AI Truths
LLM hallucinations are an architectural problem, not a compute problem. Therefore, LLM errors and hallucinations will persist no matter how much capital LLM builders throw at the next model.
History is rhyming, not repeating. No, this is not like the early days of Microsoft and Google. MSFT and GOOGL where profitable on day 1. OpenAI and Anthropic are not profitable at the model level.
AI Agent risk is real. There is enormous risk associated with autonomous AI Agents because of the first bullet point. A rogue agent may drain your bank account, or, delete your customer database as was the case with Replit back in July.
LLM returns are diminishing. A 10x investment in LLM training no longer translates to a 10x return in LLM performance.
Rise of opensource LLMs. OpenAI, Anthropic, Grok and Google/Gemini will have to deal with low-cost, high-privacy, high performance opensource models for decades to come. This obviously puts downward pricing pressure on proprietary LLMs.