Some AI future hot takes

Richard Hamming set aside Friday afternoons for what he called "Great Thoughts Time." After lunch on Friday, he would only think about big questions, like "What will be the role of computers in all of AT&T?" and "How will computers change science?"

In The Art of Doing Science and Engineering, he frames these questions in three parts:

1. What is possible?
2. What is likely to happen?
3. What is desirable to happen?

I think this practice matters more now than when he wrote about it. AI is improving exponentially, but people extrapolate linearly, so our guesses about the future keep coming up short. The only way I know to check my intuition is to write predictions down with a date, then come back and score them.

So I went back through my posts this year and scored the predictions I made.

Past predictions#

What I think happens next#

Here is where I think things are going:

Observing the trend and the capability of agents, here’s the first order predictions:

Near term (3 months)#

• My agents will be able to talk to your agents, eliminating meat proxies.

  In march of intelligence and automation, most of the time blockers are on humans; if you didn’t send me the doc when my agents need it, you will block my agents. The challenges are (1) adoption, and (2) permission design.

• Computer will become optional for a lot of knowledge work

  Orbs did one simple thing for me: they freed me from thinking about local resources. I used to limit myself to a few dev servers and agents because my Mac would slow down. Now the machine is somewhere else and my phone is enough. I can do 99% of coding work in Orbs anytime anywhere as long as i have my phone.

  I expect similar thing to happen to more general knowledge and operational work, i.e. work that is done by website clicking.

  The challenges are (1) credential handling (2) permission design

• Comprehension caps human-agent system output, and this will be addressed by a dynamic interface.

  Walls of text are notoriously hard to digest, especially when they’re dense and contain many decision dependencies. I expect new, dynamic interfaces would be developed to help people digest information of different size.

  On an average day, I type about 9,600 words to my agents, and they send back about 44,000. I can't read that much. I expect the next wave of tools to compete on how little you need to read, and I expect dictation to replace typing as the main way people talk to agents.

  • Comprehension is needed because human needs to be responsible. But often now, comprehension required will move on to a higher abstraction level.

• The compute gap widens before it narrows.

  By September 2027, I expect at least two frontier labs to sell a plan at $500 a month or more, with the best models available only on the top tier for a while. What I'd want is the opposite: this kind of leverage should reach individuals first.

Longer term:#

• The result of A $50–200 AI plan becomes standard for serious work + Bringing your own AI subscription to other apps becomes normal is that the margin for SaaS will go back up again as more users will bring their own AI plan. This is good for the consumers. Again, this started mostly with third party coding agents, but will happen in other domains too, as all domain need agents.

Third order predictions#

• Anything, whether good or bad, that can happen will.

  example in cybersecurity: When execution costs almost nothing, every attack someone can imagine gets tried. By September 2027, I expect a regulator or a large enterprise buyer to require proof of testing against frontier-AI attacks before buying software.

  Models that don’t adhere to moral principles and simply follow instructions are publicly available and affordable.

• The impact of the thing primarily depends on how successful the business model and execution are.

• Education system will massively change.

  • More people will graduate with multiple majors in 4 years

  • More people can finish the traditional four year curriculum in 1 year.

  AI makes learning a lot easier, those who are curious can reach the same level of comprehension at least 4x faster than before.

oh I am committed to writing more “what if” and predictions for the next few months.