4chan-Trained AI Outperforms Base Model as NVIDIA Backs OpenAI’s Funding Round
New Models & Research
Assistant_Pepe_8B Outperforms Base Model After 4chan Fine-Tuning: A fine-tuned model, Assistant_Pepe_8B, built on Nvidia’s Nemotron, showed unexpected performance improvements after training on 4chan data, challenging the notion of an "alignment tax" where fine-tuning typically reduces general intelligence. The model also exhibited shifts in political alignment and reasoning capabilities.
vllm-mlx Optimizes AI Inference on Apple Silicon: New research reveals that vllm-mlx achieves 21% to 87% higher throughput than llama.cpp on Apple Silicon, showcasing significant efficiency gains for AI model deployment on Apple hardware.
Industry & Investments
NVIDIA Confirms OpenAI Investment, Denies $100B Stake: NVIDIA CEO Jensen Huang confirmed the company’s participation in OpenAI’s latest funding round but clarified the investment will be far below the rumored $100B, following reports of prior tensions between the two firms.
Market & Economic Impact
AI Advancements Disrupt Software Loans: The rapid progression of AI models (e.g., Anthropic’s Claude) is rendering traditional software obsolete, triggering a loan crisis in the sector. Companies like Cloudera and Rocket Software face declining loan prices amid the AI-driven shift.
Company Strategies & Challenges
Anthropic Struggles to Balance Safety and Scaling: Anthropic is navigating tensions between its AI safety-first ethos and commercial pressures to compete with rivals like OpenAI and Google, risking slower growth despite its principled stance.