Steven Drost

If you try to always learn new things, you realise you spend a lot of time talking and thinking about things you’ve - best case - only just understood. Means: your communication with others about those things is probably pretty unpolished. Or something.

Happy to share new progress in AI for Maths @GoogleDeepMind . In extremal combinatorics, AlphaEvolve has helped establish new lower bounds for FIVE classical Ramsey numbers - a problem so challenging that even Erdős commented on its difficulty. Historically, computationally deriving these bounds required bespoke, human-designed search algorithms. For many of these bounds, the best previous results are at least a decade old. AlphaEvolve changes this by acting as a single meta-algorithm that automatically discovers the search procedures needed to find these new bounds. 📷

One of the greatest charts I have ever seen

The "There’s an app for that" era is officially over. 💀 We’ve reached Peak App Fatigue. Users don’t want to manage 50 different icons, subscriptions, and notification badges anymore. They want outcomes, not interfaces. So, why build apps at all? Because the "App" is changing from a Destination to a Data Source. The New Stack: 1. The User: Expresses intent (e.g., "Book a flight to NYC and find a gym nearby with a squat rack.") 2. The AI Agent: The new OS. It navigates the web so the user doesn't have to. 3. The App: The specialized "worker" that provides the API, the logic, and the specific utility the AI needs to fulfill the request. We aren't building for human eyes anymore; we’re building for Machine Consumption. If your app doesn't have a robust API or "Agentic" compatibility, you aren't just losing users - you’re becoming invisible to the AI they use to run their lives. The purpose of building an app today isn't to steal 10 minutes of screen time. It’s to provide the most reliable, permissionless infrastructure for an AI to get the job done. 🏗️🤖

Great analysis and framing. This piece filled many gaps in my humble thinking around AI, data centres, costs, etc.

@levie ‘Global agent’ vs ‘subagent’ divide reminds me of ‘global’ vs ‘local’ scepticism in philosophy of epistemology. Wonder if the same discourse, solutions, hacks will play out in agentic AI systems? So ‘tracking the truth across possible worlds’, etc.

We’re starting to get a clearer sign of how vast the surface area of context engineering is going to be. To build AI agents, in theory, it should be as simple as having a super powerful model, giving it a set of tools, having a really good system prompt, and giving it access to data. Maybe at some point it really will be this simple. But in practice, to make agents that work today, you’re dealing with a delicate balance of what to give to the global agent vs. a subagent. What things to make agentic vs. just a deterministic tool call. How to handle the inherent limitations of the context window. You had to figure out how to retrieve the right data for the user’s task, and how much compute to throw at the problem. How to decide what to make fast, and suffer potential quality drops, vs. slow but maybe annoying. And endless other questions. So far there’s no one right answer for any of this, and there are meaningful tradeoffs for any given approach you take. And importantly, getting this right requires a deep understanding of the domain you’re solving the problem for. Handling this problem in AI coding is different from law, which is different from healthcare. This is why there’s so much opportunity for AI agent plays right now.

AI is crazy because you deeply need to avoid being sentimental about any part of your product architecture at all times. The models are upgrading at such a fast rate that you have to constantly reevaluate the scaffolding you’ve built, and figure out what now can be solved in a better or cheaper way due to a new breakthrough. Companies will win or lose entirely by their ability to let go of something that they ostensibly got good at because the models can now solve that for them. This is generally where startups win over time because they emerge in a period when something is far easier to solve in a modern way, and the incumbent doesn’t properly adapt. The key is to ensure you have an architecture that gives you this flexibility, which means creating the right abstractions early on to benefit from these constant updates.

@roisinmurphy You are an amazing talent. Your video to Fader is deeply beautiful.

I believe I’ve blocked several million simpletons (well it feels like it) who have said “I never got the hype anyway” good for them but there was no “hype” ever.

If innovation and Scottish public services is your thing, get involved! 👇

over the next few years a majority of the population will outsource their taste, their opinions, their thoughts, their work, and ultimately their lives to language models. the great cultural norming is coming, you will be unbelievably powerful if you resist model capture

citadels law: any industry with sufficiently high stakes will end up mirroring the culture of hedge funds massive cash comp for the top performers, bid away dynamics where entire teams walk together, extreme litigation of non competes, hardo work 24/7 culture, short tenures

I used AI for this.

EVERYONE dumb (ok, naive) enough to think the west is ahead of China on AI and that we “must win” … has to watch this brilliant exchange between @altcap and @bgurley. SPOT FUCKING ON.

@YotelHQ urgently trying to get hold of your front desk at YOTEL Ginza. Please advise!

I'm at YC demo day and it's a new era. Out: "Uber for Dogs" In: "Anduril for Dogs"

Teaching ethics to undergrads in 2025 is bizarre, because: (1) They insist morality is entirely relative and culturally constructed... (2) ...while simultaneously holding unshakeable ethical convictions and viewing disagreement as moral monstrosity.

We’re an AI enabler in the Scottish tech ecosystem ⁦@drostyboy⁩ ⁦@CodeBaseTech⁩ ⁦@TheScotsman⁩ ⁦@scot_business⁩ scotsman.com/news/were-an-a…

MY MIND HAS JUST BEEN BOGGLED. I just tried the early release of OpenAI's new DeepResearch feature (rolling out later today to the $200/month Pro users). I've been working in the CRM software industry for 30+ years. It's not just that I've had court-side seats to the game, I've been on the court, doing my best to play the game. First in vertical CRM (my first startup) and now as co-founder of HubSpot. I've had some modest success and I feel like I have a pretty good handle on things in the industry. That's why OpenAI's new DeepResearch feature boggled my mind. I asked it create a detailed research report including competitive analysis, positioning, growth, product strategy and AI vision for the industry. What it produced was an 11,000 word report. With data. And citations. And tables. And genuinely great insights -- including some I hadn't really thought of before. What has me excited is not just that it can produce this kind of output (though that's pretty cool). What has me excited is that we'll be able to use this kind of output as *input* to a subsequent step in an agentic workflow. Because the future is about agent composability. Being able to pull together pieces and put them together into a larger whole. The same way we build teams to work together to tackle higher order missions and goals. This has been what I've been dreaming about for years. It's finally starting to happen. We are seeing more and more of the future -- and it's happening quickly. DISCLOSURE: I'm a small investor in OpenAI, and also a big fan.