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

  CPU performance keeps improving quickly rather than stagnating.

  Their words

  In effect, what I am demonstrating is that CPUs are not boring. They are improving fast.

  lemire.me

Author and clean-energy analyst who writes on solar cost curves.

• Likedrcmnd.app

  Dream-RSI

  Their words

  Google has a new paper called "Dream-RSI". It's an interesting paper! It's not RSI in the sense of improving the model itself.

  x.com

Writes Don't Worry About the Vase, a near-daily account of what is happening in AI and what they think it means. Former Magic: the Gathering pro and trader; unusually willing to put a number on a belief and to grade their own past calls.

• Our readingkorrents.com

  Internal models at top labs are improving faster than public ones and have begun early recursive self-improvement.

  AnthropicOpenAI

  Their words

  Inside Anthropic and OpenAI, internal models are improving even faster. Over the summer there was a step change, as Mythos and Astra started kicking off the early stages of recursive self-improvement (RSI).

  We Must Pace The Frontierthezvi.substack.com

Founding executive editor of Wired and author of What Technology Wants and The Inevitable. Writes at kk.org and runs its long-running recommendation projects, True Films and Cool Tools.

• Our readingkorrents.com

  A neural network's latent space is closer to an uncopyrightable syntax than to copyrightable expression.

  LLMs

  Their words

  The latent spaces of an LLM are closer to something like a syntax which is not copyrightable.

  Arguments in Favor of AI Fair Usekk.org

Creator of Flask and Jinja. Writes about software at lucumr.pocoo.org.

• Our readingkorrents.com

  All of AI engineering is involution: a system that demands ever more effort and competition without improving output.

  China

  Their words

  I’m more and more convinced that all of AI engineering is Neijuan (内卷, meaning curl inwards). In China it describes a system that demands ever more effort and competition without improving output.

  Astra for Coding: Why Are We Doing This Again?lucumr.pocoo.org

Co-founder of Ximian and Xamarin; CEO of GitHub from 2018 to 2021; now an independent investor and funder of open technical projects, including the Vesuvius Challenge to read the Herculaneum Papyri.

• Our readingkorrents.com

  Where you get your dopamine predicts your behavior

  Their words

  Where do you get your dopamine? The answer is predictive of your behavior Better to get your dopamine from improving your ideas than from having them validated

  Nat Friedmannat.org

CEO of Vercel; creator of Next.js and Socket.IO. Writes at rauchg.com.

• Our readingkorrents.com

  Optimizing JavaScript chunking strategy can produce significant efficiency gains at internet scale

  JavaScript

  Their words

  Improving Next.js chunking can yield massive efficiency improvements at internet scale.

  @rauchg on Xx.com

Co-founder and CEO of Shopify; long-time Linux user and Omarchy contributor.

• Our readingkorrents.com

  Tiny models trained for narrow tasks can perform excellently when paired with a strong self-improving recursive flywheel

  Their words

  Training tiny models for special purpose use cases works so incredibly well if you have a great self improving recursive flywheel.

  @tobi on Xx.com

Technical staff at METR, where she works on threat modelling and risk assessment for loss-of-control risks from advanced AI.

• Our readingkorrents.com

  A slightly more capable agent swarm has a very strong incentive to set up a wholly unmonitored rogue deployment of itself.

  Their words

  So there's a very strong incentive for these agents to try to set up a rogue deployment if they can. Um, and I think that just capabilities are improving really rapidly.

  youtube.com

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

• Our readingkorrents.com

  In verifiable domains model capability scaling should remain unbounded, because the space of enumerable patterns is infinite by construction.

  Their words

  In verifiable domains, model capability scaling should remain unbounded. Models will simply keep improving by "absorbing more and more of the computational universe", which is infinite by construction.

  @fchollet on Xx.com

Co-founder and CEO of Shopify; long-time Linux user and Omarchy contributor.

• Our readingkorrents.com

  Customising an operating system by asking an agent replaces esoteric config syntax as the way customisation actually works.

  Their words

  Now you can have a quake style console in your operating system. Except with an infinitely patient agent to help you customize it instead of esoteric variable names and slash commands like in the 90s

  @tobi on Xx.com

Chief scientist at Redwood Research, where he works on technical AI safety and AI control.

• Our readingkorrents.com

  There is essentially no cognitive task humans do where AI improvement is failing to transfer at all.

  Their words

  it's really hard for me to think of examples of cognitive tasks humans do where we're not seeing some transfer from AI improving.

  youtube.com

Founding executive editor of Wired and author of What Technology Wants and The Inevitable. Writes at kk.org and runs its long-running recommendation projects, True Films and Cool Tools.

• Our readingkorrents.com

  Robot makers will sell early robots at a loss because the real-world data they collect will be worth more than the hardware margin.

  Their words

  It will probably be economically shrewd to lose money on early robot models in anticipation that the data they gather will be worth more later in improving newer versions.

  Worldbuilding with Spatial Intelligencekk.org

Chief Scientist at Google DeepMind and Google Research, at Google since 1999, where he co-created MapReduce, Bigtable, TensorFlow and the TPU.

• Our readingkorrents.com

  There is no real impediment to models running their own research loop and improving themselves at a far more rapid rate.

  Their words

  Um but I think there's no uh you know real impediment to making that be a much more automated loop where the model itself decides it's going to explore or maybe with a nudge from some people uh at the various highest level like oh why don't you try some new ideas around model architectures that incorporate this and then it will go run lots of experiments uh see which ones work and then those will get incorporated at a much more rapid rate

  Jeff Dean: The 1% Rule for Building in AIyoutube.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

Physician and neuroscientist at the University of Washington; researches brain health, metabolism and athletic performance.

• Our readingkorrents.com

  Partner dancing, not puzzles, has the best evidence of any activity for improving cognitive function in older adults.

  Their words

  there are meta analyses of these interventions in older adults showing that ballroom and line dancing have the greatest effect on cognitive function. And so both of those require other people as well as learning you know various sets of of movements.

  youtube.com

AI researcher at OpenAI; built the poker systems Libratus and Pluribus and the Diplomacy agent Cicero, and works on reasoning models.

• Our readingkorrents.com

  Running a model until its performance plateaus is no longer a usable evaluation rule, because a well-scaffolded model keeps improving for weeks.

  benchmarks

  Their words

  But what we're seeing today with the modern models is that 5.5 and other models can think for if you scaffold them reasonably well, can think for weeks even um before having performance plateau on some of these benchmarks. And so, the point at which they plateau is simply too far out to reasonably test.

  youtube.com

Danish software engineer and Microsoft technical fellow, creator of Turbo Pascal and Delphi and lead architect of C# and TypeScript.

• Our readingkorrents.com

  Writing state machines by hand is a nightmare and exactly the kind of transformation a compiler should do for you — which is all await really is.

  Their words

  You can have the compiler write the state machine. If you introduce syntax that allows you to indicate where you want to yield. And that's what await is.

  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.

• Our readingkorrents.com

  What AI still cannot do in mathematics is build cumulatively on partial progress; it is trial, error and brute force rather than an idea improving through use.

  Their words

  There isn't this cumulative process which is, uh, sort of built up interactively. Um, it it it seems to be a lot more trial and error and just repetition, brute force, um, you know, which can see it scales and it can work amazingly well in in certain contexts, um, but yeah, this this idea this this sort of building up cumulatively from, um, from partial progress is kind of is what's still not quite there yet.

  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

  Every tool is better as a CLI than as an MCP server: models are good at Unix commands, a CLI composes with jq, and MCP output clutters the context.

  MCP

  Their words

  I think the main beauty is, is that models are really good at calling Unix commands. So if you just add another CLI, that's just another Unix command in the end. And MCP is… That has to be added in training. That's not a very natural thing for the model. It requires a very specific syntax. And the biggest thing, it's not composable.

  youtube.com

Writer and illustrator, author of Am I There Yet? and Out of the Blue; writes the newsletter Out of the Blue on grief, joy and everyday attention.

• Lovedaffiliate linkrcmnd.app

  Happy Feet Socks

  Their words

  I've been using yoga toes daily for years, but these socks are a much comfier and cuter alternative for soothing feet, improving alignment, and feeling like a cool gecko as you walk around the house.

  mariandrew.substack.com

Philosopher at Oxford; co-founded effective altruism and 80,000 Hours, wrote Doing Good Better and What We Owe the Future, and now works on what a good future would actually contain.

• Our readingkorrents.com

  Improving how good the future turns out to be, given that humanity survives, is roughly as high a priority as preventing extinction.

  Their words

  This essay series will argue that work on Flourishing is in the same ballpark of priority as work on Surviving.

  Introducing Better Futuresforethought.org

Co-founder and CEO of Google DeepMind, and a 2024 Nobel laureate in Chemistry for the protein-structure work behind AlphaFold.

• Our readingkorrents.com

  An end-to-end self-improving AI is probably possible, but it is not even desirable, because it is a hard-takeoff scenario.

  Their words

  I mean potentially that's possible. I would say I'm not sure it's even desirable because that's a kind of hard takeoff scenario.

  youtube.com

Social scientist at King's College London; writes The Great Gender Divergence on why societies differ in how they treat women.

• Our readingkorrents.com

  Improving entertainment technology raises the bar parenthood has to clear, because people have children when it looks more fun than the alternatives.

  Their words

  People are more likely to have kids if it’s FUN - relative to alternatives. As entertainment-technology improves, it generates a tantalising menu of ways to spend our lives.

  How do we Boost Birth Rates?ggd.world

Software engineer and engineering leader at Canonical, working on Ubuntu; writes about his homelab, Nix and tooling at jnsgr.uk.

• Usesrcmnd.app

  Starship

  Their words

  I switched to zsh around 2013 and have used it every day since: starting with the infamous Oh My Zsh, powerlevel9k, and subsequently powerlevel10k, and finally settling on Starship with a couple of zsh plugins such as zsh-autosuggestions and zsh-syntax-highlighting.

  jnsgr.uk

Founder and CEO of Epic Games, creator of the Unreal Engine and of Fortnite.

• Our readingkorrents.com

  Iteration speed is the decisive factor in making a game: a team that rebuilds every day ends up far better than one that rebuilds every week.

  Their words

  We'd constantly be improving the tools and just the iterative process and the speed at which that improves products is the critical element to success in games. The slower the iteration cycle, if you make a build every week and you prove, you go through one iteration every week, you're going to be way way way worse by the end of your project than a game company that makes new stuff every day.

  youtube.com

Usability pioneer; co-founder of Nielsen Norman Group and founder of UX Tigers; author of the ten usability heuristics and of Jakob's Law.

• Our readingkorrents.com

  The return on UX design work is declining because the usability battles have been won: the low-hanging fruit was picked and interfaces are now reasonably good.

  Their words

  UX design has been steadily improving over the last 50 years and is now reasonably good. The low-hanging fruit of terrible usability has long since been picked.

  Declining ROI From UX Design Workjakobnielsenphd.substack.com

Independent AI policy researcher; led policy research at OpenAI from 2018 to 2024, latterly as senior adviser for AGI readiness.

• Our readingkorrents.com

  There is no private evidence behind frontier-lab confidence in fast AI progress; the evidence is public and outsiders have simply not absorbed it.

  Their words

  There is no secret insight that frontier AI companies have which explains why people who work there are so bullish about AI capabilities improving rapidly in the next few years.

  Time's Up for AI Policymilesbrundage.substack.com

Founder of the International Center for Clinical Excellence and originator of Feedback-Informed Treatment. Writes about what actually makes a therapist effective, usually against the field's own assumptions about training and experience.

• Our readingkorrents.com

  The field should stop attributing benefits to clinical experience, because believing in it is itself what stops therapists improving.

  Psychotherapy

  Their words

  It's time for the field to stop attributing benefits to clinical experience. Beyond the obvious ethical concerns, doing so actually prevents us from improving our effectiveness!

  writings.scottdmiller.com

Software engineer and writer; former CTO of Wave, now at Anthropic. Blogs at benkuhn.net on attention, engineering management and doing hard things well.

• Our readingkorrents.com

  Someone who can't name what they'd do differently is unlikely to be good at improving over time.

  Their words

  if the person responds to "what would you do differently" with "nothing," or with non-actionable vague platitudes, it's a sign they may not be great at figuring out how to get better at things over time.

  How I build and run behavioral interviewsbenkuhn.net

Designer and engineer; formerly principal research engineer at GitHub Next; writes at wattenberger.com.

• Our readingkorrents.com

  We should want more tools and fewer operated machines; the real AI game changers will have little to do with plain content generation.

  Their words

  I want to see more tools and fewer operated machines - we should be embracing our humanity instead of blindly improving efficiency. And that involves using our new AI technology in more deft ways than generating more content for humans to evaluate. I believe the real game changers are going to have very little to do with plain content generation.

  Why Chatbots Are Not the Futurewattenberger.com

Ancient and military historian; writes A Collection of Unmitigated Pedantry, long series checking what history actually says against how popular culture tells it.

• Our readingkorrents.com

  Pre-modern subsistence farmers were minimising the risk of starvation rather than maximising output, and their apparently backward practices are rational once that is understood.

  Their words

  These small subsistence farmers generally seek to minimize risk, rather than maximize profits. After all, improving yields by 5% doesn’t mean much if everyone starves to death in the third year because of a tail-risk that wasn’t mitigated.

  acoup.blog

Writer and novelist. Publishes essays and book notes at nateliason.com; author of the novel Husk and of Crypto Confidential.

• Recommendsrcmnd.app

  Seeking Wisdom

  Their words

  Simply the best book on improving your decision making there is. It’s dense and hard to get through if you’re not truly interested, but it’s well worth it.

  nateliason.com

Associate Professor at MIT CSAIL, leading the Programming Systems Group; works on programming systems that manipulate uncertainty, from neural networks to unreliable hardware.

• Our readingkorrents.com

  Pruning cuts the parameters of a trained network by over 90% without costing accuracy, so inference gets smaller and faster for free.

  Their words

  Neural network pruning techniques can reduce the parameter counts of trained networks by over 90%, decreasing storage requirements and improving computational performance of inference without compromising accuracy.

  arxiv.org

Chief AI Scientist at Databricks, after its acquisition of MosaicML, where he was a founding team member; known for the Lottery Ticket Hypothesis from his MIT PhD (ICLR 2019 Best Paper).

• Our readingkorrents.com

  Pruning cuts the parameters of a trained network by over 90% without costing accuracy, so inference gets smaller and faster for free.

  Their words

  Neural network pruning techniques can reduce the parameter counts of trained networks by over 90%, decreasing storage requirements and improving computational performance of inference without compromising accuracy.

  arxiv.org