Julia Rozanova
@juliaroz.bsky.social
Aficionado of cat-eye spectacles. Fantasy book weirdo. Catch me on the dance floor!
|| Machine Learning Research @ Featurespace || PhD from Uni of Manchester in Compsci /Natural Language Processing || MSc Maths, Stellenbosch ||
Cambridge, UK
|| Machine Learning Research @ Featurespace || PhD from Uni of Manchester in Compsci /Natural Language Processing || MSc Maths, Stellenbosch ||
Cambridge, UK
Reposted by Julia Rozanova
I'll admit, I was skeptical when they said Gemini was just like a bunch of PhDs. But I gotta admit they nailed it.
August 17, 2025 at 1:51 PM
I'll admit, I was skeptical when they said Gemini was just like a bunch of PhDs. But I gotta admit they nailed it.
Reposted by Julia Rozanova
EVERYONE SHOULD LISTEN TO MEAL OF THORNS, the best-titled podcast in living memory
Got to meet @amalelmohtar.com in London and I was PROPER giddy to hear her chat about The River had Roots. Folklore!! Grammar!! Harps from body parts!! Secret out-of-book world-building snippets!!
Also I got to gush about her Traitor Baru Cormorant episode on @mealofthorns.bsky.social, IMPORTANT
Also I got to gush about her Traitor Baru Cormorant episode on @mealofthorns.bsky.social, IMPORTANT
April 25, 2025 at 9:06 AM
EVERYONE SHOULD LISTEN TO MEAL OF THORNS, the best-titled podcast in living memory
Got to meet @amalelmohtar.com in London and I was PROPER giddy to hear her chat about The River had Roots. Folklore!! Grammar!! Harps from body parts!! Secret out-of-book world-building snippets!!
Also I got to gush about her Traitor Baru Cormorant episode on @mealofthorns.bsky.social, IMPORTANT
Also I got to gush about her Traitor Baru Cormorant episode on @mealofthorns.bsky.social, IMPORTANT
April 25, 2025 at 8:36 AM
Got to meet @amalelmohtar.com in London and I was PROPER giddy to hear her chat about The River had Roots. Folklore!! Grammar!! Harps from body parts!! Secret out-of-book world-building snippets!!
Also I got to gush about her Traitor Baru Cormorant episode on @mealofthorns.bsky.social, IMPORTANT
Also I got to gush about her Traitor Baru Cormorant episode on @mealofthorns.bsky.social, IMPORTANT
Reposted by Julia Rozanova
Game rec for anyone interested in language / NLP:
Chants if Sennaar!
Absolute stunner of a game based on language learning, grammar, text and context.
Chants if Sennaar!
Absolute stunner of a game based on language learning, grammar, text and context.
January 4, 2025 at 11:21 AM
Game rec for anyone interested in language / NLP:
Chants if Sennaar!
Absolute stunner of a game based on language learning, grammar, text and context.
Chants if Sennaar!
Absolute stunner of a game based on language learning, grammar, text and context.
Reposted by Julia Rozanova
Reposted by Julia Rozanova
Reposted by Julia Rozanova
I built a Playstacean!
March 21, 2025 at 11:51 PM
I built a Playstacean!
Reposted by Julia Rozanova
Roses are red
Violets are blue
Violets are blue
March 20, 2025 at 5:43 PM
Roses are red
Violets are blue
Violets are blue
Reposted by Julia Rozanova
wow I do not like those Reddit lines
February 18, 2025 at 1:06 PM
wow I do not like those Reddit lines
Reposted by Julia Rozanova
When a Chinese company I've never heard of releases some good or service that upsets Sam Altman
January 27, 2025 at 3:18 PM
When a Chinese company I've never heard of releases some good or service that upsets Sam Altman
Reposted by Julia Rozanova
Friends, for something to be open source, we need to see
1. The data it was trained & evaluated on
2. The code
3. Model architecture
4. Model weights.
DeepSeek only gives 3, 4. And I'll see the day that anyone gives us #1 without being forced to do so, because all of them are stealing data.
1. The data it was trained & evaluated on
2. The code
3. Model architecture
4. Model weights.
DeepSeek only gives 3, 4. And I'll see the day that anyone gives us #1 without being forced to do so, because all of them are stealing data.
January 29, 2025 at 5:01 AM
Friends, for something to be open source, we need to see
1. The data it was trained & evaluated on
2. The code
3. Model architecture
4. Model weights.
DeepSeek only gives 3, 4. And I'll see the day that anyone gives us #1 without being forced to do so, because all of them are stealing data.
1. The data it was trained & evaluated on
2. The code
3. Model architecture
4. Model weights.
DeepSeek only gives 3, 4. And I'll see the day that anyone gives us #1 without being forced to do so, because all of them are stealing data.
Reposted by Julia Rozanova
One of my grand interpretability goals is to improve human scientific understanding by analyzing scientific discovery models, but this is the most convincing case yet that we CAN learn from model interpretation: Chess grandmasters learned new play concepts from AlphaZero's internal representations.
Bridging the Human-AI Knowledge Gap: Concept Discovery and Transfer in AlphaZero
Artificial Intelligence (AI) systems have made remarkable progress, attaining super-human performance across various domains. This presents us with an opportunity to further human knowledge and improv...
arxiv.org
January 27, 2025 at 9:43 PM
One of my grand interpretability goals is to improve human scientific understanding by analyzing scientific discovery models, but this is the most convincing case yet that we CAN learn from model interpretation: Chess grandmasters learned new play concepts from AlphaZero's internal representations.
Reposted by Julia Rozanova
Reposted by Julia Rozanova
thats ok im already fluent in anxiety
January 16, 2025 at 10:51 PM
thats ok im already fluent in anxiety
Reposted by Julia Rozanova
My poor, tortured heart is at ease any time I see a device that still has a 3.5 headphone jack 🥲
January 16, 2025 at 3:19 PM
My poor, tortured heart is at ease any time I see a device that still has a 3.5 headphone jack 🥲
Reposted by Julia Rozanova
after days of grifters and con men at CES we stumbled upon the booth for VLC. they were all dressed as wizards and told us, "we have nothing to sell, we just decided to show up". i told them I'd been using their software to pirate media for 15 years and they said "keep doing that"
January 10, 2025 at 9:24 PM
after days of grifters and con men at CES we stumbled upon the booth for VLC. they were all dressed as wizards and told us, "we have nothing to sell, we just decided to show up". i told them I'd been using their software to pirate media for 15 years and they said "keep doing that"
Reposted by Julia Rozanova
I came across this lovely poster on my walk to the Computer Lab in Cambridge this morning.
January 7, 2025 at 11:06 AM
I came across this lovely poster on my walk to the Computer Lab in Cambridge this morning.
Reposted by Julia Rozanova
incredible -- the NYT ran fluff "what i hope to see in 2025" blurbs from CEOs and economists, and then this guy
www.nytimes.com/2025/01/02/o...
www.nytimes.com/2025/01/02/o...
January 4, 2025 at 4:52 AM
incredible -- the NYT ran fluff "what i hope to see in 2025" blurbs from CEOs and economists, and then this guy
www.nytimes.com/2025/01/02/o...
www.nytimes.com/2025/01/02/o...
Game rec for anyone interested in language / NLP:
Chants if Sennaar!
Absolute stunner of a game based on language learning, grammar, text and context.
Chants if Sennaar!
Absolute stunner of a game based on language learning, grammar, text and context.
January 4, 2025 at 11:21 AM
Game rec for anyone interested in language / NLP:
Chants if Sennaar!
Absolute stunner of a game based on language learning, grammar, text and context.
Chants if Sennaar!
Absolute stunner of a game based on language learning, grammar, text and context.
Reposted by Julia Rozanova
please enjoy this "vintage movie poster" I saw in a dream which was so funny to my subconscious that I immediately woke myself up to write it down for later
December 31, 2024 at 7:27 PM
please enjoy this "vintage movie poster" I saw in a dream which was so funny to my subconscious that I immediately woke myself up to write it down for later
Reposted by Julia Rozanova
Here's my end-of-year review of things we learned out about LLMs in 2024 - we learned a LOT of things simonwillison.net/2024/Dec/31/...
Table of contents:
Table of contents:
December 31, 2024 at 6:10 PM
Here's my end-of-year review of things we learned out about LLMs in 2024 - we learned a LOT of things simonwillison.net/2024/Dec/31/...
Table of contents:
Table of contents:
Reposted by Julia Rozanova
This is an excellent illustration of the challenges presented by "alignment". Whose values and concerns are to be reflected in the alignment? It is, and always will be, a political decision. Language models are political models.
Let me say this again: the AI my W. Indian student used to correct her syntax (Claude I think) also removed every mention of Franz Fanon from her work! I suspect this is because an engineer somewhere decided Fanon's ideas are controversial? WTF!
Paging @ruha9.bsky.social & @alondra.bsky.social!!
Paging @ruha9.bsky.social & @alondra.bsky.social!!
December 28, 2024 at 7:17 PM
This is an excellent illustration of the challenges presented by "alignment". Whose values and concerns are to be reflected in the alignment? It is, and always will be, a political decision. Language models are political models.
Reposted by Julia Rozanova
STEM is so cool. Science is amazing. Kids should develop curiosity about how the universe works and proficiency in uncovering its secrets, so that they can most effectively optimize clickthrough rates on banner ads at the bottom of VC-funded website-apps.
December 27, 2024 at 3:22 AM
STEM is so cool. Science is amazing. Kids should develop curiosity about how the universe works and proficiency in uncovering its secrets, so that they can most effectively optimize clickthrough rates on banner ads at the bottom of VC-funded website-apps.
Reposted by Julia Rozanova
Good night, everyone
February 20, 2024 at 12:09 AM
Good night, everyone
Reposted by Julia Rozanova
What's in an attention head? 🤯
We present an efficient framework – MAPS – for inferring the functionality of attention heads in LLMs ✨directly from their parameters✨
A new preprint with Amit Elhelo 🧵 (1/10)
We present an efficient framework – MAPS – for inferring the functionality of attention heads in LLMs ✨directly from their parameters✨
A new preprint with Amit Elhelo 🧵 (1/10)
December 18, 2024 at 5:55 PM
What's in an attention head? 🤯
We present an efficient framework – MAPS – for inferring the functionality of attention heads in LLMs ✨directly from their parameters✨
A new preprint with Amit Elhelo 🧵 (1/10)
We present an efficient framework – MAPS – for inferring the functionality of attention heads in LLMs ✨directly from their parameters✨
A new preprint with Amit Elhelo 🧵 (1/10)