Pinned Tweet
there's lots of new folks around here lately, so a quick (re)intro 👋
im Kory, Research Scientist at @GoogleDeepMind working on generative AI models and agents
i've spent ~10yr doing live theatre comedy alongside machines (my hair is ... shorter now)
always trying to get he hands of creative people
if this resonates, say hey and tell me about what you are building...
im Kory, Research Scientist at @GoogleDeepMind working on generative AI models and agents
i've spent ~10yr doing live theatre comedy alongside machines (my hair is ... shorter now)
always trying to get he hands of creative people
if this resonates, say hey and tell me about what you are building...
12
5
99
4,127
human-in-the-loop
└ human-on-the-loop
└ human-near-the-loop
└ human-cc'd-on-the-loop
└ human-legally-liable-for-the-loop
└ human-on-the-loop
└ human-near-the-loop
└ human-cc'd-on-the-loop
└ human-legally-liable-for-the-loop
82
421
4,947
133,850
I live in @Montreal and work on artificial intelligence. So, this story sits right in my backyard on Mount Royal.
Get this --- at the bottom of the mountain in 1949, @mcgillu psychologist Donald Hebb proposed a learning rule behind modern neural networks ("cells that fire "), then ran secret sensory-deprivation studies for Canada's military...
Four hundred meters up Peel Street, inside a Victorian stone mansion called Ravenscrag, Dr. Ewen Cameron read Hebb's results and tried to wipe psychiatric patients' minds clean and rebuild them with under-pillow tape loops, funded by both the CIA (MKUltra Subproject 68... yea, for real) and Ottawa.
I've been writing lately about why artists and Canadian stories need to be part of how we build with AI, and I wanted to see if I could make a short piece of local history ....
So, here's my first attempt at directing a ~5-min historical microdrama using @GoogleDeepMind Gemini 4 Argon and agentic tools in @antigravity ...
Over 12 versions,
- I ran subagent film and sound critics
- generated dozens of archival motion shots with Gemini Omni 1.1 Flash
- cast a weathered narrator and 1949 sleep-record voices on Gemini 3.8 TTS
- wrote a Python 48 kHz stereo foley mixer so Peel Street church bells, Ampex tape splices, and under-pillow ghost loops landed cleanly ducked under the narrator's breath pauses...
Twelve iterations and 300 seconds of generated video later, the total Gemini API spend was about $35...
I present... RAVENSCRAG
Get this --- at the bottom of the mountain in 1949, @mcgillu psychologist Donald Hebb proposed a learning rule behind modern neural networks ("cells that fire "), then ran secret sensory-deprivation studies for Canada's military...
Four hundred meters up Peel Street, inside a Victorian stone mansion called Ravenscrag, Dr. Ewen Cameron read Hebb's results and tried to wipe psychiatric patients' minds clean and rebuild them with under-pillow tape loops, funded by both the CIA (MKUltra Subproject 68... yea, for real) and Ottawa.
I've been writing lately about why artists and Canadian stories need to be part of how we build with AI, and I wanted to see if I could make a short piece of local history ....
So, here's my first attempt at directing a ~5-min historical microdrama using @GoogleDeepMind Gemini 4 Argon and agentic tools in @antigravity ...
Over 12 versions,
- I ran subagent film and sound critics
- generated dozens of archival motion shots with Gemini Omni 1.1 Flash
- cast a weathered narrator and 1949 sleep-record voices on Gemini 3.8 TTS
- wrote a Python 48 kHz stereo foley mixer so Peel Street church bells, Ampex tape splices, and under-pillow ghost loops landed cleanly ducked under the narrator's breath pauses...
Twelve iterations and 300 seconds of generated video later, the total Gemini API spend was about $35...
I present... RAVENSCRAG
3
2
13
814
Full 9:16 and 16:9 cuts, the essay, and an interactive reel-to-reel tape loop player are live here:
korymathewson.com/ravenscrag/
korymathewson.com/ravenscrag/
0
0
0
156
the models get better.
and my favourite kind of coding remains
spending three late nights
building some tiny web interaction
to make two friends laugh in a group chat
and my favourite kind of coding remains
spending three late nights
building some tiny web interaction
to make two friends laugh in a group chat
2
2
14
727
Sometimes multi-agent swarms remind me of bad improv:
three models enthusiastically "Yes, And"-ing each other's hallucinations into a confusing 400,000-token rat hole
three models enthusiastically "Yes, And"-ing each other's hallucinations into a confusing 400,000-token rat hole
5
1
44
2,271
generating real human laughter remains AGI-hard
when 80% comes from spontaneous conversation? why does most AI humor research focus on written jokes?
it was an honour and privilege to publish alongside long time creative collaborators @Boydbranch, Gunter Lösel ( at @zhdk), @PiotrImprov, and @thomas_wint!
when 80% comes from spontaneous conversation? why does most AI humor research focus on written jokes?
it was an honour and privilege to publish alongside long time creative collaborators @Boydbranch, Gunter Lösel ( at @zhdk), @PiotrImprov, and @thomas_wint!
1
1
8
1,332
out now in the Studies in American Humor: "In and Out of the Circle of Expectations."
drawing on a decade of live @improbotics shows, we look at a funny paradox: early glitchy generative models were often funnier onstage than modern AI, for all kinds of reasons!
scholarlypublishingcollective.org/psup/american-…
drawing on a decade of live @improbotics shows, we look at a funny paradox: early glitchy generative models were often funnier onstage than modern AI, for all kinds of reasons!
scholarlypublishingcollective.org/psup/american-…
2
1
6
227
Fantastic read… thanks for posting
cheap (time/$) AI-augmented execution changes the software career ladder…
I see junior engineers shipping working prototypes moving faster than senior engineers who protect processes
cheap (time/$) AI-augmented execution changes the software career ladder…
I see junior engineers shipping working prototypes moving faster than senior engineers who protect processes
The coding interview is dead, at least for my team at Anthropic. I've seen enough to know that it no longer provides signal. I removed it from loops this month.
I have a spicy, optimistic take - I think there's an opportunity for junior engineers: felixrieseberg.com/hunger-beats-t…
2
2
6
1,135
I’m still amazed at what Google's Gemini 4 Argon can do...
ive not done much game dev, but it's been so fun and easy to build a simple, low-key game
Pop The Balloon 🎈
and shaping it around real player feedback.
Link to play + join the Android beta below 👇
ive not done much game dev, but it's been so fun and easy to build a simple, low-key game
Pop The Balloon 🎈
and shaping it around real player feedback.
Link to play + join the Android beta below 👇
3
0
2
665
Play on iOS or Web, or join the Android beta (apparently I need 12 testers to unlock Google Play!):
korymath.github.io/pop-the-balloo…
korymath.github.io/pop-the-balloo…
1
0
2
213
Sunday Spicy Take 🌶️
You learn more about a frontier model in five minutes of trying to get it to write a weird joke than from 40 bold numbers in a benchmark table
You learn more about a frontier model in five minutes of trying to get it to write a weird joke than from 40 bold numbers in a benchmark table
2
0
7
970
Kory Mathewson retweeted
we're on a mission to end AI slop, and we're hiring (a lot)!
looking for polymaths to join us:
- research engineers into HCI, aesthetics, evals, and post-training for non-verifiable domains - engineers who've built RL environments, and have an eye for design - AI engineers who've worked on memory, context, and observability
if you’re deeply technical but have a creative side to you, you love ambiguous, nuanced, hard problems, that require knowledge across many domains, you’ll have fun here
tastelabs.com/careers or DM me
looking for polymaths to join us:
- research engineers into HCI, aesthetics, evals, and post-training for non-verifiable domains - engineers who've built RL environments, and have an eye for design - AI engineers who've worked on memory, context, and observability
if you’re deeply technical but have a creative side to you, you love ambiguous, nuanced, hard problems, that require knowledge across many domains, you’ll have fun here
tastelabs.com/careers or DM me
I pulled the fastest growing startups by hiring velocity over the past 12 weeks:
46
26
568
55,591
> be me
> right ribs hurt, jujitsu, ya,ya
> go to doctor: you need an xray
> get an xray: you're images will be sent to your doctor
> can I have my own images?
> sure, we can burn them on a disc
> okay, wait 30 minutes
> while waiting, buy disc reader, haven't had that in ... a decade?
> read disc,... , DICOMDIR...
> "The viewer operates under the following environment: Windows XP" - nice
> CPU: Pentium 4. wut is this 1998?
> copy files to mac folder
> agy --dangerously-skip-permissions
> extract the images and tell me whats up
> good news: lung good, ribs good mostly
> bad news: fracture
> by the way kory, I made you a full dicom image viewer web app... just incase
> dope, reminds me of my time studying biomedical engineering
> cool cool, make it a funny meme
> doctor calls: confirms everything
> right ribs hurt, jujitsu, ya,ya
> go to doctor: you need an xray
> get an xray: you're images will be sent to your doctor
> can I have my own images?
> sure, we can burn them on a disc
> okay, wait 30 minutes
> while waiting, buy disc reader, haven't had that in ... a decade?
> read disc,... , DICOMDIR...
> "The viewer operates under the following environment: Windows XP" - nice
> CPU: Pentium 4. wut is this 1998?
> copy files to mac folder
> agy --dangerously-skip-permissions
> extract the images and tell me whats up
> good news: lung good, ribs good mostly
> bad news: fracture
> by the way kory, I made you a full dicom image viewer web app... just incase
> dope, reminds me of my time studying biomedical engineering
> cool cool, make it a funny meme
> doctor calls: confirms everything
5
2
52
6,841
Incredible result.
And a very inspiring workflow with generative ai Jon.
And a very inspiring workflow with generative ai Jon.
This bug was first found and fixed by Gemini Flash 3.7, and most experiments in the paper were done by Flash 3.8. Gemini 4 Argon wrote the first draft of the paper and figures, which I aggressively nitpicked until the content was solid but the writing was not. Then I fed Argon a handful of my first-author papers from the last few years and had it do a rewrite in my style (which worked astonishingly well), and then I manually did a rewrite (the hands + keyboard variety) after that. I really liked this workflow! If you like my writing style feel free to throw some of my old papers in your context window and do the same.
0
0
10
1,383
awesome.
Using a single Nano Banana image and Blender MCP, Gemini 4 Argon built a 3D Gothic cathedral in @antigravity.
It iterates on the geometry, textures, and lighting to create the finished scene.
1
1
11
1,767
Kory Mathewson retweeted
Generative AI doesn’t make you creative. It rewards you for already being metacognitive.
New study in Journal of Applied Psychology --
Those with high metacognition (planning/monitoring thought) saw large creative gains from AI.
Those without it? Zero significant gain.
The skill of 2026 isn't "using AI."
It's "managing your own mind while using AI."
New study in Journal of Applied Psychology --
Those with high metacognition (planning/monitoring thought) saw large creative gains from AI.
Those without it? Zero significant gain.
The skill of 2026 isn't "using AI."
It's "managing your own mind while using AI."
2
1
15
873