Over the last few months companies have gone from “you must use AI or you will get fired” to “use less AI or you will be fired”.
As you know, it all comes down to cost. Companies thought that they could replace some salaried developers by giving AI to the remaining staff. Now it seems a lot of companies are back peddling on that stance now that the true cost of AI is appearing.
AI has been heavily subsidised by investor money to try to get people hooked, however with AI companies not turning a profit they have to increase the price or restrict usage to keep going.
I have seen people reporting on Reddit that their companies have taken a U-turn and started restricting usage (obviously it is Reddit so take with a pinch of salt).
- Company took away access to claude
- Corporate first said that they highly encourage use of AI, now with billing in place, they want us to be mindful
- Has your company started limiting AI usage?
- Sorry to say, but I’m happy to see AI fail
I keep seeing tutorials of how to set up your AI agents to do all sorts of tasks. These “skills” the AI agents possess is just a markdown file containing instructions of what commands to run etc.
Every time an agent wants to use one of these skills it has to read in the skill file which uses up tokens and then will execute on it, mostly in an indeterministic way.
The LLMs are a black box and the big providers are constantly tweaking how they work to try and save money or squash bad behaviour. A skill file that works now might not work the same in the future.
I understand for AI coding agents you might want it to use these tools without asking but for fairly deterministic processes I feel you would be much better off with a script that doesn't use any tokens at all.
In some cases, the skills contain a bunch of code snippets telling the AI how to run things, such as this slack-gif-creator example.
It was easy when AI was practically free just to run everything through an LLM and have it control the flow. With the costs rising I think less people will be inclined to just throw a markdown file at it and see what happens.
So, what's the alternative?
The obvious solution is to convert the process into a script but in some cases there will be pieces of the operation that still need AI.
Maybe the data comes in an unstructured format, and you require AI to format the data. You could be getting data from multiple sources, and you want AI to summarise it for you.
In these cases, having an LLM in the flow is useful, but if you can minimise where it is used it will cost a lot less. For a lot of small tasks, local AI models will likely be sufficient.
Let's say you wanted to create a custom daily briefing.
Instead of writing a skills file to pull down the weather, your calendar, and the latest news, you can turn it into a script.
- Call the weather API to get today's weather and remove any parts you don't care about.
- Call the calendar API and filter to just today's events.
- Get a pre-aggregated RSS feed and look at today's stories.
- Send the parts that need summarising to a local or cheap cloud AI model to summarise.
In this case it is likely only the weather that could do with a more personal touch. Everything else could be just put in a list.
Of course working out what can be scripted and what needs to go to an LLM requires some thought and problem solving, you know, actual engineering skills.
The AI companies want you to shove everything into a markdown file because they make more money, but that isn't how software developers should be working.
While AI is still available heavily subsidised and through subscription, use it to build tools that will still work when AI isn't financially viable any more.
❤️ Picks of the Week #
I am going to be limiting the number of links I put in this newsletter to 10-15 a week. I don't want to overwhelm everyone with too much information.
🤖 AI #
Steve Wozniak cheered after telling students they have AI – actual intelligence - At least Steve seems more relatable than the other talks that have happened recently. Imagine promoting AI to graduating students when it is actively making the job market worse.
AI has a multiplying effect on existing technical skills - I think this is why developers see drastically different results from AI. If you are already a great programmer and are guiding the AI appropriately it can help but for if you are not, it can just make you worse.
I'm Tired of Talking to AI - It is so disrespectful sending AI responses to people. It is crazy that we live in a world where thinking for yourself is a superpower.
Tech CEOs are apparently suffering from AI psychosis - A CEO one-shots something in Claude and thinks that it is close enough to what developers produce that they don't need developers. Then wonder why the org isn't seeing the same productivity gains.
Various LLM Smells - I suspect we are going to see a lot of these in website designs going forward.
🛠️ Tools #
An Introduction to Meshtastic - With nations around the world introducing censorship and age verification I have been looking into things like Meshtastic and Reticulum. At the moment it seems only useful for short range messaging (a few miles) but this might improve in the future.
I'm Getting into Mesh Networks (Meshtastic, MeshCore, and Reticulum) - Some good information about the different networks. Reticulum which I have talked about before looks particular good too.
💻 Programming #
Serving a website on a Raspberry Pi Zero running in RAM - This is cool, but they aren't entirely relying on the Raspberry Pi given the TLS termination is done by a VPS.
Nobody cracks open a programming book anymore - I learnt to program from books and I think having a separate medium actually helps you retain information more. With a book you have to look for the information and decide which bits are important to you. A chatbot just gives you the information which doesn't help you learn.
🔒 Security #
A web page that shows you everything the browser told it without asking - I showed a similar page last week. It is scary how much information is leaked by your browser.
🏢 Tech Industry #
Cloudflare to cut about 20% of its workforce - More cuts. Cloudflare still has revenue of $665 million but they don't want to pay for workers.
🎮 Gaming #
SimCity 3k in 4k - I loved playing SimCity 3000 growing up. They also have SimCity 3000 unlimited on Steam for £5.
🧠 Miscellaneous #
Taking a walk may lead to more creativity than sitting, study finds (2014) - I have found if I am stuck on a programming problem that going for a walk helped. It is certainly healthier than asking an LLM.
Can we have the day off? - This is the issue I have with capitalism and corporate culture in general. If I complete my work faster I should get more time to myself, instead I just get more work. AI is just going to accelerate burnout in workers.
Wikipedia doesn't need my cash - Not long ago Wikipedia had big banner on the page asking for donations yet they are sitting on $290 million in the bank.
💬 Quote of the Week #
This is because the effects of a typical one or two-week vacation wear off almost immediately and that the whole point of most vacations is to merely take a break from work instead of resting for its own sake.
From the article “The Top 10 Career Myths We Should Stop Believing” by Paul Millerd.
Top comments (0)