Daniel Mewes
@dmewes.com
Computer scientist. Interested in technology, artificial and natural intelligence, emergent complexity, among other things. Blogging at amongai.com.
Currently at Imbue. Previously Ambient.ai, Stripe, RethinkDB, Max Planck Institute.
Currently at Imbue. Previously Ambient.ai, Stripe, RethinkDB, Max Planck Institute.
Poetiq's methodology on top of Gemini 3 & GPT 5.1 exceeds average human performance on ARC-AGI-2!
This is huge.
Only caveat is that they evaluated on the public set - it might have been used in post training of Gemini 3? Looking forward to see private eval results! poetiq.ai/posts/arcagi...
This is huge.
Only caveat is that they evaluated on the public set - it might have been used in post training of Gemini 3? Looking forward to see private eval results! poetiq.ai/posts/arcagi...
Traversing the Frontier of Superintelligence
Poetiq is proud to announce a major milestone in AI reasoning. We have established a new state-of-the-art (SOTA) on the ARC-AGI-1 & 2 benchmarks, significantly advancing both the performance and the e...
poetiq.ai
November 21, 2025 at 5:56 PM
Poetiq's methodology on top of Gemini 3 & GPT 5.1 exceeds average human performance on ARC-AGI-2!
This is huge.
Only caveat is that they evaluated on the public set - it might have been used in post training of Gemini 3? Looking forward to see private eval results! poetiq.ai/posts/arcagi...
This is huge.
Only caveat is that they evaluated on the public set - it might have been used in post training of Gemini 3? Looking forward to see private eval results! poetiq.ai/posts/arcagi...
There have been a few works that showed signs of LLM introspection since I wrote my article. Though the consensus still seems to be that it's quite unreliable. amongai.com/2024/12/24/l...
October 29, 2025 at 7:01 PM
There have been a few works that showed signs of LLM introspection since I wrote my article. Though the consensus still seems to be that it's quite unreliable. amongai.com/2024/12/24/l...
Reposted by Daniel Mewes
This was a tough but necessary decision - I posted my own notes on this here, from the perspective of a current PSF board member simonwillison.net/2025/Oct/27/...
October 27, 2025 at 8:34 PM
This was a tough but necessary decision - I posted my own notes on this here, from the perspective of a current PSF board member simonwillison.net/2025/Oct/27/...
I have to admit the mechanism behind cable impedance and impedance matching never really clicked for me, despite being a licensed radio amateur for 20 years.
...until I watched this video by AlphaPhoenix - it's such an amazing visualization of what's going on in a cable! youtu.be/RkAF3X6cJa4?...
...until I watched this video by AlphaPhoenix - it's such an amazing visualization of what's going on in a cable! youtu.be/RkAF3X6cJa4?...
What does "impedance matching" actually look like? (electricity waves)
YouTube video by BetaPhoenix
youtu.be
October 24, 2025 at 3:22 PM
I have to admit the mechanism behind cable impedance and impedance matching never really clicked for me, despite being a licensed radio amateur for 20 years.
...until I watched this video by AlphaPhoenix - it's such an amazing visualization of what's going on in a cable! youtu.be/RkAF3X6cJa4?...
...until I watched this video by AlphaPhoenix - it's such an amazing visualization of what's going on in a cable! youtu.be/RkAF3X6cJa4?...
Claude Haiku 4.5 outperforms Sonnet 4 in some coding benchmarks (such as SWE-bench Verified). This is exciting, since it's 1/3 the price.
However, in my actual use, I've found it to be a bit underwhelming compared to Sonnet 4 (not to mention Sonnet 4.5).
What has been your experience with Haiku?
However, in my actual use, I've found it to be a bit underwhelming compared to Sonnet 4 (not to mention Sonnet 4.5).
What has been your experience with Haiku?
October 23, 2025 at 5:35 PM
Claude Haiku 4.5 outperforms Sonnet 4 in some coding benchmarks (such as SWE-bench Verified). This is exciting, since it's 1/3 the price.
However, in my actual use, I've found it to be a bit underwhelming compared to Sonnet 4 (not to mention Sonnet 4.5).
What has been your experience with Haiku?
However, in my actual use, I've found it to be a bit underwhelming compared to Sonnet 4 (not to mention Sonnet 4.5).
What has been your experience with Haiku?
This is so true.
I think almost nobody is talking about symbolic GOFAI these days, so I'm not concerned about that.
But all of ML being re-branded to AI lately, while AI has simultaneously been made synonymous with generative AI in many places, has led to so much confusion.
I think almost nobody is talking about symbolic GOFAI these days, so I'm not concerned about that.
But all of ML being re-branded to AI lately, while AI has simultaneously been made synonymous with generative AI in many places, has led to so much confusion.
The fallout from the fact that data science/classical machine learning & generative AI are both called "AI" has been remarkably broad & persistent
Policy addresses the wrong harms, companies have been confused about who should lead efforts, hiring is misguided, academic discussion is often muddled.
Policy addresses the wrong harms, companies have been confused about who should lead efforts, hiring is misguided, academic discussion is often muddled.
October 22, 2025 at 5:51 PM
This is so true.
I think almost nobody is talking about symbolic GOFAI these days, so I'm not concerned about that.
But all of ML being re-branded to AI lately, while AI has simultaneously been made synonymous with generative AI in many places, has led to so much confusion.
I think almost nobody is talking about symbolic GOFAI these days, so I'm not concerned about that.
But all of ML being re-branded to AI lately, while AI has simultaneously been made synonymous with generative AI in many places, has led to so much confusion.
The age of single-serving, disposable (simple) software is here. I now frequently have LLMs write software for me that I use exactly once.
Often to convert some data or create visualizations, but also one-off new features to add to some open-source application that I only want to use once.
Often to convert some data or create visualizations, but also one-off new features to add to some open-source application that I only want to use once.
October 8, 2025 at 12:07 AM
The age of single-serving, disposable (simple) software is here. I now frequently have LLMs write software for me that I use exactly once.
Often to convert some data or create visualizations, but also one-off new features to add to some open-source application that I only want to use once.
Often to convert some data or create visualizations, but also one-off new features to add to some open-source application that I only want to use once.
Reposted by Daniel Mewes
new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process.
link in reply!
link in reply!
October 5, 2025 at 2:36 PM
new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process.
link in reply!
link in reply!
Today we're releasing Sculpture, an AI coding tool that combines async agents with the ability to collaborate with agents locally.
It also comes with built-in verifiers to automatically check the quality of AI written code. More to come! imbue.com/sculptor-ann...
It also comes with built-in verifiers to automatically check the quality of AI written code. More to come! imbue.com/sculptor-ann...
September 30, 2025 at 5:04 PM
Today we're releasing Sculpture, an AI coding tool that combines async agents with the ability to collaborate with agents locally.
It also comes with built-in verifiers to automatically check the quality of AI written code. More to come! imbue.com/sculptor-ann...
It also comes with built-in verifiers to automatically check the quality of AI written code. More to come! imbue.com/sculptor-ann...
I used AI vibe coding tools to port Anthropic's API client to Common Lisp: github.com/danielmewes/...
It was a very quick and fun mini project that taught me a thing or two about Lisp. Use it at your own risk.
It was a very quick and fun mini project that taught me a thing or two about Lisp. Use it at your own risk.
GitHub - danielmewes/anthropic-sdk-cl-port: An AI-written port of the Anthropic client SDK to Common Lisp.
An AI-written port of the Anthropic client SDK to Common Lisp. - danielmewes/anthropic-sdk-cl-port
github.com
September 22, 2025 at 4:42 PM
I used AI vibe coding tools to port Anthropic's API client to Common Lisp: github.com/danielmewes/...
It was a very quick and fun mini project that taught me a thing or two about Lisp. Use it at your own risk.
It was a very quick and fun mini project that taught me a thing or two about Lisp. Use it at your own risk.
I feel like each of Anthropic's three post mortems is missing some key explanation step in their root cause descriptions. www.anthropic.com/engineering/... 🧵
A postmortem of three recent issues
This is a technical report on three bugs that intermittently degraded responses from Claude. Below we explain what happened, why it took time to fix, and what we're changing.
www.anthropic.com
September 18, 2025 at 5:47 PM
I feel like each of Anthropic's three post mortems is missing some key explanation step in their root cause descriptions. www.anthropic.com/engineering/... 🧵
My Hulu / Disney+ subscription will be preempted indefinitely.
September 18, 2025 at 5:14 AM
My Hulu / Disney+ subscription will be preempted indefinitely.
Reposted by Daniel Mewes
As AI systems keep getting better at very hard problems while getting more opaque, the way that we work with AI is shifting shifting from being collaborators who shape the process to being supplicants who receive the output.
I discussed what that means. www.oneusefulthing.org/p/on-working...
I discussed what that means. www.oneusefulthing.org/p/on-working...
On Working with Wizards
Verifying magic on the jagged frontier
www.oneusefulthing.org
September 11, 2025 at 8:55 PM
As AI systems keep getting better at very hard problems while getting more opaque, the way that we work with AI is shifting shifting from being collaborators who shape the process to being supplicants who receive the output.
I discussed what that means. www.oneusefulthing.org/p/on-working...
I discussed what that means. www.oneusefulthing.org/p/on-working...
Reposted by Daniel Mewes
Meta trained a special “aggregator” model that learns how to combine and reconcile different answers into a more accurate final one, instead of relying on simple majority voting or reward model ranking on multiple model answers.
September 9, 2025 at 2:03 PM
Meta trained a special “aggregator” model that learns how to combine and reconcile different answers into a more accurate final one, instead of relying on simple majority voting or reward model ranking on multiple model answers.
Reposted by Daniel Mewes
The funny thing about the prediction that AI would be writing 90% of all code by now is that the prediction's failure distracts from the fact that AI adoption in code writing is actually extremely high, it was already over 30% in December, 2024 according to one measure, with large economic impact.
September 3, 2025 at 4:19 PM
The funny thing about the prediction that AI would be writing 90% of all code by now is that the prediction's failure distracts from the fact that AI adoption in code writing is actually extremely high, it was already over 30% in December, 2024 according to one measure, with large economic impact.
I remember seeing this graph and thinking it was about the influence of training data on LLM responses as well. Quite misleading.
This chart is everywhere and is being horribly misinterpreted.
This is not where the training data for AI comes from, it is a study done by a SEO firm that claims to show how often sites come up at least once in THE WEB SEARCH FUNCTION of certain AI agents when they do a web search for more info.
This is not where the training data for AI comes from, it is a study done by a SEO firm that claims to show how often sites come up at least once in THE WEB SEARCH FUNCTION of certain AI agents when they do a web search for more info.
September 2, 2025 at 3:52 AM
I remember seeing this graph and thinking it was about the influence of training data on LLM responses as well. Quite misleading.
I think we're starting to see diminishing returns from LLM pre- and post-training. The limitations of today's LLMs are unlikely to just disappear with the next bigger model.
This is not all bad: we can start focusing on how to work around those limitations and how to put current LLMs to work.
This is not all bad: we can start focusing on how to work around those limitations and how to put current LLMs to work.
August 29, 2025 at 4:25 PM
I think we're starting to see diminishing returns from LLM pre- and post-training. The limitations of today's LLMs are unlikely to just disappear with the next bigger model.
This is not all bad: we can start focusing on how to work around those limitations and how to put current LLMs to work.
This is not all bad: we can start focusing on how to work around those limitations and how to put current LLMs to work.
Just experienced phantom breaking for the first time in my Kia with HDA2 (Highway Drive Assist) and it made me sentimental about the exciting times of early Tesla Autopilot 😢
August 29, 2025 at 4:00 PM
Just experienced phantom breaking for the first time in my Kia with HDA2 (Highway Drive Assist) and it made me sentimental about the exciting times of early Tesla Autopilot 😢
Thanks to the genius of a well-thought-out tariff policy, manufacturing trucks (for the US market) in Mexico now incurs lower tariffs than manufacturing them domestically. 🤦 www.reuters.com/business/aut...
August 27, 2025 at 4:21 AM
Thanks to the genius of a well-thought-out tariff policy, manufacturing trucks (for the US market) in Mexico now incurs lower tariffs than manufacturing them domestically. 🤦 www.reuters.com/business/aut...
Reposted by Daniel Mewes
What if we could evolve AI models like organisms, letting them compete, mate, and combine their strengths to produce ever-fitter offspring?
Excited to share our new paper, “Competition and Attraction Improve Model Fusion” presented at GECCO 2025 (runner-up for best paper)!
arxiv.org/abs/2508.16204
Excited to share our new paper, “Competition and Attraction Improve Model Fusion” presented at GECCO 2025 (runner-up for best paper)!
arxiv.org/abs/2508.16204
August 25, 2025 at 2:48 AM
What if we could evolve AI models like organisms, letting them compete, mate, and combine their strengths to produce ever-fitter offspring?
Excited to share our new paper, “Competition and Attraction Improve Model Fusion” presented at GECCO 2025 (runner-up for best paper)!
arxiv.org/abs/2508.16204
Excited to share our new paper, “Competition and Attraction Improve Model Fusion” presented at GECCO 2025 (runner-up for best paper)!
arxiv.org/abs/2508.16204
Study finds that polarization and echo chambers on social media may be inherent to the medium and not due to algorithms. www.science.org/content/arti...
Don’t blame the algorithm: Polarization may be inherent in social media
In simulations, AI-generated users of stripped-down social media without content algorithms still split into polarized echo chambers
www.science.org
August 24, 2025 at 5:11 PM
Study finds that polarization and echo chambers on social media may be inherent to the medium and not due to algorithms. www.science.org/content/arti...
Reposted by Daniel Mewes
Impressive result: 99.9% on AIME 2025 with both open models and 85% fewer tokens!
The trick? Measuring confidence in the reasoning traces as they are being generated, then ending low confidence traces early.
jiaweizzhao.github.io/deepconf/
The trick? Measuring confidence in the reasoning traces as they are being generated, then ending low confidence traces early.
jiaweizzhao.github.io/deepconf/
Deep Think with Confidence
Deep Think with Confidence (DeepConf): A simple yet powerful method that significantly improves both reasoning efficiency and performance at test time by leveraging model-internal confidence signals.
jiaweizzhao.github.io
August 24, 2025 at 1:28 PM
Impressive result: 99.9% on AIME 2025 with both open models and 85% fewer tokens!
The trick? Measuring confidence in the reasoning traces as they are being generated, then ending low confidence traces early.
jiaweizzhao.github.io/deepconf/
The trick? Measuring confidence in the reasoning traces as they are being generated, then ending low confidence traces early.
jiaweizzhao.github.io/deepconf/
Reposted by Daniel Mewes
It seems like there is not enough of a policy response to the existence of self-driving cars, with 57M miles of data, Waymo’s autonomous vehicles experience 85% less serious injuries & 79% less injuries overall than cars with human drivers
2.4 million are injured & 40k killed in US accidents a year
2.4 million are injured & 40k killed in US accidents a year
August 23, 2025 at 1:45 PM
It seems like there is not enough of a policy response to the existence of self-driving cars, with 57M miles of data, Waymo’s autonomous vehicles experience 85% less serious injuries & 79% less injuries overall than cars with human drivers
2.4 million are injured & 40k killed in US accidents a year
2.4 million are injured & 40k killed in US accidents a year
One thing I've learned is that when Americans call something "European style", it's 100% not like anything I've ever seen or eaten in Europe. I think it's just an attribute used to mean "a bit different from how we usually do it in America"?
August 20, 2025 at 3:01 AM
One thing I've learned is that when Americans call something "European style", it's 100% not like anything I've ever seen or eaten in Europe. I think it's just an attribute used to mean "a bit different from how we usually do it in America"?
Reposted by Daniel Mewes
This is clearly where this is all going
August 17, 2025 at 10:00 PM
This is clearly where this is all going