new study finds that the treatment—which reprograms a patient’s own immune cells to target and destroy tumors—can double remission time in advanced cases, with some patients remaining cancer-free for over four years. What sets this apart isn’t just the improved survival rates, but the personalized nature of the approach: each therapy is custom-built from the patient’s own cells rather than relying on generic drugs. Doctors believe this could mark a turning point, with CAR-T potentially becoming standard care for a wider range of cancers.

Scientists in Japan have developed a new type of plastic that dissolves completely in seawater within hours, leaving behind no microplastic residue. The material matches the strength of conventional plastics but breaks down into harmless components that can be consumed by natural bacteria, and it also decomposes in soil within about 200 hours. Non-toxic, non-flammable, and carbon-neutral, this innovation offers a promising, environmentally friendly alternative for applications like packaging, though it is still in the early stages of commercialization.

Tools

Share with friends

Enjoying The Hillsberg Report? Share it with friends who might find it valuable!

Haven't signed up for the weekly notification?