Technical Support
January 28, 2025
Technical conditions support a tactically bullish near-term outlook. The SPX held near-term technical support at ~5959 yesterday, while the CBOE Volatility Index (VIX) backed off intraday highs to close at subdued levels <20.
The ‘Chatbot Arena’ leaderboard is supposed to offer an unbiased assessment of AI outputs (blind ranking by users based on responses). Google’s Gemini and GPT/GPTo1 models have consistently been among the top two lately with DeepSeek’s R1 technically tied for third and its V3 model in eleventh place.
The strong reasoning performance of DeepSeek’s R1 model at a lower cost is attributed to deliberately skipping the supervised fine-tuning stage. Instead of using human feedback or AI feedback, R1 utilized Reinforcement Learning (RL) in a rules-based format. According to the company, the model ‘self-evolves’ during the RL stage, improving its reasoning capabilities along the way by adjusting responses in pursuit of better accuracy in the correct format. Unfortunately, DeepSeek announced R1 training cost of ~$6M doesn’t include overhead, capex or the cost of previous runs, and it’s difficult to assess the apparent accomplishment without knowing the true all-in costs and the model’s reliance on existing LLMs for training.
Lower training and inference costs should eventually drive increased AI demand/greater proliferation, which should result in increased demand for semiconductors, networking equipment and other forms of AI infrastructure. This may seem counterintuitive in the moment, but improvements in compute efficiency/lower costs have historically driven increased adoption.
Read more |