[AINews] not much happened today • ButtondownTwitterTwitter

The AI Twitter Recap section provides insights into recent developments and discussions in the AI community on Twitter. It covers topics like AI model developments and comparisons, innovations in AI image generation, reinforcement learning and reasoning advancements, AI infrastructure and compute updates, AI in enterprises and applications, and open-source AI and API integrations. The section highlights key discussions, announcements, and critiques made by various individuals in the AI space, shedding light on trends and advancements in the field.

Theme 1: DeepSeek-R1 Runs Inference on Huawei's 910C Chips

• DeepSeek is conducting inference on Huawei's new 910C chips after training on Nvidia H800, with models like DeepSeek-R1-Distill, Qwen-14B, Qwen-32B, and Llama-8B already launched.

• Discussions highlight skepticism about the Huawei 910C chips' performance and the integration of DeepSeek models with Huawei Cloud's ModelArts Studio.

• Theme 2: DeepSeek-R1: Efficient Training Costs Explored

• The post raises questions about the $6 million cost estimate for training DeepSeek-R1, referencing claims by Alex Wang about DeepSeek's training with 50,000 H100 GPUs.

• Discussions cover the MIT License of DeepSeek, detailed cost analysis comparing training times with other models like Llama 3, and skepticism about financial strategies and infrastructure use.

• Theme 3: DeepSeek Censorship: A Comparative Analysis

• Comparisons between DeepSeek and Western censorship models are discussed, along with experiences running different models locally and the perceived advantages of OpenAI and Anthropic models.

• Theme 4: Janus Pro 1B: In-browser Multimodal AI Innovation

• Janus Pro 1B's local browser execution using WebGPU is detailed, highlighting its multimodal capabilities and supporting resources like online demos, ONNX models, and source code.

• Discussions cover the model's performance on low VRAM setups, concerns about model quality, and possibilities of using AI models for robotics applications.

The AI Discord Recap section covers various themes and developments in the AI industry discussed on different Discord channels. It includes insights on DeepSeek's impact on the AI world, challenges faced by users with tools like DeepSeek R1 and Qwen, hardware concerns, and user experiences. The section also delves into the discussions on AI reasoning models, open-source innovations, AI tools, and hardware infrastructure. From DeepSeek's cost advantages to emerging models like Qwen 2.5-Max and YuE in music generation, the recap provides a comprehensive overview of the AI landscape.

The Discord channels discussed a variety of topics ranging from workflow guide praises to credits confusion for premium subscriptions. Issues with Amazon Nova and Bedrock were quickly resolved, while deep discussions unfolded about Gemingi video integration and provider pricing topics. In Eleuther Discord, debates arose about the potential abandonment of GRPO and the cost-effective training of DeepSeek project. Gemini and Google surfaced as leaders in video and reasoning tasks, respectively. OpenRouter vs. Official API speeds were compared across Emojis, sparking discussions about free model availability and balancing token fees. Stability.ai Discord brought forwards critiques on Janus and DeepSeek models, alongside debates on AMD cards, upscalers, and tariffs. The other sections talked about NotebookLM's feedback collection, DeepSeek exposure, document dilemmas, and calls for automated citation tools. The LLM Agents section highlighted class information and updates for the MOOC, as well as discussions around YouTube edits and NotebookLM insights. The sections on Nomic.ai and Torchtune Discord covered topics like Jinja template issues, DeepSeek deployment on GPT4All, concerns about feature requests, XLSX file uploads, and the evolution of Web Search in GPT4All.

DeepSeek Delivers LlamaIndex Boost

LlamaIndex announced a first-party integration with the DeepSeek-R1 API, enabling usage of deepseek-chat and deepseek-reasoner. The recommended setup is %pip install llama-index-llms-deepseek, granting immediate access to enhanced model features.

SOFTIQ Shaves Tenders to 10 Minutes

The new SOFTIQ SaaS app uses LlamaIndex workflows to slash analysis time for public sector tenders to under 10 minutes each. This approach sharpens selection accuracy, reducing wasted work for construction companies.

LlamaReport Docs Emerge Soon

Members confirmed LlamaReport documentation is in progress and will be published soon, referencing a Twitter link for updates. They hinted at upcoming features but advised the community to stay tuned for the official doc release.

Dead Link Bites the Dust in Docs

A Pull Request removed a nonfunctional link from fine-tuning.md, which was confirmed missing from the codebase. The PR is a one-line fix that tidies up unneeded references.

RAG Retrieval & FastAPI Streams in Play

A user explored triggering RAG retrieval within reasoning model steps, citing the Search-o1 paper. Others recommended streaming with an async generator in FastAPI, then injecting retrieval results back into the ongoing response.

A member reported issues with using sonar and response_format together, resulting in invalid JSON in Markdown format. Switching to sonar-pro resolved the problem, but concerns were raised about its higher cost. The cost difference between sonar and sonar-pro was also discussed, highlighting the financial burden posed by the necessity of using sonar-pro for stable operations.

The Nous Research AI General section discusses the recent launch of Nous Psyche and its impact, confusion over DeepSeek model pricing, comparisons between AI reasoning and human logic, practical business applications of AI, and the need for a deeper understanding of AI mechanisms. Participants also dive into feeding content to models, impersonating authors, using AI for advanced search, training costs, and evaluating model responses.

The LM Studio hardware discussion covers various topics related to hardware performance and requirements for running large models. Users reported achieving expected token speeds with DeepSeek-R1 32B models and resolving GPU detection issues in LM Studio by switching to the CUDA runtime. Discussions highlighted the impact of SSD NVMe speeds on model loading and the necessary specs for running the 70B DeepSeek R1 model, including a minimum of 30GB VRAM. Users also discussed the performance limitations of memory bandwidth on Apple devices compared to dedicated GPUs like the RTX 3060 and above.

The section discusses updates related to the OpenRouter platform and Eleuther community. For OpenRouter, there are announcements regarding operational issues with Amazon Nova models and Bedrock, Gemini Video Support implementation, comparative model speed discussions, OpenRouter's flexibility with providers, and insights on free model offerings and pricing. On the Eleuther side, topics include concerns about GRPO usability, DeepSeek's affordable training strategy, discussions on LLM reasoning capabilities, finding remote job opportunities in LLM research, and exploring neuroscience insights on AI. The Eleuther section also delves into GRPO's momentum matrices, model-based reinforcement learning, Muesli method comparisons, the YuE Music Generation Model, and privileged bases in Transformers.

• Janus Pro paper and reception: A reminder for the community to manage expectations in a competitive landscape.
• Speculation on Meme Coin Launch: Anticipation for a meme coin launch based on casual confirmations.
• Qwen2.5-Max Model Launch: Qwen team introduces Qwen2.5-Max with performance improvements highlighted through benchmarks.
• Positive Recognition for Joshua: Joshua receives praise for his character, emphasizing positive community interactions.
• Interconnects (Nathan Lambert): Various discussions on DeepSeek's adoption, Qwen 2.5-Max performance, AI job market shifts, influencer impact, and tackling Groq limitations.
• Links mentioned: Include tweets and mentions related to the ongoing discussions.

This section delves into various discussions surrounding different deep AI models and their capabilities. Users express mixed reviews on the Janus model, highlighting concerns about its image generation abilities, particularly the 7B model. Recommendations for Stable Diffusion setup on AMD cards and hardware choices for AI tasks are provided. The debate on RAM vs VRAM for AI builds sparks differing opinions. There is also a mention of upscalers' consistency and preferences, along with a comparison of the Deepseek model with other LLMs. The section also includes discussions on the Deep AI model Deepseek, its pricing, and performance, as well as other AI-related topics like Rax, Qwen 2.5-Max, and Open Source AI developments.

In this section, various features and updates related to LlamaIndex are discussed, including the integration of the DeepSeek-R1 API, which enhances functionalities with models like deepseek-chat and deepseek-reasoner. Users can easily install LlamaIndex to leverage these features. Additionally, the SOFTIQ SaaS app, utilizing LlamaIndex Workflows, has significantly reduced analysis time for public sector tenders to under 10 minutes each. This efficiency improvement aids construction companies in accurate selection while saving time and effort. The section also mentions the forthcoming LlamaReport documentation, with preparations underway for its publication in the near future.

The team reported that the docs were down but are now back up and available, including the GPU package API documentation in nightly. Members expressed gratitude for the quick resolution and appreciated the update on the documentation status. A member commented that Deepseek seems to have outperformed Modular in achieving similar goals with Max and Mojo. However, others argued the two are fundamentally different, likening Modular's role to that of a tractor store supporting farmers rather than competing against them.

The epilogue section encourages readers to stay updated by subscribing to AI News. A subscription form is provided where readers can enter their email and subscribe. The footer also includes links to the AI News Twitter page and newsletter. The section is brought to you by Buttondown, a platform to start and grow newsletters.