top of page
Search

The periodic table for AI models just dropped. 

  • Writer: Anmol Shantha Ram
    Anmol Shantha Ram
  • May 1
  • 2 min read

If the chemical table analogy holds, AI futures become predictable.

This marks a big leap for forecasting machine learning models.


Last week researchers from MIT, Google and Microsoft revealed I-Con.

Their paper calls it Information Contrastive Learning (I-Con) or the "periodic table for AI models". Think Mendeleev, but for models. 


I-Con arranges more than twenty diverse algorithms in one grid.

It covers algorithms like K Means, PCA and modern contrastive techniques.

The grid even captures the maths behind today's LLMs. 


And just like the chemical table once held blank squares that scientists later filled with elements.

I-Con shows similar empty spaces today.

And each space predicts an algorithm that should exist but has not yet been discovered.


And that could reset your machine-learning roadmaps. 

Because:

- Discovery becomes deliberate

 The research team filled one gap straight away and lifted ImageNet-1K accuracy by eight percentage points

      White-space search replaces months of trial and error


- A shared map for a crowded field

     Roughly a quarter-million machine-learning papers have landed since 2020

     A two-axis grid helps teams hire faster and spot duplicate work sooner


- Shorter and cheaper build cycles

     Need bias resistance? Start in the debiasing column and blend proven tricks across rows

     Scaling to a billion items? Compare rows that keep global structure with those that focus on local detail, then borrow the best optimiser nearby

     Machine learning starts to look like engineered design rather than craft


- A fairer playing field

 Start-ups gain a map that narrows the gap with big-tech model vaults

 Investors can back teams that target high-value empty squares instead of another copy-and-paste idea


I-Con is an innovation matrix. 

Mark where your current models sit. 

Circle the empty squares that align with your data strengths. 

Form small exploration squads. 

Then move.


Credit to Google AI Perception, MIT CSAIL and the Microsoft team for creating this framework. 

ree

Links:

 
 
 

Recent Posts

See All
My playbook for building MVPs

MVPs come up a lot in my conversations. Here's my playbook for planning and building MVPs.  - Launch (likely something bad) quickly  -...

 
 
 

Comments


© 2024 by Anmol Shantha Ram

bottom of page