Emsi.Me
@emsiak.bsky.social
Making ML ideas a reality.
ML Hacker.
ML Hacker.
Pinned
Emsi.Me
@emsiak.bsky.social
· Nov 23
GitHub - emsi/gh_copilot_unofficial_openai_client: GitHub Copilot Unoffician OpenAI API Client
GitHub Copilot Unoffician OpenAI API Client. Contribute to emsi/gh_copilot_unofficial_openai_client development by creating an account on GitHub.
github.com
Turns GitHub copilot into OpenAI API endpoint with modules like GPT and Claude in one place.
We Panic About AI Hallucinations While Ignoring 94% Human Error Rates open.substack.com/pub/emsime/p...
We Panic About AI Hallucinations While Ignoring 94% Human Error Rates
This post was originally published on emsi.me
open.substack.com
August 26, 2025 at 3:23 PM
We Panic About AI Hallucinations While Ignoring 94% Human Error Rates open.substack.com/pub/emsime/p...
Many people believe that language models cannot become smarter without access to new data or external support. However, it turns out that models have the ability to self-improve by learning deeper thinking and problem analysis.
January 13, 2025 at 4:41 PM
Many people believe that language models cannot become smarter without access to new data or external support. However, it turns out that models have the ability to self-improve by learning deeper thinking and problem analysis.
Ever felt like there are more people on X than here?
Think again.
Within seconds of posting a comment on X I see some bots interacting with my it and I see a steady stream of similar "followers" on X.
A one reaction here is worth dozens reactions here (at least).
Think again.
Within seconds of posting a comment on X I see some bots interacting with my it and I see a steady stream of similar "followers" on X.
A one reaction here is worth dozens reactions here (at least).
January 4, 2025 at 4:19 PM
Ever felt like there are more people on X than here?
Think again.
Within seconds of posting a comment on X I see some bots interacting with my it and I see a steady stream of similar "followers" on X.
A one reaction here is worth dozens reactions here (at least).
Think again.
Within seconds of posting a comment on X I see some bots interacting with my it and I see a steady stream of similar "followers" on X.
A one reaction here is worth dozens reactions here (at least).
When you ask AI for spiritual advice :)
There's many things in it that I cannot comprehend.
There's many things in it that I cannot comprehend.
December 27, 2024 at 11:04 PM
When you ask AI for spiritual advice :)
There's many things in it that I cannot comprehend.
There's many things in it that I cannot comprehend.
I'm starting to get worried.
December 20, 2024 at 3:50 PM
I'm starting to get worried.
AI Achieves Superhuman Performance in Medical Reasoning - Emsi's feed
December 18, 2024 at 11:37 PM
AI Achieves Superhuman Performance in Medical Reasoning - Emsi's feed
After diving into Google's recent Willow quantum computing announcement, I wanted to share some thoughts on what makes this development interesting from a computer scientist's perspective.
December 11, 2024 at 7:22 PM
After diving into Google's recent Willow quantum computing announcement, I wanted to share some thoughts on what makes this development interesting from a computer scientist's perspective.
What if I told you that somewhere, right now, a mushroom is driving a robot? No, this isn’t a scene from a quirky sci-fi movie—this is real life.
When Mushrooms Drive: The Rise of Biohybrid Robots - Emsi's feed
What if I told you that somewhere, right now, a mushroom is driving a robot? No, this isn’t a scene from a quirky sci-fi movie—this is real life. Scientists from…
www.emsi.me
December 10, 2024 at 8:54 PM
What if I told you that somewhere, right now, a mushroom is driving a robot? No, this isn’t a scene from a quirky sci-fi movie—this is real life.
COCONUT (Chain of Continuous Thought) revolutionizes AI reasoning by allowing models to think in latent space. It mirrors human cognition, where not all reasoning is verbalized.
34.1% accuracy on GSM8k math tasks and 97% on complex planning tasks!
tinyurl.com/coconut-cot
34.1% accuracy on GSM8k math tasks and 97% on complex planning tasks!
tinyurl.com/coconut-cot
Meta's Breakthrough: Teaching Language Models to Think Outside the Box - Literally - Emsi's feed
Remember when we thought language models had to express their reasoning through words, just like humans do? Well, Meta’s researchers have just turned that assumption on its head with an…
tinyurl.com
December 10, 2024 at 6:13 PM
COCONUT (Chain of Continuous Thought) revolutionizes AI reasoning by allowing models to think in latent space. It mirrors human cognition, where not all reasoning is verbalized.
34.1% accuracy on GSM8k math tasks and 97% on complex planning tasks!
tinyurl.com/coconut-cot
34.1% accuracy on GSM8k math tasks and 97% on complex planning tasks!
tinyurl.com/coconut-cot
SWE-agent, this all-in-one AI assistant autonomously fixes GitHub bugs, performs web tasks, and identifies cybersecurity vulnerabilities through its EnIGMA mode. It seamlessly integrates with tools, offering a "plug in and play" experience.
Riding the Next Wave in Automated Software Engineering: Meet SWE-agent - Emsi's feed
Discover how SWE-agent, developed by Princeton and Stanford researchers, is revolutionizing software development by enabling AI to autonomously fix GitHub issues, detect security vulnerabilities, and ...
www.emsi.me
December 9, 2024 at 10:37 PM
SWE-agent, this all-in-one AI assistant autonomously fixes GitHub bugs, performs web tasks, and identifies cybersecurity vulnerabilities through its EnIGMA mode. It seamlessly integrates with tools, offering a "plug in and play" experience.
Finally, the feature I had been waiting for so long!
Ollama Unveils Structured Outputs for Enhanced Data Extraction in AI Models - Emsi's feed
Ollama has introduced structured outputs, enhancing the way models can generate outputs by adhering to formats defined by JSON schemas. This update, available in both Python and JavaScript libraries, ...
www.emsi.me
December 9, 2024 at 7:18 PM
Finally, the feature I had been waiting for so long!
Reposted by Emsi.Me
We could go on about how we welcome publishers, we don't demote links, we encourage independent developers to build apps and extensions on top of Bluesky's network.... but instead, we'll show you.
All thanks to the incredible community here! 🦋
All thanks to the incredible community here! 🦋
The Engagement Is Better on Bluesky - Bluesky
Bluesky is the lobby to the open web. Find and build your community here.
bsky.social
November 29, 2024 at 10:30 PM
We could go on about how we welcome publishers, we don't demote links, we encourage independent developers to build apps and extensions on top of Bluesky's network.... but instead, we'll show you.
All thanks to the incredible community here! 🦋
All thanks to the incredible community here! 🦋
Reposted by Emsi.Me
Did you know that 99% of email today is spam? Your inbox isn’t 99% spam because AI is used to filter it.
The same 99% will happen here too, but if AI researchers continue to get perma-banned for making available the datasets needed to filter it, it’s going to make this platform unusable.
The same 99% will happen here too, but if AI researchers continue to get perma-banned for making available the datasets needed to filter it, it’s going to make this platform unusable.
November 28, 2024 at 6:12 PM
Did you know that 99% of email today is spam? Your inbox isn’t 99% spam because AI is used to filter it.
The same 99% will happen here too, but if AI researchers continue to get perma-banned for making available the datasets needed to filter it, it’s going to make this platform unusable.
The same 99% will happen here too, but if AI researchers continue to get perma-banned for making available the datasets needed to filter it, it’s going to make this platform unusable.
50% model reduction without quality loss
New Compression Technique Could Significantly Reduce AI Infrastructure Costs - Emsi's feed
A promising new compression technique called ZipNN demonstrates the ability to reduce AI model sizes by up to 50% without performance loss. This development addresses a growing challenge in the…
www.emsi.me
November 28, 2024 at 2:02 PM
50% model reduction without quality loss
Marco-o1: designed for tackling complex, open-ended problems. By integrating Chain-of-Thought fine-tuning, Monte Carlo Tree Search it sets a new standard in problem-solving, enhances capabilities, offering nuanced understanding of colloquial expressions. Available for researchers and developers.
Marco-o1: Advancing AI Capabilities in Open-Ended Problem Solving - Emsi's feed
The Marco-o1 project represents a significant advancement in the field of artificial intelligence, introducing a Large Language Model (LLM) designed to address complex, open-ended problems alongside m...
www.emsi.me
November 26, 2024 at 2:59 PM
Marco-o1: designed for tackling complex, open-ended problems. By integrating Chain-of-Thought fine-tuning, Monte Carlo Tree Search it sets a new standard in problem-solving, enhances capabilities, offering nuanced understanding of colloquial expressions. Available for researchers and developers.
Reposted by Emsi.Me
I wrote this code to follow the same people someone else is following. I figured that would fix my feed esp if someone is having a good experience I can just "have what they are having"
gist.github.com/hamelsmu/fb9...
gist.github.com/hamelsmu/fb9...
"I'll have what they are having" for bluesky. The motiviation is to mimic who someone else is following who reports they are having a good experience on bluesky!
"I'll have what they are having" for bluesky. The motiviation is to mimic who someone else is following who reports they are having a good experience on bluesky! - follow_theirs.py
gist.github.com
November 24, 2024 at 3:37 PM
I wrote this code to follow the same people someone else is following. I figured that would fix my feed esp if someone is having a good experience I can just "have what they are having"
gist.github.com/hamelsmu/fb9...
gist.github.com/hamelsmu/fb9...
AI-generated ideas scored significantly higher on novelty (5.64 out of 10) compared to human expert ideas (4.84 out of 10). These differences aren’t just statistical noise – they’re significant at the p<0.01 level, meaning we can be quite confident in the results.
AI Outperforms Human Experts in Research Ideation - Emsi's feed
AI-generated ideas scored significantly higher on novelty (5.64 out of 10) compared to human expert ideas (4.84 out of 10). When human experts helped rank and filter AI's ideas, the score rose even hi...
www.emsi.me
November 24, 2024 at 3:55 PM
AI-generated ideas scored significantly higher on novelty (5.64 out of 10) compared to human expert ideas (4.84 out of 10). These differences aren’t just statistical noise – they’re significant at the p<0.01 level, meaning we can be quite confident in the results.
An interesting approach to speed up and reduce LLM memory usage at inference.
GemFilter: Streamlining Long-Context Processing for Faster LLMs - Emsi's feed
Processing long-context inputs has always been a challenge for Large Language Models (LLMs), demanding substantial computational resources and increasing latency. The new algorithm, GemFilter, offers ...
www.emsi.me
November 24, 2024 at 2:29 PM
An interesting approach to speed up and reduce LLM memory usage at inference.
Interesting idea. Better than DeBERTa-based classifiers in AUC and precision.
Defending LLMs: Using Machine Learning to Combat Prompt Injection Attacks - Emsi's feed
Large Language Models (LLMs) are widely integrated into modern organizational frameworks, celebrated for their advanced generative abilities. Yet, this integration comes with its share of vulnerabilit...
www.emsi.me
November 23, 2024 at 5:31 PM
Interesting idea. Better than DeBERTa-based classifiers in AUC and precision.
Turns GitHub copilot into OpenAI API endpoint with modules like GPT and Claude in one place.
GitHub - emsi/gh_copilot_unofficial_openai_client: GitHub Copilot Unoffician OpenAI API Client
GitHub Copilot Unoffician OpenAI API Client. Contribute to emsi/gh_copilot_unofficial_openai_client development by creating an account on GitHub.
github.com
November 23, 2024 at 5:20 PM
Turns GitHub copilot into OpenAI API endpoint with modules like GPT and Claude in one place.
Already using it on production.
When graded by gpt-4o it marginally outperforms gpt-4o itself on coding tasks.
When graded by gpt-4o it marginally outperforms gpt-4o itself on coding tasks.
Qwen2.5-Coder Series: Revolutionizing Open-Source Code Generation with Advanced LLMs - Emsi's feed
The recent unveiling of the Qwen2.5-Coder series marks a significant advancement in the field of open-source code Large Language Models (LLMs), introducing a range of models that are powerful, diverse...
www.emsi.me
November 23, 2024 at 5:13 PM
Already using it on production.
When graded by gpt-4o it marginally outperforms gpt-4o itself on coding tasks.
When graded by gpt-4o it marginally outperforms gpt-4o itself on coding tasks.
Reposted by Emsi.Me
Here is a list of ML OSS & Open Source / Science enthusiasts I found on Bluesky 🦋
go.bsky.app/8MFcfXd
Let me know if you find such people here!
I'm still new here and probably the list misses many must-add people, so let's built it together💪
go.bsky.app/8MFcfXd
Let me know if you find such people here!
I'm still new here and probably the list misses many must-add people, so let's built it together💪
November 21, 2024 at 5:19 AM
Here is a list of ML OSS & Open Source / Science enthusiasts I found on Bluesky 🦋
go.bsky.app/8MFcfXd
Let me know if you find such people here!
I'm still new here and probably the list misses many must-add people, so let's built it together💪
go.bsky.app/8MFcfXd
Let me know if you find such people here!
I'm still new here and probably the list misses many must-add people, so let's built it together💪