This paper didn’t go viral but it should have. A tiny AI model...
Shruti@heyshrutimishra
112 views
Aug 02, 2025
~2 min read
Advertisement
1
This paper didn’t go viral but it should have.
A tiny AI model called HRM just beat Claude 3.5 and Gemini.
It doesn’t even use tokens.
They said it was just a research preview.
But it might be the first real shot at AGI.
Here’s what really happened and why OpenAI should be worried: 🧵
A tiny AI model called HRM just beat Claude 3.5 and Gemini.
It doesn’t even use tokens.
They said it was just a research preview.
But it might be the first real shot at AGI.
Here’s what really happened and why OpenAI should be worried: 🧵
2
2/ Sapient Intelligence is a Singapore-based AI research startup focused on creating brain-inspired reasoning systems. They recently dropped HRM - a brain-inspired AI model that doesn’t think in tokens.
HRM (Hierarchical Reasoning Model) uses multi-timescale recurrence - a structure inspired by how humans reason, not how language models complete sentences.
One loop handles fast decisions. Another refines ideas over time. Together, they think, not just complete.
HRM (Hierarchical Reasoning Model) uses multi-timescale recurrence - a structure inspired by how humans reason, not how language models complete sentences.
One loop handles fast decisions. Another refines ideas over time. Together, they think, not just complete.
3
3/ Most AI models today "think" by writing one word at a time.
That’s called chain-of-thought.
It looks smart but if it makes a mistake early, everything after falls apart.
It’s fragile. It’s slow. And it’s not real thinking.
HRM works differently.
It doesn’t think out loud. It thinks silently like your brain.
Instead of writing words, it keeps ideas inside and improves them over time.
This is called chain-in-representation. A whole new way of reasoning.
That’s called chain-of-thought.
It looks smart but if it makes a mistake early, everything after falls apart.
It’s fragile. It’s slow. And it’s not real thinking.
HRM works differently.
It doesn’t think out loud. It thinks silently like your brain.
Instead of writing words, it keeps ideas inside and improves them over time.
This is called chain-in-representation. A whole new way of reasoning.
4
4/ How it works (inspired by your brain):
→ Fast module for low-level computations
→ Slow module for high-level plans
→ They loop over time
→ Each high-level update gives the fast module new direction
The result? Sustained, recursive, real reasoning.
→ Fast module for low-level computations
→ Slow module for high-level plans
→ They loop over time
→ Each high-level update gives the fast module new direction
The result? Sustained, recursive, real reasoning.
5
5/ Forget 175B parameter hype.
This model is 0.027B.
Trained on just 1000 examples per task.
Still:
→ 40.3% on ARC (vs Claude’s 21.2%)
→ 100% on Sudoku Extreme
→ 100% on Maze-Hard
No CoT. No pretraining. No hallucinations.
This model is 0.027B.
Trained on just 1000 examples per task.
Still:
→ 40.3% on ARC (vs Claude’s 21.2%)
→ 100% on Sudoku Extreme
→ 100% on Maze-Hard
No CoT. No pretraining. No hallucinations.
6
6/ Let that sink in:
Claude and Gemini can’t solve Sudoku Extreme.
HRM solves every single puzzle.
Using internal backtracking , like a human.
Not a prompt hack.
Actual search + memory + revision.
Claude and Gemini can’t solve Sudoku Extreme.
HRM solves every single puzzle.
Using internal backtracking , like a human.
Not a prompt hack.
Actual search + memory + revision.
7
7/ The kicker?
Its internal structure spontaneously mirrors brain patterns:
→ Different dimensionality in low vs high modules
→ Same ratios observed in neuroscience studies
→ Emergent not hardcoded
It’s not just inspired by brains. It learns like one.
Its internal structure spontaneously mirrors brain patterns:
→ Different dimensionality in low vs high modules
→ Same ratios observed in neuroscience studies
→ Emergent not hardcoded
It’s not just inspired by brains. It learns like one.
8
8/ The story behind it is wild too.
→ Built by a Gen-Z Tsinghua prodigy
→ Teamed up with ex-DeepMind researchers
→ Turned down Musk
→ Dropped HRM as a statement: “We’ll build brains, not bots.”
→ Built by a Gen-Z Tsinghua prodigy
→ Teamed up with ex-DeepMind researchers
→ Turned down Musk
→ Dropped HRM as a statement: “We’ll build brains, not bots.”
9
9/ So what’s next?
This isn’t the final answer.
But it might be the crack in the Transformer empire.
Because scaling won’t get us to AGI.
But structure just might.
This isn’t the final answer.
But it might be the crack in the Transformer empire.
Because scaling won’t get us to AGI.
But structure just might.
10
10/ AGI won’t emerge from more tokens.
It’ll come from better loops.
From models that don’t just sound smart but think smart.
The future of reasoning starts here. And it’s called HRM.
It’ll come from better loops.
From models that don’t just sound smart but think smart.
The future of reasoning starts here. And it’s called HRM.
11
12
Use AI to think faster, work smarter, and stay ahead
Join thousands of readers getting practical AI tips every week: 👉 shrutimishra.co
Join thousands of readers getting practical AI tips every week: 👉 shrutimishra.co