Matisyahu's Blog

I’ve stayed at home all today because the weather has been pretty abysmal but in the mean time I’ve sorted things out – I’ve sold my Samsung phone and watch along with the SmartTag2 and I’ll be sending it off either on Friday or Monday next week depending on when the money arrives in my bank account. I’ve copied the emails from Google Workspace and got that up and running, my Apple TV 4K is setup and working and I’ve now got an Apple Watch Ultra 3 where I find the navigation a whole lot easier than the Samsung watch along with a better battery life due to watchOS being better optimised when compared to wearOS (which is based on Android). I’ll be writing a longer blog about my experience with the Samsung and Google ecosystem but I’ll leave that to another day.

I’ve been watching the Build 2026 keynote and it is interesting to see the Nvidia RTX Spark announcement but what has made it interesting is the unreported data centre cancellations and the selling off of land that was purchased to future data centre expansion. The combination of a capable platform that can run LLMs locally, talk about OpenClaw and combine that with the data centre reversal makes me wonder whether the AI bubble, rather than bursting, will be like a a hot air balloon in a cartoon that gets a hole in it and the flies around letting the air out. The dream that AI could be like SaaS aka a ‘licence to print money’ has turned out not to be reality given that the past booms were always about “if we build it up, we’ll lose money in the sort term but once we reach economies of scale we’ll start making a profit”. The problem is that the cost of AI, both inference as well as training, is too expensive and the inaccuracy too high that even if they were to move to token based billing the problem is that the amount of wasted tokens because the model gets it wrong undermines the argument that AI is cheaper than hiring actual humans to do the job.

The interesting part are the new models that were announced that are specialised and scaled down so that they can be run locally. The impression I get, there will be a place for cloud based AI but for many end users for what they want to do then the local model will be good enough – the ability for locally run models to pull down information from the internet and run the inference on device rather than in the cloud. I think that as the hardware becomes cheaper to run models on device and the scaled down versions are good enough to to do what most people want, the desire for end users to use something like ChatGPT (let along pay for it) will disappear – I wouldn’t be surprised if in the future, assuming Anthropic and OpenAI survive, they will become enterprise and developer focused businesses that’ll be locked behind a pay wall with no free chatbots like we see today or even chatbots at all.

At the end of the day, at some point the backers of your organisation will start demanding a return on investment or at least some sort of pathway to profitability and that is damn well next to impossible if you’re giving away ChatGPT for free to 700+ million users where you’re losing money on every single inference (this is the reason why boasting amount 700+ users but only 3% paying for the service (and even then with the subscription based users there is still money being lost on every transaction) is a just a ridiculous thing to boast about. It is like saying you have the best restaurant in town because of the number of customers but ignoring the fact that people are coming to your restaurant because you’re giving food away for free.

On a side note, I’m enjoying the UniFi Dream Router 7 – got it all setup, rock solid and going reliably however I decided to move back to my ISPs DNS server and I have found it working slightly faster than using a third party one. I originally had changed it over to Cloudflare however the ISPs DNS has been rock solid so I think I’ll be sticking with that. What I found interesting is there was a message about there being IPv6 being available on my WAN connection however when I enabled it (DHCPv6 with the Prefix Delegation size set to auto) it didn’t get an IPv6 address. I checked out on Geekzone to find out what had happened and it appears that maybe Spark is testing IPv6 support to a limited number of customers or possibly just employees so I guess we’ll find out whether anything comes of it. Oh, and on a side note, I decided to stick with Skinny since the amount I pay each month for my mobile is the cheapest on offer and the internet connection is a lot cheaper as well (I’m on a 900Mbps down and 500Mbps up).