The AI burnout

Hi, and happy Tuesday.

Last week, Elon Musk got on a SpaceX earnings call and projected SpaceX would become a $1 trillion revenue company by 2030, pulled forward from 2031. (There are currently no $1 trillion revenue companies). Most of SpaceX’s revenue would come from building out a country’s worth of data centers.

This came within a tumultuous few weeks for the AI space.

• Kimi 3 - a new model out of China - was a 'second DeepSeek moment.' It showed near-frontier AI performance could still be run on (admittedly expensive) desktop hardware. Kinda awkward if you’re building a small country’s worth of data centers, right?

• Stock markets wobbled as memory stocks crashed. For a moment it looked like the bubble might be bursting.

• The AI economy increasingly depends on a remarkably small number of companies. Anthropic is becoming SpaceX’s biggest customer, for example, and SpaceX is growing to be NVidia’s biggest customer.

Back on level ground, our conversations within Prescouter, as well as with partners and clients, has increasingly led us to believe the key problem with AI is that it's unreliable. 

There are, of course, workarounds and techniques for increasing its reliability, but a light switch that works only 70% of the time leads to all kinds of complexity.

• In which 30% of cases does it not work?

• In those cases, do you need to press the switch a particular way to make it work?

• Is everyone having this problem with the light switch? 

For most people, a task that should have taken one second - switching off the light on their way out - suddenly feels like a half-day diversion of troubleshooting.

The tax shows up in many ways:

• Fixing a confident wrong answer often takes more mental energy than just doing the task yourself from scratch.

• Having to repeat basic instructions or argue with a system that acknowledges its flaw but fails to fix it - “conversational paralysis”.

• Reviewing AI generated output from colleagues, to weed out if it might contain anything useful. 

If you’re experiencing this yourself, three techniques that we’ve come to rely on are:

• Treating the output like a draft from a confused intern rather than an authoritative final product. Yes, this includes telling your colleagues that’s what it looks like!

• Verifying the output - either yourself, by asking for things that are easy for you to check (e.g. links to sources) or by having a second AI model the work.

• Codifying your feedback / corrections into an instructions file that you play back into the AI for the next time you perform that type of task, so it keeps all those in mind.

But, again, for most people - this is still too much work.

We have to remind ourselves, though, that only a few years ago, we couldn’t even get LLMs to do math.

The technology is improving at a rapid pace.

So long as it continues to do so, the projections and the data center buildouts may still prove to be prescient - even with open source model economy.

Best,

Dino