[AINews] Life after DPO (RewardBench) • ButtondownTwitterTwitter

The section discusses various updates and news in the AI space from May 28, 2024. Highlights include X.ai raising $6b, ICLR recap, and a project called LlamaFS going viral. The section also delves into terms like RLHF (Reinforcement Learning from Human Feedback), DPO (Direct Preference Optimization), and RewardBench, providing insights into their significance in language models. Future directions for alignment research are outlined, emphasizing the need for more data, improved DPO techniques, and personalized language models. The RewardBench paper is mentioned, highlighting challenging reward model benchmarks and the success of Reward Model focused Llama 3 8B models over other competitors on the leaderboard.

AI Twitter Recap

• xAI Raises $6 Billion at $24 Billion Valuation: xAI raised $6 billion, valuing the company at $24 billion. Speculations on spending and comparisons to other AI companies were discussed.
• Criticism of Elon Musk and xAI: Yann LeCun and others criticized Elon Musk and xAI for various reasons like vengeful politics, hype, and conspiracy theories.
• AI Safety and Existential Risk Debate: Discussions revolved around counterarguments to AI doomerism, rebuttals to regulation proposals, and debates on general intelligence.
• Developments in AI and Robotics: Updates on various AI and robotics developments, including Naver Labs Robot Cafe, Microsoft's AI announcements, Tokyo Robotics demo, and other innovations were shared.
• New AI Research Papers: Grokked transformers, stacking transformers, automatic data curation, Meteor, and AutoCoder were among the research topics discussed on Twitter.
• Debates and Discussions: Debates on math skills vs verbal skills, mechanical interpretability, and sentience and consciousness in AI systems were highlighted.
• Miscellaneous: Other interesting topics such as a history project using LLMs and AI, discussions on Rabbit AI, Nvidia's rise, the Metaverse hype cycle, and developments in the LLaMA ecosystem were also covered.

A summary of different discussions and developments in various AI-focused Discord channels is presented. Challenges in fine-tuning models like Llama 3 and Mistral, advancements in multimodal models and integration, open-source AI projects, new model releases, benchmarking, ethics, legislation, AI's societal impact, AI art, AI-generated content, memes, and humor were highlighted. Specific details and updates from HuggingFace, Perplexity AI, Stable Diffusion, Unsloth AI, and Nous Research AI Discord channels were discussed.

Nathan Lambert and the community delve into various AI-related topics, from tax billing insights to AI single-mindedness experiments. Discussions also touch on Google's AI missteps and dataset misconceptions, shedding light on the diverse landscape of AI developments and challenges.

The OpenAI community Discord channels were overflowing with engaging discussions and updates. From advancements in AI models like GQA in CMDR models and VRAM efficiency to scholarly achievements and community support, the Dialogues in various Discord servers revealed a dynamic exchange of ideas and insights within the AI community.

The section covers various discussions on new models like Zyphra Zamba, SD Audio 2.0, and regulatory concerns in AI companies. It also delves into the challenges members face in tech tests, like time limits and compiler crashes. Additionally, the community explores topics related to bitwise operations, research bounties, and scheduling optimization in models. The ongoing debate over AI companies' regulation and the release of a new textbook on reinforcement learning highlight the diverse range of conversations in the community.

What's the best way to prompt the Llama 2 chat models effectively? This section delves into the various conversations around LLM finetuning, including humorous comments, fulfillment of conference recording requests, utilizing Modal for PyTorch code, dataset management in Modal, documentation examples, leveraging Modal for Kaggle competitions, and distributing HF credits. Members also discuss the best model for Spanish text generation, upcoming announcements on credits distribution, and more. Various links and resources are shared throughout the conversations providing insights and guidance for fine-tuning models.

The section discusses various topics related to LLM finetuning, particularly focusing on the Axolotl configurations and clarifications. Discussions include the merging of the latest axolotl and llama 3 demo, seeking dataset templates and pre-processing issues, default config values, template-free prompt construction confusion, and the importance of debugging tools and callback functions for logging model predictions.

• Protein Dataset Gets Major Updates: Updates on a protein visualization project, examples for human hemoglobin and more.
• Transcription App with OpenAI's Whisper Rocks!: Introduction to a transcription app utilizing OpenAI's Whisper.
• Call for Feedback on Decentralized Internet Infra: Request for feedback on a project building infrastructure for a decentralized internet.
• 3D Model Visualization in Browser Challenges: Efforts to find a solution for 3D model rendering of protein structures.
• SimpleTuner Bug Fixes Improve Training: Fixing bugs in SimpleTuner enhances its training performance.

This section discusses various topics related to LLMs and AI models. A member shared updates on creating scripts using llama.cpp to handle function calls and model responses. The Hermes model was praised by the same member. Another member sought resources for running Llama3 LoRA on a 3080 GPU. Additionally, a new developer introduced themselves, expressing interest in Mistral v0.3 and seeking advice on fine-tuning models for tool-calling.

JEPA vs LLMs Spark Debate:

A discussion unfolded about JEPA's potential to lead to AGI as proposed in 'A Path Towards Autonomous Machine Intelligence'. Members criticized the model for being similar to existing models like GPT and DINO but in different domains, with skepticism about its scalability and context handling.

ROPE's Influence on Long-Term Context:

A new approach to RoPE was discussed, suggesting limitations regarding context length capabilities in LLMs. A paper proposed a novel understanding of RoPE's long-term decay properties.

Modula: A New Training Strategy:

An interesting project called Modula was shared, which introduces scalable neural network training through automatic normalization using the modular norm.

Chameleon Model Insights:

The Chameleon model, capable of multimodal tasks like text and image generation, was highlighted for its state-of-the-art performance in multiple domains.

Bitune Enhances LLM Instruction-Tuning:

Bitune, a novel approach for improving instruction-tuning in LLMs through causal and bidirectional attention, was discussed, claiming significant improvements in zero-shot performance across various reasoning tasks.

LM Studio ▷ #🎛-hardware-discussion (5 messages):</h3>

Mojo Nightlies and Bitpacking Alignment Issues:

• Discussion on frequent releases of Mojo nightlies, currently at version 2024.5.2414, with shared links to changelogs and community meetings.
• Alignment issues with bitpacking affecting storage of bool values in memory, leading to workarounds and bug documentation.

OpenAI AI Discussions:

• Conversations on running LLMs with Nvidia A40 GPUs, features of Microsoft Copilot+ PCs, AI model water consumption concerns, AI empowerment through iterative work, and comparing GPT-4 capabilities with GPT-3.5.

LangChain AI General Discussions:

• Topics include using a CSV agent in LangChain, integrating agents into Sequential Chains, customization of CSV Agent output key, adding memory to Sequential Chains, and issues with SQL agents handling multi-table queries.

LangChain AI Share Your Work:

• Showcasing advancements in AI technology on Twitter, release of everything-ai v2.0.0 with new features, Visual Agents flow engineering platform demos, and EDA GPT demo.

LAION General Discussions:

• Discussions on Pirate Bay's role in AI, Japan's stance on AI training, controversies over model techniques, human preference study with Ella-SDXL, and critique of artifacts in AI-generated images.

LAION Research Discussions:

• Share of a major research paper mapping the workings of Claude 3 Sonnet, debate on using AI concept activations as an ad product, reflections on AI model progress, and discussions on sparsity in neural networks.

• Hugging Face Leaderboard blogpost shared: A post by Clementine, running the HF OSS Leaderboard, was shared. It delves into LLM evaluation practices and the significance of leaderboards and non-regression testing (Hugging Face blog).
• Website poisoning works on Google's AI overviews: A link to a revelation by Mark Riedl about a website poisoning attack that affects Google's AI overviews (X post). This led to further discussion on using custom search engine browser bypasses to avoid such issues.
• Thomas Dohmke's TED Talk on AI in coding: Members discussed Thomas Dohmke's TED Talk on how AI is lowering the barriers to coding. There were mixed feelings about its current reliability, but acknowledgment that UX improvements allow quicker workarounds for issues.

Twinny + LM Studio blow minds as local co-pilot:

A user shared their positive experience using Twinny with LM Studio as a local co-pilot replacement. They asked about running this setup via llamafiles and received confirmation that running two llamafiles at the same time is possible by assigning different ports.

Embedding images with llama.cpp endpoint confusion solved:

A member asked if the llamafile/llama.cpp server supports images in llama embeddings and shared a command that did not work as expected. They later clarified that the /v1/embeddings endpoint does not accept image_data but using the /embedding endpoint works as expected.

Running continue.dev with llamafile performance issues:

Another user reported running continue.dev with llamafile, noting it was slow on a Mac M2 but somewhat faster on an older Nvidia GPU.

Inquiries on building and training custom LLMs:

A member sought advice on building and training a custom LLM using company documentation for internal use. They received a recommendation to use HuggingFace Transformers for training, noting that llamafile only supports inference.

This section discusses various performance and benchmark-related discussions within the Mojo community. It includes topics such as the impact of different implementations on performance metrics, issues with SIMD-based implementations, and challenges faced with larger byte cases. The community also explores improvements in performance through the use of decorators and potential memory reuse issues. Overall, these discussions offer insights into optimizing performance and addressing challenges in Mojo development.

The conversations in the 'Cohere general' channel covered various topics. Members discussed the functionality and comparisons of Aya-23, particularly in terms of multilingual capability and specialized training. Another user sought advice on developing a RAG mobile app with on-phone LLMs for privacy reasons. The channel also explored system prompts in Aya-23 and how users modified Command R prompts to work effectively. Additionally, there was clarification on the availability of Aya-23-35b for non-commercial use only. Various links were shared including GitGud and documentation for migrating from Cogenerate to Co.

Members of the OpenInterpreter community engage in various discussions related to the Large Action Model (LAM), running models on mobile devices, storage solutions for model data, installation queries for Open Interpreter, and a new markdown export feature:

Networked Llamafile Server Tips:

• Members discussed making the llamafile server available across a network with tips like adding --host <my ip> or using --host 0.0.0.0. This makes the server accessible from different machines on the same network.

Unexplained Blank Responses from Llama3-70B:

• Users reported blank responses from the llama3-70b model and sought help by sharing logs for troubleshooting. Another user stepped in but didn't have a direct solution, indicating it might require deeper investigation.

Release of Llamafile v0.8.5 and Benchmarks:

• The community celebrated the release of llamafile version 0.8.5, highlighting that it now offers fast inference for K quants on X86 CPUs. Members were encouraged to join a benchmarking club to test and share results using llamafile-bench.

Home Assistant Integration Wish List:

• Home Assistant integration feedback highlighted the need for a standardized local API similar to OpenAI’s, suggesting names like Apilla and noting features like API discoverability via DNS-SD/zeroconf and secure APIs as desirable.

Model Selection in Python Example:

• Questions arose regarding specifying models in the Python example for LLaMA_CPP integration, with users sharing snippets and seeking clarity on whether model specification is necessary when running instances like TinyLlama.

AI Stack Devs (Yoko Li) ▷ late-night-lounge (8 messages🔥):

• A member explored the SadTalker GitHub repo for animations and shared it with the community, offering help if needed. They also recommended checking Hugging Face Spaces for running SadTalker locally. Another new animation tool from Tencent AI Lab, V-Express, was discussed for generating engaging head videos.

Interconnects (Nathan Lambert) ▷ news (4 messages):

• Zyphra Zamba was released quietly, with resources like a tech report and Torch reference code provided. A comparison with OLMo 1.7 is in progress, and SD Audio 2.0 reportedly leaked on 4chan. Links to related information were shared.

Interconnects (Nathan Lambert) ▷ ml-drama (4 messages):

• Ex-OpenAI board members call for AI regulation due to concerns about profit incentives and safety protocols. Allegations against Sam Altman regarding toxic culture were discussed.

Interconnects (Nathan Lambert) ▷ lectures-and-projects (10 messages🔥):

• Discussions included uncertainty about the 224n class timeline, a reinforcement learning textbook launch, and praises for Chris Manning and Chris Potts.

tinygrad (George Hotz) ▷ general (11 messages🔥):

• Topics ranged from test time limit extensions to struggles with large expressions and switching focus to different bounties. Incompatibility of doubles and bitwise operations, interest in bounties, and PR status on tinygrad were also discussed.

tinygrad (George Hotz) ▷ learn-tinygrad (5 messages):

• Discussions covered topics like 'vin' in the UOp class, Taylor Approximation feedback requests, and post-dominator analysis for scheduling in model optimization.