Everyone is welcome to Star us on GitHub:
Distributed full-link causal learning system OpenASCE: https://github.com/Open-All-Scale-Causal-Engine/OpenASCE
Large model-driven knowledge graph OpenSPG: https://github.com/OpenSPG/openspg
Large-scale graph learning system OpenAGL: https://github.com/TuGraph-family/TuGraph-AntGraphLearning
Hot topics in this issue:
Apple launches new iPad Pro, equipped with latest AI chip
The powerful MoE language model DeepSeek-V2 is officially open source, with 1 million tokens, and its performance is close to GPT-4-turbo
OpenAI SearchGPT may be launched next week
LSTM author releases new LLM architecture intended to replace Transformer
Copilot Chat is now fully available on GitHub mobile
...
01. Apple launches new iPad Pro, equipped with the latest AI chip
Apple has launched the latest versions of its iPad Pro and iPad Air tablets, as well as a new Apple Pencil Pro. At the heart of the new iPad Pro is Apple's new custom M4 processor, which is four times more powerful than existing iPad Pro models.
M4 is another leap forward for Apple's chips. It is built on the second-generation 3nm process and has better energy efficiency. It also has a new display engine that can bring better accuracy, color and brightness to the screen.
The new neural engine equipped with the M4 chip is almost tailor-made for AI capabilities, and its computing power is increased by 10% to 15% compared to the previous generation M3. The neural network engine is mainly responsible for the machine learning part, such as face unlocking, image processing, etc., and products equipped with M4 chips will greatly improve execution efficiency in these applications.
Learn more:
https://www.cnn.com/2024/05/07/tech/apple-unveils-new-ipad-lineup/index.html
02. The powerful MoE language model DeepSeek-V2 is officially open source, with 1 million tokens, and its performance is close to GPT-4-turbo
DeepSeek has open sourced a powerful mixed expert (MoE) language model DeepSeek-V2. DeepSeek-V2 achieves stronger performance while saving 42.5% of training costs, reducing KV cache by 93.3%, and increasing maximum generation throughput by 5.76 times. The model performance of DeepSeek-V2 is very impressive: it exceeds GPT-4 on the AlignBench benchmark and is close to GPT-4-turbo; it is comparable to LLaMA3-70B and better than Mixtral 8x22B in MT-Bench; it is good at mathematics, code and reasoning .
In terms of price, the DeepSeek-V2 API is priced as follows: input per million tokens is US$0.14 (approximately RMB 1) and output is US$0.28 (approximately RMB 2, 32K context). Compared with GPT-4-Turbo pricing, the price is only Nearly one percent of the latter.
Learn more:
https://stanforddaily.com/2024/04/25/openai-ceo-sam-altman-talks-ai-development-and-society/
03. OpenAI SearchGPT may be launched next week
OpenAI is about to launch the search engine "SearchGPT", known internally as "Sonic". In addition to text search, it will also provide image search, weather, calculator, time zone and other widgets and prompts for follow-up questions. This is based on leaked screenshots and information from web developer Tibor Blaho.
"SearchGPT" will display the results in a chat window similar to ChatGPT and summarize the web page in up to 300 characters. According to Blaho, the language model used by SearchGPT is GPT-4 Lite, GPT-4 or GPT-3.5.
Learn more:
https://the-decoder.com/openais-searchgpt-might-be-shown-next-monday-together-with-gpt-4-lite/
04. The author of LSTM releases a new LLM architecture intended to replace Transformer
Sepp Hochreiter, the inventor of LSTM, released a new LLM architecture: xLSTM. The xLSTM architecture is based on the traditional LSTM and introduces exponential gating with memory hybrid and a new memory structure. It performs well in language modeling compared to state-of-the-art methods such as Transformers and State Space Models. Larger xLSTM models will become serious competitors to the large language models currently built using Transformer technology. Additionally, xLSTM has the potential to impact various other deep learning areas, including reinforcement learning, time series prediction, and physical system modeling.
Learn more:
**05.**Copilot Chat is now fully available on GitHub mobile
GitHub announced on Tuesday that Copilot Chat, its AI chat interface for asking coding-related questions and code generation, is now generally available on its mobile app. Mario Rodriguez, GitHub's newly promoted senior vice president of product, said that the mobile app is very popular and can perform tasks such as star repos and some of GitHub's social features. Developers are also using the mobile chat feature to ask questions about specific repos on the go. The problem.
Learn more:
https://techcrunch.com/2024/05/07/copilot-chat-in-githubs-mobile-app-is-now-generally-available/
Article recommendations
From attention mechanism to RLHF, a must-read list for getting started with large model technology
If there is any infringement, please contact us for deletion.
Follow us
OpenSPG:
Official website: https://spg.openkg.cn
Github: https://github.com/OpenSPG/openspg
OpenASCE:
官网：https://openasce.openfinai.org/
GitHub：[https://github.com/Open-All-Scale-Causal-Engine/OpenASCE]