Article URL: https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1 Comments URL: https://news.ycombinator.com/item?id=49525378 Points: 35 # Comments: 3
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
A public GitHub project called 'Codex with ChatGPT' separates the planning and code-review roles from actual code execution by assigning reasoning tasks to the ChatGPT web app while Codex handles file edits, shell commands, and test runs. The two tools communicate through a local C2C Bridge using structured state messages, with ChatGPT accessing repository context via nine read-only tools covering file reads, Git status, diffs, and test outputs. The project's security design keeps full file bodies and logs out of control-plane messages, and the bridge itself has no write, delete, shell, or commit capabilities, with sensitive paths blocked by default. Setup requires Git, Node.js 20 or later, and Cloudflare's tunneling tool, and uses OAuth 2.1 with one-time pairing codes for the publicly reachable endpoint. The project is an unofficial community effort and has not been endorsed or independently security-assessed by OpenAI.
EarthLink Network made AI the operational foundation of its entire company, starting from the premise that all work should run on AI rather than selectively assigning tasks to it. Rather than drawing a human-AI boundary in advance, the company handed over all work to AI first and observed what remained. Through this process, routine tasks with fixed steps — such as data gathering, text cleanup, and repetitive formatting — migrated to AI naturally. What stayed on the human side was judgment: setting priorities, deciding what to drop, and taking responsibility. The company frames this outcome not as an ideology but as a practical conclusion derived from running the experiment and measuring results by convenience and efficiency.
A developer building a semantic shell — which maps natural language commands to local filesystem tools — initially tested Needle2, a small model designed for tool calling and structured extraction. While Needle2 worked in basic scenarios, problems emerged as the tool set grew beyond a handful of options, since its retrieval-based shortlisting risked excluding the correct tool before final selection. The developer also found that small models are highly sensitive to tool descriptions, requiring precise, disambiguating language rather than brief documentation-style text. Concerns about whether confidence scores across candidate groups were globally comparable added further uncertainty. Ultimately, the need for native C++ integration without external processes led the developer to switch to llama.cpp paired with the Granite model.
Experts and researchers are raising concerns that heavy reliance on AI tools is causing skilled workers to lose their expertise over time. As users depend more on AI-generated outputs, institutional knowledge is gradually eroding rather than being preserved. This degraded human knowledge then feeds back into AI models as training input, a phenomenon researchers call 'knowledge collapse.' The cycle results in progressively worse model outputs as the quality of human-generated input declines. Multiple studies and papers, including research on semantic entropy and AI's impact on human cognition, are highlighting this feedback loop as a growing risk.
Discussion (0)
Log in to join the discussion and vote.
Log in