Model Releases and Updates
OpenAI GPT-5 removes temperature and top_p hyperparameters
OpenAI's latest GPT-5 model API has removed hyperparameters such astemperature and top_p. Anthropic also no longer supports using top_p and temperature together, and developers need to adjust their code to handle these cases.> Links: OpenAI docs | Anthropic migration guide
Anthropic releases Claude 4.5 with improved tool calling
Anthropic launched Claude 4.5 with more reliable tool calling and longer context support, while deprecating the combined use oftop_p and temperature; default parameters are recommended.> Link: Anthropic announcement
MiniMax M2 launch: better price-performance than Claude Sonnet
MiniMax released its 230B-parameter M2 model at just 8% of Claude Sonnet's price with 2x faster inference, supporting tool calling and long context. Free trials are available on OpenRouter.> Links: OpenRouter trial | MiniMax announcement
Tahoe AI releases 3B Tahoe-x1 genomics model
Tahoe AI's 3B-parameter Tahoe-x1 model, trained on 100M samples for gene, cell, and drug representations, achieved SOTA on cancer benchmarks and is open-sourced.> Link: Tahoe AI announcement
Nvidia releases Nemotron Nano 2 VL vision model
Nvidia's Nemotron Nano 2 VL supports image understanding and multimodal reasoning, ships with an 8M-sample CC-BY-4.0 dataset, and is available on Hugging Face.> Link: Nvidia announcement
Moonshot Kimi K-2 Coding Plan going international
Moonshot AI's Kimi K-2 Coding Plan, currently China-only, will launch internationally soon, supporting code generation and debugging.> Link: Moonshot Discord discussion
Company and Product News
OpenAI API biometric requirement sparks controversy
OpenAI is reportedly asking long-term API users to provide biometric identification (e.g., iris scans), raising privacy concerns; users noted Anthropic and Google have no similar requirements.> Link: OpenAI discussion
Anthropic removes Claude Code feature over MCP security issue
Anthropic removed a feature from the Claude Code web feature after discovering an MCP protocol security vulnerability; developers should watch for follow-up patches.> Link: GitHub PR
Modular Mojo goes open source, GPU support tiered
Modular announced it will open-source Mojo and MAX, with GPU support tiered (Tier 1 datacenter, Tier 3 consumer). AMD consumer cards require manual coding; Apple Silicon requires reverse engineering.> Link: Modular Discord discussion
Nextdata OS launches with Data 3.0 vision
Nextdata released Nextdata OS, a semantic-first, AI-native data operating system replacing traditional pipelines, with unified management of structured and unstructured data.> Link: Nextdata announcement
Users cancel Manus.im subscriptions, switch to Claude
Users are cancelling Manus.im subscriptions due to fast credit consumption, switching to Claude ($20 completed 3 projects). Manus is seen as suitable for users who prefer not to do research, while Anthropic offers fuller capabilities.> Link: Manus Discord discussion
Research and Tools
Triton performance gap on T4 vs A100
Users report Triton matrix multiplication examples run extremely slowly on Colab T4 (sm75) but normally on A100 (sm80), because T4 lacks Tensor Cores; autotune settings need adjustment.> Link: Triton notebook
CUDA compilation tips: nvcc -dryrun and -keep
Usenvcc -dryrun to understand the compilation process and -keep to retain intermediate files (.ptx/.cubin), which can be manually edited and recompiled to optimize performance.> Link: CUDA best practices
MCP registry confusion clarified
Users confused modelcontextprotocol/registry with github/mcp; GitHub will integrate the former as upstream and mirror content. Developers can self-publish to the OSS registry.> Link: MCP docs
tinygrad FSDP implementation and contributions
Discussion on tinygrad's FSDP bounty; newcomers are advised to start with low-value bounties and mind Pyright type checking—real type issues have been found.> Link: tinygrad bounties
DSPy beats LangChain for structured tasks
Developers report DSPy outperforms LangChain on structured tasks and model upgrades: LangChain requires rewriting prompts on model upgrades, while DSPy supports optimization. A Bay Area Meetup is upcoming.> Link: DSPy Meetup
Pyright type checking in tinygrad
Pyright surfaced real type issues in tinygrad; tasteful fixes are encouraged to improve code quality.> Link: tinygrad PRs
Community and Events
- Bay Area DSPy Meetup on November 18: an in-person meetup on model optimization and applications. Registration
- GPU MODE irl-accel-hackathon: projects include a Mini-PyTorch GPU allocator and Monarch contributions; proposals required. Form
- EleutherAI on open-source AI: users want distributed open-source AI like the internet, but technical challenges remain; the Petals project has lost momentum with new architectures. Discussion
- Nous Research on LLM ideological bias: discussion of ideological bias in Western GPT models, Claude's more "infant-like" behavior, and KBLaM vs RAG context quality research. Discussion
Hardware and Infrastructure
RTX 5090 slow with tinygrad
Users report slow tinygrad performance on RTX 5090 (12 images at 512x512); adding.contiguous() is suggested as a fix, and an issue should be filed.> Link: tinygrad issues
TinyBox hardware questions
Users asked whether the TinyBox motherboard supports 9005 CPUs, 12 DIMM slots, and 500W CPUs; no clear answer yet, plus a question on whether the Discord bot is open source.> Link: TinyBox discussion
GPU bandwidth testing and optimization
Testing found 256 threads/block with plain data types performs best; vectorized types are slower, andclearL2 should be set to false to avoid negative bandwidth readings.> Link: GPU MODE discussion
---
*Source: Easy AI educational project*