This is my uneducated (in llm training) speculation. Anthropic is proving that focusing on alignment first and foremost, baked into every step, is not only the ethical choice -- it is simply the greater product. My frustrations with codex and other models is that for long horizon multi turn tasks is that the llm loses the plot of what we're doing. Alignment is what makes good employees, good team members, good contributors, good managers, good friends etc. Species survival as a bonus, alignment + intelligence beats raw intelligence for value and benefit across the board.
This is my uneducated (in llm training) speculation. Anthropic is proving that focusing on alignment first and foremost, baked into every step, is not only the ethical choice -- it is simply the greater product. My frustrations with codex and other models is that for long horizon multi turn tasks is that the llm loses the plot of what we're doing. Alignment is what makes good employees, good team members, good contributors, good managers, good friends etc. Species survival as a bonus, alignment + intelligence beats raw intelligence for value and benefit across the board.