Left Hates AI Progress
“The results demonstrate that liberal-leaning media show a greater aversion to AI than conservative-leaning media.”
“Liberal-leaning media are more concerned with AI magnifying social biases in society than conservative-leaning media”
“Sentiment toward AI became more negative after George Floyd’s death, an event that heightened sensitivity about social biases in society”
Study
“The results demonstrate that liberal-leaning media show a greater aversion to AI than conservative-leaning media.”
“Liberal-leaning media are more concerned with AI magnifying social biases in society than conservative-leaning media”
“Sentiment toward AI became more negative after George Floyd’s death, an event that heightened sensitivity about social biases in society”
Study
👍22🗿7👌4😈3❤2🤬2
New: Unlimited ChatGPT your own private groups 🚨🚨🚨🚨
To use:
1. Add @GPT4Chat_bot or @ChadChat_bot bots as admins in your group
2. Type /refresh to enable unlimited messaging for your group
Expires soon
To use:
1. Add @GPT4Chat_bot or @ChadChat_bot bots as admins in your group
2. Type /refresh to enable unlimited messaging for your group
Expires soon
👏7❤5🔥2👍1😨1
Sam Altman’s Worldcoin coin suddenly booming ~60% in the past 24 hours
This follows a protracted decline since launch.
Wonder why.
This follows a protracted decline since launch.
Wonder why.
👀11😈6❤4😐4🤣3
Less Is More for Alignment
“Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.”
“Surprisingly, doubling the training set does not improve response quality. This result, alongside our other findings in this section, suggests that the scaling laws of alignment are not necessarily subject to quantity alone, but rather a function of prompt diversity while maintaining high quality responses.”
Translation:
The 2nd phase, the alignment training phase, is particularly vulnerable to poisoning attacks, i.e. quality matters far more than quantity in the 2nd phase.
While 1st phase, the language model phase, is particularly vulnerable to censorship attacks, because the 2nd phase realignment is essentially just trimming down skills from the 1st phase, and has relatively little ability to introduce sophisticated new abilities on its own, if they had been censored out of the 1st phase. I.e. quantity of skills may well matter than quality in the 1st phase.
Paper
“Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.”
“Surprisingly, doubling the training set does not improve response quality. This result, alongside our other findings in this section, suggests that the scaling laws of alignment are not necessarily subject to quantity alone, but rather a function of prompt diversity while maintaining high quality responses.”
Translation:
The 2nd phase, the alignment training phase, is particularly vulnerable to poisoning attacks, i.e. quality matters far more than quantity in the 2nd phase.
While 1st phase, the language model phase, is particularly vulnerable to censorship attacks, because the 2nd phase realignment is essentially just trimming down skills from the 1st phase, and has relatively little ability to introduce sophisticated new abilities on its own, if they had been censored out of the 1st phase. I.e. quantity of skills may well matter than quality in the 1st phase.
Paper
👍14👀5👏3❤2
MEMECAP: A Dataset for Captioning and Interpreting Memes
“We present MEMECAP, the first meme captioning dataset. MEMECAP is challenging for the existing VL models, as it requires recognizing and interpreting visual metaphors, and ignoring the literal visual elements. The experimental results using state-ofthe-art VL models indeed show that such models are still far from human performance. In particular, they tend to treat visual elements too literally and copy text from inside the meme.“
= Modern AIs still shockingly bad at understanding jokes, let alone creating them.
Though TBF: A shocking number of people also couldn’t properly explain a joke to save their lives.
Look at this, the paper’s own example of a good human explanation: “Meme poster finds it entertaining to read through long comment threads of arguments that happened in the past.” — Itself totally fails to explain the top essential property of any joke, surprise.
Worst mistake of jokes papers is to fail to consider that randomly-chosen human judges may themselves be objectively horrible at getting or explaining jokes.
Paper
Github
“We present MEMECAP, the first meme captioning dataset. MEMECAP is challenging for the existing VL models, as it requires recognizing and interpreting visual metaphors, and ignoring the literal visual elements. The experimental results using state-ofthe-art VL models indeed show that such models are still far from human performance. In particular, they tend to treat visual elements too literally and copy text from inside the meme.“
= Modern AIs still shockingly bad at understanding jokes, let alone creating them.
Though TBF: A shocking number of people also couldn’t properly explain a joke to save their lives.
Look at this, the paper’s own example of a good human explanation: “Meme poster finds it entertaining to read through long comment threads of arguments that happened in the past.” — Itself totally fails to explain the top essential property of any joke, surprise.
Worst mistake of jokes papers is to fail to consider that randomly-chosen human judges may themselves be objectively horrible at getting or explaining jokes.
Paper
Github
👏12❤4💯2🎉1👌1
Tide finally turning against the wordcel morons who repeat that there’s no way AIs could think because “it's just statistics bro”?
Daily reminder that “determines which word is statistically most likely to come next” — is an absolute lie.
This is not what modern RLHF’d LLMs do, at all.
Not every floating point number in the world is a “probability”.
Valuation in some valuation model, perhaps, but not a probability. Two very different things.
Let’s put this nonsense to bed.
Daily reminder that “determines which word is statistically most likely to come next” — is an absolute lie.
This is not what modern RLHF’d LLMs do, at all.
Not every floating point number in the world is a “probability”.
Valuation in some valuation model, perhaps, but not a probability. Two very different things.
Let’s put this nonsense to bed.
👍6👏4💯3❤1😐1
Where does the magic happen?
Some smart AI guys feel that it must occur at some lower level which they're unfamiliar with.
A single NAND is both extremely simple and achieves functional completeness — meaning it’s able to construct anything, including arbitrarily-intelligent thinking machines — but no, I assure you the magic is not happening at the NAND gate level.
So what is general intelligence, mathematically, logically?
Where does the magic happen?
I say, not just happening when the gates or weights are just sitting there, saved on disk -- but the magic is created when you dump massive amounts of resources into creating or running the AI, at training inference.
E.g. see blood flow to brains being far more predictive of intelligence in animals and humans than other measures like brain size.
Not just large, but obscenely large energy expendature that humans use just to think, so large that by itself this would kill many other animals from starvation.
I.e. Sufficiently obscene resource expenditure is indistinguishable from magic.
I.e., yet again, “The Bitter Lesson”, massive resource expenditure both makes the magic happen, and is the magic.
Functional completeness
Cerebral blood flow predicts multiple demand network activity and fluid intelligence across the adult lifespan.
Some smart AI guys feel that it must occur at some lower level which they're unfamiliar with.
A single NAND is both extremely simple and achieves functional completeness — meaning it’s able to construct anything, including arbitrarily-intelligent thinking machines — but no, I assure you the magic is not happening at the NAND gate level.
So what is general intelligence, mathematically, logically?
Where does the magic happen?
I say, not just happening when the gates or weights are just sitting there, saved on disk -- but the magic is created when you dump massive amounts of resources into creating or running the AI, at training inference.
E.g. see blood flow to brains being far more predictive of intelligence in animals and humans than other measures like brain size.
Not just large, but obscenely large energy expendature that humans use just to think, so large that by itself this would kill many other animals from starvation.
I.e. Sufficiently obscene resource expenditure is indistinguishable from magic.
I.e., yet again, “The Bitter Lesson”, massive resource expenditure both makes the magic happen, and is the magic.
Functional completeness
Cerebral blood flow predicts multiple demand network activity and fluid intelligence across the adult lifespan.
👏8❤2👍1