I really am the last person who should be initiating a post on AI, BUT, recently I built a website using Claude Ghostty and I cannot tell you how easy it was and is to change and update and it costs just my time. In previous business I have paid thousands to web developers to build inferior sites and it was no fun. Now I really enjoy it and it helps develop my business model. There are of course infinite other things Claude can do and we should all know something about AI, if only to work out who is faking etc… So, here it is… go for it.
Cheers Paul. I have almost zero knowledge... if we could share a list of what tool is capable of doing what it would be a great start hopefully. Midjourney - image creation through descriptive prompts
https://ghostty.org/ https://github.com/ https://vercel.com/ You need these for a website or app. Instructions go into Ghostty, Github stores photos, videos etc… Vercel publishes to the internet. Using Ghostty is very conversational, like working with a colleague in the same room. Write very clear instructions, it checks with you if it has understood correctly, then “push to the internet” and see if it is what you want. I am sure it will do infinite other tasks that I have yet to discover.
35 minutes of possible enlightenment, albeit short-lived, I cannot vouch for the veracity of this but it is certainly compelling enough to seriously consider. ... ai agi asi (gameover) scary as fuck, like us mere mortals ever had a say
It’s a race between our leaders and something else to end the world, it always has been and it always will be. Anyway, back to useful things you can do with Claude
Our household is very deep in the machine learning world. With my previous roles in healthcare machine learning, specifically MRI/CT machine learning early cancer detection visual to text models (takes CT scan images, pin points cancer nodules at a tiny scale and very accurate). I recently built a financial model using broker supplied data using only local hardware. No external models or such. It's been a project I've been working on now for ~3 years, with hands on building for 1.5 years and recently started the closing phases. But having the hardware to run local models make the subscription costs for Claude and the likes seem very inflated. If you have the hardware. As you can essentially 'download' and run models similar to Claude/chatGPT if you want a workhorse on your computer for coding, designing, task creation/management etc. Or even your own little LLM like chatgpt to just waffle bollocks to, even with its own web MCP (more on that below). As for running local models, the open source space on this is getting very close, to the point the average gaming GPU can run models around the same 'intelligence' as those provided by services around a year ago. With the top end open models competing, if not beating some of the main stream services, again, if you have the hardware at home. Regarding how the LLM models are actually built, the layering around their response structure and hallucination management, once you understand the embedding space structure and the MoE (ministry of experts) approach for fast response, you realise that what seems intelligent, is simply a very good predictor at what the next word/sentence should be given the trillions of lines of information it has already been shown to 'align' it's embeddings (memory) to. Hence conversations are only ever going to be historically fed on information, any 'today' information is simply called through something called a MCP, essentially a little 'tool' the LLM can use to ask questions to and get responses back -- 'Web MCP, what is the weather today?' "30C" 'repond 30c to user'. I'm fascinated by the subject as you can see as it's allowed me to step away from a career under company to career by my own pace. And I'm loving it These types of videos are interesting, albeit, a bit tin foily. Yes, models exist that are very capable, but that is because they have been trained on the specific task needed. For example, a farming model to predict ideal planting windows for crop types exist and were built on satelite data collected from the 70's. It's tensor mathmatics and has been around for decades. What people are seeing is the tensor mathmatics now capable of running on modern hardware to a scale of not only conversation speed response, but memory capacity has expanded which allows for context windows to increase for the model to re-read each response without having to trim that context off to save computational memory (I'll not dive into context quantisation caching...). So again, the models are indeed 'evolving', but nothing like people believe. True artificial intelligence is as far away in a dream as it was 10 years ago. Creative and trainable knowledge tools are available today and have been for decades, retail models however, that's the recent boom. And unfortunately, as in most business, the retail mind is shaded from a lot of function and ability hidden for enterprise hands. For the interested, I have a 5090, 14900k and 96gb ram. I run the Qwen 27B Dense model when I want to talk to a local LLM for tasks that I'd rather not pay a service (Claude/GPT) for and to execute work I have written up, such as project tasks. My financial model is black box and purely built from the ground up and having the hardware to develop it on really unlocked the subject of home development.
Just to add another perspective to the subject, ai and job loss. It's almost attached to the hip of the ai subject and I myself, ironically, fell subject to layoffs. The stance from my self is simple, the majority of leaders peddling AI into their businesses simply do not understand the tool. And when you ask these decision makers to show material value in end results directly related to shelling out incredible amounts of cash for tokens, the pitch for AI quickly falls flat on its face. Put simply, and as I challenged my peers with - if the models are that impressive, why are Anthropic and it's cohorts charging a subscription fee rather than share value in their enterprise accounts. Because they know long term, there is no value added to the table other than a fast agentic work doing menial tasks under strict instructions, of which were written by a human. Anything requiring judgment or instinct instantly collapse any models into hallucinations and loops of thinking in one way or another. So by firing the skill and expecting the horse to do the farming, the land owner just shot them selves in the foot. As for site design, I've got family building similar things using Claude, it's very capable when you have the goal in mind. I used Claude code for the majority of the grunt work on my work. I truly could waffle on about this subject for days.
I’ve never used ai until five months ago (we get Claude and Gemini logins at work) but can safely say if you prompt it with your knowledge it’s fucking amazing how good it can make you look without spending 5 days pulling a power point presentation together. I’ve been leading a discovery project for my current firms future roadmap who are an industry leading translation firm (amongst other products). Ai/llm’s are obviously hugely impacting the industry. It’s really interesting, but the firm is suffering from change fatigue in a big way. It takes its toll on everyone eventually. Thats the real fight imo.
Agreed, amazing when the track is structured and purposeful. As for the fatigue, it's the rate of work increasing, every answer is now available instantly and if you're unable to provide one, you become the question. It puts a lot of weight on keeping up with the pace without feeling it because most retail models makes every step feel as though you just broke the mold. I tend to crunch through my AI workload for a few days and then spend a week or so working away from it, else I'd lose my sanity, or what's left of it.
They’ve just got rid of three people with 60 years experience at the firm between them with no handover cos a Gemini presentation told them they were “the bottle necks” etc…. Scary times for everyone as you say.
