Generative AI

On LLMs, Privacy and Inadequate Silicon

I’ll admit it feels trite writing a post here — especially in a world where people rely on LLMs so heavily that parsing whether something is an original thought seems almost pointless. That said, I’ve grown quite attached to ChatGPT, Claude, and Manus AI over the last few months. There’s real appeal in using them as thinking or conversation partners rather than merely answer engines. They’ve been enormously useful for pressure-testing models and interrogating assumptions: from crafting mental frameworks for everyday problems to building out complex financial models (‘how much more will I need if I decide to swap out the 4% rule for a 3% rule?’).

The more I use these tools, the guiltier I feel about feeding so much data into them. Naturally, I’ve spent time experimenting with local models like Mistral 7B and LLaMA 3 8B. Both are impressive and get the job done, but they’re not as smart or fast as ChatGPT — or as strong at coding as Claude (yes, guilty of vibe-coding).

Using Ollama and Open WebUI abstracts much of the complexity, but RAG (retrieval-augmented generation) pipelines have been necessary to make local setups even slightly multi-modal. The real trick now is figuring out what I do locally and what runs through ChatGPT or Claude.

Working locally isn’t just about privacy — it changes your relationship with the tool. Some lessons learned from running local models:

Speed matters — an M1 MacBook Air doesn’t quite cut it when you need speed or performance.

Context windows are a bigger limitation than parameter counts.

Local control forces discipline: understanding memory usage, disk space, prompt structure, and retrieval mechanics.

All of this is to say that the last couple of years have been a lot of fun. It genuinely feels like it’s accelerated my sense of wonder and pace of learning.

PS - I was overusing em dashes way before LLMs.

(Meredith Whittaker) On Signal, Encryption and AI

Wired has an interview with Meredith Whittaker from Signal - her stances on Surveillance Capitalism, Signal’s not-for-profit structure and AI make for very interesting reading.

Yeah. I don’t think anyone else at Signal has ever tried, at least so vocally, to emphasize this definition of Signal as the opposite of everything else in the tech industry, the only major communications platform that is not a for-profit business.

Yeah, I mean, we don’t have a party line at Signal. But I think we should be proud of who we are and let people know that there are clear differences that matter to them. It’s not for nothing that WhatsApp is spending millions of dollars on billboards calling itself private, with the load-bearing privacy infrastructure having been created by the Signal protocol that WhatsApp uses.

Now, we’re happy that WhatsApp integrated that, but let’s be real. It’s not by accident that WhatsApp and Apple are spending billions of dollars defining themselves as private. Because privacy is incredibly valuable. And who’s the gold standard for privacy? It’s Signal.

I think people need to reframe their understanding of the tech industry, understanding how surveillance is so critical to its business model. And then understand how Signal stands apart, and recognize that we need to expand the space for that model to grow. Because having 70 percent of the global market for cloud in the hands of three companies globally is simply not safe. It’s Microsoft and CrowdStrike taking down half of the critical infrastructure in the world, because CrowdStrike cut corners on QA for a fucking kernel update. Are you kidding me? That’s totally insane, if you think about it, in terms of actually stewarding these infrastructures.

So you’re saying that AI and surveillance are self-perpetuating: You get the materials to create what we call AI from surveillance, and you use it for more surveillance. But there are forms of AI that ought to be more benevolent than that, right? Like finding tumors in medical scans.

I guess, yeah, although a lot of the claims end up being way overhyped when they’re compared to their utility within clinical settings.

What I’m not saying is that pattern matching across large sets of robust data is not useful. That is totally useful. What I’m talking about is the business model it’s contained in.

OK, say we have radiological detection that actually is robust. But then it gets released into a health care system where it’s not used to treat people, where it’s used by insurance companies to exclude people from coverage—because that’s a business model. Or it’s used by hospital chains to turn patients away. How is this actually going to be used, given the cost of training, given the cost of infrastructure, given the actors who control those things?

AI is constituted by this mass Big Tech surveillance business model. And it’s also entrenching it. The more we trust these companies to become the nervous systems of our governments and institutions, the more power they accrue, the harder it is to create alternatives that actually honor certain missions.

Just seeing your Twitter commentary, it seems like you’re calling AI a bubble. Is it going to self-correct by imploding at some point?

I mean, the dotcom bubble imploded, and we still got the Big Tech surveillance business model. I think this generative AI moment is definitely a bubble. You cannot spend a billion dollars per training run when you need to do multiple training runs and then launch a fucking email-writing engine. Something is wrong there.

But you’re looking at an industry that is not going to go away. So I don’t have a clear prediction on that. I do think you’re going to see a market drawdown. Nvidia’s market cap is going to die for a second.

On 'Simulated Worlds'

OpenAI’s video generation model Sora, in its current iteration, is incredible even though it’ll undoubtedly get better in the coming months. The post they have on their website makes for fascinating reading.

Extending generated videos. Sora is also capable of extending videos, either forward or backward in time.

Long-range coherence and object permanence. A significant challenge for video generation systems has been maintaining temporal consistency when sampling long videos. We find that Sora is often, though not always, able to effectively model both short- and long-range dependencies. For example, our model can persist people, animals and objects even when they are occluded or leave the frame. Likewise, it can generate multiple shots of the same character in a single sample, maintaining their appearance throughout the video.

Interacting with the world. Sora can sometimes simulate actions that affect the state of the world in simple ways. For example, a painter can leave new strokes along a canvas that persist over time, or a man can eat a burger and leave bite marks.

These advancements, alongside how far LLMs and other transformer-based technologies have come in the past few months, have been quite something to behold. While equal parts exciting and terrifying, it’s hard not to think about how and how much they’ll impact industries and society at large. It will likely become harder (and more time-consuming) to sift through what is a genuine advancement and not just another grift (NFTs, anyone?). Art, music, technology, video games, programming, editing, writing, law, disinformation, misinformation, capital markets, cybersecurity, democracy, and medicine will all invariably see some impact. A small part of me thinks that, as amazing as all this is right now, ‘AI’ (quotes intentional) may not be immune to enshittification—not just from the pressure to monetise but also from the unstoppable deluge of low-quality and unimaginative generated content.