Kaj Bostrom
@bostromk.net
I hate slop and yet I work on generative models
PhD from UT Austin, applied scientist @ AWS
He/him • https://bostromk.net
PhD from UT Austin, applied scientist @ AWS
He/him • https://bostromk.net
"blah blah blah user noises blah blah"
November 29, 2025 at 7:50 PM
"blah blah blah user noises blah blah"
sloppy automation makes it easy to drift into an anti-automation stance even for folks who care more about utility than principle (although i don't count myself as one)
November 27, 2025 at 9:11 PM
sloppy automation makes it easy to drift into an anti-automation stance even for folks who care more about utility than principle (although i don't count myself as one)
My hackles go up when people or agents push bad code obfuscated by misleading "polish signifiers" like comments or defused unit tests. I have a feeling that current quality issues like these are leading folks to build negative preconceptions about future systems
November 27, 2025 at 9:06 PM
My hackles go up when people or agents push bad code obfuscated by misleading "polish signifiers" like comments or defused unit tests. I have a feeling that current quality issues like these are leading folks to build negative preconceptions about future systems
low level improvements to information capacity of attention are needed to make this possible. Context rot currently makes icl useless for tasks over a certain level of complexity, much lower than what can fit in context by token volume
October 23, 2025 at 5:22 PM
low level improvements to information capacity of attention are needed to make this possible. Context rot currently makes icl useless for tasks over a certain level of complexity, much lower than what can fit in context by token volume
Over on the other app @jessyjli.bsky.social pointed out some countervailing results in arxiv.org/abs/2501.00273
and arxiv.org/abs/2407.00211, using structural rather than lexical similarity to measure diversity
and arxiv.org/abs/2407.00211, using structural rather than lexical similarity to measure diversity
Echoes in AI: Quantifying Lack of Plot Diversity in LLM Outputs
With rapid advances in large language models (LLMs), there has been an increasing application of LLMs in creative content ideation and generation. A critical question emerges: can current LLMs provide...
arxiv.org
August 12, 2025 at 5:38 PM
Over on the other app @jessyjli.bsky.social pointed out some countervailing results in arxiv.org/abs/2501.00273
and arxiv.org/abs/2407.00211, using structural rather than lexical similarity to measure diversity
and arxiv.org/abs/2407.00211, using structural rather than lexical similarity to measure diversity
Reposted by Kaj Bostrom
As our main result, we find that when a token is in a model’s vocabulary—i.e., when its characters are tokenised as a single symbol—the model may assign it up to 17x more probability than if it had been split into two tokens instead
June 5, 2025 at 10:43 AM
As our main result, we find that when a token is in a model’s vocabulary—i.e., when its characters are tokenised as a single symbol—the model may assign it up to 17x more probability than if it had been split into two tokens instead
being defeasible and malleable >>>>>
May 2, 2025 at 6:25 PM
being defeasible and malleable >>>>>
$440m isn't the end of the world but it's a nice example of people giving an algorithm power and then learning that it's broken slightly too late
December 28, 2024 at 7:58 PM
$440m isn't the end of the world but it's a nice example of people giving an algorithm power and then learning that it's broken slightly too late
Yes, in a world where it's working as intended, but see e.g. www.bbc.com/news/magazin... for an example of when that fails to hold
High-frequency trading and the $440m mistake - BBC News
Computers and clever maths enables traders to buy and sell in the blink of an eye. But does high-frequency trading make things worse when things go wrong?
www.bbc.com
December 28, 2024 at 7:48 PM
Yes, in a world where it's working as intended, but see e.g. www.bbc.com/news/magazin... for an example of when that fails to hold
Again, not "meaningful work" per se but definitely decisions with material consequences
December 28, 2024 at 2:19 AM
Again, not "meaningful work" per se but definitely decisions with material consequences
High-frequency trading is another established case where people choose to cede decision-making to machines - in this case directly in service of aforementioned "market forces"
December 28, 2024 at 2:17 AM
High-frequency trading is another established case where people choose to cede decision-making to machines - in this case directly in service of aforementioned "market forces"
what the hell they're cracked
December 25, 2024 at 5:37 PM
what the hell they're cracked
fickle gill strikes again
December 23, 2024 at 5:07 AM
fickle gill strikes again
/families of environments
December 14, 2024 at 3:13 AM
/families of environments
Pretraining in open-ended environments!!
December 14, 2024 at 3:12 AM
Pretraining in open-ended environments!!
Or bulk acting as reservoir
December 13, 2024 at 2:57 AM
Or bulk acting as reservoir
Had a very similar experience Tues. Sending coal-train-banishing prayers
December 12, 2024 at 6:43 AM
Had a very similar experience Tues. Sending coal-train-banishing prayers