alacris.bsky.social
@alacris.bsky.social
Reposted by alacris.bsky.social
In the Niger Delta, child mortality in oil-producing areas is literally *double* that of children further away.
www.theguardian.com/global-devel...
The fact that people attack EVs - and *by consequence, defend oil* - for environmental and moral standards, is frankly, patently absurd.
www.theguardian.com/global-devel...
The fact that people attack EVs - and *by consequence, defend oil* - for environmental and moral standards, is frankly, patently absurd.
April 6, 2025 at 11:51 PM
In the Niger Delta, child mortality in oil-producing areas is literally *double* that of children further away.
www.theguardian.com/global-devel...
The fact that people attack EVs - and *by consequence, defend oil* - for environmental and moral standards, is frankly, patently absurd.
www.theguardian.com/global-devel...
The fact that people attack EVs - and *by consequence, defend oil* - for environmental and moral standards, is frankly, patently absurd.
Reposted by alacris.bsky.social
Meta's Multi-Token Attention
- To make it possible to use the similarities of multiple vector pairs to determine where attention must focus, instead of just single vector pair.
- They add convolutions for keys, queries, and attention heads to allow conditioning on neighboring tokens!
- To make it possible to use the similarities of multiple vector pairs to determine where attention must focus, instead of just single vector pair.
- They add convolutions for keys, queries, and attention heads to allow conditioning on neighboring tokens!
April 4, 2025 at 5:11 AM
Meta's Multi-Token Attention
- To make it possible to use the similarities of multiple vector pairs to determine where attention must focus, instead of just single vector pair.
- They add convolutions for keys, queries, and attention heads to allow conditioning on neighboring tokens!
- To make it possible to use the similarities of multiple vector pairs to determine where attention must focus, instead of just single vector pair.
- They add convolutions for keys, queries, and attention heads to allow conditioning on neighboring tokens!
Reposted by alacris.bsky.social
she finally answered
www.youtube.com/shorts/x2grE...
www.youtube.com/shorts/x2grE...
Hornet, when does Silksong come out #playframe #shorts #hollowknight #silksong
YouTube video by PlayFrame
www.youtube.com
April 2, 2025 at 3:58 PM
she finally answered
www.youtube.com/shorts/x2grE...
www.youtube.com/shorts/x2grE...
Reposted by alacris.bsky.social
This research article came out a few days ago:
transformer-circuits.pub/2025/attribu...
... and while nothing should shock those who follow the research on how LLMs "think", it does such a good job laying it out that deserves to be highlighted.
Let's dig in (🧵)!
transformer-circuits.pub/2025/attribu...
... and while nothing should shock those who follow the research on how LLMs "think", it does such a good job laying it out that deserves to be highlighted.
Let's dig in (🧵)!
April 2, 2025 at 12:25 PM
This research article came out a few days ago:
transformer-circuits.pub/2025/attribu...
... and while nothing should shock those who follow the research on how LLMs "think", it does such a good job laying it out that deserves to be highlighted.
Let's dig in (🧵)!
transformer-circuits.pub/2025/attribu...
... and while nothing should shock those who follow the research on how LLMs "think", it does such a good job laying it out that deserves to be highlighted.
Let's dig in (🧵)!
Reposted by alacris.bsky.social
My ideal city plan involves two types of superblocks (say 3-5 normal city blocks on each axis) - car-free and car-separated. Both share in common that there's adequate parking and good road conditions at the periphery - so you can get *to or from* them with ease - but from there it's on foot / bike.
November 26, 2024 at 11:58 AM
My ideal city plan involves two types of superblocks (say 3-5 normal city blocks on each axis) - car-free and car-separated. Both share in common that there's adequate parking and good road conditions at the periphery - so you can get *to or from* them with ease - but from there it's on foot / bike.
Reposted by alacris.bsky.social
Reposted by alacris.bsky.social
False. LLMs are fuzzy logic engines, built on neurons / nodes, which are themselves also fuzzy logic circuits.
Just as a demonstration, here's Claude answering a logic question I made up on the spot, about fictional concepts, with distraction sentences, and ordering shuffled.
Just as a demonstration, here's Claude answering a logic question I made up on the spot, about fictional concepts, with distraction sentences, and ordering shuffled.
September 12, 2024 at 3:58 PM
False. LLMs are fuzzy logic engines, built on neurons / nodes, which are themselves also fuzzy logic circuits.
Just as a demonstration, here's Claude answering a logic question I made up on the spot, about fictional concepts, with distraction sentences, and ordering shuffled.
Just as a demonstration, here's Claude answering a logic question I made up on the spot, about fictional concepts, with distraction sentences, and ordering shuffled.
Reposted by alacris.bsky.social
Nous Research has independently replicated the claims in "The Era of 1-bit LLMs." If this scales beyond the 1B size, it suggests a path to reduce memory consumption by an order of magnitude, and reduce end-to-end energy consumption by a factor of 20x to 70x. #MLSky 🤖
huggingface.co/1bitLLM/bitn...
huggingface.co/1bitLLM/bitn...
March 30, 2024 at 7:58 PM
Nous Research has independently replicated the claims in "The Era of 1-bit LLMs." If this scales beyond the 1B size, it suggests a path to reduce memory consumption by an order of magnitude, and reduce end-to-end energy consumption by a factor of 20x to 70x. #MLSky 🤖
huggingface.co/1bitLLM/bitn...
huggingface.co/1bitLLM/bitn...
Reposted by alacris.bsky.social
Interesting argument made in Musk v. Altman: #OpenAI 's business model is basically a tax cheat. Start out as a non-profit; do your development using tax-deductible investments; then roll that tech into a for-profit. Investors in effect pay half as much, cheating the govt, but retain the full upside
March 2, 2024 at 10:38 AM
Interesting argument made in Musk v. Altman: #OpenAI 's business model is basically a tax cheat. Start out as a non-profit; do your development using tax-deductible investments; then roll that tech into a for-profit. Investors in effect pay half as much, cheating the govt, but retain the full upside
Reposted by alacris.bsky.social
ChatGPT is definitely weaker at creative writing than competitors - I'm not sure why. (I haven't tested GPT-4). Compare, for example, vs. Bard/Gemini (ignoring Gemini Pro), which is in most regards inferior to ChatGPT. But IMHO definitely better at writing.
1. Prompt
2. ChatGPT
3. Bard
1. Prompt
2. ChatGPT
3. Bard
February 18, 2024 at 1:11 PM
ChatGPT is definitely weaker at creative writing than competitors - I'm not sure why. (I haven't tested GPT-4). Compare, for example, vs. Bard/Gemini (ignoring Gemini Pro), which is in most regards inferior to ChatGPT. But IMHO definitely better at writing.
1. Prompt
2. ChatGPT
3. Bard
1. Prompt
2. ChatGPT
3. Bard
Reposted by alacris.bsky.social
Here's what diffusion models are doing under the hood.
Basically, rather than storing images, they learn what concepts look "like". What sort of shapes and colours and textures & relative sizes & the like they tend to have. So they can enhance e.g. how "egglike" the egg is, how "nestlike" the nest.
Basically, rather than storing images, they learn what concepts look "like". What sort of shapes and colours and textures & relative sizes & the like they tend to have. So they can enhance e.g. how "egglike" the egg is, how "nestlike" the nest.
February 18, 2024 at 7:35 PM
Here's what diffusion models are doing under the hood.
Basically, rather than storing images, they learn what concepts look "like". What sort of shapes and colours and textures & relative sizes & the like they tend to have. So they can enhance e.g. how "egglike" the egg is, how "nestlike" the nest.
Basically, rather than storing images, they learn what concepts look "like". What sort of shapes and colours and textures & relative sizes & the like they tend to have. So they can enhance e.g. how "egglike" the egg is, how "nestlike" the nest.
Reposted by alacris.bsky.social
How Large Language Models learn and map space and time:
arxiv.org/pdf/2310.022...
They actually develop a sort of "temporal lobe", with specific clusters of neurons dedicated to spatiotemporal tasks. The paper trained "probes" to predict time/coords from neuron activations to make the maps.
arxiv.org/pdf/2310.022...
They actually develop a sort of "temporal lobe", with specific clusters of neurons dedicated to spatiotemporal tasks. The paper trained "probes" to predict time/coords from neuron activations to make the maps.
October 12, 2023 at 2:08 PM
How Large Language Models learn and map space and time:
arxiv.org/pdf/2310.022...
They actually develop a sort of "temporal lobe", with specific clusters of neurons dedicated to spatiotemporal tasks. The paper trained "probes" to predict time/coords from neuron activations to make the maps.
arxiv.org/pdf/2310.022...
They actually develop a sort of "temporal lobe", with specific clusters of neurons dedicated to spatiotemporal tasks. The paper trained "probes" to predict time/coords from neuron activations to make the maps.
Reposted by alacris.bsky.social
While some neural networks are easy to study (such as image recognition), LLMs are difficult due to the high level of neuron polymorphism. A recent study from Anthropic defeats this - and created a cool atlas of functions - by mapping clusters of neurons.
transformer-circuits.pub/2023/monosem...
transformer-circuits.pub/2023/monosem...
October 13, 2023 at 3:02 PM
While some neural networks are easy to study (such as image recognition), LLMs are difficult due to the high level of neuron polymorphism. A recent study from Anthropic defeats this - and created a cool atlas of functions - by mapping clusters of neurons.
transformer-circuits.pub/2023/monosem...
transformer-circuits.pub/2023/monosem...
Reposted by alacris.bsky.social
Nature: So, what shade of blue would you like to be?
Splendid Fairy-wren: Yes
#urbanNature #wildOz
🪶🌏🐦
Splendid Fairy-wren: Yes
#urbanNature #wildOz
🪶🌏🐦
September 30, 2023 at 4:21 AM
Nature: So, what shade of blue would you like to be?
Splendid Fairy-wren: Yes
#urbanNature #wildOz
🪶🌏🐦
Splendid Fairy-wren: Yes
#urbanNature #wildOz
🪶🌏🐦
Reposted by alacris.bsky.social
Most AI creative writing is… bad, so firms are on a spree hiring poets, novelists, and other humanities experts to try to fix it
Why Silicon Valley’s biggest AI developers are hiring poets
Training data companies are grabbing writers of fiction, drama, poetry, and also general humanities experts to improve AI creative writing.
restofworld.org
September 21, 2023 at 12:16 AM
Most AI creative writing is… bad, so firms are on a spree hiring poets, novelists, and other humanities experts to try to fix it
Reposted by alacris.bsky.social
You would never do this.
Warn your friends to never do this. Never ever.
The best ppl warn everybody.
Warn your friends to never do this. Never ever.
The best ppl warn everybody.
August 21, 2023 at 1:53 PM
You would never do this.
Warn your friends to never do this. Never ever.
The best ppl warn everybody.
Warn your friends to never do this. Never ever.
The best ppl warn everybody.
Reposted by alacris.bsky.social
... said "I went to the stork", the fact that it's a low-probability word means that extra attention should be drawn to it and its context.
The thoughts themselves are never in words. But the brain loves to translate between thoughts and language, and in some people, it becomes an inner monologue.
The thoughts themselves are never in words. But the brain loves to translate between thoughts and language, and in some people, it becomes an inner monologue.
September 19, 2023 at 11:27 AM
... said "I went to the stork", the fact that it's a low-probability word means that extra attention should be drawn to it and its context.
The thoughts themselves are never in words. But the brain loves to translate between thoughts and language, and in some people, it becomes an inner monologue.
The thoughts themselves are never in words. But the brain loves to translate between thoughts and language, and in some people, it becomes an inner monologue.
Reposted by alacris.bsky.social
... the speech of others, based on the known context. For example, if you heard (A) "I went to the..." or (B) "I saw a..." and the next word was unclear whether it was "store" or "stork", a probabilistic prediction based on context is useful for clarifying it. Furthermore, if the person actually...
September 19, 2023 at 11:25 AM
... the speech of others, based on the known context. For example, if you heard (A) "I went to the..." or (B) "I saw a..." and the next word was unclear whether it was "store" or "stork", a probabilistic prediction based on context is useful for clarifying it. Furthermore, if the person actually...
Reposted by alacris.bsky.social
A lot of people with an inner monologue (about 30-50% of the population) mistakenly think that the monologue is their thoughts themselves. It's actually the brain's "next-word prediction" circuitry translating thoughts into words. Our brains predict words - both for our own speech & interpreting...
September 19, 2023 at 11:23 AM
A lot of people with an inner monologue (about 30-50% of the population) mistakenly think that the monologue is their thoughts themselves. It's actually the brain's "next-word prediction" circuitry translating thoughts into words. Our brains predict words - both for our own speech & interpreting...
Reposted by alacris.bsky.social
Gravitational waves were first detected by LIGO OTD in 2015, 100 years after they were predicted by Einstein.
The discovery, announced the following February, marked the beginning of gravitational wave astronomy. 🧪 🔭
Figures: LIGO
The discovery, announced the following February, marked the beginning of gravitational wave astronomy. 🧪 🔭
Figures: LIGO
September 14, 2023 at 1:31 PM
Gravitational waves were first detected by LIGO OTD in 2015, 100 years after they were predicted by Einstein.
The discovery, announced the following February, marked the beginning of gravitational wave astronomy. 🧪 🔭
Figures: LIGO
The discovery, announced the following February, marked the beginning of gravitational wave astronomy. 🧪 🔭
Figures: LIGO
Reposted by alacris.bsky.social
About the Likes tab:
Likes are public on Bluesky. We’ve added a Likes tab to your own profile so you can quickly find past posts you’ve liked, but we chose not to show that tab on other users’ profiles.
Those likes can still be accessed by APIs, so other apps may show them.
Likes are public on Bluesky. We’ve added a Likes tab to your own profile so you can quickly find past posts you’ve liked, but we chose not to show that tab on other users’ profiles.
Those likes can still be accessed by APIs, so other apps may show them.
August 25, 2023 at 5:00 PM
About the Likes tab:
Likes are public on Bluesky. We’ve added a Likes tab to your own profile so you can quickly find past posts you’ve liked, but we chose not to show that tab on other users’ profiles.
Those likes can still be accessed by APIs, so other apps may show them.
Likes are public on Bluesky. We’ve added a Likes tab to your own profile so you can quickly find past posts you’ve liked, but we chose not to show that tab on other users’ profiles.
Those likes can still be accessed by APIs, so other apps may show them.
Reposted by alacris.bsky.social
Reposted by alacris.bsky.social
I just want to say Discord is so amazing. Blessing to the world.
April 22, 2023 at 10:01 PM
I just want to say Discord is so amazing. Blessing to the world.
Reposted by alacris.bsky.social
one feature i really enjoy on twitter is muted keywords, would love to see this added to bluesky.
(ironically it’s because i want to mute the keyword “twitter” on here, among other keywords)
(ironically it’s because i want to mute the keyword “twitter” on here, among other keywords)
May 30, 2023 at 7:46 PM
one feature i really enjoy on twitter is muted keywords, would love to see this added to bluesky.
(ironically it’s because i want to mute the keyword “twitter” on here, among other keywords)
(ironically it’s because i want to mute the keyword “twitter” on here, among other keywords)