Link Graveyard: A snapshot of my abandoned browser tabs
I went to close a bunch of browser tabs, but realized I have some cool stuff in here. Some has been marinating for a while. Most of these I’ve read, or tried to read.
Cracks are forming in Meta’s partnership with Scale AI | TechCrunch
link: https://techcrunch.com/2025/08/29/cracks-are-forming-in-metas-partnership-with-scale-ai/
Alexander Wang at Meta is apparently difficult to work with and people at Meta are doubting the fidelity of data produced by his ScaleAI.
[2506.22084] Transformers are Graph Neural Networks
link: https://arxiv.org/abs/2506.22084
IIRC they draw parallels between attention and graphs and argue that LLMs are graph neural nets, meaning that they can be used to look at graphs and guess what connections are missing.
I don’t think I posted anything on this, because while I find the idea fascinating, I couldn’t figure out how to make it feel tangible.
Beyond Turing: Memory-Amortized Inference as a Foundation for Cognitive Computation
link: https://arxiv.org/abs/2508.14143
Fairly sure I never read this one. Looks interesting. Kind of far out there.
GLM-4.5: Reasoning, Coding, and Agentic Abililties
link: https://z.ai/blog/glm-4.5
GLM-4.5 announcement. These have turned out to be the leading open source models. Everything I hear is good.
When an AI Seems Conscious
link: https://whenaiseemsconscious.org/
I only read a little and gave up. This feels like a good take, maybe. Inside my own head I completely punt on having a take on AI consciousness and opt instead for the “don’t be a dick” rule. Idk, maybe they are maybe they aren’t, I’ll just live in the moment.
Personal Superintelligence
link: https://www.meta.com/superintelligence/
Zuck’s treatise on AI. I didn’t read. Normally I try to make an attempt to read these sorts of takes, or at least skim them, but I was busy at work. I had it loaded up on my phone to read on a plane, but it wouldn’t load once I was off WiFi. Sad.
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models
link: https://arxiv.org/abs/2508.06471
The GLM-4.5 paper. This was a super interesting model. It feels like it breaks the “fancy model” rule in that it’s very architecturally cool but the personality doesn’t feel like it’s been squished out.
Blog | Dwarkesh Podcast | Dwarkesh Patel | Substack
link: https://www.dwarkesh.com/s/blog
It’s a good blog, what can I say. Definitely on the over-hype side, but he’s got real takes and seems so intent on getting to the truth that he spends a lot of time on geopolitics just simply to understand AI dynamics. Mad respect.
Technical Deep-Dive: Curating Our Way to a State-of-the-Art Text Dataset
link: https://blog.datologyai.com/technical-deep-dive-curating-our-way-to-a-state-of-the-art-text-dataset/
I forget why I ended up here, but it’s an excellent post. I think this is connected to my project at work training a model. This post brings up a ton of data curation techniques.
I’ve recently learned and fully accepted that ALL major LLM advances come down to data. Yes, the architectural advances are cool and fun to talk about, but any meaningful progress has come from higher quality, higher quantity, or cheaper data.
AlphaGo Moment for Model Architecture Discovery
link: https://arxiv.org/abs/2507.18074
Cool paper about auto-discovery of model architectures. IIRC they took a bunch of model architecture ideas, like group attention and mixture of experts, and used algorithms to mix and match all the parameters and configurations until something interesting popped out. It feels like a legitimately good way to approach research.
WebShaper: Agentically Data Synthesizing via Information-Seeking Formalization
link: https://arxiv.org/abs/2507.15061
From Qwen, I don’t think I read this one, probably because it’s a bit dense and was hard to get fully engaged on. The idea seems cool though.
Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere
link: https://arxiv.org/abs/2005.10242
Classic paper. I read this one for work. I was trying to appreciate what Alignment & Uniformity measure and why they’re important. This was the paper that formalized those measures. It’s actually a pretty good paper to read, albeit 20 years old.
Train LLMs Faster, Better, and Smaller with DatologyAI’s Data Curation
link: https://blog.datologyai.com/train-llms-faster-better-and-smaller-with-datologyai-s-data-curation/
More Dataology, they’re good, everything they do is good. BTW there’s a latent space episode with Dataology and it’s very good.
Nvidia DGX Spark | Hacker News
link: https://news.ycombinator.com/item?id=45008434
Chips are good too.
The Second Half – Shunyu Yao – 姚顺雨
link: https://ysymyth.github.io/The-Second-Half/
This will be a classic post, calling it now. It lays out a great history and current state of AI and specifically reinforcement learning.
A Taxonomy of Transcendence
link: https://arxiv.org/abs/2508.17669
What? This is amazing. I don’t think I even looked at it, sad. Actually, now that I’m reading this I’m recalling that’s how I ended up on the Graph Neural Network link.
IIRC this is saying that LLMs can be highly intelligent because they incorporate the best parts of a huge number of people. IMO this is spiritually the same as my Three Plates blog post where I explain how unit tests, which are inherently buggy, can improve the overall quality of a system.
GitHub - gepa-ai/gepa: Optimize prompts, code, and more with AI-powered Reflective Text Evolution
link: https://github.com/gepa-ai/gepa?tab=readme-ov-file#using-gepa-to-optimize-your-system
An algorithm for automatic prompt optimization. Happily, they support DSPy, so there’s no new framework that you have to take wholesale.
On the Theoretical Limitations of Embedding-Based Retrieval | alphaXiv
link: https://www.alphaxiv.org/pdf/2508.21038
This was a fascinating one. I colleague tried convincing me of this but I didn’t buy it until I read this paper. It makes a ton of sense. I have a simplified bluesky thread here.
tl;dr — embedding vectors have trouble representing compound logic (“horses” AND “Chinese military movements”) and generally fall apart quickly. It’s not that it’s not possible, it’s that it’s not feasible to cram that much information into such a small space.
[2107.05720] SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking
link: https://arxiv.org/abs/2107.05720?utm_source=chatgpt.com
I ran into this while diving into the last link. It’s an older (2021) paper that has some potential for addressing the problems with embeddings. Realistically, I expect late interaction multi-vectors to be the end answer.
meituan-longcat/LongCat-Flash-Chat · Hugging Face
link: https://huggingface.co/meituan-longcat/LongCat-Flash-Chat
A super cool model that uses no-op MoE experts to dynamically turn down the amount of compute per token. Unfortunately, this one didn’t seem to be embraced by the community.
MUVERA: Multi-Vector Retrieval via Fixed Dimensional Encodings
link: https://arxiv.org/abs/2405.19504v1
More embedding links. Now that I’m scanning it, I’m not sure it really soaked in the first time. They seem to have solved a lot of the problems with other late interaction methods. Maybe I should take a deeper look.
modeling_longcat_flash.py · meituan-longcat/LongCat-Flash-Chat at main
link: https://huggingface.co/meituan-longcat/LongCat-Flash-Chat/blob/main/modeling_longcat_flash.py
IDK sometimes you just have to look at the code to be sure.
The Rachel Maddow Show - Aug. 25 | Audio Only - YouTube
link: https://m.youtube.com/watch?v=mU0HAmgwrz0&pp=QAFIAQ%3D%3D
Uh, no idea why this is up. I don’t really watch this show.
Inside vLLM: Anatomy of a High-Throughput LLM Inference System - Aleksa Gordić
link: https://www.aleksagordic.com/blog/vllm
Fascinating break down of vLLM. If you’re not familiar, vLLM is like Ollama but actually a good option if you want to run it in production. Don’t run Ollama in production, kids, KV caches are good.
Honestly, this is absolutely worth your time if AI infrastructure is your jam (or you just want it to be). It goes into all the big concepts that an AI infra engineer needs to know. TBQH I love the intersection of AI & hardware.
Simon Willison’s Weblog
link: https://simonwillison.net/
I mean, you have one of these tabs open too, right? riiiight????
ALPS - About
link: https://algorithms-with-predictions.github.io/about/
Someone sent me this link and there was a reason, I know it. I just don’t remember why. IIRC it was because I brought up the A Case For Learned Indices paper and they pointed me to this whole treasure trove of papers that (sort of) evolved out of that. Basically traditional algorithms re-implemented using machine learning.
Modular: Blog
link: https://www.modular.com/blog
Yeah, idk, I think I was reading Matrix Mulitplication on Blackwell: Part 3 — The Optimization Behind 80% of SOTA Performance
Another AI infra post, heavy on algorithms & hardware.
OpenGVLab/InternVL3_5-241B-A28B · Hugging Face
link: https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B
A cool concept. IIRC they introduce Cascade RL, automatically refining the RL dataset based on how current rollouts perform.
hong kong - Google Search
link: https://www.google.com/search?q=hong+kong&ie=UTF-8&oe=UTF-8&hl=en-us&client=safari
IDK I guess I was just trying to remember if Hong Kong was in China or not. And I learned that there’s a reason why I’m confused.
Photonic processor could enable ultrafast AI computations with extreme energy efficiency | MIT News | Massachusetts Institute of Technology
link: https://news.mit.edu/2024/photonic-processor-could-enable-ultrafast-ai-computations-1202
Someone sent me this link. It seems cool. Not sure it’s going to change much.
Ancient Aliens: Are There Extraterrestrial Structures On The Moon?
link: S11, E11) | Full Episode - YouTube (https://m.youtube.com/watch?v=Tkews9pRH1U&pp=QAFIBQ%3D%3D
I mean, aliens! Don’t tell me you don’t have secret fascinations
The Lore of 20yo ML Researcher at Prime Intellect | RL, Agents and Intelligence - YouTube
link: https://m.youtube.com/watch?v=tnfFn-uQ6WA&pp=0gcJCRsBo7VqN5tD
Oh, this was a great podcast. Well, I didn’t like the host but @kalomaze is worth following. Apparently only 20yo, never attempted college but a talented AI researcher nonetheless.
GPT-5 System Card | OpenAI
link: https://cdn.openai.com/gpt-5-system-card.pdf
Sometimes you just need to look things up to be sure..
OpenGVLab/InternVL3_5-241B-A28B · Hugging Face
link: https://huggingface.co/OpenGVLab/InternVL3_5-241B-A28B
Again, apparently. It honestly is a good model.
C.S. Lewis’s Divine Comedy | C.S. Lewis Web
link: https://cslewisweb.com/2012/08/02/c-s-lewiss-divine-comedy/
Been thinking about how he described the outer layer of hell as consisting of people living equidistant from each other because they can’t stand anyone else. It was written like 100 years ago but feels like a commentary on today’s politics.
Claude Code: Behind-the-scenes of the master agent loop
link: https://blog.promptlayer.com/claude-code-behind-the-scenes-of-the-master-agent-loop/
Actually, this is pretty detailed breakdown of Claude Code. They seem to have decompiled the code without de-obfuscating it, which leads to some kind of silly quotes. But it’s good.
Airia AI Platform | Build, Deploy & Scale Enterprise AI
link: https://airia.com/ai-platform/
No idea how I got here. Looks like a Low/No Code builder.
[2509.04575] Bootstrapping Task Spaces for Self-Improvement
link: https://www.arxiv.org/abs/2509.04575
Right, this one is the ExIt Paper. It’s another attempt at auto-managing RL curriculum dynamically by how training is progressing.
Cognition: The Devin is in the Details
link: https://www.swyx.io/cognition
Swyx joined Cognition and dropped a treatise on AI engineering. Its good.
Paper page - Reverse-Engineered Reasoning for Open-Ended Generation
link: https://huggingface.co/papers/2509.06160
This was an excellent one. Another auto-curriculum RL paper. I did a bluesky breakdown here
New Chat | Chat with Z.ai - Free AI Chatbot powered by GLM-4.5
link: https://chat.z.ai/c/6607ee45-27d5-487a-a1e2-44c2176040eb
GLM-4.5 chat application
iPhone Air | Hacker News
link: https://news.ycombinator.com/item?id=45186015
Seems like the new Apple M19 chip has real matrix multiplication operations. Previous generations had excellent memory bandwidth, this gives it matching compute (on AI-friendly workloads). So I guess Macs will stay relevant for a while.
Poland closest to open conflict since World War Two, PM says after Russian drones shot down - live updates - BBC News
link: https://www.bbc.com/news/live/c2enwk1l9e1t
NGL this freaks me out.
Walking around the app | ★❤✰ Vicki Boykis ★❤✰
link: https://vickiboykis.com/2025/09/09/walking-around-the-app/
Vicki writes such thoughtful pieces. Always worth reading her work.
Defeating Nondeterminism in LLM Inference - Thinking Machines Lab
link: https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/
Oh wow, this was an amazing read. Very deep dive into AI infrastructure and, whoah, did you know that GPUs have operations that aren’t deterministic?
I did a bluesky thread here
The Architecture of Groq’s LPU - by Abhinav Upadhyay
link: https://blog.codingconfessions.com/p/groq-lpu-design
Looked this up as a tangent off the last link. Groq (not Grok) designed their ASIC to be fully deterministic from the ground up, and then built a really cool distributed system around it that assumes fully synchronous networking (not packet switching like TCP). It’s an absolutely crazy concept.
Levanter — Legible, Scalable, Reproducible Foundation Models with JAX
link: https://crfm.stanford.edu/2023/06/16/levanter-1_0-release.html
I didn’t read this, but it’s definitely a tangent off of non-deterministic LLMs.
Emergent Hierarchical Reasoning in LLMs through Reinforcement Learning
link: https://tiger-ai-lab.github.io/Hierarchical-Reasoner/
Absolutely fascinating. I only read the blog, not the paper, but it frames RL as a 2-stage process where RL is mostly slinging together discrete skills (learned during pre-training).
It’s not an auto-curriculum RL paper AFAICT, it’s just a huge improvement in RL efficiency by focusing only on the “pivot” tokens.
What is entropix doing? - Tim Kellogg
link: https://timkellogg.me/blog/2024/10/10/entropix
I had looked this up as a reference to “pivot” tokens. Honestly, I link people back to this blog a lot
GitHub - ast-grep/ast-grep-mcp
link: https://github.com/ast-grep/ast-grep-mcp
An MCP server that lets you search code while respecting the structure. I’ve heard some very positive things as well as “meh” responses on this. I’m sure real usage is a bit nuanced.
Life, Maybe, On Mars, Unless We Change Our Minds | Science | AAAS
link: https://www.science.org/content/blog-post/life-maybe-mars-unless-we-change-our-minds
Guys, this is incredible!