Martin Singh-Blom
@singhblom.com
Head of ML at Embark Studios. Working on RL for physics based animation, LLM applications in games, generative models, etc.
I like it when ML makes what previously was impossible possible. Previously web dev, genetics, mathematics, physics.
I like it when ML makes what previously was impossible possible. Previously web dev, genetics, mathematics, physics.
Here: arxiv.org/abs/2510.07591
IASC: Interactive Agentic System for ConLangs
We present a system that uses LLMs as a tool in the development of Constructed Languages. The system is modular in that one first creates a target phonology for the language using an agentic approach ...
arxiv.org
October 11, 2025 at 10:03 AM
Here: arxiv.org/abs/2510.07591
The authors of the paper highlight the understanding of LLM aspect. It was much better at some more common types of grammatical structures than others.
They also mention that it might be useful for translation into low resource languages in the future but the results aren't good enough yet.
They also mention that it might be useful for translation into low resource languages in the future but the results aren't good enough yet.
October 11, 2025 at 9:49 AM
The authors of the paper highlight the understanding of LLM aspect. It was much better at some more common types of grammatical structures than others.
They also mention that it might be useful for translation into low resource languages in the future but the results aren't good enough yet.
They also mention that it might be useful for translation into low resource languages in the future but the results aren't good enough yet.
Not all of the rest of America. Alaska is still free not part of Greater Saskatchewan.
August 14, 2025 at 12:04 PM
Not all of the rest of America. Alaska is still free not part of Greater Saskatchewan.
But if you ask ChatGPT to introspect about its thinking process most of the answer will be drawn from the knowledge in the LLM not from the prompt.
August 6, 2025 at 9:26 PM
But if you ask ChatGPT to introspect about its thinking process most of the answer will be drawn from the knowledge in the LLM not from the prompt.
But Masnick's post is true about LLMs. They are in some sense just really fancy autocompletes. Puck's objection is an objection about how an LLM based chat system works, with system prompts and memories. If you ask ChatGPT about its system prompt you can get good answers.
August 6, 2025 at 9:23 PM
But Masnick's post is true about LLMs. They are in some sense just really fancy autocompletes. Puck's objection is an objection about how an LLM based chat system works, with system prompts and memories. If you ask ChatGPT about its system prompt you can get good answers.
If you are arguing about whether LLM introspection can tell you something about how LLMs work you are definitely in nerd territory 😁 And I think I agree with you.
August 6, 2025 at 9:12 PM
If you are arguing about whether LLM introspection can tell you something about how LLMs work you are definitely in nerd territory 😁 And I think I agree with you.
LLMs don't remember anything, they just model language. Systems like ChatGPT have other memory components though. LLMs will be able to produce good answers on how LLMs work though, since that is public info that would be part of the training data. ChatGPT will also use results from Google these days
August 6, 2025 at 8:38 PM
LLMs don't remember anything, they just model language. Systems like ChatGPT have other memory components though. LLMs will be able to produce good answers on how LLMs work though, since that is public info that would be part of the training data. ChatGPT will also use results from Google these days
Reposted by Martin Singh-Blom
We've seen this story a lot. Personal computers were niche in the 80s. Hard to use, it seemed obvs that they would remain toys for nerds.
Cell phones were niche, and for the rich.
I think the downside prediction for widespread on-device AI is "it takes a decade." "Can't happen" is not an outcome.
Cell phones were niche, and for the rich.
I think the downside prediction for widespread on-device AI is "it takes a decade." "Can't happen" is not an outcome.
August 2, 2025 at 7:01 PM
We've seen this story a lot. Personal computers were niche in the 80s. Hard to use, it seemed obvs that they would remain toys for nerds.
Cell phones were niche, and for the rich.
I think the downside prediction for widespread on-device AI is "it takes a decade." "Can't happen" is not an outcome.
Cell phones were niche, and for the rich.
I think the downside prediction for widespread on-device AI is "it takes a decade." "Can't happen" is not an outcome.
Reposted by Martin Singh-Blom
Bro, just one more humanoid robot, bro please, this one will be the true general purpose technology bro, this one will fill any labor shortage bro, our nanas and papas will let them into their homes to help them age in place bro, it’s the future bro, I won’t put guns on them I swear bro, bro
March 6, 2025 at 5:18 PM
Bro, just one more humanoid robot, bro please, this one will be the true general purpose technology bro, this one will fill any labor shortage bro, our nanas and papas will let them into their homes to help them age in place bro, it’s the future bro, I won’t put guns on them I swear bro, bro
Reposted by Martin Singh-Blom
If you have infinite data and compute, 1-nearest neighbors is also optimal.
July 31, 2025 at 4:47 AM
If you have infinite data and compute, 1-nearest neighbors is also optimal.
Reposted by Martin Singh-Blom
/
* yes, NWT means capital tax is lower the higher-return is the investment - ie incentivising better investments/faster growth, relative to other ways of taxing capital. It's a bug not a feature. Favour entrepreneurs over rentiers and heirs.
on.ft.com/44FZKhk
/
* yes, NWT means capital tax is lower the higher-return is the investment - ie incentivising better investments/faster growth, relative to other ways of taxing capital. It's a bug not a feature. Favour entrepreneurs over rentiers and heirs.
on.ft.com/44FZKhk
/
Why a wealth tax is capitalism’s handmaiden
[FREE TO READ] Taxing capital holdings boosts rewards for investing well
on.ft.com
July 22, 2025 at 9:58 AM
/
* yes, NWT means capital tax is lower the higher-return is the investment - ie incentivising better investments/faster growth, relative to other ways of taxing capital. It's a bug not a feature. Favour entrepreneurs over rentiers and heirs.
on.ft.com/44FZKhk
/
* yes, NWT means capital tax is lower the higher-return is the investment - ie incentivising better investments/faster growth, relative to other ways of taxing capital. It's a bug not a feature. Favour entrepreneurs over rentiers and heirs.
on.ft.com/44FZKhk
/