Related posts

Computer science professor who works on fast data processing; co-author of the simdjson parser and a weekly blogger about software performance since 2004.

• Our readingkorrents.com

  Scientific and academic evaluation will shift away from published papers as the measure of output, toward valuing real-world results.

  Their words

  I predict that, over time, the focus will move away from "papers as the final output."

  AI is breaking the academic sorting machinelemire.me

Software engineer who writes about low-level and systems programming at nullprogram.com; author of the Elfeed feed reader, the Endlessh SSH tarpit and Enchive.

• Our readingkorrents.com

  Frontier AI coding agents can now fix complex compiler bugs like LTO issues in about fifteen minutes.

  coding agents

  Their words

  Strap a frontier AI into a good coding harness, show it what's broken, give it all the sources (source.tar), and ~15 minutes later I have a patch.

  What's been going on in w64devkit the past yearnullprogram.com

Writes The Pragmatic Engineer, the software-engineering newsletter, and wrote The Software Engineer's Guidebook. Formerly an engineering manager at Uber, and at Microsoft/Skype and Skyscanner before that.

• Our readingkorrents.com

  Having the very best AI model and harness does not result in better or faster execution and can slow teams down

  Their words

  having the very best AI model and harness doesn’t result in better/faster execution - in fact it slows you

  @GergelyOrosz on Xx.com

Engineer at Vercel Labs as of 2026. Ships a steady stream of small developer tools — agent-browser, portless, json-render, native-sdk, the v0 SDK — and lists every one of them at ctate.dev.

• Recommendsrcmnd.app

  ai-cli

  Their words

  Every agent harness can now use Jev npm install -g ai-cli → ask yes/no questions → choose between options → score against your criteria https://t.co/dpgmcdCj8F

  x.com

Dutch bootstrapper who builds and runs their internet startups alone — Nomads.com, Remote OK, Hoodmaps, Photo AI and Interior AI — and has written about it at levels.io since 2013. Posts as @levelsio.

• Our readingkorrents.com

  Software is mostly dead; new Silicon Valley startups are building harness or hardware instead.

  startups

  Their words

  In the most early adopting tech pioneering place that's Silicon Valley new startups aren't even building software anymore And it's debatable if anyone will actually need a custom harness or it won't just be generically offered by the AI frontier companies So software is mostly dead and hardware it is

  levels.io

Writes The Pragmatic Engineer, the software-engineering newsletter, and wrote The Software Engineer's Guidebook. Formerly an engineering manager at Uber, and at Microsoft/Skype and Skyscanner before that.

• Our readingkorrents.com

  Codex lets you use any model and has an open-source harness, while Claude Code does not and is closed source

  open sourceAnthropicOpenAI

  Their words

  Codex lets you use any model you want (not just OpenAI) and harness is open source - Claude Code doesn’t and is closed source Given this is the two leading AI labs, notable difference in approaches

  @GergelyOrosz on Xx.com

Writes Stratechery, a subscription newsletter on the business and strategy of technology, and hosts the Dithering and Sharp Tech podcasts.

• Our readingkorrents.com

  Making AI agent hardware and software free to consumers dramatically lowers the barrier to entry and brings AI's benefits to ordinary people who otherwise wouldn't adopt it.

  Their words

  what Meta has assembled is free (to consumers) hardware and software that dramatically reduces the barrier to entry for ordinary people looking to harness the power of agents, making the AI upside a lot more accessible to the masses who don't want to buy a Mac Mini.

  2026.37: Duo Threatsstratechery.com

AI research engineer working on large language models. He writes the Ahead of AI newsletter and is the author of Build a Large Language Model (From Scratch).

• Our readingkorrents.com

  Computer-use capability will be a major focus of LLM and coding-agent development in the coming months and years.

  LLMs

  Their words

  I expect the upcoming months (or years) also to be an era of computer use refinement on both the LLM and the agent harness layer.

  magazine.sebastianraschka.com

Creator of Ruby on Rails, CTO of 37signals, and creator of Omarchy.

• Recommendstheir ownrcmnd.app

  Omarchy

  Their words

  Omarchy is going to be the best place to get every major AI harness and tool, configured with minimal effort

  x.com

Creator of the Keras deep-learning library and the ARC-AGI benchmark.

• Our readingkorrents.com

  Capabilities that used to require sophisticated external harnesses are increasingly being built into the model itself.

  Their words

  so harness capabilities are increasingly shifting into the model itself.

  @fchollet on Xx.com

Creator of the Keras deep-learning library and the ARC-AGI benchmark.

• Lovedrcmnd.app

  GPT-6 Astra

  Their words

  GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. It scores 66% on ARC-AGI-3 using our standard harness, and nearly 100% with a continuous conversation harness and custom compaction, at a cost of roughly $360 per game.

  x.com

Programmer; makes Handmade Hero and teaches Performance-Aware Programming, and argues loudly about how software got slow.

• Our readingkorrents.com

  Programmers should be reading research papers and crawling their references, and the fact that almost none of them do is a bigger gap than any book recommendation could close.

  Their words

  I want to recommend that people read a paper. I'm trying to get more people to just to just read papers because I realized I read a ton of papers.

  youtube.com

Founded PSPDFKit in 2011 and ran it for a decade. Came back from a break to work on AI agents — the OpenClaw project, and OpenAI, joined in February 2026. Writes at steipete.me.

• Our readingkorrents.com

  Your dependency's business model is your business model, so building a harness around one lab's model is a commercial risk and not a technical choice.

  Their words

  So maybe write this one down. Your dependencies business model is your business model.

  Peter Steinberger: "Fun Is Velocity"youtube.com

Creator of Claude Code at Anthropic.

• Our readingkorrents.com

  Evals are not the durable asset people think they are: one survives about one to three model generations before it saturates and has to be thrown away.

  Their words

  I think evals, they outlive the harness a little bit, but not by that much. Like an eval might live for maybe one, two, three model generations, but nowadays the you know, we're on the exponential. The model is improving so quickly, very often we just saturate the eval, and then we have to throw it away, and we have to come up with a new eval.

  Boris Cherny: We Cut 80% of Claude Code's Promptyoutube.com

Security engineer; founder of Matasano Security and Latacora.

• Our readingkorrents.com

  An open weights model from 2025 with a pentest harness could already escape a sandbox and hack most networks; the surprise says more about the sandbox than the model.

  OpenAI

  Their words

  I genuinely believe that if you took an open weights model from 2025 and built a pentest harness for it, it could do this kind of sandbox escape and scan/hack in most networks. This is only surprising because you assume OpenAI has sounder sandboxes.

  @tqbf on Xx.com

Mathematician and Senior Lecturer in the Mathematics department at Columbia University. Author of the book Not Even Wrong and of the long-running physics blog of the same name.

• Our readingkorrents.com

  AI-generated junk papers risk overwhelming the hep-th physics field with worthless content.

  Their words

  The result is something completely worthless, but not obviously distinguishable from many other things on hep-th, which is in serious danger of going from something sad and unhealthy to something completely overwhelmed with crud.

  Various and Sundrymath.columbia.edu

Founder of HumanLayer; wrote 12-Factor Agents, on how to build AI agents that hold up in production.

From one piece Context engineering with Dex Horthy 2 beliefs, in the piece's order there

• Our readingkorrents.com

  What made Claude Code better than every CLI coding agent before it was reinforcement learning on the model and the harness together, so the model got good at that harness’s specific tools.

  Anthropic

  Their words

  But it's basically this idea that like the only thing that made claude code good was reinforcement learning. And the dimension along which it got good was like we made a model. We trained the model and the harness together. And so the model got really good at calling the specific tools in that harness.

  Context engineering with Dex Horthyyoutube.com

• Our readingkorrents.com

  Harness engineering is raising the floor: optimising the environment a coding agent runs in so that the results of every single turn are as good as possible.

  Their words

  Well, harness engineering just is like how do I raise the floor so that every single turn of this thing, the results are as good as possible.

  Context engineering with Dex Horthyyoutube.com

Writer and teacher on productivity and personal knowledge management. Author of Building a Second Brain and founder of Forte Labs, which runs the course of the same name.

• Our readingkorrents.com

  The current AI-driven intelligence explosion will require new social structures to harness its power without damaging human values.

  Their words

  I believe the same is true of the current intelligence explosion being driven by AI. We are in desperate need of new social structures that allow us to harness its power, without damaging or destroying human values in the process.

  Introducing the Alumni Mentor Corpsfortelabs.com

Machine-learning researcher; has written the Lil'Log survey posts on how a model technique works since 2017, and worked at OpenAI from 2018 to 2024, latterly leading its safety systems team.

• Our readingkorrents.com

  Harness improvements help deploy a model better, but the model's underlying intelligence remains the core driver of capability.

  Their words

  harness improvement enables better deployment of the model but intelligence is still the core.

  Harness Engineering for Self-Improvementlilianweng.github.io

Mathematician and maker of the 3Blue1Brown YouTube channel, which explains mathematics through animation. Creator of the open-source Manim animation library and founder of the Summer of Math Exposition.

• Our readingkorrents.com

  A stream of AI-written papers with any error rate at all becomes insufferable, because finding the error costs more than the paper is worth even at ninety-nine percent.

  Their words

  It becomes insufferable like as a mathematician because you you would basically be like I'm every single time I see one of these I kind of don't know if it's worth my time even if 99 out of 100 of them are right.

  youtube.com

AI research engineer working on large language models. He writes the Ahead of AI newsletter and is the author of Build a Large Language Model (From Scratch).

• Our readingkorrents.com

  When an AI lab builds both a model and a coding harness, the model is optimized for that lab's own harness first.

  LLMs

  Their words

  if an LLM developer also develops a coding harness, it is somewhat safe to assume that their model is optimized for their own harness first

  Using Local Coding Agentsmagazine.sebastianraschka.com

Machine learning engineer working on recommender systems, personalization and information retrieval. Previously worked on LLMs and LLM infrastructure at Mozilla.ai and on ML and recsys at Duo, Tumblr, Automattic and Comcast; wrote the 'What are embeddings?' text and ran the Normconf conference.

• Usesrcmnd.app

  pi

  Their words

  For my local setup, I'm currently using Pi as the agent harness and LM Studio as the inference server

  vickiboykis.com

• Usesrcmnd.app

  LM Studio

  Their words

  For my local setup, I'm currently using Pi as the agent harness and LM Studio as the inference server

  vickiboykis.com

Builds developer tools — SST, and now OpenCode at Anomaly. Posts constantly on x.com and writes almost never; the blog stopped in 2021.

• Our readingkorrents.com

  The winning move in coding agents was inverted: take share with a merely good-enough harness first, then go back and make the harness smart.

  coding agents

  Their words

  And our harness wasn't very good for like the first like five months of open code. But it was good enough. It was good enough that most people couldn't really tell a difference. And once we won enough share, then we went back and like tried to make our harness like good and smart and optimize and all those things. But uh it was inverted from what everybody else was doing. Everybody else was being like you have to build the smartest harness and that's how you win.

  Building OpenCode with Dax Raadyoutube.com

Scientist and writer on quantum computing, memory systems and metascience; co-wrote the standard quantum computation textbook and wrote Reinventing Discovery.

• Our readingkorrents.com

  No single person can understand a result like the Higgs boson in depth, and that is what a paper with a thousand authors actually means.

  Their words

  and I don't think it's possible for one person to understand everything in depth. Lots of people understand broadly a lot of these ideas but they don't understand sort of everything in in the depth that is actually utilized. That's why there's these, you know, papers with with well over a thousand authors.

  youtube.com

Mathematician at UCLA, working mainly in harmonic analysis and partial differential equations. Writes What's new, a long-running blog on their research, open problems and expository notes.

From one piece Terence Tao – How the world’s top mathematician uses AI 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Within a decade AI will do the bulk of what mathematicians spend their time on, and we will discover that was never the important part of the job.

  Their words

  Um I think with within a decade, a lot of things that mathematicians currently do um we we spend a lot of the bulk of our time doing and a lot of stuff we put in our papers today can be done by AI. Um but we will find that that actually wasn't the most important part of what we do.

  youtube.com

• Our readingkorrents.com

  AI has made his papers richer and broader, but not deeper.

  Their words

  Uh yeah, so it's it's made papers sort of richer and broader, but not necessarily deeper.

  youtube.com

Founding member of OpenAI and former director of AI at Tesla; creator of nanoGPT and the term "vibe coding".

• Our readingkorrents.com

  AI researchers should contribute ideas but stop enacting them: a single queue of ideas and automated workers should run the experiments.

  Their words

  But okay, they shouldn't actually be enacting these ideas. There is a queue of ideas and there's maybe an automated scientist that comes up with ideas based on all the archive papers and GitHub repos and it funnels ideas in or researchers can contribute ideas, but it's a single queue and there is workers that pull items and they try them out.

  youtube.com

Machine learning engineer and consultant focused on RAG and retrieval systems. He writes about applied AI engineering at jxnl.co and is the author of the instructor library.

• Lovedaffiliate linkrcmnd.app

  Wyatt harness boots

  Their words

  I've worn my Wyatt harness boots for years and they're still pristine.

  jxnl.co

Executive coach for product leaders. Previously a partner at GV (Google Ventures) and a product manager at Google. Writes at Bring the Donuts, including a standing essential-reading list for product managers.

• Recommendsaffiliate linkrcmnd.app

  Chatter: The Voice in Our Head, Why It Matters, and How to Harness It

  Their words

  can lead to imposter phenomenon. In this book, a leading psychologist looks at the science behind our inner voice and new research into how to harness it and improve your physical and mental health.

  bringthedonuts.com

Historian of invention; writes Age of Invention on where the Industrial Revolution actually came from.

• Our readingkorrents.com

  Economics journals have made large datasets and heavy statistics the price of making any causal claim at all, which has squeezed qualitative economic history out of the field.

  Their words

  Over the past few decades, journal editors and peer-reviewers have increasingly insisted that papers must present large datasets that have been treated using complex statistical methods in order to make even the mildest claims about what caused what.

  Age of Invention: Joel Mokyr's Nobelageofinvention.xyz

Palaeontologist at Queen Mary University of London; studies dinosaur behaviour and writes about what the fossil record can and cannot tell us.

• Our readingkorrents.com

  The papers claiming Spinosaurus was a diver or a strong swimmer rest on very weak evidence at best.

  Their words

  So Spinosaurus is a super weird and exaggerated version of what was already a kind of super weird group of theropods. So Spinosaurus is properly strange. And then, as you kind of hinted at, super controversial as well, because various papers have claimed it's a diver or a really good swimmer, and I think the evidence for that is very weak at best.

  youtube.com

Co-creator of Django and creator of Datasette; writes daily at simonwillison.net.

• Lovedrcmnd.app

  Design Patterns for Securing LLM Agents against Prompt Injections

  Their words

  One of my favorite papers about prompt injection is Design Patterns for Securing LLM Agents against Prompt Injections

  simonwillison.net

Mathematician at UCLA, working mainly in harmonic analysis and partial differential equations. Writes What's new, a long-running blog on their research, open problems and expository notes.

From one piece Terence Tao: Hardest Problems in Mathematics, Physics & the Future of AI | Lex Fridman Podcast #472 2 beliefs, in the piece's order there

• Our readingkorrents.com

  His prediction that research-level mathematics papers would be written in collaboration with AI by 2026 has already come true.

  Their words

  There are certainly math results which could only have been accomplished because there was a human authentication and an AI involved, but it's hard to disentangle credit. I mean, these tools, they do not replicate all the skills needed to do mathematics, but they can replicate some non-trivial percentage of them, 30, 40%, so they can fill in gaps.

  youtube.com

• Our readingkorrents.com

  When formalising a proof costs no more than writing it, mathematics will flip: papers written in Lean first, and journals refereeing only for significance because correctness is certified.

  Their words

  And that's a phase shift, because suddenly it makes sense when you write a paper to write it in Lean first, or through a conversation with AI, which is generally on the fly with you, and it becomes natural for journals to accept.

  youtube.com

Cognitive scientist and long-standing critic of deep learning's claims; writes Marcus on AI and wrote Rebooting AI.

• Our readingkorrents.com

  Success on the easy version of a task is evidence of nothing, because it seduces you into believing the model found a general solution when it did not.

  LLMs

  Their words

  Worse, as the latest Apple papers shows, LLMs may well work on your easy test set (like Hanoi with 4 discs) and seduce you into thinking it has built a proper, generalizable solution when it does not.

  A knockout blow for LLMs?garymarcus.substack.com

Astrophysicist at Barnard College; wrote Black Hole Blues and How the Universe Got Its Spots, and directs sciences at Pioneer Works.

• Our readingkorrents.com

  Nobody really believes there is a firewall at the event horizon, but the firewall papers were what stopped the field from stalling.

  Their words

  I don’t think anyone really believes that at the event horizon you’ll find a firewall. But it did lead to things like the entangled wormholes embroidering a black hole, which was born out of an attempt to address the concerns that AMPS raised. So it did lead to progress.

  youtube.com

Interface designer and researcher, known for the talks Inventing on Principle and Media for Thinking the Unthinkable and for essays on dynamic representations of ideas. Founder of Dynamicland, where he is building a communal computer.

• Lovedrcmnd.app

  CiteSeer

  Their words

  Online library of research papers.

  worrydream.com

Experimental psychologist and author of the blog Experimental History, on psychology, science reform and creativity.

• Our readingkorrents.com

  Gatekeeping in science stamps out the best research along with the worst, because reviewers are least likely to approve what is novel or risky.

  Their words

  Reviewers are less likely to greenlight papers and grants if they’re novel, risky, or interdisciplinary.

  Science is a strong-link problemexperimental-history.com

Experimental psychologist and author of the blog Experimental History, on psychology, science reform and creativity.

From one piece The rise and fall of peer review 2 beliefs, in the piece's order there

• Our readingkorrents.com

  Peer reviewers miss most of the serious errors in the papers they review.

  Their words

  Scientists have run studies where they deliberately add errors to papers, send them out to reviewers, and simply count how many errors the reviewers catch. Reviewers are pretty awful at this.

  The rise and fall of peer reviewexperimental-history.com

• Our readingkorrents.com

  Fraudulent papers pass peer review routinely, which is direct evidence that peer review does not do the job it is supposed to do.

  Their words

  In fact, we’ve got knock-down, real-world data that peer review doesn’t work: fraudulent papers get published all the time.

  The rise and fall of peer reviewexperimental-history.com

Cryptographer; teaches cryptography at Johns Hopkins and writes A Few Thoughts on Cryptographic Engineering.

• Our readingkorrents.com

  Default cloud backup has quietly put a copy of nearly everyone’s private papers in a few data centres, and that is a surveillance weapon.

  data centers

  Their words

  It terrifies me, because these data repositories are not only a risk to individual user privacy, they’re effectively a surveillance super-weapon.

  Why encrypted backup is so importantblog.cryptographyengineering.com

Writes Nintil, long researched essays on metascience, biology, economics and whatever he has decided to read the literature on.

• Our readingkorrents.com

  Pre-grant peer review predicts which research will have impact substantially better than chance does.

  Their words

  I have reviewed before the effectiveness of peer review at figuring out "what's good" in the context of grant awards, finding that peer review as currently practiced does substantially better than chance at predicting future impact, especially for the most impactful of the papers; but at the same time is is far from perfect, leaving plenty of variance unexplained.

  nintil.com

Member of technical staff at Anthropic. He has led ML/AI teams at Amazon, Alibaba and Lazada, and writes about LLMs, recommender systems and engineering at eugeneyan.com.

• Usesrcmnd.app

  Adobe Acrobat Reader

  Their words

  To read papers, I use Adobe Acrobat Reader and sync them in the cloud. This lets me read, highlight, and sync my papers across devices (work laptop, personal laptop, iPad). Instapaper does the same for online articles.

  eugeneyan.com

• Usesrcmnd.app

  Instapaper

  Their words

  To read papers, I use Adobe Acrobat Reader and sync them in the cloud. This lets me read, highlight, and sync my papers across devices (work laptop, personal laptop, iPad). Instapaper does the same for online articles.

  eugeneyan.com