Anthropic has released their research roadmap for 2021/2023-2025, which is not what they intend to publish, but rather a public document that needs to be provided to the court due to copyright litigation. You can take a look at their predictions about the development of AI a few years ago. The internal document of Anthropic, "Concise Research Roadmap," written by Jared Kaplan in July 2021, elaborates on Anthropic's research priorities and long-term goals for mid-2021. The other document is an internal Anthropic document written by Matt Bell in June 2023, titled 'Product Visions for Early 2025'. Document from: https://www. (lesswrong.com)/posts/6eaKqtmDuo6CmkqMK/loops-s-shortform? commentId=7C9iqsDCdxC2PfKDf July 2021 Consensus Research Roadmap: Proof of Concept: Creating a universal AI assistant that can collaborate with users in a conversational manner to complete tasks such as creating, editing, and interpreting software. The most ideal form should possess all the abilities of a software engineer colleague (answering general questions, writing functions, finding bugs, designing APIs, discussing code styles and best practices, etc.). To achieve this goal, it is necessary to: -Large scale generative models -Long memories -Fine tuning through RL (reinforcement learning) combined with automated rewards and human feedback Product Visions for Early 2025, June 23: Core content of the document: Three product visions -Claude the Virtual Employee/Claude the Robot Claude is able to independently undertake clear roles within the company (software development, marketing, operations, etc.), using the same interface as humans or directly calling APIs to independently complete tasks -Claude the Personal Assistant/Claude the Cyborg (Personal Assistant) Claude is an extension of user will, similar to an 'AI Chief of Staff', greatly amplifying personal abilities. It helps with coding, email writing, research, tax processing, etc., but by default, it has low autonomy and aims to be a sub goal of the user, which will be confirmed with the user before key actions. -Claude as Intelligent Org Infrastructure Claude becomes the proactive collaborative infrastructure, making the entire organization more efficient. It can read and understand massive amounts of information such as Slack, Docs, meetings, etc., actively helping to coordinate, connect people, resolve conflicts, reduce information overload and goal misalignment. This includes discussions among several of their planning and drafting personnel: Safe path preference: Ethan Perez and others tend to start with Cyborg (safer, human in the loop) and gradually transition to Robot to have more time to solve alignment problems. Directly charging the robot is considered to have a higher risk. Software complexity vs model capability: Jared Kaplan tends to minimize the complexity of non AI software, allowing models to naturally integrate into human roles; Ben Mann is concerned that the model may not be reliable and that software "hacks" can go further ahead. UBI and its social impact: Some people are concerned that the Robot pathway may trigger unemployment panic and need to be accompanied by UBI; But some people (such as Michael Sellitto and Matt Bell) have pointed out that UBI is not omnipotent, as work also provides dignity and identity, and pure UBI may bring social problems. Competition and Market: Tom Brown believes that competitors will prioritize pursuing robots because they can create economic value the fastest; Some people are also concerned that the Robot narrative is too scary and not conducive to public acceptance. Progressive path: Most opinions believe that Cyborg or Org Infrastructure is a natural transition to Robot, and it is more realistic to first create a safer version.