[AINews] not much happened today • ButtondownTwitterTwitter

This section provides a detailed recap of the recent activities and updates in the world of AI, as shared on Twitter and Reddit. It includes updates on AI model developments, research findings, tools and platforms, safety and regulation discussions, as well as some humorous takes on AI and software development practices. The Twitter recap covers various topics such as advancements in math models, pricing updates from Google AI, bug bounty programs at Anthropic, and tutorials on fine-tuning AI models. On the other hand, the Reddit recap discusses specialized AI models for mathematics, challenges faced in implementing certain functions, and the use of YAML over JSON for better efficiency. Overall, the summaries capture the essential discussions happening in the AI community.

Discussions continue around advancements in AI models and performance benchmarks. The community explores the capabilities of different models and their potential limitations. Notable topics include Gemma 2 being compared to LLaMA and Mistral, the release of Replete-LLM-Qwen2-7b with impressive capabilities, debates on model benchmarking accuracy, discussions on continuous batching for model optimization, and concerns about Flash Attention 3 compatibility. The AI community remains focused on improving and understanding the capabilities of various AI models.

In the Stability.ai (Stable Diffusion) section, users discuss optimizing VRAM without downgrading models, recommending face swapping tools, observing variable performance in Stable Diffusion, commissioning custom Lora models securely, and exploring live preview settings. These discussions highlight practical tips, tool recommendations, and ongoing challenges in utilizing Stable Diffusion technology.

In this section, numerous updates and discussions from different channels within the Nous Research AI Discord are highlighted. These include the introduction of the CRAB benchmark framework for multimodal agents, upcoming presentations on AI techniques, sharing of unique dinner menus, implementations of AI models in Excel spreadsheets, and discussions on model performance comparisons, SOTA claims, and benchmarking. Additionally, new model releases like Replete-LLM Qwen2-7b and positive feedback on Hermes 2 Pro are also mentioned, showcasing the active and diverse engagement within the community.

The section discusses various challenges and debates related to model performance and benchmarking. Users expressed enthusiasm about models but were skeptical about claims, emphasizing the need for benchmark validation. A debate arose over the reliability of hand testing versus standard benchmarks in assessing model performance. Discussions revolved around issues like reliability of benchmarks, multi-GPU setups, and new model releases. The importance of personal testing over relying solely on claims was highlighted, alongside considerations for power supply demands and features of new models like Qwen2-Audio.

Dreambooth LoRA Scripts Released: The team announced the release of Dreambooth LoRA training scripts for FLUX.1, which includes support for text encoder training of CLIP. They cautioned that memory requirements are quite high and urged users to check the README.### Broken Link in README: A member pointed out that the guide link from @bghira is broken in the README, prompting a quick acknowledgment of the issue. The team responded: 'thanks for the catch! will pr a fix.'### Training LoRA in bf16 Recommended: There was a discussion about whether LoRA should be trained in bf16 for consistency with the base models. One member confirmed, 'yeah I'd stick to bf16', but also noted it might work fine with fp16, referencing a GitHub fix.### Request for Model Distribution Support: A user expressed appreciation for the balanced mode in diffusers that enables running native Flux on 2x 16GB and inquired about model distribution support for Lora training. 'I know I'm stretching it,' they added, highlighting the long-term vision for the feature.### Runtime Error Encountered: A user reported encountering a RuntimeError related to shape sizing while running the training script for Dreambooth

Members discussed various aspects of LoRA training and model management. One member suggested focusing on training LoRAs instead of full models, highlighting minimal benefits of training larger architectures. Concerns were raised about loading Flux for inference and the space required for training LoRAs. Additionally, discussions included the efficient aggregation of VRAM with CUDA setup, challenges in splitting models across GPUs, exploring ONNX for model optimization, and encountering errors in device mapping and model loading. Suggestions were shared on using proper techniques with multiple GPUs and quantizing models to improve efficiency when transitioning to ONNX. The conversation emphasized the importance of model optimization, resource management, and effective device mapping in model loading.

OpenAI's Strawberry Model sparks interest:

OpenAI's new model, 'Strawberry', aims to enhance AI reasoning capabilities and tackle complex research tasks, generating significant buzz within the AI community. Sam Altman's social media hint about strawberries was interpreted as a clue towards this innovative project, igniting excitement among enthusiasts.

Comparing 3.33 and 3.4 decimals:

The comparison shows that 3.4 is greater than 3.33, emphasizing the importance of aligning decimal points for accurate assessments. This method aids in precise measurements relevant in fields like science and finance, where even small differences hold significance.

Anduril achieves a $14B valuation:

Defense tech startup Anduril Industries has raised $1.5 billion, now boasting a valuation of $14 billion, marking a significant jump from its previous $8.5 billion valuation. The company doubled its revenue to approximately $500 million, fueled by government contracts and investments from major firms.

Stuck Astronauts' return delayed:

NASA officials announced that two astronauts stuck at the International Space Station since June 2024 may not return to Earth until February 2025. The delay is due to mechanical failures with the Boeing Starliner capsule, which has raised safety concerns regarding the astronauts' journey home.

AI tools transforming medical advocacy:

Innovative companies are developing AI tools to assist with medical note analysis and help individuals manage their health. These advancements provide essential support for women dealing with breast implant illness, enhancing their understanding and healthcare experience.

Low VRAM Mode:

• Experimenting with model options can help optimize performance.

Face Swapping Tools Comparison:

• Members recommend using Rope for face swapping as it is easier to install than Roop.

Stable Diffusion Performance Factors:

• Users reported variances in sampling speeds (s/it) and concerns over performance consistency when changing model sizes.

Commissioning Custom Lora Models:

• Participants discuss finding reliable commissions for creating custom pony lora models and suggest using Civitai's bounty system for secure transactions.

Live Preview Settings Queries:

• Community members inquire about optimizing live preview settings in A1111 for image generation workflows.

DALL·E 3 Available for Free Users

• ChatGPT Free users can now generate up to two images per day using DALL·E 3.

  • This update enables users to create images for various applications, such as slide decks or personalized cards.

Image Creation Simplified

• Users can simply ask ChatGPT to create an image according to their needs, enhancing user convenience.

  • This feature makes it easier to visualize concepts and enhance personal communications.

A member expressed confusion about LangChain's capability to provide a uniform API for all LLMs, noting it works for OpenAI but not for Anthropic. Another member confirmed that while function calls may be similar, prompt modifications are necessary due to differences across LLMs. Concerns were raised regarding LangChain's struggles with LLM feature consistency and the declining community support. Members highlighted issues with Anthropic's Claude 3.5 downtime, the disconnect between Discord discussions and official product announcements, and frustration with LangChain's tool and documentation. The user seeks better insights for implementation strategies and commercial support for assistance with the platform.

Modular's License Permissiveness Under Scrutiny:

A member highlighted that Modular's license for using max/mojo is permissive unless you're attempting to commercialize an AI infrastructure platform. Members are questioning what happens if Modular expands into other domains, such as robotics or AI labeling platforms.

Non-Competitive Software Could Become Competitive:

Discussion revealed that if software is not competitive today, but becomes competitive in the future, it remains non-competitive under Modular's licensing agreement. However, questions arose about whether development on such software must be frozen once it turns competitive.

Call for Triton Lang Custom Kernel Users:

A request was made for Triton lang users who have written a custom kernel to participate in a one-on-one conversation with the product team. Incentives include receiving some Mojo swag for their contributions.

Initial Awareness of Triton Language:

A member expressed curiosity, noting it was their first time hearing about Triton. This indicates a potential interest in expanding knowledge about newer languages and technologies within the community.