Matty
@cheungmatty.bsky.social
Reposted by Matty
Motif 2.6B — compact model with long context
unique: trained on AMD GPUs
focus is on long context & low hallucination rate — imo this is a growing genre of LLM that enables new search patterns
huggingface.co/Motif-Techno...
unique: trained on AMD GPUs
focus is on long context & low hallucination rate — imo this is a growing genre of LLM that enables new search patterns
huggingface.co/Motif-Techno...
Motif-Technologies/Motif-2.6B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
August 24, 2025 at 2:07 PM
Motif 2.6B — compact model with long context
unique: trained on AMD GPUs
focus is on long context & low hallucination rate — imo this is a growing genre of LLM that enables new search patterns
huggingface.co/Motif-Techno...
unique: trained on AMD GPUs
focus is on long context & low hallucination rate — imo this is a growing genre of LLM that enables new search patterns
huggingface.co/Motif-Techno...
Reposted by Matty
A look into GPT-5’s confirmed new router
best model price-per-dollar, by far
if GPT-4 brought mixture of experts (MoE), GPT-5 brought mixture of models (the router)
open.substack.com/pub/swyx/p/g...
best model price-per-dollar, by far
if GPT-4 brought mixture of experts (MoE), GPT-5 brought mixture of models (the router)
open.substack.com/pub/swyx/p/g...
GPT-5's Router: how it works and why Frontier Labs are now targeting the Pareto Frontier
The big reveal of GPT-5 was entirely unexpected but is welcome nonetheless - there's a router!
open.substack.com
August 22, 2025 at 8:47 PM
A look into GPT-5’s confirmed new router
best model price-per-dollar, by far
if GPT-4 brought mixture of experts (MoE), GPT-5 brought mixture of models (the router)
open.substack.com/pub/swyx/p/g...
best model price-per-dollar, by far
if GPT-4 brought mixture of experts (MoE), GPT-5 brought mixture of models (the router)
open.substack.com/pub/swyx/p/g...
Reposted by Matty
Persona Vectors
brb 👀👀👀👀👀👀
Anthropic just dropped this paper. They can steer models quite effectively, and even detect training data that elicits a certain (e.g. evil) persona
arxiv.org/abs/2507.21509
brb 👀👀👀👀👀👀
Anthropic just dropped this paper. They can steer models quite effectively, and even detect training data that elicits a certain (e.g. evil) persona
arxiv.org/abs/2507.21509
Persona Vectors: Monitoring and Controlling Character Traits in Language Models
Large language models interact with users through a simulated 'Assistant' persona. While the Assistant is typically trained to be helpful, harmless, and honest, it sometimes deviates from these ideals...
arxiv.org
August 1, 2025 at 5:30 PM
Persona Vectors
brb 👀👀👀👀👀👀
Anthropic just dropped this paper. They can steer models quite effectively, and even detect training data that elicits a certain (e.g. evil) persona
arxiv.org/abs/2507.21509
brb 👀👀👀👀👀👀
Anthropic just dropped this paper. They can steer models quite effectively, and even detect training data that elicits a certain (e.g. evil) persona
arxiv.org/abs/2507.21509
Reposted by Matty
Hinton nails it:
"when they [linguists] say things like, "These things don't understand anything, they're just a statistical trick," they don't actually have a model of what understanding is...if you ask what's the best model we have of understanding, it's these large language models."
"when they [linguists] say things like, "These things don't understand anything, they're just a statistical trick," they don't actually have a model of what understanding is...if you ask what's the best model we have of understanding, it's these large language models."
July 26, 2025 at 11:41 AM
Hinton nails it:
"when they [linguists] say things like, "These things don't understand anything, they're just a statistical trick," they don't actually have a model of what understanding is...if you ask what's the best model we have of understanding, it's these large language models."
"when they [linguists] say things like, "These things don't understand anything, they're just a statistical trick," they don't actually have a model of what understanding is...if you ask what's the best model we have of understanding, it's these large language models."
Reposted by Matty
My latest research on LLM grooming, co-authored with @ninajankowicz.com and @garymarcus.bsky.social, found that OpenAI's so-called "reasoning" o3 model cited Pravda network content 26.8% of the time in response to specific prompts.
This research got cited today in a column by the Financial Times:
This research got cited today in a column by the Financial Times:
Disinformation warriors are ‘grooming’ chatbots - on.ft.com/46hV6qW via @FT
Disinformation warriors are ‘grooming’ chatbots
Russian hackers are exploring ways to inject propaganda into the training data of generative AI models
on.ft.com
July 18, 2025 at 4:56 PM
My latest research on LLM grooming, co-authored with @ninajankowicz.com and @garymarcus.bsky.social, found that OpenAI's so-called "reasoning" o3 model cited Pravda network content 26.8% of the time in response to specific prompts.
This research got cited today in a column by the Financial Times:
This research got cited today in a column by the Financial Times:
Reposted by Matty
A nice 15-min video introducing the recent thought-provoking paper: “Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis”
By Akarsh Kumar, Jeff Clune, Joel Lehman, Ken Stanley
Paper: arxiv.org/abs/2505.11581
www.youtube.com/watch?v=o1q6...
By Akarsh Kumar, Jeff Clune, Joel Lehman, Ken Stanley
Paper: arxiv.org/abs/2505.11581
www.youtube.com/watch?v=o1q6...
AI doesn't work the way you think it does
YouTube video by Machine Learning Street Talk
www.youtube.com
July 5, 2025 at 6:58 AM
A nice 15-min video introducing the recent thought-provoking paper: “Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis”
By Akarsh Kumar, Jeff Clune, Joel Lehman, Ken Stanley
Paper: arxiv.org/abs/2505.11581
www.youtube.com/watch?v=o1q6...
By Akarsh Kumar, Jeff Clune, Joel Lehman, Ken Stanley
Paper: arxiv.org/abs/2505.11581
www.youtube.com/watch?v=o1q6...
Reposted by Matty
‘Dying
Is an art, like everything else.
I do it exceptionally well.’
From 'Lady Lazarus', which appears in Sylvia Plath's Ariel.
Is an art, like everything else.
I do it exceptionally well.’
From 'Lady Lazarus', which appears in Sylvia Plath's Ariel.
July 5, 2025 at 12:01 PM
‘Dying
Is an art, like everything else.
I do it exceptionally well.’
From 'Lady Lazarus', which appears in Sylvia Plath's Ariel.
Is an art, like everything else.
I do it exceptionally well.’
From 'Lady Lazarus', which appears in Sylvia Plath's Ariel.
Reposted by Matty
Ancient Rome was a rigidly hierarchical society where the ruling elite stigmatized everyone who had to work with their hands.
Yet Roman workers still found ways to resist exploitation through strikes and other forms of collective action.
Yet Roman workers still found ways to resist exploitation through strikes and other forms of collective action.
The Hidden History of Class Struggle in the Roman Empire
Ancient Rome was a rigidly hierarchical society where the ruling elite stigmatized everyone who had to work with their hands. Yet Roman workers still found ways to resist exploitation through strikes ...
jacobin.com
July 2, 2025 at 6:51 PM
Ancient Rome was a rigidly hierarchical society where the ruling elite stigmatized everyone who had to work with their hands.
Yet Roman workers still found ways to resist exploitation through strikes and other forms of collective action.
Yet Roman workers still found ways to resist exploitation through strikes and other forms of collective action.
Reposted by Matty
Some street photographers on Instagram are merely mindlessly taking advantage of their camera without polished photographic visions and abilities to discern beauty…
June 22, 2025 at 8:29 AM
Some street photographers on Instagram are merely mindlessly taking advantage of their camera without polished photographic visions and abilities to discern beauty…
Reposted by Matty
Five things to remember about war:
June 22, 2025 at 4:28 AM
Five things to remember about war:
Reposted by Matty
o3-pro summary by @latent.space
“The plan o3 gave us was plausible, reasonable; but the plan o3 Pro gave us was specific and rooted enough that it actually changed how we are thinking about our future.”
this is a very good article, read it
www.latent.space/p/o3-pro
“The plan o3 gave us was plausible, reasonable; but the plan o3 Pro gave us was specific and rooted enough that it actually changed how we are thinking about our future.”
this is a very good article, read it
www.latent.space/p/o3-pro
God is hungry for Context: First thoughts on o3 pro
OpenAI dropped o3 pricing 80% today and launched o3-pro. Ben Hylak of Raindrop.ai returns with the world's first early review.
www.latent.space
June 11, 2025 at 10:59 AM
o3-pro summary by @latent.space
“The plan o3 gave us was plausible, reasonable; but the plan o3 Pro gave us was specific and rooted enough that it actually changed how we are thinking about our future.”
this is a very good article, read it
www.latent.space/p/o3-pro
“The plan o3 gave us was plausible, reasonable; but the plan o3 Pro gave us was specific and rooted enough that it actually changed how we are thinking about our future.”
this is a very good article, read it
www.latent.space/p/o3-pro
Reposted by Matty
I put together an annotated version of the new Claude 4 system prompt, covering both the prompt Anthropic published and the missing, leaked sections that describe its various tools
It's basically the secret missing manual for Claude 4, it's fascinating!
simonwillison.net/2025/May/25/...
It's basically the secret missing manual for Claude 4, it's fascinating!
simonwillison.net/2025/May/25/...
May 25, 2025 at 1:51 PM
I put together an annotated version of the new Claude 4 system prompt, covering both the prompt Anthropic published and the missing, leaked sections that describe its various tools
It's basically the secret missing manual for Claude 4, it's fascinating!
simonwillison.net/2025/May/25/...
It's basically the secret missing manual for Claude 4, it's fascinating!
simonwillison.net/2025/May/25/...
Reposted by Matty
Introducing Bases, a new core plugin that lets you turn any set of notes into a powerful database. With Bases you can organize everything from projects to travel plans, reading lists, and more.
Bases are now available in Obsidian 1.9.0 for early access users.
Bases are now available in Obsidian 1.9.0 for early access users.
May 21, 2025 at 3:29 PM
Introducing Bases, a new core plugin that lets you turn any set of notes into a powerful database. With Bases you can organize everything from projects to travel plans, reading lists, and more.
Bases are now available in Obsidian 1.9.0 for early access users.
Bases are now available in Obsidian 1.9.0 for early access users.
Reposted by Matty
The medial #PFC is involved in economic & social #DecisionMaking in humans. @zhilinsu.bsky.social @sdnl.bsky.social &co show that dorsal mPFC is causally implicated in susceptibility to social influence, while ventral mPFC is involved in temporal discounting @plosbiology.org 🧪 plos.io/3GwXqQ7
April 29, 2025 at 1:27 PM
The medial #PFC is involved in economic & social #DecisionMaking in humans. @zhilinsu.bsky.social @sdnl.bsky.social &co show that dorsal mPFC is causally implicated in susceptibility to social influence, while ventral mPFC is involved in temporal discounting @plosbiology.org 🧪 plos.io/3GwXqQ7
Reposted by Matty
The poster for Wes Anderson's The Phoenician Scheme 🛀
The film will premiere at this year's Cannes Film Festival and be released in theaters on May 30.
The film will premiere at this year's Cannes Film Festival and be released in theaters on May 30.
April 22, 2025 at 5:11 PM
The poster for Wes Anderson's The Phoenician Scheme 🛀
The film will premiere at this year's Cannes Film Festival and be released in theaters on May 30.
The film will premiere at this year's Cannes Film Festival and be released in theaters on May 30.
Reposted by Matty
The Washington Post covered LLM grooming today, also citing my quant analysis for @americansunlight.org on the Pravda network's publishing rate: 10,000 articles of pro-Russia disinfo targeting LLMs every day.
OpenAI, Anthropic, and Perplexity did not respond to interview requests. I wonder why?
OpenAI, Anthropic, and Perplexity did not respond to interview requests. I wonder why?
Russia seeds chatbots with lies. Any bad actor could game AI the same way.
In their race to push out new versions with more capability, AI companies leave users vulnerable to “LLM grooming” efforts that promote bogus information.
www.washingtonpost.com
April 17, 2025 at 2:21 PM
The Washington Post covered LLM grooming today, also citing my quant analysis for @americansunlight.org on the Pravda network's publishing rate: 10,000 articles of pro-Russia disinfo targeting LLMs every day.
OpenAI, Anthropic, and Perplexity did not respond to interview requests. I wonder why?
OpenAI, Anthropic, and Perplexity did not respond to interview requests. I wonder why?