I'm just waiting to see the outcome of Devin, especially the "reliance" on a AI model for everything from planning, building, all the way through to testing. If you use AI for personal needs, you may have noticed how same models get "d-mber" upon the release of a new model and so forth.. They aren't always up, they have downtime. They have output issues, you won't get the same answer 10 times in a row. It'll rearrange it. All these AI models Citi is using, and building it's practices on are 3rd party. We are paying millions to a company per year or depending on usage/# of users for some - in order for that 3rd party company to rent compute power and license us their "AI" model - which is just either GPT/Claude with a face lift and a folder of custom prompts. I looked up Devin's history and apparently "Cognition that makes Devin bought the scraps that were left after Google bought developers and ip from Windsurf". Plus you can just google devin and see exactly what it's build. "Browserbase MCP, GitHub MCP, MCP Shell Server and Desktop Commander MCP. You can find Devin’s exact prompt on GitHub, this is important for directing tool usage. Devin uses Qwen32B-R1 as its tool execution LLM and it uses open AI frontier models for all other uses."
I can see next year they hike the contract up by 40-50%, then more the year after and so on... All the companies we rent AI tools from will do that, it's basic business. Then higher execs will ask themselves, why did we integrated this in every process and at every step.. Can't wait for the AI bubble to pop.