Rohit
@strangeloopcanon.com
I think, I write, sometimes I bleep? skeet? bleat? |
Read essays here https://www.strangeloopcanon.com |
Read the book here http://amazon.com/dp/B0CJ9F327M |
Read essays here https://www.strangeloopcanon.com |
Read the book here http://amazon.com/dp/B0CJ9F327M |
who has collated the best compendium of the types of AI errors, mistakes and reward hacks that we've seen happen in the last couple years?
November 8, 2025 at 4:49 AM
who has collated the best compendium of the types of AI errors, mistakes and reward hacks that we've seen happen in the last couple years?
You know, in any real Jurassic Park, nobody would be able to touch grass
October 31, 2025 at 7:52 PM
You know, in any real Jurassic Park, nobody would be able to touch grass
LLMs are cool but they're made of words, which means if you change the words you can change its proclivities.
I was wondering about this, and turns out you can "poison" a model by just rewriting text that it learns from, steered by a simple persona vector.
www.strangeloopcanon.com/p/poisoned-p...
I was wondering about this, and turns out you can "poison" a model by just rewriting text that it learns from, steered by a simple persona vector.
www.strangeloopcanon.com/p/poisoned-p...
Poisoned prose
on semantic trojan horses in LLMs
www.strangeloopcanon.com
October 27, 2025 at 5:21 PM
LLMs are cool but they're made of words, which means if you change the words you can change its proclivities.
I was wondering about this, and turns out you can "poison" a model by just rewriting text that it learns from, steered by a simple persona vector.
www.strangeloopcanon.com/p/poisoned-p...
I was wondering about this, and turns out you can "poison" a model by just rewriting text that it learns from, steered by a simple persona vector.
www.strangeloopcanon.com/p/poisoned-p...
sora is getting quite good with standup comedy
October 25, 2025 at 7:31 PM
sora is getting quite good with standup comedy
Kids love categorisation. 8yo has been recently been thinking about how evolution happens and he has decided that every island where separate species evolves, a la Island speciation, should be called its own continent.
October 15, 2025 at 4:59 AM
Kids love categorisation. 8yo has been recently been thinking about how evolution happens and he has decided that every island where separate species evolves, a la Island speciation, should be called its own continent.
Thank god this man was born in the era of social media.
October 13, 2025 at 1:31 AM
Thank god this man was born in the era of social media.
Reposted by Rohit
> It is 2025 BC. I am a soldier from one of the outlying provinces called to defend the pyramid in Memphis
> It is 2025 AD. I am a soldier from one of the outlying provinces called to defend the pyramid in Memphis
> It is 2025 AD. I am a soldier from one of the outlying provinces called to defend the pyramid in Memphis
October 11, 2025 at 5:32 PM
> It is 2025 BC. I am a soldier from one of the outlying provinces called to defend the pyramid in Memphis
> It is 2025 AD. I am a soldier from one of the outlying provinces called to defend the pyramid in Memphis
> It is 2025 AD. I am a soldier from one of the outlying provinces called to defend the pyramid in Memphis
Google's Nobels are another indication that Peter Thiel was right with his theory that monopolies are good. And that natural monopolies like Google might even be better than regulated monopolies like Bell Labs.
October 7, 2025 at 6:30 PM
Google's Nobels are another indication that Peter Thiel was right with his theory that monopolies are good. And that natural monopolies like Google might even be better than regulated monopolies like Bell Labs.
Most LLM prose is flat. But, there's no logical reason it should be! So I built “Horace” to see if measuring cadence can help steer models toward better writing, by measuring rhythm and surprise.
www.strangeloopcanon.com/p/can-we-get...
www.strangeloopcanon.com/p/can-we-get...
Can we get an AI to write better?
A small step
www.strangeloopcanon.com
October 7, 2025 at 6:21 PM
Most LLM prose is flat. But, there's no logical reason it should be! So I built “Horace” to see if measuring cadence can help steer models toward better writing, by measuring rhythm and surprise.
www.strangeloopcanon.com/p/can-we-get...
www.strangeloopcanon.com/p/can-we-get...
Reposted by Rohit
it may feel a little like homework, but i genuinely think people should read the lincoln-douglas debates to see what the practical politics of anti-slavery looked like in an environment where the public was neutral to hostile to the question of emancipation
September 28, 2025 at 4:11 PM
it may feel a little like homework, but i genuinely think people should read the lincoln-douglas debates to see what the practical politics of anti-slavery looked like in an environment where the public was neutral to hostile to the question of emancipation
Another update to mlx-genkit. Fixed a bunch more options incl doing structured adherence and support for streaming. Happy Saturday!
pip install mlx-genkit
github.com/strangeloopc...
pip install mlx-genkit
github.com/strangeloopc...
GitHub - strangeloopcanon/mlx-genkit: Torch feature parity for MLX
Torch feature parity for MLX. Contribute to strangeloopcanon/mlx-genkit development by creating an account on GitHub.
github.com
September 26, 2025 at 11:19 PM
Another update to mlx-genkit. Fixed a bunch more options incl doing structured adherence and support for streaming. Happy Saturday!
pip install mlx-genkit
github.com/strangeloopc...
pip install mlx-genkit
github.com/strangeloopc...
This values TikTok the same as Snapchat
September 26, 2025 at 5:31 AM
This values TikTok the same as Snapchat
Figured the best way to learn how to do distributed KV cache is to build one. Ideal for RL rollouts, synthetic data gen, or large scale evals.
Think y'all might like it, why let chinese labs have all the fun. BCache..
github.com/strangeloopc...
Think y'all might like it, why let chinese labs have all the fun. BCache..
github.com/strangeloopc...
GitHub - strangeloopcanon/BCache: High Performance Hierarchical KV Cache Planner
High Performance Hierarchical KV Cache Planner. Contribute to strangeloopcanon/BCache development by creating an account on GitHub.
github.com
September 25, 2025 at 11:30 PM
Figured the best way to learn how to do distributed KV cache is to build one. Ideal for RL rollouts, synthetic data gen, or large scale evals.
Think y'all might like it, why let chinese labs have all the fun. BCache..
github.com/strangeloopc...
Think y'all might like it, why let chinese labs have all the fun. BCache..
github.com/strangeloopc...
To not let MoonshotAI have all the fun, tried building a homegrown library to hot-swap your production LLM weights, fast, with versioned model updates across a GPU cluster.
Not my expertise, so feedback sought and welcome!
github.com/strangeloopc...
Not my expertise, so feedback sought and welcome!
github.com/strangeloopc...
GitHub - strangeloopcanon/hotweights: Fast, versioned weight updates for LLM serving and training
Fast, versioned weight updates for LLM serving and training - strangeloopcanon/hotweights
github.com
September 23, 2025 at 5:06 PM
To not let MoonshotAI have all the fun, tried building a homegrown library to hot-swap your production LLM weights, fast, with versioned model updates across a GPU cluster.
Not my expertise, so feedback sought and welcome!
github.com/strangeloopc...
Not my expertise, so feedback sought and welcome!
github.com/strangeloopc...
Tylenol is poison. It's why I only take paracetamol.
September 22, 2025 at 10:08 PM
Tylenol is poison. It's why I only take paracetamol.
$100k surcharge on every H1B application. Solve for the equilibrium.
September 20, 2025 at 3:52 AM
$100k surcharge on every H1B application. Solve for the equilibrium.
Reposted by Rohit
In case any of you guys like playing around with AI on Apple silicon, an attempt to get MLX usage to parity with torch
github.com/strangeloopc...
github.com/strangeloopc...
GitHub - strangeloopcanon/mlx-gen-parity: Torch feature parity for MLX
Torch feature parity for MLX. Contribute to strangeloopcanon/mlx-gen-parity development by creating an account on GitHub.
github.com
September 13, 2025 at 12:45 AM
In case any of you guys like playing around with AI on Apple silicon, an attempt to get MLX usage to parity with torch
github.com/strangeloopc...
github.com/strangeloopc...
In case any of you guys like playing around with AI on Apple silicon, an attempt to get MLX usage to parity with torch
github.com/strangeloopc...
github.com/strangeloopc...
GitHub - strangeloopcanon/mlx-gen-parity: Torch feature parity for MLX
Torch feature parity for MLX. Contribute to strangeloopcanon/mlx-gen-parity development by creating an account on GitHub.
github.com
September 13, 2025 at 12:45 AM
In case any of you guys like playing around with AI on Apple silicon, an attempt to get MLX usage to parity with torch
github.com/strangeloopc...
github.com/strangeloopc...
mlx-gen-parity: made training utilities work better, model helpers, use mlx.nn.tree_map when available ... made it easier to use MLX to run like GRPO type training runs.
github.com/strangeloopc...
github.com/strangeloopc...
GitHub - strangeloopcanon/mlx-gen-parity
Contribute to strangeloopcanon/mlx-gen-parity development by creating an account on GitHub.
github.com
September 12, 2025 at 4:45 PM
mlx-gen-parity: made training utilities work better, model helpers, use mlx.nn.tree_map when available ... made it easier to use MLX to run like GRPO type training runs.
github.com/strangeloopc...
github.com/strangeloopc...
Pretty spectacular news! Sorely needed.
September 10, 2025 at 11:42 PM
Pretty spectacular news! Sorely needed.
At the broadest level there is an epidemic of unhappiness and disillusionment that is sweeping across every nation and much of the problems we see seem downstream of that fact.
September 10, 2025 at 12:55 AM
At the broadest level there is an epidemic of unhappiness and disillusionment that is sweeping across every nation and much of the problems we see seem downstream of that fact.
We should rename pre training and post training to nature and nurture.
September 5, 2025 at 8:17 PM
We should rename pre training and post training to nature and nurture.
The year is 2045. Imagine AI kinda plateaus or only grows normally (so no godlike ASI). What are your best guesses about what the world is going to be like, geopolitically, economically, socially?
August 30, 2025 at 1:22 AM
The year is 2045. Imagine AI kinda plateaus or only grows normally (so no godlike ASI). What are your best guesses about what the world is going to be like, geopolitically, economically, socially?
AI has graduated from PhD student to advisor
August 25, 2025 at 6:16 AM
AI has graduated from PhD student to advisor