Concordia is a library for generative agent-based modeling that works like a table-top role-playing game.
It's open source and model agnostic.
Try it today!
github.com/google-deepm...
Thinking about doing the same but from other reason.
While the cascade feels smarter than cursor it often introduces frustrating changes. Cursor might not be that smart but seems to be much more predictable.
dumber but predictable > smarter but frustrating
Still cooking my aibricks project.
Simple example of what I mean by "configuration driven":
From the same people (lmarena.ai) who brought you Chatbot Arena, they are introducing WebDev Arena.
Leaderboard: web.lmarena.ai/leaderboard
LEXICO defining the Pareto frontier of KV cache compression
Textual 1.0 has been released. 🥳
Three years in the making. A TUI framework that is bigger than the terminal.
To celebrate, I want to give away some trade secrets. Because I am appalling at keeping secrets.
Tell me what you think of the diagrams...
textual.textualize.io/blog/2024/12...
1000x inference cost reduction by converting qwen/llama to rwkv architecture without retraining from scratch.
Big if true/without-any-major-issues. Definitely worth checking.
huggingface.co/recursal/QRW...
Q-RWKV-6 32B Instruct Preview
substack.recursal.ai/p/q-rwkv-6-3...
I was also thinking about microrepos or rather increased modularization. It might improve issues with “collateral damage” but I’m anticipating more problems with proper module api usage. It seems solvable though with proper module interface docs.
Well... just today I went full caps "DO NOT TOUCH X ..." and started few sentences with "How TF ...". It was all good in the end but the road was very bumpy.
I'm actually laughing out loud 😂
Same here. Lots of tests, frequent testing, frequent commits. But it’s frustrating it wasn’t like this before.
I think that in the last ~10 days something changed in the agent part that applies the patches. The diffs can look good but applying them often results in totally different changes as if they were “interpreted” and not just applied. I often observe damage to the logic that was not in the diff.
Is this Windsurf?
The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities
arxiv.org/abs/2411.04986
entrapix — your LLM should raise a ConfusedAgentError when it doesn’t know
Why? Because as an application developer, I have a lot of things I can do to give the LLM more information
github.com/tkellogg/oll...
Leaked Windsurf #codegen prompt: www.reddit.com/r/LocalLLaMA...
Paper📰: Evaluating Language Models as Synthetic Data Generators ( arxiv.org/abs/2412.03679 )
Code💻: github.com/neulab/data-...
Data: huggingface.co/Data-Agora
It’s excellent! For me your account is top-of-mind when it comes to AI/ML news.
Please don’t mind any unfollow from my side - I’m experimenting with the atproto api.
www.bnnbloomberg.ca/business/tec...
One of the best pieces on the topic I've read this year.
Endless ruminations on this: @josiecox.bsky.social
I agree with most of this but I think that Steve is not extrapolating enough.
The future is not just about ‘better models’ it’s also about better workflow on the technical side (better agents) and on the process side (better code structure / work division).
We are still early in this.
This is how happiness looks like.
Now cooking the config layer and streaming.
Middleware for streaming might require some trial and error but hey, thats the fun part!
The results might be nice - imagine streaming but line by line instead by random number of tokens - the latency is still low and you can easily act on produced text!
I’m getting a deja vu of binary vs text formats discussions.
This 1h free course is an excellent intro to memgpt like agent architecture. No frameworks - just doing fun things from scratch.
I highly recommend this one even to people not interested in AI powered games - agentic internals are very similar everywhere.
www.deeplearning.ai/short-course...
