Vijay Sathish

contrarian; systems thinker; cross pollinator; AI/ML @blockinc @stanford @nvidia;
check highlights for artificial × biological intelligence deep dives 👇

[Vijay Sathish @Vslowsprinter]

Robotics stole hierarchical control from the brain.

But a significant gap still exists.

Read thread to find out 👉 https://t.co/soora3cBFG

[Vijay Sathish @Vslowsprinter]

Robotics stole hierarchical control from the brain. But a significant gap still exists. Read on to see what’s missing 🧵 👇🏽

(Follow me (@Vslowsprinter) for regular AI x bio insights; only images are AI gen; thoughts my own)

[Vijay Sathish @Vslowsprinter]

@jaya11235 Philosophically, they are both optimization algorithms. I will post about the brain vs neural net training too in the near future. The brain does not have backprop either. It uses local hebbian plasticity learning and spike timing dependent plasticity.

[Vijay Sathish @Vslowsprinter]

Evolution is a vastly inefficient optimization algorithm compared to backprop+gradient descent; yet it has 3 tricks up its sleeve

1 the timescale (2 billion years)
2 massive parallel search
3 compounding by retaining & stacking successful edits.

For unique bio x AI takes 👇

[Vijay Sathish @Vslowsprinter]

@gregisenberg A good majority of these ideas just seem like work about the work (tools for monitoring, correcting, facilitating, trusting agents etc).

[Vijay Sathish @Vslowsprinter]

@RExpunged It very much is. I recommend 'the blind watchmaker' or 'the selfish gene' by Richard Dawkins if you want to learn more. And it's truly magnificent.

[Vijay Sathish @Vslowsprinter]

@HerbertPotdevin Thanks for the really detailed and thoughtful response Herbert.
Agree on evolution not chasing a specific loss function. Evolution's 'loss function' of 'survive and replicate' allows for vastly more exploration and diversity, albeit at the cost of efficiency.

[Vijay Sathish @Vslowsprinter]

@_____Peter_ Haha, evolution has the opportunity to explore more via it's algorithm and produces more diversity for sure. Backprop has no exploration component, only extreme optimization efficiency.

[Vijay Sathish @Vslowsprinter]

The spike gets the spotlight. The pump pays the bill.
Neuron signaling is 'free' because its like discharging a battery. The brain's energy (ATP) bill comes from restoring the charge.

The life lesson: the real work always happens behind the scenes, away from the spotlight.

Follow @Vslowsprinter for unique AI x bio takes and deep dives:

[Vijay Sathish @Vslowsprinter]

This is age old principle that is going to propagate up the entire stack - when constraints start to matter - chips, models, routers, data, agents, tools, apps and everything in between get specialized and the system complexity increases. https://t.co/FVTlN3aFU2

[a16z @a16z]

In the '90s, general-purpose chips wiped out the specialists. The original Thinking Machines, Cray Computer, and a wave of others shut down, and their best ideas got absorbed into other chips.

Zain Asgar of @gimletlabs argues the conditions that made generalists win have changed. Dennard scaling (shrinking transistors made chips faster at the same power) ended around 2005, transistor costs stopped falling around 2012, and memory bandwidth never kept up with compute.

Meanwhile, token demand is projected to grow 24x by 2030 while available power grows about 2x. When power is the bottleneck, performance per watt wins, and that's where specialized chips shine.

Full piece from Zain on why "the time of specialization is upon us": https://t.co/dFA8McPid5

[Vijay Sathish @Vslowsprinter]

@a16z @gimletlabs This is age old principle that is going to propagate up the entire stack - when constraints start to matter - chips, models, routers, data, agents, tools, apps and everything in between get specialized and the system complexity increases:

[Vijay Sathish @Vslowsprinter]

The brain has a key design principle - "complicate". You heard that right, complicate, not simplify. Why? Because, the brain has to optimize for efficiency: under extreme space, energy and compute constraints. AI's constraints will ultimately boil down to similar constraints so that it can be unit economically profitable in the long term to all parties involved. Therefore, I predict that future super intelligence architecture will look nothing like today's gen AI and it's wasteful compute strategies.

[Vijay Sathish @Vslowsprinter]

@subyroy Not sure I follow your comment. What is the reference to html here in the context of this post?

[Vijay Sathish @Vslowsprinter]

Robotics stole hierarchical control from the brain.

But a significant gap still exists. Answer pinned to my profile 👇

[Vijay Sathish @Vslowsprinter]

@MEmelianchenko Care to share the link?

Caching generally works when the new request is EXACTLY identical as before.

Ex: pick up a cup is the same task in robotics, but cup size, weight, shape, orientation, material changes with each task instance, but is still handled efficiently in brain.

[Vijay Sathish @Vslowsprinter]

Remember Kahneman's Thinking, Fast and Slow?

The brain decides when to rely on learned patterns (system 1-fast/Cerebellum) vs. deliberate thinking (system 2-slow/Prefrontal Cortex). Energy and compute constraints necessitates extreme efficiency.

Where is the System 1 for RL?

[Vijay Sathish @Vslowsprinter]

@GauravBrief Yes, exactly.

[Vijay Sathish @Vslowsprinter]

Robotics stole hierarchical control from the brain.

But a significant gap still exists.

Read thread to find out 👉 https://t.co/soora3cBFG

[Vijay Sathish @Vslowsprinter]

Robotics stole hierarchical control from the brain. But a significant gap still exists. Read on to see what’s missing 🧵 👇🏽

(Follow me (@Vslowsprinter) for regular AI x bio insights; only images are AI gen; thoughts my own)

[Vijay Sathish @Vslowsprinter]

Today’s hard problem should become tomorrow’s easy skill.
Follow me (@Vslowsprinter) for regular AI x bio insights)

[Vijay Sathish @Vslowsprinter]

8/9 Google’s recently demoed using reasoning to generate training experience (i.e. policy distillation).

The bigger gap, in my view, is making this loop continuous and online. Can a robot turn effortful deliberation and repeated execution into reliable, reusable skills?

[Vijay Sathish @Vslowsprinter]

7/9 Robotics also mirrors this hierarchical control

In Gemini Robotics, the Embodied Reasoning model plans, reasons and breaks complex goals into tasks. The Vision Language Action model senses and executes them, while the reasoning model monitors progress. The hierarchy exists.

[Vijay Sathish @Vslowsprinter]

Robotics stole hierarchical control from the brain. But a significant gap still exists. Read on to see what’s missing 🧵 👇🏽

(Follow me (@Vslowsprinter) for regular AI x bio insights; only images are AI gen; thoughts my own)

[Vijay Sathish @Vslowsprinter]

6/9 The cerebellum predicts the consequences of actions.

Comparing predictions with sensory feedback generates an error signal that is used to adjust motor system in real time. Fine control doesn’t require replanning the entire task.

[Vijay Sathish @Vslowsprinter]

5/9 But repeatable skills can’t be completely hardcoded.

Imagine picking up a cup. Its position, weight, orientation and distance may be different from the last time. Mastery means adapting a learned skill to the current situation.

[Vijay Sathish @Vslowsprinter]

4/9 'Policy distillation' amortizes the costs for frequent tasks in the brain.

When deliberation and refinement lead to mastery via repetition, the action policy becomes a repeatable skill requiring less conscious supervision. Practice makes execution cheaper.

[Vijay Sathish @Vslowsprinter]

3/9 Therefore, deliberation is effortful and computationally expensive.

It makes economic sense to recruit this analytical machinery only when a novel situation, complexity or high stakes demand it.

[Vijay Sathish @Vslowsprinter]

2/9 Planning and deliberation require recruiting diverse experts, simulating trajectories and evaluating counterfactuals.

Cortical, memory and basal ganglia brain regions interact throughout. Longer horizon simulations are even more expensive.

[Vijay Sathish @Vslowsprinter]

1/9 The brain evolved hierarchical planning and control. Why?

Because necessity is the mother of invention. And because the brain is energetically and compute constrained necessitating extreme efficiency.