The big stick of AI efficiency, the AI capability/utility paradox and how to vibecode a Chrome extension
AI video of the week: influencers get biblical
Thought of the week: the big stick of AI efficiency
Another week, another gloomy forecast for white-collar workers. This time, researchers from Anthropic, creators of the Claude LLM, warned we might become little more than ‘meat robots’.
Speaking on Silicon Valley’s favourite AI podcast show, they outlined:
"A really scary future… in which AIs can do everything except physical robotic tasks. Humans might end up controlled by AI, equipped with AirPods and smart glasses, guided step-by-step through their daily tasks."
While undoubtedly extreme, this vision comes off the back of recent comments from Anthropic's CEO, Dario Amodei, who predicted the decimation of entry-level office workers in the next few years as they fail to compete with AI agents.
While I’ve been warning of the job AI-pocalypse for months now, for me the bigger and more immediate worry is the growing employer perception that workers should now be able to do far more in less time due to AI. Consider the results from a landmark UK government trial involving over 20,000 civil servants. The trial claimed generative AI tools saved participants nearly two working weeks per year.
As an academic, I'm always sceptical of studies that double as advertorial. Digging deeper, this ‘trial’ was essentially Microsoft giving free access to its Copilot AI for three months. The researchers themselves admitted that time savings were self-reported estimates, ranging from "less than five minutes a day" to "over an hour". In other words, we have no reliable measurement of actual efficiency gains – just subjective guesses. Also, there was nothing in the data on the time wasted dealing with Copilot’s AI-generated inaccuracies and hallucinations.
Also, if you interrogate the data more closely, the highest AI adoption rate was for Teams, Microsoft’s virtual meeting app, because getting an AI to auto-transcribe meeting minutes is genuinely useful. More telling, though, in the adoption rate graph above is that, after the initial enthusiasm of playing with a shiny new toy, usage of AI declined across all apps. This research, therefore, feels more like a thinly-veiled justification for the expensive roll-out of half-baked AI, funded by taxpayers, rather than credible evidence of genuine efficiency improvements.
And herein lies the biggest issue with the roll-out of AI. While AI isn't necessarily causing immediate job losses, it’s intensifying pressure on existing employees to beat the bot and, critically, giving employers an excuse to cut jobs.
While the mass ‘AI-pocalypse’ might not be imminent, we are now entering an era of relentless pressure on workers to deliver more, with or without AI assistance.
Why AI can be a colossal waste of your time
One of my favourite books of recent years was Four Thousand Weeks by Oliver Burkeman. If you’re unfamiliar with it, the core message is that we should ditch our relentless attempts at time management and productivity hacks, and instead embrace the journey and process of life itself. Especially profound for someone who starts his day with multiple task lists on different devices is Burkeman’s advice on to-do lists. He argues that each completed task only creates space for new tasks, perpetuating a never-ending cycle of busyness without meaningful progress.
In short, we should prioritise doing the right things rather than merely getting things done.
The problem with AI, though, is that it suddenly gives people superpowers that a few months ago they could only have dreamed of. For example, in one day, I:
vibecoded a functional data visualisation tool (chart maker) with bolt.new
created a soundtrack on Udio
generated three videos “in the style of famous directors” to test out Google’s new Veo 3 image generator
researched our family summer holiday with Manus, the AI agent system
had a conversation with my Danish AI tutor in ChatGPT
started vibecoding a task management app before quickly realising that Google Tasks does a pretty good job of this already.
To cap it all, I ironically vibecoded a time-tracking Chrome extension to allow me to see what I spend my time on!
At first glance, it would appear that on that day I was extremely productive: all of the things I created would, without AI, have taken me months. However, none of the tasks progressed my business, nor on further reflection did the AI produce any truly useful content. Moreover, by the end of the day, I realised that my carefully prioritised to-do list remained annoyingly intact because I was seduced by the ability to produce ‘stuff’.
And that’s the problem. AI technology continually tempts us with potential. The tech whispers seductively, much like Frodo’s precious ring in The Lord of the Rings: "Fola, I'm here – the biggest brain in the world, ready to make your dreams come true." Just try out any of the latest ChatGPT models and you’ll find that they’re increasingly being programmed to maximise ‘engagement’. Ask it a question and not only will it answer but will also suggest a follow-on activity to keep you on the platform.
Now, given that I’m an AI researcher, I can be forgiven for ‘wasting’ time playing with AI, but I suspect that many of us are doing lots of things with AI because we can, rather than because we should. This highlights a critical tension, which I call the ‘capability/utility paradox’: just because technology – especially AI – enables us to do something, it doesn't automatically mean we should.
The fundamental issue here is that as we all become increasingly reliant on AI, we forget that it’s clever but not smart. It can suggest endless tasks and facilitate their execution, but it’s incapable of telling us what we should be doing with our precious time.
It’s critical that everyone understands AI and its impact and, where possible, experiments with how to apply it. However, ultimately, it's up to us – not AI – to determine how we squeeze the most out of our precious 4,000 weeks.
How to vibecode a Chrome extension
Google Chrome extensions sit within your browser and are incredibly easy to code because there is copious documentation on how to build them and the code structure is basic.
To code an extension:
Go to Google AI Studio. You can get one million free tokens, which is more than enough to build a simple extension.
In the chatbox, describe as clearly as possible what you want to build. I usually add a screenshot so that the bot gets the styling right.
It will then give you a bunch of code that you need to paste into a code editor (my preferred one is the default notepad in Windows – a handy shortcut to access this is Windows key + R and then type “notepad”).
Create a folder for these files.
Go to “Manage extensions” in Chrome or type “chrome://extensions/” in the URL box.
Make sure you have developer mode turned on and then click “Load unpacked” to load the folder of files you vibecoded.
If all goes well, your extension should now have been loaded and you can view it by pinning it to the extension panel.
What we’re reading this week
OpenAI launched o3-pro, an AI model that the company claims is its most capable yet. O3-pro is a version of OpenAI’s o3, a reasoning model that the start-up launched earlier this year. As opposed to conventional AI models, reasoning models work through problems step by step, enabling them to perform more reliably in domains like physics, maths and coding.
Meta’s AI App Lets You ‘Discover’ People’s Bizarrely Personal Chats. Launched in April, the Meta AI platform offers a “discover” feed that includes user queries containing medical, legal and other seemingly sensitive information.
Researchers from the University of Geneva and the University of Bern found that ChatGPT and other AI systems beat humans on emotional intelligence tests, indicating AI may be better at reading emotions and responding than people.
Tools we’re playing with
Perplexity Labs. Testing the claim that it can accomplish in 10 minutes what would previously have taken days of work, tedious research and coordination of many different skills.
n8n. We’re experimenting with building a series of agents and n8n is a workflow automation app that lets you do just that.
Flow – Veo 3’s video generation tool. Impressive on the surface but still tricky to stitch together something coherent.
That's all for now. Subscribe for the latest innovations and developments with AI.
So you don't miss a thing, follow our Instagram and X pages for more creative content and insights into our work and what we do.