AIR#331 - AGI Experiment Fails, Microrobots Show Swarm Intelligence & DeepSeek v3 Outperforms Claude 3.5!
Hey there!
Here's the latest AI news for today. Enjoy!
Today's top stories
🔥 My failed attempt at AGI on the Tokio Runtime
An attempt to create AGI using a biologically inspired neural network on the Tokio Runtime fails to achieve meaningful results.
Magnetic swarm intelligence of mass-produced, programmable microrobot assemblies
Mass-produced magnetic microrobots exhibit swarm intelligence for versatile tasks, enabling autonomous cooperation and advanced functionalities.
DeepSeek v3 beats Claude sonnet 3.5 and way cheaper
DeepSeek-V3, a cost-effective 671B parameter model, outperforms Claude 3.5 in benchmarks while ensuring efficient training.
ChatGPT Saved Chats: Save and organize your important conversations in ChatGPT
ChatGPT Saved Chats extension lets users easily save and organize important conversations for quick access.
DeepSeek-v3 Technical Report [pdf]
DeepSeek-V3 technical report available on GitHub, detailing updates and features of the AI tool.
Inverse Design of Complex Nanoparticle Heterostructures via DL on Graphs
Researchers leverage deep learning on graphs to optimize the design of upconverting nanoparticles, enhancing their emission properties.
Fine-tune classifier with ModernBERT in 2025
Learn to fine-tune ModernBERT for efficient LLM routing, achieving faster processing and improved classification performance.
Leaked Documents Show OpenAI Has a Clear Definition of 'AGI'
Leaked documents reveal OpenAI and Microsoft define AGI as AI generating $100 billion in profit, sparking debate on its true meaning.
Colab Notebook – RAG on Your Unstructured Data
Explore a Colab Notebook for implementing Retrieval-Augmented Generation (RAG) on unstructured data.
Boox devices now ship with a Chinese propaganda AI assistant
Boox devices now include a Chinese propaganda AI assistant, raising concerns about censorship and the implications of Chinese electronics.
Roasted Christmas Spam from Muhu AI
Muhu.ai's unsolicited holiday email "roasting" open source developers highlights AI's misunderstanding and exploitation of tech workers.
Azure GPT-3 got replaced by a creepy Chinese state censored model by ByteDance
Azure GPT-3 has been replaced by a censored AI model from ByteDance, sparking controversy in the tech community.
Show HN: I built an AI-powered website builder–no code required
RapidSite offers an AI website builder that creates custom sites without coding, allowing users to launch in minutes.
Show HN: Quixotic – a tool for wasting bot and LLM scraper time
Quixotic generates fake content to mislead bots and LLM scrapers, using a Markov Chain text generator for static websites.
DeepSeek-V3, ultra-large open-source AI, outperforms Llama and Qwen on launch
DeepSeek-V3, a new open-source AI model, surpasses Llama and Qwen in performance, marking a significant advancement in AI technology.
Deepseek released a new frontier LLM with 11x lesser compute than Meta's largest
DeepSeek's new LLM achieves frontier capabilities with 11x less compute than competitors, showcasing efficient resource use.
Deliberation in Latent Space via Differentiable Cache Augmentation [pdf]
A new method enhances LLMs by augmenting their cache with latent embeddings, improving reasoning task performance and reducing latency.
Nvidia's Christmas Present: GB300 and B300 – Reasoning Inference, Amazon, Memory
Nvidia unveils GB300 and B300 GPUs, enhancing reasoning model inference and reshaping the supply chain for hyperscalers.
ChatGPT search vs. Google: A deep dive analysis of 62 queries
A comprehensive analysis of 62 queries reveals Google outperforms ChatGPT search in most categories, especially local and commercial queries.
Show HN: I Built an Email Agent That Learns Your Context to Filter Emails
Ping is an email agent that learns your context to help filter important emails effectively.
Show HN: Web RAG to generate perplexity like answers from your docs [in browser]
Akira Docs introduces a web tool for generating perplexity-like answers from documents directly in the browser.
BLT: Byte Latent Transformer
The Byte Latent Transformer (BLT) innovates LLMs by using dynamic byte-level patching instead of fixed tokenization for improved efficiency.