[Перевод] О странной фаллоцентричности модели GPT-J

TL;DR Статья посвящена находкам, описанным в моих постах Mapping the Semantic Void, часть I и II . Создав специальный эмбеддинг в центроиде токенов (векторе средних значений всех 50257 эмбеддингов токенов GPT-J ), при помощи промта приказав модели определить его и учтя логиты, можно создать «дерево определений» состоящее в подавляющем большинстве из туманных сформулированных неопределённостей. Это вряд ли может удивлять, ведь модели GPT-J, по сути, дают задачу определить «что-то среднее». Однако наиболее вероятная ветвь в дереве, дающая определение, содержащее что-то конкретное, определяет «призрачный токен» (ghost token) в центроиде как «мужской пенис» ( a man's penis ). Снизив уровень отсечки кумулятивной вероятности, чтобы создать длинные списки возможных определений, мы выясним, что почти все ветви, предоставляющие определения, касающиеся чего-то конкретного, связаны с сексом/деторождением, и среди них лишь время от времени встречаются связанные со статусом. Как обычно, я понятия не имею, что всё это значит, но буду рад вашим предположениям!

https://habr.com/ru/articles/796289/

#gptj #chatgpt #большие_языковые_модели #llm

О странной фаллоцентричности модели GPT-J

TL;DR  Статья посвящена находкам, описанным в моих постах Mapping the Semantic Void, часть I  и  II . Создав специальный эмбеддинг в центроиде токенов (векторе средних значений всех...

Хабр

The future of generative AI is niche, not generalized

By better bridging the gap between generative AI and more specific and niche datasets, over time people should build a subtly different relationship with the technology. It will lose its mystique as something that ostensibly knows everything, and it will instead become embedded in our context.

#artificialintelligence #AI #GenerativeAI #LLM #GPTJ #GPTNeo #technology #tech #innovation

https://www.technologyreview.com/2023/04/27/1072102/the-future-of-generative-ai-is-niche-not-generalized/

The future of generative AI is niche, not generalized

ChatGPT has sparked speculation about artificial general intelligence. But the next real phase of AI will be in specific domains and contexts.

MIT Technology Review

@oxtyped
You can play with my bot on GPT3.5-T. #GPT4 is far more impressive but it's not totally public yet - I'm coding 5x faster with it and troubleshooting stuffs quicker than having to check in a traditional search engine. Concerning #OpenAI I'm not really happy with the decisions they are taking today and was following them since the Beta of #GPT3. I'm interested about other models that I can run offline in case something happens like #gptj or the infamous leaked #llama. 😉

@kikobar @eob

#Cerebras-#GPT vs #LLaMA #AI Model Comparison | #LunaTrace

"Is Cerebras-GPT worth using?
Based on the data above it's not really better than any existing #OSS models so it's hard to say if it's a better choice than #GPTJ, #GPTNeoX, or other AI models for any tasks."

https://www.lunasec.io/docs/blog/cerebras-gpt-vs-llama-ai-model-comparison/

Cerebras-GPT vs LLaMA AI Model Comparison | LunaTrace

Comparing the real-world performance of GPT-like AI models with public weights

How to use GPT-3, GPT-J and GPT-NeoX, with few-shot learning

GPT-3, GPT-J and GPT-NeoX are very powerful AI models. We're showing you here how to effectively use these models thanks to few-shot learning.

GPT-J is a self-hosted open-source analog of GPT-3: how to run in Docker...

Learn how to setup open-source GPT-J model on custom cheapest servers with GPU. Try to run the text generation AI model of the future and talk to it right now!

Default Editor

The technique works with any causal LM, as long as it was trained to accept arbitrary text fragments (not necessarily starting at sentence or document boundary), which happens to be how large #GPT-like models (#GPT2, #GPT3, #GPTJ, ...) are usually trained.

The main trick is in realizing that the necessary probabilities can be computed efficiently by running the model along a sliding window. 🧵3/4

GPT3, chatGPT, content seems to be the holy grail of so many agencies nowadays. And with all the uncertainties how Google and other search engines might react - how are you guys checking if the content (a new) text agency delivered is 100% of human mind, 50% assisted by AI or even totally AI created? Which tools, scripts, ideas are you using to ensure highest possible quality you get for the good money you paid?

Any suggestions?

#gpt3 #chatgpt #ai #aicontent #GLaM #bloom #gptj #goose #jasper

Really great announcement from Together. It is amazing how small models can outperform gigantic ones on specific tasks with a bit of training. #GPTJ might also be a little bit special in this regard.

https://www.together.xyz/blog/releasing-v1-of-gpt-jt-powered-by-open-source-ai

#MachineLearning #ml #deeplearning

Releasing v1 of GPT-JT powered by open-source AI — TOGETHER

With a new decentralized training algorithm, we fine-tuned GPT-J (6B) on 3.53 billion tokens, resulting in GPT-JT (6B), a model that outperforms many 100B+ parameter models on classification benchmarks.

TOGETHER
In diesem Jahr fand die Ars Electronica nach der Pandemie-Pause wieder in persona statt – ein Rundgang über das elektronische Kunst-Festival in Österreich.
Missing Link: Welcome to Planet B - Ars Electronica, ein Festival in Bewegung
Missing Link: Welcome to Planet B - Ars Electronica, ein Festival in Bewegung

In diesem Jahr fand die Ars Electronica nach der Pandemie-Pause wieder in persona statt – ein Rundgang über das elektronische Kunst-Festival in Österreich.

heise online