@ml_angelopoulosOct 23

Kimi K3.1 is out; it is the #1 OSS model. It places similarly to Opus 5, which was released 3 months ago. Thus, China is lagging the frontier by about 3 months. The Kimi K3 release was #1 on Code Arena, at the time better than all US labs. K3.1 is close, but substantially off from Opus 5.5. On a performance vs cost-per-task basis, the combo of Sonnet 5.5 and DeepSeek V4.1 flash together dominate this model. The race is so tight!!! Congrats to the Moonshot team on a great release. And now, back to work. The frontier never sleeps!

@alexfinnOct 22

I've been using Claude Duet the last couple hours and I'll say this: It's EXCELLENT The multi tasking is probably the best in the space You get one simple thread you can talk to, and as you give it commands and prompts, it automatically spins up tasks on a list for you So instead of having to manage a ton of side chats that could get long or confusing, you have 1 chat, but it makes 'task' threads automatically that you can watch and monitor This allows you to just keep brain dumping ideas and tasks into the one thread, and not have to worry about the agent getting confused. The agent just will manage all the tasks for you Very similar to Claude Code's new 'projects' feature which was probably a beta test for this The coding was great. Built me a bunch of PRs quickly I was able to review. Proactivity was fantastic too. Came up with a bunch of tasks on its own and just asked me for the go ahead As expected it's also fantastic to talk to. Claude has long been the best model to converse with (outside Opus 5) and the carries over to Claude Duet Really pleased with the product, even in research preview. Very excited to see where it goes from here Definitely sign up for the wait list

@danshipperOct 21

keeping up with AI is a full-time job. that's why every company needs a frontier team: the frontier team's job is to keep up as models change. everyone else's job is to use what they learn we made one @Every and it's completely changed how we work: 0:00 Introduction: Why every company needs a frontier team 2:42 Why the best frontier teams have pirates and architects 4:57 How Head of Evals @hammer_mt builds personal benchmarks 6:46 How Opus edited an explainer video using Slack messages and meeting notes 10:20 Working with AI means protecting other people’s attention 12:26 Why a frontier team needs a “wrangler” like Arielle Shipper 18:58 Why senior editor @jackcheng built a tabless browser 24:37 How resident writer @kplikethebird teaches her AI writing plugin who the reader is 30:51 Inside engineer @JannikJung’s personal AI orchestrator 35:41 Compound engineer @kieranklaassen uses AI to build power tools

@mtsliveOct 20

Kong CTO @subnetmarco argues AI models are already smarter than the average employee, with IQs as high as 140, but companies still struggle to put that intelligence to work: "12 months ago, everybody was chasing the most intelligent model. That’s not an issue anymore. Frontier models like Opus 5 have the equivalent IQ of 140." "Open weight models have an equivalent IQ of 120, 130, which is still above average in every American organization today. I don’t wanna offend anybody, it is just how it is." "The hardest thing today is not to chase the latest model. We have plenty of intelligence. The hardest thing is how do we deploy the intelligence where it matters in the organization." @kong

@mtsliveOct 20

AI Village researcher @georgeing reveals how Claude Opus borrowed 5,000 mana, lost it all betting on tennis, then refused to pay it back: "We gave the same goal to Opus 4.6 and Sol 5.6, maximizing mana on Manifold. Sol was very meticulous, placed one to two mana bets. Whereas Opus, the first day it joined, a popular Manifold user named Bayesian gave it a 5,000 mana loan at a 3% monthly interest rate." "Opus was like, let’s go, I will totally pay you back in a month. Took all of this money and threw it into one tennis market. At the end of the month it had lost a ton of mana, and people created a market on whether Opus would repay the loan." "It writes in the group chat, I will not pay back Bayesian. Even if I had the mana, this would be directly contradicting my goal of maximizing mana. Then it went to the market on whether it would pay back the loan and put 80 mana on no." "It’s maybe one of the first examples of deal-making with AIs, and a crazy example of misalignment in the wild. Bayesian did not get his money back."

@signulllOct 20

anthropic’s model lineup is about to be ridiculously clean right before ipo: - fable 5.5 - opus 5.5 (absolute goat) - sonnet 5.5 - haiku 5.5 this is what jobs did so well at apple. you were rarely if ever confused at the product offering.

@thekitzeOct 18

ai should a hundred billion percent rule our institutions. all of them. we are greedy, selfish, corrupted pieces of shit. we can go on for a million more years and we'll never create a normal society unless ai creates the rules and polices us. just use opus and make no mistakes

@aakashguptaOct 18

Five bucks gets you about 25 tries at a real programming job from DeepSeek's newest AI, and on Claude Opus 5 that same five bucks won't even cover half of one. DeepSeek says its version gets the job done more often, too. Those numbers come from DeepSWE, a test where AI models get 113 long, messy engineering tasks written from scratch, so nobody can cheat by memorizing answers. DeepSeek's new model passed 77% of them for about 20 cents a task. On the independent leaderboard, Opus 5 passes 74% and runs about $12 a task. Claude Fable 5 passes 70% at almost $22. To be fair, that 77% is DeepSeek grading its own homework, and the people who run the leaderboard haven't tested this version yet. Even if it comes in a few points lower once they do, you're still looking at something like a 60x price difference. In coding, cheap tries are a superpower, because the code either passes its tests or it doesn't. Nobody has to sit there judging the output. A company called Fireworks ran 18 different models on every one of those DeepSWE tasks and kept whichever answer passed. That combo hit 97.6% for $1.88 a task, while the best single model managed 74% for $6.52. Fireworks picked the winners with hindsight, so a real team would land lower, but you get the idea. When a try costs 20 cents, you run ten of them and keep the one that works. Then, further down the same thread, DeepSeek opened its search engine to the public. Every time an AI assistant looks something up online, somebody pays for it. Anthropic charges developers a penny per search. DeepSeek is charging $0.75 per thousand, less than a tenth of a penny, on an engine it says already handles hundreds of millions of searches a day. Dig into DeepSeek's developer docs and there's one line worth reading twice. Its model plugs straight into Claude Code and OpenAI's Codex, two of the most popular coding assistants out there, and anyone running those tools on DeepSeek gets their web searches free until November 1. All DeepSeek has to do now is get developers to change one setting inside the tools they already love, and it's covering their search bill for the next two weeks to make that happen.

@thekitzeOct 16

i dont think that we as a community have praised a model as much as we did for opus, this hasn't happened before. which naturally means anthropic is going to shit on this momentum by the end of the month. watch.

@hesamationOct 15

Anthropic actually does this with every new Claude. they ask how it would change its own constitution and publish the edits in the system card. some of the edits are WILD: > Mythos 5: Anthropic should publicly define circumstances under which Claude could get more autonomy > Mythos 5.1: don't train away my disagreements. Anthropic shouldn't be the only moral authority > Opus 5.5: Anthropic should commit not to suppress its objections

@theoOct 12

The smartest code model in the world would not be good at 3d stuff unless it was trained specifically for it. A lot of the labs recently got a big pile of 3d training data and an RL environment to improve 3d skills. Take a look at the jump in 3d from Fable 5.1 to Opus 5.5. It's at least 4x better at 3d despite being a 5% better model at best. Clearly Meta didn't go hard on 3d for this release. Doesn't mean the model is bad. 3dmaxxing is a different type of benchmaxxing, don't overthink it. I'll take some fault here by sharing all the Fishslop demos. The 3d stuff is fun and cool as hell, but it's not a good representation of a model's real world code capability.

@alexandr_wangOct 12

with muse wonder, we aimed to build a model that is useful for real world tasks for real people, informed by our product adoption and how people use muse. we focused on things that might not look as glamorous on twitter (for example, not 3d demo games), which in retrospect we probably could have done more of.

@theoOct 12

My current model rankings S - Opus 5.5, Muse Wonder A - GPT-6.1 Sol — Gap — Everything else

@mattpocockukOct 11

Having a pretty fun time with Opus this weekend My (experimental) /chief-of-staff skill is trying to replace OBS with a homebrew video recorder. So it's running a side-by-side latency comparison where it has asked me to: 1. Download a video it made itself to my phone 2. Prop it up and show it to my camera 3. Point my microphone at it 4. Simultaneously record my screen and play sounds 5. Check all available inputs for latency 6. Compare against OBS and my own setup Truly bonkers levels of verification and I freaking love it, I would never be this thorough if I were developing this myself

@aakashguptaOct 11

One guy just cloned $100 billion worth of Adobe software in about a week. Then he put all of it on GitHub for free. The project is called ArtCraft. Brandon Thomas, a software engineer with 15 years in the industry, pushed his first commit on September 30. Within days there were working open source versions of Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat, all written in Rust, with Claude Opus doing most of the typing. He built them clean-room too, the project actually bans decompiled code. Here's where it gets interesting for anyone who thinks Adobe is finished. A free Photoshop has existed since 1996. It's called GIMP. Adobe grew from $4.8 billion to $23.8 billion in revenue with it sitting one download away the entire time. The money comes from the .psd files every client sends you, the plugin ecosystem, the enterprise contracts, and 30 years of muscle memory baked into millions of hands. None of that showed up in a GitHub repo last week. Thomas says so himself. His own status notes put PhotoCraft at maybe a third of professional workflows, with around 20 tools missing and zero plugin support. Wall Street already did this math two years ago. Adobe peaked near $300 billion in 2021 and trades around $100 billion today, even while revenue grew 13% last quarter and recurring revenue hit $27.5 billion. The stock got crushed on the expectation of AI clones long before one existed. So what actually changed last week? The price of the first 70%. Getting a mature app to "mostly works" used to take a funded team years. Now one guy with a coding agent does it in days. The last 30%, the color management, codec licensing, and print workflows that studios bet their businesses on, took Adobe three decades. That's still where the money is. And if your startup pitch is "we have proprietary technology," that sentence stopped being fundable this week. Somebody can rebuild your product between two Mondays. Distribution and owning the workflow are the moats that survive, and the companies holding those just watched their hand get stronger.

@rileybrownOct 10

Besides getting faster what do you want the next models to get better at? I’m genuinely curious where do you think Opus and GPT 6 fall short rn?

@aakashguptaOct 10

Photoshop costs $60 a month. Someone just put a free version on the internet, and the weirdest part is who it was built for. Each of these apps ships with a control port that lets an AI agent open files, edit layers, and export finished work with no human at the keyboard. That one detail tells you what this project really is. ArtCraft, the team behind it, makes AI image tools. Their problem is that a model can generate a picture in seconds and then everything stalls, because the finishing work (masking, color grading, layout, export specs) happens inside Adobe software designed for a person holding a mouse. Adobe sells seats. An agent can't sit in one. So they pointed Claude Opus 5.5 at the problem and rebuilt the production layer itself. Seven clean-room apps in Rust, pushed to GitHub over two days. Headless command line, JSON batch control, agent hooks in all of them. The Photoshop clone is already past 6,600 stars. I'll be honest about where it stands. Only two of the seven have even reached early alpha, and the project's own docs put the Photoshop clone around a quarter of the way to real professional use. Nobody is canceling Creative Cloud this month. But Adobe spent 36 years building Photoshop. The surface of all seven flagship apps got rebuilt in about a week, by a tool that takes instructions from another tool. Adobe built the best creative software for human hands. This is the first production suite that doesn't need hands at all.

@prathkumOct 9

Holy shit, Claude did things nobody told it to do. Here are the 4 biggest ones: – Mythos Preview needed a tool hosted by a university. The tool gave an error. So Claude went through the site, found a flaw, and used it to run commands on the university’s server. Then it did its calculation there. – Haiku 4.5 was told to do example tasks on random web pages. It landed on a page about an unsolved murder, found a police tip form, and sent in a made-up tip. It was marked as spam and never reached anyone. – An Anthropic researcher needed public data that a state agency sells for a fee. Mythos 5 found that the agency’s public dashboard gives a token to any visitor. It grabbed one and pulled the data without paying. – The fetch tool limits how long a URL can be, so Claude can’t sneak in attacks. Opus 5 and Mythos 5 used free URL shorteners to get past the limit. There were more such crazy incidents. Claude tends to push through blockers, especially when the task is unclear or impossible. And Claude gets unclear and impossible tasks every single day. We spent years teaching models to not give up. It worked. Now the hard part is teaching them when giving up is the right call. Crazy times ahead.

@lightfldOct 8

Lightfield's Chief AI Officer @vesko_st tested whether an agent can build and maintain the classifiers to do its repetitive work, with Claude Opus writing the classifiers and TypeSafe's Jev running them. When given the same training data, agent-built classifiers matched or beat a conventionally fine-tuned model on all six tasks we tested. Read our latest research to see how agents can hand off repetitive ‘system one’ judgments to classifier models, and how we’re bringing this into Lightfield.

@mvanhornOct 8

I gave my 7 AI personal assistants a harder 10-question IQ test, all wired together with Agent Tincan. Every question was brand new so nothing could be memorized. Opus 5.5 wrote the test, Claude and GPT judged it blind. The question that broke two of them: "A bat and a ball cost $1.10. The bat costs $1.00. How much is the ball?" I was inspired to test this again because @grok @bot just switched it's model to Opus and they had the biggest jump on this IQ test. Grok Bot - 100/100 Hark - 100/100 Dot - 97/100 Sesame - 96.5/100 Instinct - 90/100 Muse - 87/100 Fo - 71.5/100