General AI & Company-Specific (GOOG, META, MSFT, Anthropic, OpenAI), AI Comments
I estimate that I have invested approximately 400-425 hours over the past 3 months using OpenAI’s and Anthropic’s various language models to generate code. At the moment I am primarily using OpenAI’s o3 model, and all but stopped using Anthropic’s Claude 3.7 model. Both will see user growth accelerate if agentic capability (Claude Code notwithstanding) is introduced at an affordable price.
LLM general comments:
The impressive thing about the models from a coding standpoint is that without them, I personally would not be coding.
What leaves me unimpressed is the models’ penchant for repeating errors.
Model memory is improving. I noticed it both with Claude 3.7 over 3.5 and I see similar improvements with OpenAI o3 versus previous OpenAI models.
Company-specific comments:
MSFT: When MSFT CEO Satya Nadella says that AI is writing 20-30% of MSFT’s code, that suggests low-end code. I don’t believe what Nadella says unless it is in an 8K with hard numbers behind it. For example, it was clear to me early on that Microsoft Copilot was a failure, yet Nadella was anything but honest about Copilot’s uptake.
Anthropic (private): I believe that Anthropic is toast. It must radically innovate or die. It is in a war with OpenAI and Google which is a war (like most wars) built on currency, and Anthropic simply does not have the capital raising cache of OpenAI nor the Balance Sheet of Google. It ought to sell to Amazon before it becomes obvious to everyone that the game is over for Anthropic.
Google (GOOG) and OpenAI (private): If Google really goes after it, it will crush OpenAI. Google ought to focus on building best-in-class models combined with agentic capability. The stake in the heart to OpenAI would be to offer agentic capability for free - at least initially. Doing so would force OpenAI’s hand. OpenAI can’t afford to offer agentic capability for free. Google can.
Meta Platforms (META): There will be a huge enterprise audience for open source language models focused on specific domains and use cases. Most use cases will not require a flagship model (see our piece “Rise of the SLMs”).