#allenai — Public Fediverse posts

@swift I mostly agree with your political view. I believe, however, that LLMs are not intrinsically 'evil'. Ethical and truly open-source LLMs are possible (and indeed, are being developed #allenai #eurollm #aperture). 1/2

@jzb I fully agree: it is certainly not the way today, with the current ecosystem dominated by a handful of large US and Chinese companies. But LLMs, as a science and technology, are not intrinsically bound to these companies. Up to us to train and develop ethical solutions. And it is not a theoretical possibility: see #allenai #EuroLLM #Aperture

Most AI is not truly open-source, even if they say they are "open", because they do not let you change the original training data, which is secret.

#Tülu and #OLMo are AI models more open than most, with available for download and modification: model weights, code, and (for OLMo) much of the training data.

They offer free AI chat, which for my coding did well, but for asking about a spiritual author it hallucinated badly, so like all #AI, is not perfect.

https://playground.allenai.org/

#AllenAI

During my intro at #Hack4Socialgood this morning, I used 🤗 Hugging Face playground to quickly show #Apertus prediction for the hackathon compared to the Olmo model from #AllenAI and #EuroLLM

Prompt and response in the ALT. Try it yourself at https://huggingface.co/playground?modelId=swiss-ai/Apertus-70B-Instruct-2509

Olmo Hybrid offers the same performance as Olmo 3 but with 49% fewer training tokens by combining transformers and linear recurrent layers, may offer better expressivity and scalability than pure models.
Released by @allenai with open weights, checkpoints, and code for community use and research #allenai #olmohybrid
https://allenai.org/blog/olmohybrid

You might have noticed that I’m rather skeptical of many current applications of “AI” (including the shift of the term to primarily mean “Generative AI”). But there is one application that I find really intriguing: The #PaperFinder https://paperfinder.allen.ai from #AllenAI 1/

#OlmOCR Runs on #macOS with #LMStudio: Simple #PDF Text Extraction 📄

#OlmOCR by #AllenAI now works on regular #macOS systems without specialized #GPU requirements using #LMStudio as inference server

🧵 👇 #OCR #ai #llm

https://arxiv.org/abs/2312.17172

Interesting new multimodal #LLM

#unified-io-2 #allenai

The session "Strategies for Matching Affiliation Strings to ROR IDs" brought together a panel of users (#OpenAlex #AllenAI #DOE-OSTI @crossref) who are working to develop innovative automated approaches to reconciling affiliation data: https://youtu.be/Tx5y7lX030U