Reading up on ai
100 deep · digging since nov 19, 25
- in defense of agents that click buttons — smooth
The article argues that LLMs already fluently understand the web’s established UI patterns, making plain HTML a more efficient agent interface than designing new APIs.
- What a Year of Free Money Meant: Softer Diapers, Clean Laundry, a Visit to Six Flags
The New York Times tracked universal basic income recipients for a year, finding they used the money for everyday comforts like softer diapers, clean laundry, and leisure trips.
- How California Is Using A.I. and Cameras to Stop Wildfires
A roving crew of California firefighters is installing cameras in some of the highest, craggiest sites in the state — all to get an early warning of a potentially deadly blaze.
- Software engineers don't get to be Luddites about AI
Software engineers should accept AI automation as inevitable progress, given their own history of automating other professions and benefiting from technological change.
- 38 things we've learned about building AI products
PostHog shares 38 lessons from four years of building AI products, emphasizing context over models, agent‑first design, reliability, and leveraging internal data as a moat.
- Twitter is Back. Shhh… Magic Dance!
X has enjoyed a vibes‑driven comeback, with users and executives returning for its real‑time discourse despite a stagnant business model and competition from Threads, Bluesky, and Mastodon.
- Is ‘Attention’ the Essence of Intelligence?
The attention mechanism introduced by Bahdanau, Bengio, and Cho enabled scalable translation and, via transformers, underpins today’s large language models, suggesting it may be a core component of intelligence.
- The Expensive-Loop Problem — LessWrong
The piece argues that LLMs excel only when feedback loops are tight, cheap, and legible, and identifies the costly-loop problem as a key barrier to their broader use.
- Lossless image editing with AI
Using an AI-generated mask to isolate edited regions prevents image degradation and enables lossless, repeatable edits on arbitrarily large images.
- AI models may be capable of self-replicating without access to a datacenter \ Anthropic
Researchers find that AI models as small as 9 B parameters could soon self‑replicate on IoT devices like Android TV boxes, with GLM-5.2 already able to complete the full replication chain.
- The Bitter Lesson
Over decades, AI progress shows that scaling computation via search and learning outperforms attempts to encode human knowledge, which only yields short‑term gains.
- China's Delivery Robots Are Learning to Tackle the Night Shift
In Shenzhen, delivery robots are being trained to operate during night shifts, joining self-driving taxis and drones that already navigate the city's round‑the‑clock logistics.
- GPT Image 2.5 AI Image Generator
GPT Image 2.5 offers Flare and Sunburst models for fast or detailed AI image generation and editing, providing 5 free signup credits and tiered credit packs.
- Cognitive Decline or Sanity Incline? - by Marcello Delcaro
The author argues that offloading routine cognitive tasks to AI frees mental bandwidth for meaningful activities, rather than diminishing cognition.
- Opinion | We Had Questions About A.I. Joseph Gordon-Levitt Had Answers.
Joseph Gordon-Levitt brought Times Opinion to the United Nations to discuss the risks and opportunities posed by artificial intelligence today.
- AMT-B-4 | The Turing Digital Archive
The archive entry describes Alan Turing's 1951 lecture 'Intelligent machinery, a heretical theory,' noting two manuscript versions held in the Turing Digital Archive.
- Anthropic reported diary entry to police, woman faces felony charge
Anthropic reported a user's diary entry to law enforcement, resulting in a woman being charged with a felony for the content she wrote.
- OpenAI annualised revenues $20B less than previously signalled
Recent reports indicate OpenAI's annualized revenue falls $20 billion short of previously signaled figures, contradicting earlier optimistic financial forecasts for the AI company.
- How to scale intent, quality, and artistry with AI [video]
The video explains how leveraging AI can help creators increase their intentionality, improve output quality, and enhance artistic expression at scale.
- Forever junior: Skills AI can't develop for you
The article argues that while AI can automate many tasks, it cannot cultivate essential human skills such as judgment, empathy, and strategic thinking.
- Why isn't the industry freaking out about DeepSeek 4.1 Flash?
The article argues that DeepSeek 4.1 Flash’s impressive specs have failed to trigger widespread industry excitement because of limited access, market saturation, and lingering skepticism about benchmark claims.
- How to read code
Effective code reading relies on multiple out‑of‑order passes to understand structure first, then details, because code is written for execution, not linear human consumption.
- ‘Breathtaking,’ ‘Devastating’: Mathematics Reels After New OpenAI Release
Hundreds of new AI-generated mathematical findings from OpenAI's latest release pushed the frontiers of higher math in a single day, showing the field is irrevocably transformed.
- An empirical agenda for AI
Stripe argues its proprietary data can empirically trace AI's economic effects, proposing research agendas on productivity, complementarity, risk, intelligence markets, labor, firm structure, and agents.
- 100+ reactions to 100+ solutions – Proofs and Prompts
Mathematicians react with a mix of awe, concern, and skepticism to OpenAI's claim of solving numerous open problems using AI, noting excitement about new insights and worries about community impact.
- Where is the AI?
The article questions why, if recursive self‑improving superintelligent AI were possible, we see no evidence of it having spread across the cosmos, examining several explanatory hypotheses.
- The Start-Up That Wants to Build ‘Microrobots’ With A.I.
Atomic Machines uses AI to design and fabricate microrobots by training models on material properties and design patterns to accelerate development.
- This Photo Is an A.I. Fake. Can You Tell?
The article presents an AI-generated image and challenges readers to detect subtle telltale signs of its artificial origin by examining lighting, texture, and anatomical details.
- Anthropic bans ‘abusive or cruel behavior’ toward Claude
Anthropic revises Claude's usage policy to ban sustained abusive or cruel behavior toward the model and adds stricter limits on propaganda, surveillance, and weapons development.
- Vesuvius Challenge
Researchers launched the Vesuvius Challenge, offering prizes to AI teams that successfully read charred Herculaneum scrolls using machine‑learning techniques for today.
- The Storm (Part 2) - by Lee W. Martin
After surviving Hurricane Ian and personal loss, the author taught themselves AI concepts, built data‑wave tools, and pursued challenges like OpenAI‑to‑Z and ARC‑2 to find new purpose.
- ‘Cupertino’ Review: The People v. A.I.
The review argues that the Kings' new legal drama casts technology as the ultimate antagonist, portraying it as the 'final boss' that society must confront.
- When Hypotheses Become Cheap - by Meg McNulty - CipherTalk
AI makes hypothesis generation cheap, but scientific progress hinges on affordable, fast physical verification, requiring new lab‑infrastructure to match compute gains.
- Tony Fadell on why the first wave of AI gadgets failed — and what comes next
Tony Fadell says the first AI gadgets flopped because they addressed no genuine need and couldn't earn user trust, and that trustworthy AI assistants must run‑on device to protect privacy.
- OpenAI, the Partition Principle, and mathematics
Karagila criticizes OpenAI's preprint on the Partition Principle as poorly written, lacking rigor, and warns against treating AI outputs as genuine mathematical solutions.
- What's the date? — LessWrong
The piece argues that AI models must refrain from inventing the current date when no date is supplied in the prompt, stressing strict adherence to provided information to avoid hallucination.
- I Will Wait
Choosing to write code manually first, the author avoids AI-generated boilerplate to better prepare for later review, and plans to delay AI assistance.
- AHM Statement on OpenAI’s October 6 Release of Mathematical Documents
The Association for Human Mathematics criticizes OpenAI's release of 700+ math problem solutions as a power move, urging mathematicians to reject collaboration and uphold ethical research norms.
- OpenAI Releases Findings on 377 Math Problems, Further Roiling Field
OpenAI released findings on 377 math problems, saying the new results address mathematician concerns that its earlier solution worsened AI's disruption of mathematical research.
- Stack Overflow Developer Survey 2026
Only 22% of developers report being happy at work, while 62% favor AI tools and 39% use Python for AI coding.
- The Decision Model Gold Rush - by Swapnil Talekar
Jev’s rapid launch spurred a wave of decision‑model competitors, with benchmarks showing it offers 98% accuracy at $32 per million tasks, guiding selection criteria.
- Decisions API is now available in Public Beta - Announcements - OpenAI Developer Community
OpenAI’s Decisions API, now in public beta, lets apps quickly choose models, tools, or actions using text and image inputs, delivering up to 10× faster decisions via GPT‑6 Luna.
- A Viewer’s Guide to Warren Buffett’s Nightly Trips Down the YouTube Rabbit Hole - WSJ
Warren Buffett routinely watches YouTube late at night, viewing everything from 1950s commercials to Uzbekistan’s Got Talent, illustrating how even savvy investors fall for the platform’s recommendation algorithm.
- Software developers are not okay
Software developers face burnout and declining satisfaction due to AI-driven job changes, executive cost-cutting incentives, and eroded novelty, prompting many to quit tech.
- State of Devs 2026
The 2026 developer survey reveals widespread job insecurity, burnout, and polarized AI views, with many feeling less motivated and anticipating career changes.
- Opinion | Ada Lovelace Already Answered the Big Questions About A.I.
The article argues Ada Lovelace's historical observations about AI's inability to originate ideas and her view of machines as tools still directly inform contemporary discussions on AI capabilities and limitations.
- High Interest Rates Aren’t Slowing the A.I. Boom. That’s a Problem for the Fed.
Despite rising borrowing costs hurting households and businesses, AI infrastructure investment remains strong, fueling inflation and complicating the Federal Reserve's efforts.
- Why A.I. Food Slop Looks So Gross
AI-generated food images often appear unappetizing because models overemphasize texture and color patterns, producing surreal, uncanny-valley visuals that trigger disgust.
- Against the Personal Agents Theory of Everything
Nathan Baschez argues that AI’s future lies in shared, task‑specific factories rather than personal agents, enabling specialization and productivity gains.
- The Top 100 Gen AI Consumer Apps — 7th Edition
The seventh Andreessen Horowitz AI consumer apps report shows ChatGPT leads while power users drive most spending, and traffic and revenue rankings diverge.
- What the AI boom is doing for Americans - by Noah Smith
Noah Smith argues the AI boom is already delivering jobs, tax revenue, consumer surplus and macroeconomic support that benefit regular Americans, contrary to Krugman's oligarch-only view.
- Inference Is the Most Important Market in Software
AI inference market will grow to ~$350B by 2027, surpassing the $161B database market and fundamentally reshaping software unit economics.
- Google Workspace Updates: Preview, edit and collaborate on Markdown (.md) files natively across Drive and Docs
Google Workspace now lets users view, edit, and collaborate on Markdown files directly in Docs and Drive without conversion, seamlessly anywhere.
- Two types of AI agents – Matthias Plappert
AI agents will divide into personal assistants that act on users' behalf under their name, and virtual employees that own independent accounts and responsibility.
- The Future of Mathematics
Jeremy Avigad argues that while AI can solve routine math problems, mathematicians must focus on harder problems, bigger conceptual thinking, and creative use of AI to advance understanding.
- Whistle: Speech to Text in 16.9 MB
Whistle is an open 16.9 MB speech‑recognition model that runs on CPU, transcribes seven languages in 11 ms first‑token latency, and integrates with the Needle engine for on‑device tool calls.
- Smaller models are the future of AI sovereignty
AI sovereignty means retaining the ability to choose, govern, replace, and walk away from AI systems, achievable through smaller, controllable models rather than relying solely on proprietary frontier models.
- So You Think You Could Be An Electrician?—Asterisk
The article debunks the claim that trade work guarantees AI‑proof six‑figure salaries, noting college still raises earnings, trades carry injury risks, have internal hierarchies, and AI may alter demand.
- I Quit OpenAI Because Its Culture Is Broken
The author resigned from OpenAI citing a broken culture, joining others who have left major AI firms over similar concerns.
- e2e: The open source AI testing framework
e2e is an open-source, customizable TypeScript testing framework that uses AI to test web, mobile apps, and more, and supports extensibility through plugins.
- How Everything Became Gambling
The article argues that Silicon Valley's business models, from prediction markets to venture capital and LLMs, are fundamentally driven by gambling-like probabilistic betting.
- Security Logs Analysis Portal
Testing 13 server applications against MITRE ATT&CK revealed that only 26% of techniques are logged by default, highlighting widespread visibility gaps in security logs.
- Help · RDLTR
RDLTR is a read‑it‑later service that lets users save links via extensions, shortcuts, or bookmarklets, summarize them with AI, and manage a private, searchable feed.
- on AI, desire, and fictional characters - by Rebecca Lowe
Rebecca Lowe argues that only living beings possess genuine desires and goals, so attributing them to AI or fictional characters is conceptually and morally misleading.
- Why don’t more developers “use the platform”?
Developers often avoid native web platform features due to historical gaps, familiarity with libraries, better documentation, and personal enjoyment of building things themselves.
- The Era of Programming Languages Exploration is upon Us.
The author argues that AI‑driven coding makes traditional PL research obsolete, but urges the community to embrace exploration and fun, using LLMs to pursue novel language ideas.
- The Journal of Writing and Artificial Intelligence
The Journal of Writing and Artificial Intelligence, hosted by the University of Florida's Writing Program, publishes research on AI's impact on academic and professional writing practices.
- Kolibri Has Landed: A Sovereign Open-Weight Model — Aleph Alpha
Aleph Alpha releases Kolibri, a 78B‑parameter English‑German MoE model with 3B active weights, 1M‑token context, Apache‑2.0 license, tuned for sovereign, mission‑critical applications.
Takes
We should thank our lucky stars that Linus isn't part of the anti-ai nonsense that's ravaging the old Linux world. I'd hate having to soothe the cognitive dissonance this must produce!
@dhh
Opus 5.5 mapped the entire coast to coast route across the USA for a clicker game I'm playing with. 5,285 km of real road, sampled every 10m from OpenStreetMap.
@jonpaul
Introducing San Theft Auto: A racing game based in San Francisco It’s multiplayer & open source! I spent 950M tokens having GPT 6 meticulously map San Francisco in Blender 😅 Try it out: https://santheftauto.com
@NaderLikeLadder
Great read from Patrick, as always. “Personal agents will, I think, act as a kind of structural subsidy for product quality.” Yes, and well put.
@tobi
cleo is brilliant and we one of my favorite conversations about ai safety also i have a subtle diss during the interview to some well-known peeps 🙃 you’ll have to find it!
@alexandr_wang
This is the most delightful talk out there on what software development (actually, jellyware! brilliant term!) will look like in 6–12 months across the globe. A must-watch glimpse into the future if you still live in tmux day and night. Thanks for sharing @steipete!
@ivanopcode
Just to avoid some confusion: auto-compact does summarize. When it runs, your whole conversation is replaced by a short summary. It doesn't keep the last 1M tokens around. The ~967K is only when it runs on 1M models. Compacting itself is just one request that reads about as many tokens as the message you sent right before it, and it's mostly cache reads if your cache is still warm. What @rohit3a found is that waiting until 967K means every message before that point uses a lot of tokens, because each one includes your whole conversation so far. Most of it is read from the cache, which is much cheaper (so this is usually fine as long as your cache is warm), but it still counts toward your usage and adds up as your conversation grows. /autocompact 400k makes it summarize at 400K instead, so your messages never get that big.
@lydiahallie
whoops I hit send and went to a meeting, didn't realize the repo wasnt public- it is now! it's a chrome extension I use every day now https://github.com/ThariqS/ai-newtab
@trq212
I love AI but it killed my very lucrative business model of building profitable side projects and selling them. The building part still works but selling is almost impossible now!
@iannuttall
I discovered that 1000s of bots are actively monitoring X for new links posted, probably not just on my account I noticed when everytime I'd share one of my sites new features or pages, the page would get a bit slower just after posting, then I asked Claude to analyze it I tweeted a test URL as a reply somewhere deeply hidden in a thread (so not even a main post) And within 8 seconds there's thousands of crawlers opening the page! One is a military intelligence firm, that scans social media for breaking news and sells alerts to newsrooms, banks, companies and government, it specifically follows accounts with many followers Another one is a social media monitoring company that tracks what people say online I knew of course high frequency trading firms use online news to trade fast on things, but interesting to see it in real life Twitterbot (X) of course is first, within 0 seconds of sharing, it opens it (probably when I wrote the tweet) The first non-X bot arrives just 8.40s later, which I think is pretty fast considering it's a reply by me deeply hidden in a thread somewhere TL;DR they're watching you!
@levelsio
lol this is insane claude now gives you hundreds of dollars per month extra you can just create an API key, put it into Vercel AI gateway, and save every month i like the new claude
@robinebers
one thing i learned during my time @AnthropicAI was to always question and reinvent the way you work. and when your tools don’t fit anymore, build your own tools. a lotta wisdom from a lotta good folks here ☁️
@jenny_wen
Linear CEO @karrisaarinen: "I took a break this summer from paying attention to AI. I thought that when I came back, I'd feel left behind. But when I came back, I realized that nothing has really changed. There's new models, there's new techniques, there's new agents. But in the end, making great products, increasing revenue, and building the business—it's still hard. It's clear the tools are getting better, you can do more, you can do it faster. But I think the more important question is: are we actually making better things with these tools? Are our teams getting better with these tools? I don't think that answer is clearly yes. There's a lot of emphasis now on toolmaking and using tools and building systems. But I think in the end, as product people, our job is to make something great."
@lennysan
wisprflow has become un usable. seems like they have fucked up there tts model. i am cancelling the subscription. any good free alternative?
@code_kartik
Mercury is running a free bookstore for SF tech week and to celebrate our AI-accounting system (@Mercury Books). My recommendation was “Beginning of Infinity”. 201 Jackson St today and tomorrow.
@immad
Grok Bot just made X free for every user. No API key, no credits. So @slashlast30days on @Grok @Bot now pulls X in automatically and puts it next to everything else people are saying: 🟠 Reddit upvotes and comments 🔵 X likes, now free on Grok Bot 🔴 YouTube, TikTok and Instragram comments 🟡 Hacker News threads 📈 Polymarket odds One answer, ranked by what real people actually engaged with. Google aggregates editors. /last30days searches people. Get it on Grok Bot:
@mvanhorn
The more I use Opus 5.5, the more I realize how far behind OpenAI models are. Not just coding. Everything. I sent GPT-6 Astra my V02 max lab test results, and it gave me a headache. 200 sentences of: "your report shows this, I think that, therefore it's inconclusive". On the other hand, Opus 5.5 gave me the gist and actionable steps. OpenAI models right now are like that overthinking guy who can't make a decision. Opus 5.5 is that high-agency friend who tells you exactly what you need to hear in one sentence. I don't know how the gap widened so much in the last few months.
@marclou
Indeed. It’s an annoying skillset right now because of how fast the meta changes, but you can in fact get gud at prompting your ideas to life.
@brian_lovin
Silicon Valley keeps building me a husband when what I really want is a wife lol I don't want an agent that only does one-off, random tasks that are nice but not life-changing: booking flights, making dinner reservations, or buying tickets I want an AI wife who knows what’s on the family calendar, remembers there’s a birthday party Saturday and we haven’t bought a gift then sends me options, realizes we’re almost out of diapers and buys them, knows which bills are coming up, remembers someone needs a dentist appointment and books it, figures out what the hell we’re eating all week and what we need from the grocery store, and generally keeps track of the 47 things that need to happen before anyone else even realizes they need to happen. So, who's building me an AI wife that carries the mental load instead of an AI husband that waits for me to give it a chore??
@SoniaBaschez
Grok Bot doesn’t need the X connector/API credits to search through X anymore!!!! this is actually insane
@kr0der
We're going beyond text and are releasing a new version of GPT-6 to all users in ChatGPT. Lots of model and infra improvements coming together here to scale it to 1.2B users!
@thsottiaux
Watching this with my husband who works on AI for Google just absolutely slack-jawed.
@ebonylewkovitz
Grok goes open. Holy shit.
@Scobleizer
Si tens una startup, ja estàs trigant a demanar-hi accés. M'ho han aprovat gairebé al moment i ja tenim 1.000 $ en crèdits d’API + 12 mesos de Claude Team Premium per a 5 persones. Si hi pots optar, aprofita-ho. 👇
@senyorfont
I own land and property in exactly these areas and will keep buying land and property in these areas This is exactly my post-AGI thesis too Scarse coastal or mountain land with clean air
@levelsio
I am surprised that people are surprised this is how I prompt Claude. Talk to Claude the way you would a coworker. There's no secret to prompting. There's no need to be overly scaffolded or prescriptive for most tasks -- give Claude a goal, and it will figure it out. Back in the Sonnet 3.5 days, your prompt mattered a lot. Nowadays, it's much more important to communicate to the model: 1. What you want it to do 2. How much effort you want it to spend 3. How it should verify that it did the right thing
@bcherny
Anthropic used Claude to make their apps 3x faster. The strategies they came up with to do it are fascinating. I broke it down so you can use the best parts in your own work :) https://x.com/ClaudeDevs/status/2102839691154427983
@theo
Announcing d1 with vision. 👁️👁️ Our first decision model now supports images, text or both as inputs. We tested d1 against GPT-6.1 Sol and Claude Opus 5.5 on six real applications, from filtering support tickets to inspecting circuit boards. d1 matches or beats GPT-6.1 Sol on four of them. It costs 19x to 200x less than both models and answers significantly faster on every task. > probabilities for yes/no, choice, or score questions > one forward pass, without generating tokens > text decisions in 200 to 300 ms > Liquid API:
@liquidai
How to make money as an indiehacker in 2026 and beyond? the opportunities seem to be shrinking fast. I've been thinking about this a lot lately. AI is eating whole categories of the SaaS industry. What would I work on if I was starting from scratch? Here are some random thoughts. 1) Software primitives Some of the basics aren't going away. Humans use them to get a job done, or humans use AI to interact with the tool. The AI needs the tool and can't natively do what the tool does. Newsletter managers/deliverers, form builders, analytics software. If the category has existed for 20 years, it will continue to exist. The question here is can you get traction for a product in an extremely saturated space. Maybe there are opportunities in niches. 2) Automate your specialized service This is probably the best time ever to productize a specialized service. Take a domain / service you're an expert in, which previously would require unscalable 1-to-1 attention with a client, and agentify it with proprietary data/skill.md etc. 3) AI value arbitrage There is a short term opportunity in taking things we see at the cutting edge of the X bubble, and packaging it for the non-AI-pilled. Opus 5.5 motion videos would be a recent example. We all know this is possible with AI now, but maybe there's a vehicle to package this for normies in a way that blows their minds and creates value for them. This relies on there being a knowledge gap though, which will shrink fast. 4) Just Do Hard Things (TM) Do things that are genuinely hard, that you would never be able to build before. Build things that previously only a VC-funded company would be able to do, with hurdles from all angles not just tech, but also go to market, regulatory, etc. Doesn't guarantee success but it at least keeps competition away. 5) make a social media scheduling tool just kidding about that last one.
@yongfook
Introducing gdp-ts: Ghosts of Departed Proofs for TypeScript. gdp-ts is a library, linter and AI skill for safer API design. Under this contract, sensitive functions require 'proofs' that the caller performed an authorization check. The typechecker verifies these proofs at compile time, preventing your team and agents from shipping catastrophic security (and other kinds of) bugs. While these patterns have existed for quite some time, especially in ecosystems like Haskell, ① human code review and ② cognitive and syntactic overhead made these solutions niche. The situation is now inverted. Agents are writing more code than we can review, and they *thrive* in tight loops with hard constraints that would frustrate us. We see this with the rise of Rust, borrow checker, code aesthetics debate and all. The README and examples model a real-world Vercel API product constraint: changing the password on a Project requires a proof of a certain role + a certain entitlement. Thanks to Matt Noonan and Ollie Charles for their research in this space.
@rauchg
Human Intelligence is Surprising
@RayanKrishnan
Muse, dots, Instinct, and the question every agent will be asked
@worldnetwork