Wrote (let gpt write) a tiny tool that scans my repos for prompt generators and allows me to quickly generate custom prompts for new apps. For example “api docs for x, 3 example files, 2 doc pages”. Really useful for 50 lines of code.

#llms

Oh geeze the #LLMs are going to be digesting their own text in publications even sooner.

https://www.nature.com/articles/d41586-023-02477-w

well i guess there’s #falcon 180b if you have an extra **400 gb** of GPU memory #llms #ai

Question for #LLMs prompt engineers (anyone):

What are the most important metrics for measuring LLM completion?

I start with some basics "system" variables:

1. Latency (or response time)
measures the elapsed time of a completion

2. Tokens Throughput = tokens/latency

3. Prompt/Completion tokens ratio

4. there are many more cross ratios maybe useful, involving LLM settings (max_tokens, temperature, etc.)

The final goal is to define some list of common vars to evaluate foundational LLMs.

Berlin, here I come!

It's now four years since the last in-person #SWIB conference. #swib23 may well be the last one, so let's try to cherish the opportunity to meet in one place!

I will be an instructor for the #Annif tutorial on Monday, moderate a session & hopefully give a Lightning talk on Tuesday. Looking forward to great talks & ofc dinner and coffee breaks are often the most interesting!

Feel free to stop by if you want to talk about #Annif, #Skosmos, #LLMs, #metadata, #Fennica etc.

Next was an amazing pair of talks by @armedchile (an ingenious method for combining high and low resource language data to build #LLMs that significantly improves translation performance) and Orevaoghene Ahia (subword tokenization effects on LLM costs/performance in different languages) at #Indaba2023. I'll be thinking about both talks for a long time - they have profound implications for the design of #GenerativeAI tech and business models. Highly recommend https://www.youtube.com/watch?v=EVi9qB_1Ccw (3/11) #AI

Another one — why not just use gzip? This paper uses tiny compression algorithms that run great on even embedded devices, and the performance comes close to where #LLMs are at. That would be a massive game changer from the current state https://www.hendrik-erz.de/post/why-gzip-just-beat-a-large-language-model

Now that #NeuralNetworks have had repeated big successes over the last 15 years, we are starting to look for better ways to implement them. Some new ones for me:

#Groq notes that NNs are bandwidth-bound from memory to GPU. They built a LPU specifically designed for #LLMs
https://groq.com/

A wild one — exchange the silicon for moving parts, good old Newtonian physics. Dramatic drop in power utilization and maps to most NN architectures (h/t @FMarquardtGroup)

https://idw-online.de/de/news820323

Furthermore, #llms are absolute monsters to slice and refactor legacy code, thus improving the longevity of already long-living software that might otherwise be painfully phased out.

2/

summarizing a train of thought from a conversation with @promovicz yesterday, which helped me formulate some of my ideas.

I think #llms allow us to write *less code*. Because they make it easy to generate the boilerplate that ensures longevity and sustainability of software projects: documentation, clean commits, unit tests, tooling, they allow projects to actually live on.

1/

I don’t think mathematica is a Great target language, nor the does prompting in the Mathematica chat enabled seem very good, and it churns through tokens like the king of Spain. My third attempt to get something going yesterday was just as much of a mess (trying to get hilbert curves computed and displayed with @defn ) as my first sessions (doing some geospatial computation / dataset computation).

Not fun.

#llms

Anthropic \ Introducing Claude Pro https://www.anthropic.com/index/claude-pro ($20/mo alternative to ChatGPT Plus) #AI #LLMs #Claude

This week's net.wars, "Small data", summarizes the talk I gave with Jon Crowcroft at this year's #gikii, arguing that large language models will ultimately prove to be a distraction: https://netwars.pelicancrossing.net/2023/09/08/small-data/ #LLMs #AI #NetWars

a while back i recall there being some tool for exploring a database of embeddings that lets you visualize and locate duplicates, etc. anyone know what it's called? #llm #llms #ai #llama2

I literally have thousands conversations like these now, from designing video game physics to transaction protocols to monad design patterns for music sequencing to numerous DSLs for everything that strikes my fancy (zine layout? comic book storyboard generation? worldbuilding CMS?).

#llms

chatgpt trick when applied to programming, don't ask "does X do Y" or "is X Y", instead ask for "give me a test program to show that X does Y".

So for example, if you wonder about the size of a struct in memory, ask it to write the program to compute or measure struct sizes.

always think meta.

#llms

For a medical & caretaking project, I experimented with combining symbolic #logic with #LLMs to mitigate their tendency to nondeterministic behavior and #hallucinations. Still, it leaves a lot of work to be done, but it's a promising approach for situations requiring higher reliability.

#ai #artificialintelligence #llm #logic #chatgpt

https://medium.com/9elements/using-symbolic-logic-for-mitigating-nondeterministic-behavior-and-hallucinations-of-llms-32aa6d10ec3c

Artificial Language Models Teach Us Nothing About #Language | Psychology Today

#Linguistics #LLMs

https://www.psychologytoday.com/intl/blog/language-and-its-place-in-nature/202309/artificial-language-models-teach-us-nothing-about

I don't know what I expected. #LLMs #AI #SALAMI #StephenHawking

Timeline History of Large Language Models - Voicebot.ai https://voicebot.ai/large-language-models-history-timeline/ (nice!) #AI #LLMs #history