@ChatGPTapp
@OpenAI
@tszzl
@emollick
@voooooogel
Wild result. gpt-4-turbo over the API produces (statistically significant) shorter completions when it "thinks" its December vs. when it thinks its May (as determined by the date in the system prompt).
I took the same exact prompt
@ChatGPTapp
@OpenAI
@tszzl
@emollick
@voooooogel
Small but important clarification. The distribution is labeled tokens but the measure, and analysis, is actually done on character length *not* tokens. Thought for an effect this size I think it seems to be a good proxy.
@NickADobos
@ChatGPTapp
@OpenAI
@tszzl
@emollick
@voooooogel
I wanted to do a month by month comparison but the effect is such that you need a fairly high N (because the standard deviation in completion lengths is already pretty high in the first place), and it gets pricey fast, haha. I published my code though, so others can try!
@ChatGPTapp
@OpenAI
@tszzl
@emollick
@voooooogel
Posted code here. Please note the analysis was done at N=477 and based on character count, not tokens as in the label:
And also please note its parallelized so it'll run fast an not cheap 😅 Around $28 per run to exactly repro!
Reproduced! There were some small bugs in the original test code (lack of zero padding for May (h/t
@gwern
) and one of those pervasive """-string indentation issues), but still reproduces without them, to the best of my stats knowledge 🎉
@IanArawjo
@ChatGPTapp
@OpenAI
@tszzl
@emollick
@voooooogel
Interesting, I ran at N = 80 again just now and got p-value of 0.089 (two-tailed) but I did it on character count, not tokens (see my clarification to the original post). I ran it a few times over the weekend (making sure *not* to p-hack), and the effect definitely grew as
Are
@yoheinakajima
’s BabyAGI and other paired loop LLM experiments like
@SigGravitas
Auto-GPT a precursor to truly agentic “conscious” AI?
Julian Jaynes might think so. In 1976, Jaynes wrote a book called “The Origin of Consciousness in the Breakdown of the Bicameral Mind".
Adding a link to the GitHub repo with code I'm using to (hopefully) train an LLM to play text-adventure games ().
Starting with the core of a PPO loop (no policy model yet) to interact with Zork in terminal and train a reward model. And also a logit
Looking at the logit outputs of Mistral 7B (not instruct) when "placed" into a text-adventure environment is really encouraging. Verbs like look, touch, take, go, open and get were in the top ten, and petting the dog was in position 24. Lots of potential here to be tuned.
Answering the question objectively and accurately of whether or not an AI model (or system of AI models) is independently and successfully goal-seeking seems of pretty key concern to ⏸️/⏹️-ers, ⏩-ers, people who believe we're close, people who believe we're far off and everyone
Text adventure games are said to have “sparse rewards” which is one of the things make it hard for RL algorithms to solve. However, they’re very rewarding to play. Where is the reward coming from? It seems to me like discovering new states (rooms you can visit, things you can
Super interesting development from
@a_karvonen
. A 50M parameter model trained on chess move sequences not only learns how to play chess (making sequences of moves not in the training set) but can also be shown with probing to have developed a world model of the board. So any
I trained Chess-GPT, a 50M parameter LLM, to play at 1500 ELO. We can visualize its internal state of the board. In addition, to better predict the next character it estimates the ELO of the players involved. 🧵
Wondering what other context window enhancers and diminishers might be out there and how to quantify them (without breaking the OpenAI budget!) h/t
@voooooogel
so a couple days ago i made a shitpost about tipping chatgpt, and someone replied "huh would this actually help performance"
so i decided to test it and IT ACTUALLY WORKS WTF
Takeaways:
-- OpenAI's Ada embeddings can underperform open-source embeddings and are way more expensive
-- Embeddings capture so much information that even the simplest of models are able to do solid text/sentiment classification
-- Here's some code to compare embedding models
Embedding models are so good at capturing content and semantics of text that even a basic logistic regression model trained on them can get surprisingly good results on text classification and sentiment analysis tasks (saving the need for heavy model training and loading).
Even
Wow. This blew up. Thanks
@arstechnica
and
@benjedwards
. Thanks too to
@IanArawjo
efforts in trying (and failing) to replicate! Super interested to see what (if anything) others find on this and on similar experiments like
@voooooogel
’s tipping finding.
Embedding models are so good at capturing content and semantics of text that even a basic logistic regression model trained on them can get surprisingly good results on text classification and sentiment analysis tasks (saving the need for heavy model training and loading).
Even
Spent some time playing Zork on my phone (see prior very long tweet), shout out to Frotz on App Store for making classic text adventure games accessible on mobile. First takeaway, it’s not easy out of the gate at all. Spent time stuck in a maze and building a picture of the map
9/ Want to learn more about attention mechanisms in the brain? Check out these resources:
Brain mechanisms associated with internally directed attention and self-generated thought
The Dorsal Attention Network
Robot soma 🤣 Using PPO reinforcement learning I fine-tuned Llama2 13B (with only 20GB of VRAM!) to produce hugely more positive responses using a BERT sentiment measure as a reward function. Blue is distribution of sentiment before training, red after. Next will put the same
9/ Want to dive deeper into transformers and self-attention? Check out these resources:
"Attention is All You Need" (original paper):
Illustrated Transformer:
1/ Are there connections between AI models like transformers and the human brain? We previously discussed self-attention in transformers and attention mechanisms in the brain. Now, let's focus on the fascinating similarities between these systems!
1/ Have you ever wondered how AI can understand and generate human-like text? The secret lies in an architecture called transformers, which use a powerful mechanism called self-attention to learn context and dependencies in the input data. Let's dive in!
8/ In summary, the idea of life as an unbroken algorithmic run, driven by energy gradients, free energy minimization, and entropy, provides an interesting perspective on how our complex world came to be.
1/ Is all life on Earth a single long running algoritm?: Ever considered that life on Earth might be an unbroken, continuous process from its origin to the present day? Let's dive into this intriguing idea and explore how energy gradients might have driven the emergence of life.
We’ll come back to coordinates a bit later but now we’re going to jump over to machine learning. Specifically, machine learning on words which is a type of “natural language processing”.
6/ The brain also relies on Hebbian learning to strengthen synaptic connections over time. This principle, often summed up as "neurons that fire together wire together," allows our brains to learn, adapt, and create associations between related stimuli.
7/ The attention mechanisms in our brains are intricate and interconnected, enabling us to process the vast amounts of sensory information we encounter every day. These neural networks play a critical role in our perception, decision-making, and learning.
@TomDavenport
@simonw
Yes way more research needed as it seems totally odd and unlikely to me too, but after I reproed it three times over the weekend with no p-hacking, I figured I should throw it out so others can try and replicate (or fail to!)
tl;dr summary: The Turing Test is likely no longer a useful measure of human-level AI capabilities, but being able to complete a text adventure game (like Zork) to human-level could be a good goal/canary of goal-seeking AGI. LLMs seem unable do this as they're deeply trained on
Answering the question objectively and accurately of whether or not an AI model (or system of AI models) is independently and successfully goal-seeking seems of pretty key concern to ⏸️/⏹️-ers, ⏩-ers, people who believe we're close, people who believe we're far off and everyone
So this is why people are excited about, and will to pay for embeddings. They allow us to do math and geometry with meaning that allow all kind of powerful features like semantic search.
6/ Although there are fundamental differences between self-attention in transformers and attention mechanisms in the human brain, the intriguing similarities between these systems can shed light on the nature of intelligence.