Andy Gill
@andygillku.bsky.social
SVP of Product Development at Kodamai. True Scotsman. Former professor at the University of Kansas. Former Googler. Jayhawk fan. Long time functional programmer.
To whoever called the SIGPLAN podcast "current continuation". Brilliant name!
May 22, 2025 at 11:41 PM
To whoever called the SIGPLAN podcast "current continuation". Brilliant name!
Reposted by Andy Gill
Turns out the TypeScript **types** are DOOM-complete. www.youtube.com/watch?v=0mCs...
TypeScript types can run DOOM
YouTube video by Michigan TypeScript
www.youtube.com
February 27, 2025 at 5:21 PM
Turns out the TypeScript **types** are DOOM-complete. www.youtube.com/watch?v=0mCs...
Haverscript (github.com/andygill/hav...) now supports Agents!
Haverscript is a lightweight Python library designed for working with LLMs and has recently introduced agents. The library offers a readable API for crafting prompts, chaining LLM calls, and integrating new AI-driven features.
Haverscript is a lightweight Python library designed for working with LLMs and has recently introduced agents. The library offers a readable API for crafting prompts, chaining LLM calls, and integrating new AI-driven features.
GitHub - andygill/haverscript: Library for building agents and managing LLM interactions
Library for building agents and managing LLM interactions - andygill/haverscript
github.com
February 27, 2025 at 3:36 AM
Haverscript (github.com/andygill/hav...) now supports Agents!
Haverscript is a lightweight Python library designed for working with LLMs and has recently introduced agents. The library offers a readable API for crafting prompts, chaining LLM calls, and integrating new AI-driven features.
Haverscript is a lightweight Python library designed for working with LLMs and has recently introduced agents. The library offers a readable API for crafting prompts, chaining LLM calls, and integrating new AI-driven features.
Gradio.app is a disappointing platform. We know how to make better looking and more usable widgets. The underlying programming model is not composable in any real sense. After spending hundreds of hours debugging gradio applications I'm actively looking for something better and perhaps LLM friendly.
Gradio
Build & Share Delightful Machine Learning Apps
Gradio.app
February 25, 2025 at 3:47 PM
Gradio.app is a disappointing platform. We know how to make better looking and more usable widgets. The underlying programming model is not composable in any real sense. After spending hundreds of hours debugging gradio applications I'm actively looking for something better and perhaps LLM friendly.
Spent the morning adding monads to Haverscript.
def bind(self, completion: Callable[[Any], Reply]) -> Reply: ...
The unit tests are the monads laws.
def bind(self, completion: Callable[[Any], Reply]) -> Reply: ...
The unit tests are the monads laws.
February 16, 2025 at 4:14 PM
Spent the morning adding monads to Haverscript.
def bind(self, completion: Callable[[Any], Reply]) -> Reply: ...
The unit tests are the monads laws.
def bind(self, completion: Callable[[Any], Reply]) -> Reply: ...
The unit tests are the monads laws.
Has anyone taken the various ICFP programing content specifications, and given them to LLMs? There are >25 challenging programming problems to input. Yes, some of the solutions are on the web already, but in various languages, and fragmented formats.
www.icfpconference.org/contest.html
www.icfpconference.org/contest.html
The International Conference on Functional
Programming
The conference provides a forum for researchers and developers to hear about the latest work on the design, implementations, principles, and uses of functional programming. The conference covers the e...
www.icfpconference.org
February 12, 2025 at 8:01 PM
Has anyone taken the various ICFP programing content specifications, and given them to LLMs? There are >25 challenging programming problems to input. Yes, some of the solutions are on the web already, but in various languages, and fragmented formats.
www.icfpconference.org/contest.html
www.icfpconference.org/contest.html
Here is my monthly bill from together.ai. Inference is cheap. This was from running unit tests for Haverscript, and cost a fraction of what booting up a stand-alone cloud machine for larger models would cost.
February 10, 2025 at 4:28 PM
Here is my monthly bill from together.ai. Inference is cheap. This was from running unit tests for Haverscript, and cost a fraction of what booting up a stand-alone cloud machine for larger models would cost.
Thank you Erik for the shout out!
January 27, 2025 at 8:10 PM
Thank you Erik for the shout out!
I gave a talk last week at the University of Kansas I2S Student Research Symposium about Haverscript and LLMs. I think the talk was well received and I got good questions. Here are the slides if anyone is interested.
docs.google.com/presentation...
docs.google.com/presentation...
Haverscript I2S
Haverscript and LLMs Andy Gill
docs.google.com
January 27, 2025 at 2:39 AM
I gave a talk last week at the University of Kansas I2S Student Research Symposium about Haverscript and LLMs. I think the talk was well received and I got good questions. Here are the slides if anyone is interested.
docs.google.com/presentation...
docs.google.com/presentation...
Took me a long time to internalize AI agents. Like monads, the idea are simple once you have it. **An AI agent is class with access to an LLM** Thats it. Three things follow. An agent has identity. An agent (and its "knowledge") can be cloned. An agent can choose to be forgetful of its history.
January 11, 2025 at 3:55 AM
Took me a long time to internalize AI agents. Like monads, the idea are simple once you have it. **An AI agent is class with access to an LLM** Thats it. Three things follow. An agent has identity. An agent (and its "knowledge") can be cloned. An agent can choose to be forgetful of its history.
I finally signed up for GitHub CoPilot. Good but not great. Knows my debugging patterns. A couple of times it's spooky when I think for a few minutes, type a variable name and it suggests exactly what I decided to do. Make some rookie mistakes. I hope I can learn better ways using it over time.
January 5, 2025 at 1:12 AM
I finally signed up for GitHub CoPilot. Good but not great. Knows my debugging patterns. A couple of times it's spooky when I think for a few minutes, type a variable name and it suggests exactly what I decided to do. Make some rookie mistakes. I hope I can learn better ways using it over time.
The reply from together.ai.
"Seed does not guarantee the same outputs. Temp=0 and consistent seed should get you close, but we have observed variance between different machines running the model. This is the same issue that OpenAI has with deterministic outputs."
So how do others test OpenAI APIs?
"Seed does not guarantee the same outputs. Temp=0 and consistent seed should get you close, but we have observed variance between different machines running the model. This is the same issue that OpenAI has with deterministic outputs."
So how do others test OpenAI APIs?
Together AI – The AI Acceleration Cloud - Fast Inference, Fine-Tuning & Training
Run and fine-tune generative AI models with easy-to-use APIs and highly scalable infrastructure. Train & deploy models at scale on our AI Acceleration Cloud and scalable GPU clusters. Optimize perform...
together.ai
December 31, 2024 at 5:44 PM
The reply from together.ai.
"Seed does not guarantee the same outputs. Temp=0 and consistent seed should get you close, but we have observed variance between different machines running the model. This is the same issue that OpenAI has with deterministic outputs."
So how do others test OpenAI APIs?
"Seed does not guarantee the same outputs. Temp=0 and consistent seed should get you close, but we have observed variance between different machines running the model. This is the same issue that OpenAI has with deterministic outputs."
So how do others test OpenAI APIs?
The second release of Haverscript is out! This release adds Haverscript middleware (aka lenses) that sit between the prompt and the LLM. There are around a dozen middleware options provided in this release, including echo, caching, and validation.
ai-andy-gill.hashnode.dev/haverscript-...
ai-andy-gill.hashnode.dev/haverscript-...
Haverscript 0.2.1 is out
Haverscript, a Domain Specific Language designed for building LLM-based agents, continues to evolve with new features and improvements aimed at enhancing its flexibility and usability. With the recent...
ai-andy-gill.hashnode.dev
December 31, 2024 at 2:03 AM
The second release of Haverscript is out! This release adds Haverscript middleware (aka lenses) that sit between the prompt and the LLM. There are around a dozen middleware options provided in this release, including echo, caching, and validation.
ai-andy-gill.hashnode.dev/haverscript-...
ai-andy-gill.hashnode.dev/haverscript-...
Could not get seed working on together.ai's API. So an LLM call was utilized to compare the output with the golden file, and structured output formatting was used to obtain a actionable similarity measure. Not ideal, but works.
ai-andy-gill.hashnode.dev/how-do-you-c...
ai-andy-gill.hashnode.dev/how-do-you-c...
How do you check LLM output for correctness?
While doing testing of the together.ai API for haverscript, I came across a problem. Together’s API seem to be ignoring the seed argument. The seed argument provides reproducibility, which is essentia...
ai-andy-gill.hashnode.dev
December 23, 2024 at 5:48 AM
Could not get seed working on together.ai's API. So an LLM call was utilized to compare the output with the golden file, and structured output formatting was used to obtain a actionable similarity measure. Not ideal, but works.
ai-andy-gill.hashnode.dev/how-do-you-c...
ai-andy-gill.hashnode.dev/how-do-you-c...
Today I found out that the sigmoid function is equivalent to softmax when applied to two inputs, with one input fixed to 0. With almost a decade of experience in machine learning, and earlier work in compute for signal processing, I had never stumbled over this factoid.
December 17, 2024 at 1:20 AM
Today I found out that the sigmoid function is equivalent to softmax when applied to two inputs, with one input fixed to 0. With almost a decade of experience in machine learning, and earlier work in compute for signal processing, I had never stumbled over this factoid.
In some code today, I reused the most perfect comment, taken from a compiler in the 90s.
# The ice is thin here but it holds.
Apologies to the original author.
# The ice is thin here but it holds.
Apologies to the original author.
December 10, 2024 at 1:38 AM
In some code today, I reused the most perfect comment, taken from a compiler in the 90s.
# The ice is thin here but it holds.
Apologies to the original author.
# The ice is thin here but it holds.
Apologies to the original author.
youtu.be/R-wbfP1pmVw?... Nice hobby project. It's always fun as a maker tinkering with such things.
DOOM on a 3D-printed mechanical TV
YouTube video by bitluni
youtu.be
December 6, 2024 at 5:40 AM
youtu.be/R-wbfP1pmVw?... Nice hobby project. It's always fun as a maker tinkering with such things.
Writing some Python trying to find the right abstraction. Wait a minute. This is a textbook example of the Expression Problem!
December 5, 2024 at 12:18 AM
Writing some Python trying to find the right abstraction. Wait a minute. This is a textbook example of the Expression Problem!
Had a go at writing a custom plugin for ComfyUI today. It was remarkably easy to do so, and I love the quick boot time which made debugging easy. I pulled out some old Sunflower Reality labs code and ported it to project equirectangular to rectilinear with sliders. Feels like a reusable component.
November 30, 2024 at 12:28 AM
Had a go at writing a custom plugin for ComfyUI today. It was remarkably easy to do so, and I love the quick boot time which made debugging easy. I pulled out some old Sunflower Reality labs code and ported it to project equirectangular to rectilinear with sliders. Feels like a reusable component.
I'm advocating for using markdown as a simple file format to recording and replaying LLM sessions, specifically using bird tracks before prompts. There are lots of ways to render markdown. HaverScript uses this format already for the echo mode. Thoughts?
ai-andy-gill.hashnode.dev/markdown-as-...
ai-andy-gill.hashnode.dev/markdown-as-...
Markdown as Context Format for LLMs
In HaverScript we use Markdown as an output format. We represent prompts by prefixing >, sometimes called bird-tracks, before the prompt, and the responses are normal markdown text. This gives both a readable output “script” of what happened, as well...
ai-andy-gill.hashnode.dev
November 17, 2024 at 11:59 PM
I'm advocating for using markdown as a simple file format to recording and replaying LLM sessions, specifically using bird tracks before prompts. There are lots of ways to render markdown. HaverScript uses this format already for the echo mode. Thoughts?
ai-andy-gill.hashnode.dev/markdown-as-...
ai-andy-gill.hashnode.dev/markdown-as-...
I've created a blog to ramble on about Domain Specific Languages, Functional Programming, AI and LLMs, and other things. Pretty sparse right now, but I hope to populate it over time.
ai-andy-gill.hashnode.dev
ai-andy-gill.hashnode.dev
Rambling about Computers and AI
Rambling about Computers and AI
ai-andy-gill.hashnode.dev
November 16, 2024 at 5:28 PM
I've created a blog to ramble on about Domain Specific Languages, Functional Programming, AI and LLMs, and other things. Pretty sparse right now, but I hope to populate it over time.
ai-andy-gill.hashnode.dev
ai-andy-gill.hashnode.dev