AI 2027

1) useful training data available in the internet 2) number of humans creating more training data ”manually” 3) parameter scaling 4) ”easy” algorithmic inventions 5) available+buildable compute

”Just” needing a few more algorithmic inventions to keep the graphs exponential is a cop out. It is already obvious that just scaling parameters and compute is not enough.

I personally predict that scaling LLMs for solving all physical tasks (eg cleaning robots) or intellectual pursuits (they suck at multiplication) will not work out.

We will get better specialized tools by collecting data from specific, high economic value, constrained tasks, and automating them, but scaling a (multimodal) LLM to solve everything in a single model will not be economically viable. We will get more natural interfaces for many tasks.

This is how I think right now as a ML researcher, will be interesting to see how wrong was I in 2 years.

EDIT: addition about latest algorithmic advances:

- Deepseek style GRPO requires a ladder of scored problems progressively more difficult and appropriate to get useful gradients. For open-ended problems (like most interesting ones are) we have no ladders for, and it doesn’t work. In particular, learning to generate code for leetcode problems with a good number of well made unit tests is what it is good for.

- Test-time inference is just adding an insane amount of more compute after training to brute-force double-check the sanity of answers

Neither will keep the graphs exponential.

Large corporations, governments, institutionalized churches, political parties, and other “corporate” institutions are very much like a hypothetical AGI in many ways: they are immortal, sleepless, distributed, omnipresent, and possess beyond human levels of combined intelligence, wealth, and power. They are mechanical Turk AGIs more or less. Look at how humans cycle in, out, and through them, often without changing them much, because they have an existence and a weird kind of will independent of their members.

A whole lot, perhaps all, of what we need to do to prepare for a hypothetical AGI that may or may not be aligned consists of things we should be doing to restrain and ensure alignment of the mechanical Turk variety. If we can’t do that we have no chance against something faster and smarter.

What we have done over the past 50 years is the opposite: not just unchain them but drop any notion that they should be aligned.

Are we sure the AI alignment discourse isn’t just “occulted” progressive political discourse? Back when they burned witches philosophers would encrypt possibly heretical ideas in the form of impenetrable nonsense, which is where what we call occultism comes from. You don’t get burned for suggesting steps to align corporate power, but a huge effort has been made to marginalize such discourse.

Consider a potential future AGI. Imagine it has a cult of followers around it, which it probably would, and champions that act like present day politicians or CEOs for it, which it probably would. If it did not get humans to do these things for it, it would have analogous functions or parts of itself.

Now consider a corporation or other corporate entity that has all those things but replace the AGI digital brain with a committee or shareholders.

What, really, is the difference? Both can be dangerously unaligned.

Other than perhaps in magnitude? The real digital AGI might be smarter and faster but that’s the only difference I see.

There is a non-zero chance that the ineffable quantum foam will cause a mature hippopotamus to materialize above your bed tonight, and you’ll be crushed. It is incredibly, amazingly, limits-of-math unlikely. Still a non-zero risk.

Better to think of “no risk” as meaning “negligible risk”. But I’m with you that climate change is not a negligible risk; maybe way up in the 20% range IMO. And I wouldn’t be sleeping in my bed tonight if sudden hippos over beds were 20% risks.