codelion

7 ms·

91.
by codelion 2y ago
It's a shame when potentially interesting papers don't hold up in practice. I've seen a few cases where the real-world performance didn't match the initial claims.
92.
by codelion 2y ago
Interesting analysis! I hadn't dug into the specific patch details like that. It's a good reminder that "correctness" isn't always the only dimension to evaluate these AI-generated patches – readability and idiomati
93.
by codelion 2y ago
That's a pretty cool migration story! I've been meaning to give podman a more serious look. The OCI image format issue is good to know about – hadn't considered that compatibility angle. I'm curious, did you notice any p
94.
by codelion 2y ago
Not a stupid question at all! Imo, you can definitely dive deep into CUDA and GPU architecture without needing to be a math whiz. Think of it like this: you can be a great car mechanic without being the engineer who designed the engine. Sta
95.
by codelion 2y ago
The DAG feature for subjective metrics sounds really promising. I've been struggling with the same "good email" problem. Most of the existing benchmarks are too rigid for nuanced evaluations like that. Looking forward to seei
96.
by codelion 2y ago
Models already have hidden latent CoT style reasoning within them, GRPO would help induce that behavior. For instance see https://x.com/asankhaya/status/1838375748165628053 where a sampling technique (CoT decoding
97.
by codelion 2y ago
Yeah, the 20-year limit is definitely a thing. Maybe the 2004 patent was for some specific improvement to the original pulse oximeter tech? Patent law is tricky like that.
98.
by codelion 2y ago
I think there's a balance to be struck. While users don't directly care about the specific tech, they do care about the results – speed, reliability, features. So, the stack is indirectly important. Picking the right tools (even
99.
by codelion 2y ago
Yeah, the difficulty of tracking is a huge factor. Plus, with torrenting, the "making available" part is pretty blatant. With Usenet or direct downloads, it's a grayer area unless you're running the server. I've alw
100.
by codelion 2y ago
Yeah, I agree. The "O1 preview" naming feels a bit misleading. It sets an expectation of broader coverage than just those specific benchmarks. It's cool to see cost reductions, but the marketing could be more transparent abou
101.
by codelion 2y ago
It's definitely possible to focus on the CUDA/GPU side without diving deep into the math. Understanding parallel computing principles and memory optimization is key. I've found that focusing on specific use cases, like optimi
102.
by codelion 2y ago
This is great to see! Open-sourcing infrastructure tools can really accelerate innovation in the AI space. I've found that having access to well-documented repos makes it much easier to experiment and build on existing work. Are there
103.
by codelion 2y ago
Interesting point. It's almost a form of "design by verifiability" – prioritizing architectures that lend themselves to easier reasoning, even if other architectures might offer marginal performance gains.
104.
by codelion 2y ago
Interesting work. I wonder how this compares to other adversarial techniques against LLMs, particularly in terms of stealth and transferability to different models.
105.
Show HN: OptiLLMBench – Test how inference optimization tricks scale up LLMs
2 points by codelion 2y ago|0 comments
106.
by codelion 2y ago
This is great, can this be used to implement a sampler based on entropy like entropix (implemented in optillm here - https://github.com/codelion/optillm/blob/main/optillm/entrop... )
107.
Adaptive Classification for Automatic LLM Temperature Optimization
1 points by codelion 2y ago|0 comments
108.
by codelion 2y ago
Some benchmarks showing the advantage over traditional approaches: Traditional classifier adding a new class: - Requires full retraining (~30-60 minutes on typical dataset) - Needs all historical data - Uses 2-3x more memory during training
109.
Show HN: A classifier that learns new categories without retraining from scratch
4 points by codelion 2y ago|1 comments
110.
Adaptive Classifier: Dynamic Text Classification with Continuous Learning(github.com)
2 points by codelion 2y ago|0 comments
111.
by codelion 2y ago
Some technical details about how it works: The core architecture combines a transformer model for embeddings with a prototype memory system and an adaptive neural head. When adding new classes, it uses Elastic Weight Consolidation (EWC) to
112.
Show HN: Adaptive-classifier – text classification with continuous learning
5 points by codelion 2y ago|1 comments
113.
by codelion 2y ago
The full dataset is here - https://huggingface.co/datasets/AI-MO/aimo-validation-aime you can use the eval script I have in optillm to benchmark on it - https://github.com/codelion/optillm&#x2
114.
by codelion 2y ago
We have tired a few different ways to convey what we do but it is hard. We want to refer to all software development activities that happen after the developer commits the code into a source control system. “Post-code” seemed a good way to
115.
by codelion 2y ago
>How are you dealing with structured outputs? The models have gotten much better at generating them with just the prompt. I have not implemented strict support for structured output or JSON generation yet. The response from the proxy are
116.
by codelion 2y ago
Optillm - https://github.com/codelion/optillm optillm is an OpenAI API compatible optimizing inference proxy which implements several state-of-the-art techniques that can improve the accuracy and performance of LLMs. T
117.
Entropy Decoding in Optillm and Early Results on GSM8k
1 points by codelion 2y ago|0 comments
118.
by codelion 2y ago
Just realised that I missed adding the link to the example - https://github.com/unclecode/crawl4ai/issues/126
119.
by codelion 2y ago
Correct but this is a good starting point for code that is written after the cut off of language models training data as you cannot otherwise debate accurate code form then for the newer versions of the library.
120.
by codelion 2y ago
With LLMs it is now quite easy to generate docs as needed. In fact we built a service to do just that - https://docs.codes/ Here is an example of how it is very useful especially for newer libraries.

More ›