Robin Allenson's Linkedin Analytics

Someone sent me ‘pic 1’ — so I asked ChatGPT to make ‘pic 2’. When someone says, "LLMs are just predicting tokens” or more generally “something is just something else”, that’s rarely the full story. We understand the logic of predicting tokens and how it does what it does, but there are lots of emergent behaviours which are interesting. At our core, humans are all made from the same stuff — but there’s a bunch of physical and chemical processes (biology) that miraculously turns us into amazingly complex life forms. Knowing how something is built does not mean knowing the complexity that emerges in what it's capable of. Reductionism can be useful until it’s pushed so far that it becomes absurd.

“Imagine a world where, for decades, everything–housing, education, food, clothing, etc.–became half as expensive every two years. It sounds utopian, but it’s something technology can deliver (and in some cases already has).” Sam Altman said the above in a blog post almost 4 years ago. Titled ‘Moore’s Law for Everything’, he goes into detail about a system that prioritises taxing companies and land to reduce wealth inequality. He then also highlights 3 crucial consequences of a technology revolution that is apparently already underway: 1️⃣ The price of labour will drop close to $0 as AI ‘joins the workforce’. 2️⃣ Drastic changes in policy will be a must. 3️⃣ If both of the above are handled well, standard of living will be better for everyone than ever before. 🤔 This a pretty cool, unemotional analysis at an economic level. And it seems plausible but I think it will take a lot of time in practice before we get there. It could be that as soon as enough people see that it is happening, a lot of the impact has already happened. Having some kind of sovereign wealth fund or universal basic income seems like a great idea to me, but I suspect the devil’s in the details. A lot of ideas sound like great ideas but when you live in a society that runs on ‘great ideas’ – it’s a mess. Humans have great capacity for screwing up loads of these things.

Serious question: Would you have an issue with most customer service roles being replaced by AI? We’ve all been in a situation where you’re yelling at a robot through the phone or trying to get a chatbot to understand your problem and getting nowhere. There are few things in life that are more frustrating. But in these scenarios, the problem isn’t that it’s a robot helping – it’s that it’s not helping at all. So if AI was able to make quick decisions and problem solve, I don’t see any good reason why these roles shouldn’t be replaced by AI. Right?

GPT-4.5 is objectively better than 4o, but most users are struggling to notice differences in everyday use. The improvements are subtle: slightly more Claude-like in its emotional intelligence, with marginally more nuanced writing capabilities.  But this is nothing like the shift we saw from 3.5 to 4.0. With each 0.5 release, the development was exponential – the model is 10-fold larger – yet it now appears that train-time compute scaling is slowing down. In reality, it's a lot more nuanced than that. Progress now requires balancing both horizontal scaling (train-time compute that builds broader world knowledge) and vertical scaling (inference-time compute that enables deeper reasoning). Most low-hanging fruit is in inference optimization, but to make the next big leap, we'll need breakthroughs in training approaches as well. Sooner or later, you'll need to improve train-time compute in order to further scale instance-time compute. I’ve noticed an important shift in how we perceive AI progress. We've gone from groundbreaking model releases to debating minute differences in nuance and capability. GPT-5 will likely involve combining 4.5's broad base with o3's nuanced reasoning - but the bigger question is whether that will feel like the kind of transformative leap we've come to expect 🤔

A recent study by The World Bank has proven that AI is revolutionising access to education, and it’s made me a bit frustrated. The study (linked in the comments) showed how AI-driven tools can act as a tutor and assist in addressing some significant challenges like staff shortages and limited resources in Nigeria’s education system. The tools were able to offer personalised learning experiences which led to significant improvements in performance across the board. What frustrates me about this is that this is exactly the kind of positive impact AI can have on society, yet it’s not something we hear about enough. Too often, conversations around AI focus on the risks or the "what-ifs," – but stories like this remind me of the incredible potential it has to transform lives for the better. I would love for conversations around AI to center more around these kinds of topics. But I guess that’s not what sells…

Is cognition a spectrum? Is intelligence? This might be a confusing topic but stay with me. Michael Levin is a biologist known for studying something called ‘diverse intelligence’. He looked at really, really simple systems – like groups of cells or basic sorting algorithms – and found that they exhibited memory and problem-solving abilities. Both of which are hallmarks of cognition. And that there are layers of problem-solving systems all the way up. But these things don’t have a ‘brain’, which leads me to ask… If these abilities can emerge from simple systems, what does this mean for how we define cognition or intelligence? After all, aren’t all the problem-solving systems we know just made of the same stuff?

What REALLY matters in software startups? Especially as AI continues to shake things up? Paul Graham’s point was this: most software service companies aren’t hired just for their software. They’re hired because they understand the problem. Even if AI becomes insanely good at writing code (to the point where you barely need programmers), the core of starting a startup won’t change. Why? Because the hardest part isn’t building the software — it’s figuring out what to build in the first place. And the best way to understand a problem is to live it, iterate on it, and learn from it over time. For example, at Similar.ai, we’ve spent years refining our software. We know what’s hard, what’s easy, and what nuances matter. And we’ve learnt what to build from years of seeing how our customers’ needs evolve. If someone asked AI to “just build something that does this,” it might spit out code, but it wouldn’t capture the depth of understanding we’ve built over time… For now, that is.

Throughout my life, I’ve come to learn that not knowing is a superpower. And it seems AI is learning that too. In some recent models, they’ve been able to teach the model to say “sorry I don’t know the answer to that” when it’s unsure about something as an antidote to hallucination. If you get wildly different answers to the same question, you can train the model to understand what uncertainty feels like. This doesn’t go all of the way to fixing hallucinations but AI recognising uncertainty and having the introspection to be able to say “no, I don’t know the answer to this” is super interesting to me. And certainly something some humans can learn from 😅

The way we think about money and the value of everything is going to be fundamentally disrupted by AI. The problem with most tech is that the pricing of tech per unit has been going down exponentially. But tech is evolving to also potentially include intelligence. Which means the price of ‘expertise’ or ‘smarts’ or other synonyms of ‘intelligence’ will go down as AI gets increasingly intelligent and more accessible. We’re used to TVs getting cheaper and healthcare getting more expensive – but what if everything got cheaper? The only things that wouldn’t are things that have some kind of brand value, or some physical friction. As far as I can imagine anyway. Do you have a better idea of what that world might look like?