Interesting Stuff #3

AI won’t replace jobs - Enabled humans will

https://www.ft.com/content/47f4d549-4560-4830-bf55-47774a9057bc

It's not about whether machines can do this job. But whether humans can do without it.

Modern capitalism moved activitiy from the household to the market - domestic production converted into paid specialisation. Technology is reversing that, leading us toward a self-service economy where we absorb the work ourselves.

The washing maching didn’t automate the laundress’s job, but gave the customers a means to do without her - one can now wash clothes themselves. Online brokerages made trading terminals obsolete for the layman. Similarly, self-checkout handed scanning and bagging to the customer.

With AI, some manual trades - which were once thought defensible - are at risk too. A homeowner who can ask their AI why their boiler is losing pressure will no longer require a heating engineer. And patients can decode lab test results before they visit their doctors.

Work that shifts to the consumer isn’t reflected in the economy that statisticians measure. GDP doesn’t account for customers bagging their own groceries. But these are very real services - the action has simply shifted to the consumer. People who rely on these indicators to judge if AI is delivering benefits will miss this deeper shift.

So, the question is not whether machines can do this job, but whether customers can be enabled to do it themselves. And by doing so, customers are indirectly taking away jobs.

I find this take interesting. There are many existing explanations of why technological advancements will fail to produce mass unemployment. Jevons paradox is one of them - where a more efficient service leads to cheaper prices hence driving up demand. But if we account for self-service, where people do the work themselves, then demand for the service collapses.

Teaching AI

https://www.economist.com/business/2026/06/25/teaching-ai-how-people-work-is-fraught-with-problems

It’s hard to teach AI how people work, when we can’t even express why we do things in certain ways.

Tacit knowledge is the kind of insight that we carry in our heads; it works, but is very hard to write down or put into a manual. Perhaps because it is based on our gut feel and personal habit, which we get through personal experience.

But AI is able to identify patterns in data that humans cannot describe. Give it a marketing copy for a specific brand, and it can uncover habits the organisation is unknowingly doing.

This gets tricky when we try to capture human intuition - which require us to obtain data of humans doing task. Intensive monitoring is sensitive. As we’ve seen in Meta’s Model Capability Initiative, where keystrokes and mouseclicks of employees are recorded to train AI, it can be controversial.

Moreover, the most intrusive monitoring regime will still miss out on the thought process going through people’s heads.

The field of AI exploded because we had better compute, algorithms, and data. For each of these areas, new problems will arise as AI develops and its usage grows. How do we have more energy-efficient chips, algorithms that deal with long-context, and collect data in human-centric ways? By forecasting these future problems, we can work backward to create those products today.

On Dolls…

UBTech’s U1 dolls.

Sexual desire has always been a powerful force in how new technologies are commercialized. The internet and streaming technologies enabled a new wave of pornographic consumption - from magazines to digital videos. Adult producers were early adopters of consumer VR. Sexual deepfakes are a major downstream abuse of image generation models. And LLM-powered boyfriends and girlfriends were one of the first few markets that emerged after the ChatGPT boom in 2022.

So AI-powered humanoid companions were only a matter of time. UBTech says the U1’s capabilities do not extend to the bedroom “for now.” But this is only be a matter of time. The premise of having an always loyal, unconditionally-love-you companion has been the subject of many romance shows we consume (we call these shows “unrealistic”), and it is only a matter of time before it becomes a reality.

I suspect there will be a “black market” offering modifications to the current batch of robots - similar to the “black market” of meta smart glass stickers that block peripheral light, allowing people to record others surreptitiously.

At the face of lower marriage rates and declining birthrates. Always-agreeable machine companions will probably make the real thing feel less worth attempting.

Offload, don’t Surrender.

https://www.economist.com/business/2026/04/30/ai-and-the-danger-of-cognitive-surrender

Cognitive offloading is a deliberate decision to delegate a specific task to technology, as we see in calculator and GPS usage. We don’t manually calculate mental sums, or pull out physical maps anymore - and they improve our mathematical performance and reduce our chances of being lost.

Yet, high use of GPS navigation was associated with worse spatial memory. Calculators removed the need for students to verify for themselves - undergraduates didn’t pick up purposefully inaccurate calculations, and some even accepted obviously absurd ones.

In online search, this is coined the “Google effect,” where people have worse recall when they can find it online.

AI accentuates these trade-offs because it is no longer offloading single tasks, but entire objectives. People stop thinking for themselves. This is cognitive surrender. We expound critical thinking, but that is also the very skill that is at risk when we overuse AI. (I even know of some friends who use AI to draft replies of extremely simple emails/messages…)

So we have to strike a fine balance. To offload, but not entirely surrender.

Token price vs Task price

https://stratechery.com/2026/whos-afraid-of-chinese-models/

“Tokens are not a commodity. A commodity is fungible: a gallon of oil is a gallon of oil. A token from one model, however, is not the same as a token from another model. Different models need different amounts of reasoning tokens to arrive at the right answer. Kimi, for example, reportedly uses significantly more tokens than Sol, rendering its price advantage moot. Moreover, some models need less tokens to execute agentic workflows.

What is fungible is intelligence - the end product that’s contructed from tokens. If both LLMs generated the right answer, then that answer is fungible; the difference in tokens generated to get to that right answer is a contributor to a difference in COGS.”

What I find interesting is that token pricing isn’t enough to estimate the costs involved - lower prices doesn’t necessarily mean its cheaper. The real metric is cost per completed task - how efficiently a model uses its tokens, and how much those tokens cost.