/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

OpenAI details CriticGPT, a GPT-4 model fine-tuned to catch errors in ChatGPT's code output, assisting human trainers tasked with assessing and spotting errors

Having humans rate a language model's outputs produced clever chatbots.  OpenAI says adding AI to the loop could help make them even smarter and more reliable.

Wired Will Knight

Context & Ripple Effects

ChatGPT’s rollout made model mistakes a visible product constraint, while subsequent coverage stressed that the system could not reliably separate fact from fiction. CriticGPT narrows that broader reliability problem to code review by adding a specialized model to the human-feedback process.

The effort extends OpenAI’s progression from the initial ChatGPT demo through GPT-4’s claimed reasoning advance: improving output quality now includes improving the process used to judge outputs, not only the base model.

First-order effects

  • OpenAI’s human trainers can use CriticGPT to surface possible flaws in ChatGPT-generated code, changing review from wholly manual assessment to AI-assisted evaluation.
  • The immediate target is code-output reliability; CriticGPT is an aid to trainers rather than evidence that ChatGPT code is automatically dependable.

Second-order effects

  • Model developers building coding assistants face greater pressure to invest in evaluator models and test workflows, rather than competing only on generation quality.
  • For users, the value of coding models increasingly depends on whether providers can demonstrate dependable review and correction processes, a concern already raised by scrutiny of ChatGPT’s limits in distinguishing fact from fiction.

Third-order effects

  • If AI critics consistently help humans identify failures, model development may shift toward layered generator-and-evaluator systems, with evaluation capacity becoming a core capability alongside model training.
  • This also makes the quality and bias of the judging model a governance issue: scaling feedback can improve coverage, but can also scale systematic blind spots unless human oversight remains effective.

The trend: AI products are moving from standalone generation toward operational assurance systems that use models to help evaluate, monitor, and improve other models’ outputs.

Discussion

  • @openai @openai on x
    We've trained a model, CriticGPT, to catch bugs in GPT-4's code. We're starting to integrate such models into our RLHF alignment pipeline to help humans supervise AI on difficult tasks: https://openai.com/...
  • @emollick Ethan Mollick on x
    Two big lessons in the new OpenAI paper on training AI to detect AI bugs, 1) Cyborgs rule: AI detected more bugs than humans alone, but humans & AI working together had lower hallucination rates... 2)...for now: human error rates were also high. And read the highlighted conclusio…
  • @chris_j_paxton Chris Paxton on x
    Important for AI agents to be able to criticize their own mistakes. cool new work from OpenAI
  • @janleike Jan Leike on x
    Very exciting that this is out now (from my time at OpenAI): We trained an LLM critic to find bugs in code, and this helps humans find flaws on real-world production tasks that they would have missed otherwise. A promising sign for scalable oversight! https://openai.com/... [imag…
  • @luke_metro @luke_metro on x
    GPT➡️CriticGPT➡ ️ManInTheArenaGPT
  • @main_horse @main_horse on x
    @OpenAI wow, that's pessimistic [image]
  • @nickadobos Nick Dobos on x
    CriticGPT A whole gpt for fixing code bugs. Perfect, I have a lot of those
  • @spencerkschiff Spencer Schiff on x
    This chart speaks volumes. CriticGPT alone has more hallucinated critiques than the human + CriticGPT team, so having a human in the loop is still useful, but that will change soon as the models improve dramatically. Very excited for RLHF to be replaced by RLAIF [image]
  • @andrewcurran_ Andrew Curran on x
    @OpenAI ‘I used the stones to destroy the stones’ [image]
  • @gdb Greg Brockman on x
    GPT-4 for catching GPT-4's mistakes: