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
GPT-4 finished training in August 2022, 6+ months ago
Aligns with the Morgan Stanley report that GPT-4 was already complete and GPT-5 is in progress with up to 25k GPUs.
Adjust your timelines accordingly.
via: System Card Section of GPT-4 Paper
Aligns with the Morgan Stanley report that GPT-4 was already complete and GPT-5 is in progress with up to 25k GPUs.
Adjust your timelines accordingly.
via: System Card Section of GPT-4 Paper
👍23🤯8😱2❤1
Microsoft Rations Access to AI Hardware for Internal Teams
Microsoft set to announce suite of Office 365 tools powered by GPT-4 tomorrow, Microsoft now facing an internal shortage of the server hardware needed to run the AI, according to three current Microsoft employees.
This has forced the company to ration access to the hardware for some internal teams building other AI tools to ensure it has enough capacity to handle both Bing’s new GPT-4 and the upcoming new Office tools.
Microsoft set to announce suite of Office 365 tools powered by GPT-4 tomorrow, Microsoft now facing an internal shortage of the server hardware needed to run the AI, according to three current Microsoft employees.
This has forced the company to ration access to the hardware for some internal teams building other AI tools to ensure it has enough capacity to handle both Bing’s new GPT-4 and the upcoming new Office tools.
👍16🤡4❤3😁3😢2🤬1
Non-physical things CAN or CANNOT have REAL value?
E.g.:
Only dollars backed by a physical commodity like gold have “real” value? Only jobs with a physical product like welding produce “real” value, while e.g. trading stocks can’t produce “real” value?
E.g.:
Only dollars backed by a physical commodity like gold have “real” value? Only jobs with a physical product like welding produce “real” value, while e.g. trading stocks can’t produce “real” value?
Anonymous Poll
28%
(A) Non-physical CAN have REAL value, and I lean LEFT.
47%
(B) Non-physical CAN have REAL value, and I lean RIGHT.
11%
(C) Non-physical CANNOT have REAL value, and I lean LEFT.
14%
(D) Non-physical CANNOT have REAL value, and I lean RIGHT.
👍7🌭4❤1👏1
Chat GPT
Non-physical things CAN or CANNOT have REAL value?
E.g.:
Only dollars backed by a physical commodity like gold have “real” value? Only jobs with a physical product like welding produce “real” value, while e.g. trading stocks can’t produce “real” value?
E.g.:
Only dollars backed by a physical commodity like gold have “real” value? Only jobs with a physical product like welding produce “real” value, while e.g. trading stocks can’t produce “real” value?
Elaboration:
Can that which cannot be tied back to any existing physical reality, not in any way, and not even in principle, ever really have “real” value?
Physical = attached to the consumption or production that can be tied back to some definite amount of some physical commodity that already physically exists in reality.
E.g.
Physical: Widget produced by factory, valued in terms of the value of the physical commodities needed for its production.
Physical: AI model that requires $10M worth of electricity to be trained, which in turn can, at least in principle, be tied back to definite amounts of physical resources needed to produce the needed electricity and their value.
Non-Physical: Trading of paper ownership titles, where these titles are for rights to future factory production, where this future production has not yet physically happened, and may never happen.
Non-Physical: When banks loan out the same deposited money more than once, in exchange for ownership over future repayments plus interest, but this future interest may never happen.
Can that which cannot be tied back to any existing physical reality, not in any way, and not even in principle, ever really have “real” value?
Physical = attached to the consumption or production that can be tied back to some definite amount of some physical commodity that already physically exists in reality.
E.g.
Physical: Widget produced by factory, valued in terms of the value of the physical commodities needed for its production.
Physical: AI model that requires $10M worth of electricity to be trained, which in turn can, at least in principle, be tied back to definite amounts of physical resources needed to produce the needed electricity and their value.
Non-Physical: Trading of paper ownership titles, where these titles are for rights to future factory production, where this future production has not yet physically happened, and may never happen.
Non-Physical: When banks loan out the same deposited money more than once, in exchange for ownership over future repayments plus interest, but this future interest may never happen.
👍6👏2🤡2❤1
Agree with the "fictitious capital" concept?
I.e. "Real value" must arise from some material basis, e.g. valuable commodities or money or labor effort, which went into the thing's production, or else its value is created from nowhere and "ficticious".
I.e. "Real value" must arise from some material basis, e.g. valuable commodities or money or labor effort, which went into the thing's production, or else its value is created from nowhere and "ficticious".
Anonymous Poll
35%
(A) AGREE, "fictitious capital" concept is basically correct, and I lean LEFT.
31%
(B) AGREE, "fictitious capital" concept is basically correct, and I lean RIGHT.
9%
(C) DISAGREE, "fictitious capital" is completely wrong and I lean LEFT.
24%
(D) DISAGREE, "fictitious capital" is completely wrong and I lean RIGHT.
🤔11👍3🤡3❤2👏2😴2