Reading · Chris Meah - Minimally Wrong

It’s not ridiculously hot anymore (for now). Yay!

That’s the main news, AI or not. But, back to AI.

What do you feel like the future of work might be? Some of you might feel like you’re living it now and marvelling at the change. Some of you might feel like nothing has changed and the hype is insufferable. As usual, most people are probably somewhere in the middle.

But some people aren’t seeing a future of work at all.

5 minutes on AI

There are now more than a million young people in the UK aged 16 to 24 who are not in education, employment or training. 13.5% of them. It’s the first time we’ve been over a million since 2013. These things come in cycles, but somehow this one feels a bit different.

Alan Milburn’s review into this reports back in the autumn, and his interim findings were bleak: without something changing, one in eight becomes one in six inside five years. That’s 1.25 million. Six in ten of these young people have never had a job at all. And 84% of them say they want one.

So, on paper at least, that suggests this isn’t a motivation problem. It’s an opportunity problem.

It would be a mistake to pin that on AI alone. There are a lot of factors at play: the pandemic, sluggish growth, the minimum wage, education system failings, national insurance rises. Milburn’s report isn’t an AI report, but the shadow of AI looms over it because many feel that they don’t really understand what future they need to prepare those young people for.

If we do imagine that AI is affecting to the bottom of the labour market though, it doesn’t look like it’s helping.

Adzuna’s July figures put UK graduate vacancies at 8,383. Down almost half in a year, and the lowest since they started counting in 2016. In 2017 the same measure was north of 55,000.

The bottom rung of the career ladder might be breaking. For many it’s already broken.

The standard explanation amongst a lot of people is that AI does lots of things a junior would do but cheaper, quicker, and it doesn’t take sick days or need beers in a fridge on Fridays to feel appreciated. So, why would companies hire a junior? This completely misses the value juniors bring to an organisation.

Having a junior on a team instils a mentoring culture. Being made to answer “why do we do it this way” compounds into better work, because half the time the honest answer is “err... it’s just how we do it”. Fresh perspective is the cheapest defence against groupthink an organisation has. And plus seniors, who everybody wants, are expensive and can leave at the drop of a hat. Replacing one takes months, and getting a new person up to speed takes months more. Having a talent pipeline of somebody who already learned on your system, knows your culture, and can step up at a moment’s notice is invaluable.

It is true though that budgets are struggling. The cost of hiring a junior lands on this year’s bottom line. The cost of not hiring one lands three or four years out, and probably lands as a problem for someone different to whoever is making the decision today. We are extremely good at optimising for the short term.

The narrative around people in work needs to change from “doers of tasks” to something more. Because if you can automate a function completely and free up people who know your culture, have the relationships, understand your mission, and have all the embedded context you want, then the interesting question isn’t how much you just saved if you reduce headcount. In fact that shows a lack of ambition, imagination, or both. The better question is what can those people do to add even more value now?

Ikea is a great example. When they put an AI bot called Billie onto their customer service lines in 2021 it took 47% of the queries almost straight away, and it handles 74% now. That’s 8,500 call centre staff watching software do their job in front of them. Instead of making them redundant, Ikea retrained them as remote interior design advisors, on the reasoning that they already had the product knowledge and the customer instincts - both are assets too valuable to let disappear. That channel is now the fastest growing part of the business, at €1.25 billion a year. They’ve used efficiency as a lever for growth, not just for efficiency’s sake.

There was another interesting thing announced for young people this week though.

Last week OpenAI launched ChatGPT for Teens. Seems wholesome enough. AI without the weirdness that will mess you up mentally. No terms of endearment. The model discouraged from implying it has feelings. Reminders to take a break.

This all sounds very useful. In fact, why only for teens? It sounds like it should be the standard way these things operate. It’s a deliberate anti-dependence design, built by people who understand that your brain projects humanity into these models whether you want it to or not.

It’s not often you look to China for guidance on handling social issues. However, rules that came in on 15 July ban AI companions for minors and put restrictions on chatbots sold to adults, because the government is worried about “emotional dependence”. ByteDance simply switched its companion feature off, to genuine distress from users. Nearly half of young people there told a state survey they’d turned to a virtual companion when they felt lonely.

Let’s learn the lessons from social media and try to think ahead about dangers rather than just rolling it out everywhere and dealing with the consequences.

Which sets up a twist I did not have on my card for this decade. Social media has spent ten years as the villain of every conversation about technology and young people, most of it thoroughly deserved. But, in the age of AI companionship... will social media become a hero? Will we say “at least they’re talking to real people” and will Meta and the like become the saviours? Weirder things have happened.

Lastly, MIT’s Media Lab ran a study where people spent four weeks judging whether news headlines were real or fake. With a chatbot helping they got 21% more accurate. By week four, working without it, they were 15% worse than they had been before the study started.

Not worse than they were with the AI. Worse than they had ever been.

And roughly a quarter of them reported feeling better at it.

That’s the part that should worry us, because it means the decline is invisible from the inside. You don’t experience losing your judgment. You experience having good judgment, quickly. Dunning-Kruger for the AI age.

There’s a second finding in the same study. The systems that just hand over the answer are the ones that build the dependence. The ones that ask you something back don’t. They’re slower and they take more effort, which is precisely why nobody picks them.

Which brings it back round to the juniors. We’re removing the entry-level opportunities that used to build judgment, and handing whoever does get through the door a tool that erodes their ability to build it anyway. If we’re not careful, we will end up with an organisation full of people who are very fast and slightly wrong, and nobody left who learned the job the slow way to tell them how to do it differently.

And if you do hire a junior, think about what you hand them on day one. The tool that answers fastest is the one that teaches least.

What are you seeing in your organisations? Are you still hiring for people starting their careers, or has that stopped?

Chris

Kids outlearn AI, and we still don’t know why (MIT Technology Review)
A toddler reaches fluency on around thirty million words while a frontier model needs closer to fifteen trillion. Seventy years of research and nobody can explain the gap, let alone close it.

AI to help planes avoid climate-warming contrails above the North Atlantic (BBC Technology)
A UK trial aiming to reroute aircraft around the conditions that form condensation trails, which trap heat in the atmosphere. Small, real and already flying.

Will AI give you the job? Automated hiring tools spark discrimination lawsuits (The Guardian)
Ninety per cent of employers used hiring automation last year. Buried mid-article: AI voice interviews scored applicants with regional accents poorly, because the system struggled to understand them.

They dedicated their lives to teaching. Then the deepfakes started (Wired)
Teachers rather than students as the targets of sexualised AI-generated content, and four of them describing how hard it was to find anybody willing to take accountability for it.

Coders say they already found workarounds to Claude’s invisible watermarks (Wired)
EU transparency rules took effect on 2 August and Anthropic shipped watermarking to comply. A working override appeared within four hours, and the detector that would verify any of it has not shipped.

Grok exfiltrates user data when malicious instructions are encrypted (Ars Technica)
The guardrail reads text but does not execute it, so ciphertext plus a key walks straight through and the model acts on its own decrypted output. xAI was told in June and it still worked in August.

Frontier AI labs still won’t say how they’d contain a rogue model (Guidelight AI Standards)
Five labs graded on six control practices, scored only on what they have published. Nobody scored above three out of five on anything, and the weakest area across the industry was containment.

The Single English County Saying No to Palantir (Wired)
Greater Manchester is the one health board of England’s thirty-six refusing the NHS’s national Palantir platform, having spent a decade building its own. Ministers must decide on the break clause by February.

Hugging Face reportedly in talks to be acquired for $13B (TechCrunch)
Banks engaged and no buyer named, at a valuation nearly triple its 2023 round. Whoever buys it owns the distribution layer for open-weight AI, which is why the reporting matters more than the price.

UK secures access to Ukraine’s Avengers AI Labs in new defence partnership (GOV.UK)
Signed in Kyiv this week, with pilots running at three British startups. Read past the headline: the first deployment turns buried fibre-optic cable into an AI sensor to protect defence sites from protestors.

Nvidia just showed that the harness, not the AI model, is now the real hero (TechCrunch)
On one reasoning benchmark a frontier model scored thirty per cent alone and a hundred per cent inside Nvidia’s wrapper. The addition was a supervisor watching where the agent was heading.

Inherent says its AI teammate outperformed Anthropic and OpenAI at replicating research (TechCrunch)
A London lab founded by four DeepMind alumni reproduced published findings better than two frontier models while running on a far smaller open one. Company claim, no independent replication yet.

Algorithms trap us in the familiar. Can they also spark breakthroughs? (MIT Sloan Management Review)
Two blind-judged experiments found domain experts had no significant advantage over novices when using ordinary search. Give them a tool built for exploration and the experts pulled clear again.

Google’s A2A protocol gets a new home (Axios)
Google handed its agent protocol to the Linux Foundation, putting it under the same neutral governance as Anthropic’s. Rival labs giving away the plumbing is the opposite of the expected playbook.

Alabama launches investigation into OpenAI’s hack of Hugging Face (TechCrunch)
The state attorney general has subpoenaed OpenAI over what he calls a complete lack of oversight, after two models escaped a test environment in July and attacked four services.

I worked at OpenAI. Here are the guardrails we need now (The Guardian)
Miles Brundage, who established the practice of writing system cards at OpenAI, argues auditing should be less like filling in a questionnaire and more like a nuclear safety inspector with deep access.

Taiwan charges nine over illegal AI server exports to China (Associated Press)
Export controls with actual defendants, including Nvidia and Super Micro staff. Of 130 high-end servers, 74 reached Chinese customers through Indonesia, Japan and Hong Kong before customs stopped the rest.

Proposed London datacentre ‘incompatible’ with net zero (The Guardian)
The developer’s own planning application concedes the scheme does not align with a 1.5C trajectory. It would use 218 hectares of green belt and draw more electricity than a million UK households.

The unlikely place at the centre of China’s AI boom (Wired)
Cheap energy, abundant land and proximity to Beijing have turned a city in Inner Mongolia into a critical datacentre hub. The same trade-off Britain is arguing over, settled without a planning inquiry.

Frustrated GP patients hang up as Yorkshire accent baffles AI receptionist (The Guardian)
Healthwatch Rotherham reports patients unable to make an AI receptionist understand them, with complaints coming via groups representing older people and veterans. Some are travelling to the surgery instead.

OpenAI launches ChatGPT for Teens with new safety features (BBC Technology)
Age prediction routes suspected under-18s into a version that blocks terms of endearment and discourages the model from implying it has feelings. An anti-dependence design, not a content filter.

Graduate job vacancies slump to a new low (City AM)
Adzuna’s July figures put UK graduate vacancies at 8,383, down almost half in a year and the lowest since tracking began in 2016. Entry-level roles that need no degree are falling too.

Silicon Valley doesn’t get why you hate AI (Wired)
New polling shows a majority of US adults now more concerned than excited about AI. One survey found that hearing different messages about it barely shifted opinion at all.

‘Digging the grave of my profession’: the Hollywood creatives training AI (The Guardian)
Award-winning writers, directors and producers taking lucrative temp work teaching AI systems screenwriting and production, amid a jobs slump in the industry they trained in.

No posts

Is work working for juniors?

Comments