Eli Bressert
@elibressert.com
Nearly two decades of AI/ML and product experience at companies like Apple, Netflix, Stitch Fix, and Scenset. I love breaking down cutting-edge technologies into actionable products for the technically curious.
January 24, 2025 at 7:08 AM
Yesterday, I presented a deep dive session on reasoning AI at @southparkcommons.bsky.social. We discussed at length @fchollet.bsky.social's paper on The Measure of Intelligence (2019). His work fostered the ARC-AGI benchmark challenge and is helping steer AI development towards general reasoning.
January 23, 2025 at 5:46 PM
Yesterday, I presented a deep dive session on reasoning AI at @southparkcommons.bsky.social. We discussed at length @fchollet.bsky.social's paper on The Measure of Intelligence (2019). His work fostered the ARC-AGI benchmark challenge and is helping steer AI development towards general reasoning.
DeepSeek's R1 model is impressive. What I find particularly interesting is the following charts. The bar chart shows how much better R1 is compared to other SOTA models and then even the 32b param model is nearly as good as the 70b param R1 model.
Sources below
Sources below
January 21, 2025 at 12:34 AM
DeepSeek's R1 model is impressive. What I find particularly interesting is the following charts. The bar chart shows how much better R1 is compared to other SOTA models and then even the 32b param model is nearly as good as the 70b param R1 model.
Sources below
Sources below
This came in from Pete Skomoroch on Twitter which is essentially answers my questions: x.com/peteskomoroc...
The 32b param model is very, very close to the performance of the 70b model and is generally better than 01-mini.
The 32b param model is very, very close to the performance of the 70b model and is generally better than 01-mini.
January 21, 2025 at 12:28 AM
This came in from Pete Skomoroch on Twitter which is essentially answers my questions: x.com/peteskomoroc...
The 32b param model is very, very close to the performance of the 70b model and is generally better than 01-mini.
The 32b param model is very, very close to the performance of the 70b model and is generally better than 01-mini.
Monday morning activity
January 20, 2025 at 7:54 PM
Monday morning activity
A new paper from Deepmind leverages genetic algorithms (search, specifically) + LLMs to get some very impressive results in planning tasks. It outperforms o1, Best-of-N, and similar methods of by a large margin.
arxiv.org/abs/2501.09891
arxiv.org/abs/2501.09891
January 20, 2025 at 7:25 PM
A new paper from Deepmind leverages genetic algorithms (search, specifically) + LLMs to get some very impressive results in planning tasks. It outperforms o1, Best-of-N, and similar methods of by a large margin.
arxiv.org/abs/2501.09891
arxiv.org/abs/2501.09891
I wonder how many times this default macOS feature makes people think they are losing their minds until they figure what is **really** going on.
January 13, 2025 at 6:11 AM
I wonder how many times this default macOS feature makes people think they are losing their minds until they figure what is **really** going on.
A great & very thoughtful post about "internet for the mind" in regards to how information access has and will change.
wattenberger.com/thoughts/the...
My fav line from it: "But how we choose to use these tools will shape not just our understanding of the world, but our place within it."
wattenberger.com/thoughts/the...
My fav line from it: "But how we choose to use these tools will shape not just our understanding of the world, but our place within it."
January 12, 2025 at 4:52 AM
A great & very thoughtful post about "internet for the mind" in regards to how information access has and will change.
wattenberger.com/thoughts/the...
My fav line from it: "But how we choose to use these tools will shape not just our understanding of the world, but our place within it."
wattenberger.com/thoughts/the...
My fav line from it: "But how we choose to use these tools will shape not just our understanding of the world, but our place within it."
Our cats must be related 😂
December 23, 2024 at 3:48 AM
Our cats must be related 😂
Now we're talking! Researchers have developed and released Genesis, a very fast, capable universal physics engine where you can use natural language to render physics simulations.
100% Pythonic interface too.
Release page: genesis-embodied-ai.github.io
100% Pythonic interface too.
Release page: genesis-embodied-ai.github.io
December 20, 2024 at 10:25 PM
Now we're talking! Researchers have developed and released Genesis, a very fast, capable universal physics engine where you can use natural language to render physics simulations.
100% Pythonic interface too.
Release page: genesis-embodied-ai.github.io
100% Pythonic interface too.
Release page: genesis-embodied-ai.github.io
I love how David MacKay discusses visually how to best use his book on Information Theory, Inference, and Learning Algorithms.
Doing visual representation of the Contents section for different use cases is the way to go.
www.inference.org.uk/itprnn/book....
Doing visual representation of the Contents section for different use cases is the way to go.
www.inference.org.uk/itprnn/book....
December 9, 2024 at 10:14 PM
I love how David MacKay discusses visually how to best use his book on Information Theory, Inference, and Learning Algorithms.
Doing visual representation of the Contents section for different use cases is the way to go.
www.inference.org.uk/itprnn/book....
Doing visual representation of the Contents section for different use cases is the way to go.
www.inference.org.uk/itprnn/book....
1/2
Today the ARC-AGI team assessed the most capable intelligent AI models and shared a great summary.
Despite the challenge being open for 5 years, no one has hit the 85% score on intelligence (hint GPT-4o scored at 5%).
2020: 20.6%
2021: 28.5%
2022: 31.4%
2023: 34.3%
2024: 55.5%
Today the ARC-AGI team assessed the most capable intelligent AI models and shared a great summary.
Despite the challenge being open for 5 years, no one has hit the 85% score on intelligence (hint GPT-4o scored at 5%).
2020: 20.6%
2021: 28.5%
2022: 31.4%
2023: 34.3%
2024: 55.5%
December 7, 2024 at 1:19 AM
1/2
Today the ARC-AGI team assessed the most capable intelligent AI models and shared a great summary.
Despite the challenge being open for 5 years, no one has hit the 85% score on intelligence (hint GPT-4o scored at 5%).
2020: 20.6%
2021: 28.5%
2022: 31.4%
2023: 34.3%
2024: 55.5%
Today the ARC-AGI team assessed the most capable intelligent AI models and shared a great summary.
Despite the challenge being open for 5 years, no one has hit the 85% score on intelligence (hint GPT-4o scored at 5%).
2020: 20.6%
2021: 28.5%
2022: 31.4%
2023: 34.3%
2024: 55.5%
You know that uv is going mainstream with Anthropic suggests to use it to install their latest update, MCP!
November 26, 2024 at 1:42 AM
You know that uv is going mainstream with Anthropic suggests to use it to install their latest update, MCP!
Being more specific, at least a set of emojis like this.
November 24, 2024 at 2:31 AM
Being more specific, at least a set of emojis like this.
One of my favorite things to do when reading technical papers is mindmapping it with LLMs. It really helps at a high level understand how the author(s) present their work.
November 22, 2024 at 12:50 AM
One of my favorite things to do when reading technical papers is mindmapping it with LLMs. It really helps at a high level understand how the author(s) present their work.
Hello Jupyter Book 2.0
November 21, 2024 at 5:29 PM
Hello Jupyter Book 2.0
Jupyter NB is truly indispensable. Seeing Jupyter Book's 2.0a evolution & the amazing features it sports is mind blowing. My three fav features are:
MyST markdown (so, so good)
Clean and beautiful PDF output
New fast search capability
blog.jupyterbook.org/posts/2024-1...
(via @choldgraf.com)
MyST markdown (so, so good)
Clean and beautiful PDF output
New fast search capability
blog.jupyterbook.org/posts/2024-1...
(via @choldgraf.com)
November 21, 2024 at 5:20 PM
Jupyter NB is truly indispensable. Seeing Jupyter Book's 2.0a evolution & the amazing features it sports is mind blowing. My three fav features are:
MyST markdown (so, so good)
Clean and beautiful PDF output
New fast search capability
blog.jupyterbook.org/posts/2024-1...
(via @choldgraf.com)
MyST markdown (so, so good)
Clean and beautiful PDF output
New fast search capability
blog.jupyterbook.org/posts/2024-1...
(via @choldgraf.com)
Good morning California 🌞🌈
November 21, 2024 at 3:02 PM
Good morning California 🌞🌈
There's a fun trend where folks are asking ChatGPT
"Based on what you know about me, draw me a picture of what you think my current life looks like"
One of my kids walked up and said, "yeah, that's 100% you" 😂
"Based on what you know about me, draw me a picture of what you think my current life looks like"
One of my kids walked up and said, "yeah, that's 100% you" 😂
November 19, 2024 at 2:41 AM
There's a fun trend where folks are asking ChatGPT
"Based on what you know about me, draw me a picture of what you think my current life looks like"
One of my kids walked up and said, "yeah, that's 100% you" 😂
"Based on what you know about me, draw me a picture of what you think my current life looks like"
One of my kids walked up and said, "yeah, that's 100% you" 😂
There's no such thing as a free lunch?
uv proves otherwise. This Python package manager gives you:
- 10x-1000x faster installs
- Zero workflow changes
- Cleaner environments
100% focus on your AI/ML Python dev work vs. package and environment management.
open.substack.com/pub/elisalgo...
uv proves otherwise. This Python package manager gives you:
- 10x-1000x faster installs
- Zero workflow changes
- Cleaner environments
100% focus on your AI/ML Python dev work vs. package and environment management.
open.substack.com/pub/elisalgo...
November 16, 2024 at 7:16 AM
There's no such thing as a free lunch?
uv proves otherwise. This Python package manager gives you:
- 10x-1000x faster installs
- Zero workflow changes
- Cleaner environments
100% focus on your AI/ML Python dev work vs. package and environment management.
open.substack.com/pub/elisalgo...
uv proves otherwise. This Python package manager gives you:
- 10x-1000x faster installs
- Zero workflow changes
- Cleaner environments
100% focus on your AI/ML Python dev work vs. package and environment management.
open.substack.com/pub/elisalgo...