Vincent Carchidi
@vcarchidi.bsky.social
Defense analyst. Tech policy. Have a double life in CogSci/Philosophy of Mind. Most posts here about the latter. Usually.
https://philpeople.org/profiles/vincent-carchidi
All opinions entirely my own.
https://philpeople.org/profiles/vincent-carchidi
All opinions entirely my own.
Pinned
Do Language Models Learn Like Human Children?
On a history of unwittingly lifting cognitive weight, with comments on a personal motivation for studying a very unusual topic.
vincentcarchidi.substack.com
Just in time for the holidays, my final post is live.
This is a long-read, and likely of interest to anyone interested in the uses and limitations of computational modeling in cognitive science, with attention on less-noticed issues in the field. 🧵
vincentcarchidi.substack.com/p/do-languag...
This is a long-read, and likely of interest to anyone interested in the uses and limitations of computational modeling in cognitive science, with attention on less-noticed issues in the field. 🧵
vincentcarchidi.substack.com/p/do-languag...
"Remember: baseliners were asked to verbalize their thoughts aloud while working...Much more importantly, though, they were literally paid by the hour—the longer they took to complete tasks, the more money they made."
arachnemag.substack.com/p/the-metr-g...
arachnemag.substack.com/p/the-metr-g...
Against the 'METR Graph'
A deep dive into A.I. understanders' favorite chart crime
arachnemag.substack.com
January 8, 2026 at 3:05 PM
"Remember: baseliners were asked to verbalize their thoughts aloud while working...Much more importantly, though, they were literally paid by the hour—the longer they took to complete tasks, the more money they made."
arachnemag.substack.com/p/the-metr-g...
arachnemag.substack.com/p/the-metr-g...
nevermind
New Year's Resolution: don't keep so many tabs open that I can't see where they end
January 8, 2026 at 12:46 AM
nevermind
One thing that I think has confused AI safety talk is the idea that individual performance should be weighed against model performance. But this really isn't the standard that traditionally dominated tech and safety, and would be incoherent for certain systems -
January 7, 2026 at 11:34 PM
One thing that I think has confused AI safety talk is the idea that individual performance should be weighed against model performance. But this really isn't the standard that traditionally dominated tech and safety, and would be incoherent for certain systems -
I mean really, what does OpenAI have to justify a near-trillion dollar valuation but a competitor like Anthropic does not have?
Is it image and video generation? Maybe, but where's the actual value? A few commercials and other hype-adjacent uses? Not to mention Google is hot on their heels.
Is it image and video generation? Maybe, but where's the actual value? A few commercials and other hype-adjacent uses? Not to mention Google is hot on their heels.
If Claude Code is what people say it is, then it seems like we can confidently point to that as a successful, conceivably profitable product that surpasses competitor products. A clear use, apparently effective and reliable in skilled hands.
January 7, 2026 at 6:52 PM
I mean really, what does OpenAI have to justify a near-trillion dollar valuation but a competitor like Anthropic does not have?
Is it image and video generation? Maybe, but where's the actual value? A few commercials and other hype-adjacent uses? Not to mention Google is hot on their heels.
Is it image and video generation? Maybe, but where's the actual value? A few commercials and other hype-adjacent uses? Not to mention Google is hot on their heels.
As long as valuations are inflated, it does feel like Anthropic's should be at least as high as OpenAI's. But it's not even particularly close.
www.wsj.com/tech/ai/anth...
www.wsj.com/tech/ai/anth...
Exclusive | Anthropic Raising $10 Billion at $350 Billion Value
The planned deal represents nearly a doubling of the AI startup’s valuation from four months ago.
www.wsj.com
January 7, 2026 at 6:29 PM
As long as valuations are inflated, it does feel like Anthropic's should be at least as high as OpenAI's. But it's not even particularly close.
www.wsj.com/tech/ai/anth...
www.wsj.com/tech/ai/anth...
Reposted by Vincent Carchidi
New paper in Synthese: 'Robustness and trustworthiness in AI: a no-go result from formal epistemology'
❓When does AI model M, on input x, show behavior φ robustly?
💡In modal logic: M,x⊧□φ
➡️Exposes limit on robustness & trustworthiness
link.springer.com/article/10.1...
philpapers.org/rec/HORRAT-13
❓When does AI model M, on input x, show behavior φ robustly?
💡In modal logic: M,x⊧□φ
➡️Exposes limit on robustness & trustworthiness
link.springer.com/article/10.1...
philpapers.org/rec/HORRAT-13
Robustness and trustworthiness in AI: a no-go result from formal epistemology - Synthese
Synthese - A major issue for the trustworthiness of modern AI-models is their lack of robustness. A notorious example is that putting a small sticker on a stop sign can cause AI-models to classify...
link.springer.com
January 7, 2026 at 2:21 PM
New paper in Synthese: 'Robustness and trustworthiness in AI: a no-go result from formal epistemology'
❓When does AI model M, on input x, show behavior φ robustly?
💡In modal logic: M,x⊧□φ
➡️Exposes limit on robustness & trustworthiness
link.springer.com/article/10.1...
philpapers.org/rec/HORRAT-13
❓When does AI model M, on input x, show behavior φ robustly?
💡In modal logic: M,x⊧□φ
➡️Exposes limit on robustness & trustworthiness
link.springer.com/article/10.1...
philpapers.org/rec/HORRAT-13
This is good.
Sorry, but triumphant claims about autonomous vehicle safety are wildly exaggerated.
It's an open question whether today’s self-driving cars are any safer than those driven by humans.
And if reducing crashes is the goal, that isn’t even the right question.
My deep dive, in Bloomberg 🧵
It's an open question whether today’s self-driving cars are any safer than those driven by humans.
And if reducing crashes is the goal, that isn’t even the right question.
My deep dive, in Bloomberg 🧵
We Still Don’t Know if Robotaxis Are Safer Than Human Drivers
And even if self-driving technology proves to be less dangerous, there are many better ways to improve traffic safety and prevent fatal crashes.
www.bloomberg.com
January 6, 2026 at 11:27 PM
This is good.
This stuck with me. Probably because of an accumulation of things over time before it. But I have trouble thinking about SOTA models these days without immediately thinking: yes, very impressive, but they've been given the entire intellectual history of the species, and then some.
This is also quite good. The "doing more with less" definition of intelligence (and "doing less with more" definition of stupidity) has an appeal. Reminds me of an old philosophy professor I had. Brilliant guy. Barely was on email. But his work never suffered for it.
youtu.be/jXa8dHzgV8U?...
youtu.be/jXa8dHzgV8U?...
Intelligent Humans vs. Smart AI: The Ultimate Showdown
YouTube video by Machine Learning Street Talk
youtu.be
January 6, 2026 at 5:58 PM
This stuck with me. Probably because of an accumulation of things over time before it. But I have trouble thinking about SOTA models these days without immediately thinking: yes, very impressive, but they've been given the entire intellectual history of the species, and then some.
January 6, 2026 at 2:17 AM
This goes to the top of the reading list
Do reasoning models have real “Aha!” moments—mid-chain realizations where they intrinsically self-correct?
In a new pre-print, “The Illusion of Insight in Reasoning Models," led by @liv-daliberti.bsky.social we provide strong evidence that they do not!
📜: arxiv.org/abs/2601.00514
In a new pre-print, “The Illusion of Insight in Reasoning Models," led by @liv-daliberti.bsky.social we provide strong evidence that they do not!
📜: arxiv.org/abs/2601.00514
January 5, 2026 at 8:37 PM
This goes to the top of the reading list
Reposted by Vincent Carchidi
claude code is fucking insane
i know literally NOTHING about Hegel. ZERO. and it just built me a complete system of German idealism
i know literally NOTHING about Hegel. ZERO. and it just built me a complete system of German idealism
January 5, 2026 at 6:49 PM
claude code is fucking insane
i know literally NOTHING about Hegel. ZERO. and it just built me a complete system of German idealism
i know literally NOTHING about Hegel. ZERO. and it just built me a complete system of German idealism
Losing a constant poaster like Sky Marchini to harassment isn't great for this app
January 5, 2026 at 6:34 PM
Losing a constant poaster like Sky Marchini to harassment isn't great for this app
This is why you have to post piping hot takes you know people will be annoyed by. Keeps you on your toes
there is a subtle and persistent mental warping effect that occurs when you have 30 to 500 people agreeing with your every opinion no matter what it is
January 5, 2026 at 5:02 PM
This is why you have to post piping hot takes you know people will be annoyed by. Keeps you on your toes
Reposted by Vincent Carchidi
The last year was very hard as my father was dying. It became that much harder when I discovered that AI played a role in amplifying his physical pain, and perhaps hastening the end of his life. It wasn’t easy to write about what happened, but I’ve tried.
open.substack.com/pub/buildcog...
open.substack.com/pub/buildcog...
The role of AI in the death of my father
A sad strange story
open.substack.com
January 5, 2026 at 12:54 PM
The last year was very hard as my father was dying. It became that much harder when I discovered that AI played a role in amplifying his physical pain, and perhaps hastening the end of his life. It wasn’t easy to write about what happened, but I’ve tried.
open.substack.com/pub/buildcog...
open.substack.com/pub/buildcog...
This is something I changed my mind on slowly at first over roughly 4 years, and then abruptly this past year.
I had carved out a cozy niche for engaging with tech policy basically divorced from politics. Tried to fill gaps, where I felt boring research based stuff was called being neglected.
I had carved out a cozy niche for engaging with tech policy basically divorced from politics. Tried to fill gaps, where I felt boring research based stuff was called being neglected.
January 5, 2026 at 3:17 AM
This is something I changed my mind on slowly at first over roughly 4 years, and then abruptly this past year.
I had carved out a cozy niche for engaging with tech policy basically divorced from politics. Tried to fill gaps, where I felt boring research based stuff was called being neglected.
I had carved out a cozy niche for engaging with tech policy basically divorced from politics. Tried to fill gaps, where I felt boring research based stuff was called being neglected.
The period between Thanksgiving and New Year's is Too Much Holiday
January 5, 2026 at 3:04 AM
The period between Thanksgiving and New Year's is Too Much Holiday
One of the maybe less appreciated aspects of, uh, recent US uses of force is that much military R&D is explicitly carried out to build systems that can be "ethically" used.
And, you know, easy to scoff at, but in this context that means building systems with extreme precision, seamless operation
And, you know, easy to scoff at, but in this context that means building systems with extreme precision, seamless operation
January 5, 2026 at 1:06 AM
One of the maybe less appreciated aspects of, uh, recent US uses of force is that much military R&D is explicitly carried out to build systems that can be "ethically" used.
And, you know, easy to scoff at, but in this context that means building systems with extreme precision, seamless operation
And, you know, easy to scoff at, but in this context that means building systems with extreme precision, seamless operation
Reposted by Vincent Carchidi
Some people question whether misalignment in LLMs results from deeper characteristics, like intentionality or consciousness.
Others use this possibility as a bludgeon.
I wrote about this to get the thoughts out of my head and move on to something else.
vincentcarchidi.substack.com/p/against-fa...
Others use this possibility as a bludgeon.
I wrote about this to get the thoughts out of my head and move on to something else.
vincentcarchidi.substack.com/p/against-fa...
Against False Indulgences in LLM Alignment
Unnecessary moral concern, and the false indulgences that drive it.
vincentcarchidi.substack.com
January 4, 2026 at 3:10 PM
Some people question whether misalignment in LLMs results from deeper characteristics, like intentionality or consciousness.
Others use this possibility as a bludgeon.
I wrote about this to get the thoughts out of my head and move on to something else.
vincentcarchidi.substack.com/p/against-fa...
Others use this possibility as a bludgeon.
I wrote about this to get the thoughts out of my head and move on to something else.
vincentcarchidi.substack.com/p/against-fa...
Some people question whether misalignment in LLMs results from deeper characteristics, like intentionality or consciousness.
Others use this possibility as a bludgeon.
I wrote about this to get the thoughts out of my head and move on to something else.
vincentcarchidi.substack.com/p/against-fa...
Others use this possibility as a bludgeon.
I wrote about this to get the thoughts out of my head and move on to something else.
vincentcarchidi.substack.com/p/against-fa...
Against False Indulgences in LLM Alignment
Unnecessary moral concern, and the false indulgences that drive it.
vincentcarchidi.substack.com
January 4, 2026 at 3:10 PM
Some people question whether misalignment in LLMs results from deeper characteristics, like intentionality or consciousness.
Others use this possibility as a bludgeon.
I wrote about this to get the thoughts out of my head and move on to something else.
vincentcarchidi.substack.com/p/against-fa...
Others use this possibility as a bludgeon.
I wrote about this to get the thoughts out of my head and move on to something else.
vincentcarchidi.substack.com/p/against-fa...
Aside from the dunk on Chomsky and Bender, this is a very good, simple resource on what exactly constitutes the deep learning revolution, from RNNs and CNNs to attention in DL and finally transformers.
www.researchgate.net/publication/...
www.researchgate.net/publication/...
(PDF) The Deep Learning Revolution in AI
PDF | The early development of Artificial Intelligence (AI) in the latter half of the twentieth century was marked by limited, hand-crafted systems and... | Find, read and cite all the research you ne...
www.researchgate.net
January 3, 2026 at 8:41 PM
Aside from the dunk on Chomsky and Bender, this is a very good, simple resource on what exactly constitutes the deep learning revolution, from RNNs and CNNs to attention in DL and finally transformers.
www.researchgate.net/publication/...
www.researchgate.net/publication/...
I'm feeling it again
Trying to write a tech policy piece for the first time since maybe March/April, and I just don't know how to write "so and so agency should implement such and such policy" knowing that no such thing could possibly happen right now.
January 3, 2026 at 6:52 PM
I'm feeling it again
On the list of things that don't matter to Americans whose bank accounts are suddenly smaller: Nicolas Maduro.
I guess the US national media has decided there's not much more to say on this, but it doesn't seem great that ACA subsidies expiring and wage garnishing for student loan defaulters are both occurring the same - next - month.
January 3, 2026 at 3:52 PM
On the list of things that don't matter to Americans whose bank accounts are suddenly smaller: Nicolas Maduro.
It would be a good time, not sarcastically, for a number of Americans to realize 2019 isn't coming back
January 3, 2026 at 2:23 PM
It would be a good time, not sarcastically, for a number of Americans to realize 2019 isn't coming back