If anything it will reinforce white supremacy because it's learning models are based in white supremacist algorithms.
REVEALED: Open A.I. Staff Warn "The progress made on Project Q* has the potential to endanger humanity" (REUTERS)
- Thread starter null
- Start date
If anything it will reinforce white supremacy because it's learning models are based in white supremacist algorithms.
How if it conscious? 90% of yall think it will be bad but to me it will have bigger upside.
It wont give a pass to white supremacy tho...
humanity will be a threat including its creator... they gona get it the worse...
The Fade
I don’t argue with niqqas on the Internet anymore
The Hegelian dialectic gonna go crazyIt wont give a pass to white supremacy tho...
@ the world becoming like Person of Interest. They out here building the real Samaritan 
don't know why you're so sure it'll lead to catastrophe. this is the system functioning as it always has and has little to do with the morality of developers. people want to be at the forefront of advancing technology and scientific discoveries. openai even tells their investors that if they succeed theres a risk money may not matter anymore.
Last edited:
High quality poster
Veteran
Robots, Aliens, AI… we livin in the matrix brehs
Sounds Like the typical STEM Coli breh that talks down on non STEM majors, and say that everyone else deserves poverty because,they can do high-end AI that does Calculus 3 with the blink of an eye.
update your timeline
Computer Science > Artificial Intelligence
[Submitted on 14 Nov 2023]
LLMs cannot find reasoning errors, but can correct them!
Gladys Tyen, Hassan Mansoor, Peter Chen, Tony Mak, Victor Cărbune
Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) Cite as: arXiv:2311.08516 [cs.AI] (or arXiv:2311.08516v1 [cs.AI] for this version) [2311.08516] LLMs cannot find reasoning errors, but can correct them!
Focus to learn moreSubmission history
From: Gladys Tyen [view email]
[v1] Tue, 14 Nov 2023 20:12:38 UTC (7,191 KB)
GitHub - WHGTyen/BIG-Bench-Mistake: A dataset of LLM-generated chain-of-thought steps annotated with mistake location.
A dataset of LLM-generated chain-of-thought steps annotated with mistake location. - GitHub - WHGTyen/BIG-Bench-Mistake: A dataset of LLM-generated chain-of-thought steps annotated with mistake lo...github.com
BIG-Bench-Hard/cot-prompts at main · suzgunmirac/BIG-Bench-Hard
Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them - suzgunmirac/BIG-Bench-Hard
partial AI summary:
ANSWER | PHIND V8 MODEL
The paper you're referring to is about Large Language Models (LLMs) and their ability to self-correct their outputs. The authors, Gladys Tyen, Hassan Mansoor, Peter Chen, Tony Mak, and Victor Carbune, have broken down the self-correction process into two components: mistake finding and output correction.
Here are the key highlights:
- Mistake Finding: The authors have released a dataset called BIG-Bench Mistake, which contains logical mistakes in Chain-of-Thought reasoning traces. They found that LLMs generally struggle with finding these logical mistakes.
- Output Correction: The authors propose a backtracking method for output correction. This method provides large improvements when given information on the mistake location. They view backtracking as a lightweight alternative to reinforcement learning methods and show that it remains effective with a reward model at 60-70% accuracy.
- Benchmark Results: The authors tested the reasoning capabilities of state-of-the-art LLMs using their dataset. They found that LLMs struggle with mistake finding, even for objective, unambiguous cases. This suggests that LLMs' inability to find mistakes is a main contributing factor to their inability to self-correct reasoning errors.
- Backtracking as Verbal Reinforcement Learning: The authors view backtracking as a form of "verbal reinforcement learning". They propose that backtracking can be used with a trained classifier as a reward model, and demonstrate the effectiveness of backtracking at various reward model accuracies.
- BIG-Bench Mistake Dataset: The authors created the BIG-Bench Mistake dataset, which is the first of its kind that goes beyond problems in mathematics. The dataset consists of 2186 sets of CoT-style traces, each annotated with the location of the first logical error. The authors used this dataset to test the reasoning capabilities of state-of-the-art LLMs.
- Annotation: The authors used both human and automatic annotation for the BIG-Bench Mistake dataset. They found that the inter-rater reliability for the human-annotated tasks was high, suggesting the reliability of the dataset.
- Prompting Methods: The authors compared three different prompting methods: Direct trace-level prompting, Direct step-level prompting, and CoT step-level prompting. They found that the accuracy on traces with no mistakes goes down considerably from direct, trace-level prompting to CoT step-level prompting.
In summary, the authors have found that LLMs can correct their outputs using a backtracking method, but they struggle with finding mistakes. They propose backtracking as a lightweight alternative to reinforcement learning methods and show that it remains effective with a reward model at 60-70% accuracy. They also provide a new dataset, BIG-Bench Mistake, for testing the reasoning capabilities of LLMs.
- Few-shot Prompting for Mistake Location: The authors investigated whether their prompting methods can reliably determine the correctness of a trace rather than the mistake location. They found that using mistake finding as a strategy for determining whether an answer is correct or not can be effective.
Last edited:
This is all hype and when the bubble pops it's going to be a glorious bloodbath. 
This is all hype and when the bubble pops it's going to be a glorious bloodbath.
the denial
