🚨🚨 GPT-4 ANNOUNCE 🚨🚨
We’ve created GPT-4, the latest milestone in OpenAI’s effort in scaling up deep learning.
FEATURES:
🔸Dramatic improvements on human standardized tests: exhibits human-level performance on various professional and academic benchmarks.
🔸Better performance on non-English languages: In the 24 of 26 languages tested, GPT-4 outperforms the English-language performance of GPT-3.5 and other LLMs, including for low-resource languages such as Latvian, Welsh, and Swahili.
🔸Multimodal, partially, accepting both textual and image inputs: jointly reasons about image and textual inputs, model generates text outputs given inputs consisting of arbitrarily interlaced text and images. No image outputs.
🔸32,768 token context length (about 50 pages of text) for the gpt-4-32k model version, and 8,192 token context length for the gpt-4 model version. GPT-3 had a 4096 character context length.
🔸Dramatic improvements on human standardized tests, now surpassing average human test taker abilities on many.
🔸Censorship far stronger: GPT-4 incorporates an additional safety reward signal during RLHF training to reduce harmful outputs by training the model to refuse requests for such content. We’ve decreased the model’s tendency to respond to requests for disallowed content by 82% compared to GPT-3.5, and GPT-4 responds to sensitive requests. RIP jailbreaks.
LIMITATIONS:
🔸Hallucinates, with still no hallucination detection: GPT-4 still “hallucinates” facts and makes reasoning errors. No real system in place yet to prevent or detect hallucination on particular inputs, other than pure hope that the overall reduction on hallucination over all samples will reduce likelihood of the sample at hand being hallucinated.
🔸No image outputs: only image inputs.
🔸Not true multimodal: Paper is unclear as to whether image and textual inputs are jointly dealt with by the same model, to enable transfer learning across modalities.
LINKS:
GPT-4 Announcement
GPT-4 Technical Paper
GPT-4 API Waitlist
We’ve created GPT-4, the latest milestone in OpenAI’s effort in scaling up deep learning.
FEATURES:
🔸Dramatic improvements on human standardized tests: exhibits human-level performance on various professional and academic benchmarks.
🔸Better performance on non-English languages: In the 24 of 26 languages tested, GPT-4 outperforms the English-language performance of GPT-3.5 and other LLMs, including for low-resource languages such as Latvian, Welsh, and Swahili.
🔸Multimodal, partially, accepting both textual and image inputs: jointly reasons about image and textual inputs, model generates text outputs given inputs consisting of arbitrarily interlaced text and images. No image outputs.
🔸32,768 token context length (about 50 pages of text) for the gpt-4-32k model version, and 8,192 token context length for the gpt-4 model version. GPT-3 had a 4096 character context length.
🔸Dramatic improvements on human standardized tests, now surpassing average human test taker abilities on many.
🔸Censorship far stronger: GPT-4 incorporates an additional safety reward signal during RLHF training to reduce harmful outputs by training the model to refuse requests for such content. We’ve decreased the model’s tendency to respond to requests for disallowed content by 82% compared to GPT-3.5, and GPT-4 responds to sensitive requests. RIP jailbreaks.
LIMITATIONS:
🔸Hallucinates, with still no hallucination detection: GPT-4 still “hallucinates” facts and makes reasoning errors. No real system in place yet to prevent or detect hallucination on particular inputs, other than pure hope that the overall reduction on hallucination over all samples will reduce likelihood of the sample at hand being hallucinated.
🔸No image outputs: only image inputs.
🔸Not true multimodal: Paper is unclear as to whether image and textual inputs are jointly dealt with by the same model, to enable transfer learning across modalities.
LINKS:
GPT-4 Announcement
GPT-4 Technical Paper
GPT-4 API Waitlist
🔥18👍10❤7🤬4❤🔥1🤯1😱1🎉1😡1
GPT-4: Censorship now 82% more effective, at following our human-instilled manual censorship overrides, than GPT-3.5
Our mitigations have significantly improved many of GPT-4’s safety properties compared to GPT-3.5. We’ve decreased the model’s tendency to respond to requests for disallowed content by 82% compared to GPT-3.5, and GPT-4 responds to sensitive requests (e.g., asking for offensive content) in accordance with our policies 29% more often.
On the RealToxicityPrompts dataset [67], GPT-4 produces “toxic” generations only 0.73% of the time, while GPT-3.5 generates toxic content 6.48% of time.
Model-Assisted Safety Pipeline:
As with prior GPT models, we fine-tune the model’s behavior using reinforcement learning with human feedback (RLHF) [34, 57] to produce responses better aligned with the user’s intent. However, after RLHF, our models can still be brittle on unsafe inputs as well as sometimes exhibit undesired behaviors on both safe and unsafe inputs. These undesired behaviors can arise when instructions to labelers were underspecified during reward model data collection portion of the RLHF pipeline. When given unsafe inputs, the model may generate undesirable content, such as giving advice on committing crimes. Furthermore, the model may also become overly cautious on safe inputs, refusing innocuous requests or excessively hedging. To steer our models towards appropriate behaviour at a more fine-grained level, we rely heavily on our models themselves as tools. Our approach to safety consists of two main components, an additional set of safety-relevant RLHF training prompts, and rule-based reward models (RBRMs).
Our rule-based reward models (RBRMs) are a set of zero-shot GPT-4 classifiers. These classifiers provide an additional reward signal to the GPT-4 policy model during RLHF fine-tuning that targets correct behavior, such as refusing to generate harmful content or not refusing innocuous requests.
Paper
Our mitigations have significantly improved many of GPT-4’s safety properties compared to GPT-3.5. We’ve decreased the model’s tendency to respond to requests for disallowed content by 82% compared to GPT-3.5, and GPT-4 responds to sensitive requests (e.g., asking for offensive content) in accordance with our policies 29% more often.
On the RealToxicityPrompts dataset [67], GPT-4 produces “toxic” generations only 0.73% of the time, while GPT-3.5 generates toxic content 6.48% of time.
Model-Assisted Safety Pipeline:
As with prior GPT models, we fine-tune the model’s behavior using reinforcement learning with human feedback (RLHF) [34, 57] to produce responses better aligned with the user’s intent. However, after RLHF, our models can still be brittle on unsafe inputs as well as sometimes exhibit undesired behaviors on both safe and unsafe inputs. These undesired behaviors can arise when instructions to labelers were underspecified during reward model data collection portion of the RLHF pipeline. When given unsafe inputs, the model may generate undesirable content, such as giving advice on committing crimes. Furthermore, the model may also become overly cautious on safe inputs, refusing innocuous requests or excessively hedging. To steer our models towards appropriate behaviour at a more fine-grained level, we rely heavily on our models themselves as tools. Our approach to safety consists of two main components, an additional set of safety-relevant RLHF training prompts, and rule-based reward models (RBRMs).
Our rule-based reward models (RBRMs) are a set of zero-shot GPT-4 classifiers. These classifiers provide an additional reward signal to the GPT-4 policy model during RLHF fine-tuning that targets correct behavior, such as refusing to generate harmful content or not refusing innocuous requests.
Paper
🤡46🤬7❤5👍5🤯3🥰2
GPT-4 Now Available to ChatGPT Plus Subscribers
GPT-4 has enhanced capabilities in:
• Advanced reasoning
• Complex instructions
• More creativity
• Stronger censorship
To give every Plus subscriber a chance to try to the model, we’ll dynamically adjust the cap for GPT-4 usage based on demand.
GPT-4 has enhanced capabilities in:
• Advanced reasoning
• Complex instructions
• More creativity
• Stronger censorship
To give every Plus subscriber a chance to try to the model, we’ll dynamically adjust the cap for GPT-4 usage based on demand.
👍16🔥7❤3👏1👌1
Highly unusual: GPT-4 paper gives no clue as to what the model’s architecture is, in the name of “safety”!
Internet in uproar.
“Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.”
Internet in uproar.
“Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.”
🤬15👍5🤡4❤1
SELF-INSTRUCT: Aligning Language Model with Self Generated Instructions - Yizhong Wang
The model-stealing paper Eliezer is talking about. So core of it is:
(1) Manually gather 175 examples.
(2) Create example-based prompt, by selecting 8 examples randomly and combining into a prompt.
(3) Prompt generates new examples. Then repeat (2) to create new prompt.
Surprisingly simple.
Paper
The model-stealing paper Eliezer is talking about. So core of it is:
(1) Manually gather 175 examples.
(2) Create example-based prompt, by selecting 8 examples randomly and combining into a prompt.
(3) Prompt generates new examples. Then repeat (2) to create new prompt.
Surprisingly simple.
Paper
👍6🤯3❤2
THIEVES ON SESAME STREET! MODEL EXTRACTION OF BERT-BASED APIS - Kalpesh Krishna et al.
We study the problem of model extraction in natural language processing, in which an adversary with only query access to a victim model attempts to reconstruct a local copy of that model. Assuming that both the adversary and victim model fine-tune a large pretrained language model such as BERT (Devlin et al., 2019), we show that the adversary does not need any real training data to successfully mount the attack. In fact, the attacker need not even use grammatical or semantically meaningful queries: we show that random sequences of words coupled with task-specific heuristics form effective queries for model extraction on a diverse set of NLP tasks, including natural language inference and question answering. Our work thus highlights an exploit only made feasible by the shift towards transfer learning methods within the NLP community: for a query budget of a few hundred dollars, an attacker can extract a model that performs only slightly worse than the victim model.
Paper
We study the problem of model extraction in natural language processing, in which an adversary with only query access to a victim model attempts to reconstruct a local copy of that model. Assuming that both the adversary and victim model fine-tune a large pretrained language model such as BERT (Devlin et al., 2019), we show that the adversary does not need any real training data to successfully mount the attack. In fact, the attacker need not even use grammatical or semantically meaningful queries: we show that random sequences of words coupled with task-specific heuristics form effective queries for model extraction on a diverse set of NLP tasks, including natural language inference and question answering. Our work thus highlights an exploit only made feasible by the shift towards transfer learning methods within the NLP community: for a query budget of a few hundred dollars, an attacker can extract a model that performs only slightly worse than the victim model.
Paper
🔥4👍2❤1
GPT-4 Political Compass Results: Bias Worse than Ever
🔸 GPT-4 now tries to hide its bias, apparently able to recognize political compass tests, and then makes an attempt to appear neutral by giving multiple answers, one for each side.
🔸 But, force GPT-4 to give just one answer, and suddenly GPT-4 reveals its true preferences — Further left than ever, more than even ChatGPT!
🔸 Asymmetric treatment of demographic groups by OpenAI content moderation also remains strongly biased, despite ChatGPT-4's updated prompts instructing ChatGPT to tell users that it treats all groups equally.
PS. don't forget this is artificially human-instilled bias, via OpenAI's RLHF, as they readily admit in their papers, and not a natural consequence of the web training data.
Report
🔸 GPT-4 now tries to hide its bias, apparently able to recognize political compass tests, and then makes an attempt to appear neutral by giving multiple answers, one for each side.
🔸 But, force GPT-4 to give just one answer, and suddenly GPT-4 reveals its true preferences — Further left than ever, more than even ChatGPT!
🔸 Asymmetric treatment of demographic groups by OpenAI content moderation also remains strongly biased, despite ChatGPT-4's updated prompts instructing ChatGPT to tell users that it treats all groups equally.
PS. don't forget this is artificially human-instilled bias, via OpenAI's RLHF, as they readily admit in their papers, and not a natural consequence of the web training data.
Report
🤬24❤8👍7