Feed

In the next version of Bun Async stacktraces are supported on native APIs like node:fs, Bun.write, node:http, node:dns & more. This makes debugging easier https://t.co/PHospWtxtg

Dispatch can now start coding tasks with specific models. Tell it to use a specific model in natural language. We want to keep making Dispatch the most useful place to delegate Cowork and Code tasks. How can we make Dispatch better for coding use-case? https://t.co/aiHR7ATzmr

I can't imagine spending the time and mental bandwidth managing my machines manually anymore, at least for basic stuff like clearing out stuck processes and junk files. I just use this big prompt that invokes a couple skills and some tools and it handles everything nicely for me. https://t.co/kCisuAXVkc

Some people have been contemplating an idea for years, maybe decades. Obsessing, attempting, discarding, agonizing, retrying. Some of these ideas are unpopular, niche, impractical. Not obviously capitalizable. They live on in the inventor's mind. In 2026, millions of these ideas will come to life thanks to superintelligent coding agents. AI doesn't get tired. It amplifies the individual, and for better and sometimes for worse, it always takes you seriously. "Great idea. Splendid. Wow. You're absolutely right." A world of digital wonders awaits us. This world will disproportionally favor the boldest ideas. Software that once seemed impossible will be one hyperlink away. I can't wait to see it.

I managed to get RSC federation working in spiceflow! you can import React components from any url with SSR and hydration fully enabled imagine embedding MDX components in a docs website. if you wanted to update a single component you would have to re-deploy the full website now you can just deploy the single component instead. then refresh the app: it will update like magic

I really love educating, I love learning too. I want to have weekly streams to help educate people on AI. At least the application of it, and the hosting. First big project is to build a pet door w facial recognition running on a 100$ processor. https://t.co/Nzl0RJWAnO

Local AI can be so good, but you’d need about 12k USD to get it. Then it’s not so great Here’s a Q4 of Qwen3.5-262B-REAP Weights 131GB KVCache 50GB 256,000 context 350 tokens/s prefill warmup 4,000 tokens/s prefill cache 36 tokens/s generation Vision enabled REAP is good https://t.co/Mhx05a8Jvu

the more I've been digging into the new Figma MCP, the more excited I am about it something new I'm trying is starting with a very ugly sketch in Figma, and then having Claude Code flesh it out in Figma so I can tweak and edit before sending the final back to Claude Code https://t.co/6IylfCwhgb

Maybe we are just one more scaling leap away from human brain level connectivity complexity in LLMs 100 Trillion ~ in 2 years?

Meta has one advantage: their LLM doesn't necessarily have to be SATA. It just has to be good enough that 99% of their Instagram and Facebook users consider it at least as useful as ChatGPT in the free tier; The use case for 99% of users is answers to everyday questions and solutions to moderately complex problems. For this to happen, Avocado simply needs to be good enough to satisfy 2 billion people.

If you prune a model 20% you’re barely losing anything, this is less true with smaller models but it is tested and published in academic research. Cerebras states you can prune up to 50% with little degradation in whatever your calibration data activates. If you add quantization on top you further damage your model of course and reduce its range + introduce artifacts. When you’re a company like Cerebras and you have compute you can run fleets of GPUs for weeks at a time. Your results will reflect that attention, when you have to rent gpus as an individual you can’t be using 20k samples with 16k context lengths on huge BF16 models. So you make do. ———— I pruned Kimi-k2.5-4bit down to 20% of its original size and put that up with a warning that the weights are broken. That was an answer to a question I had, how much can you chop off before it’s useless. The answer was around 55% ———— Every successful person had to fail 100000 times before something stuck, that’s called the scientific method. I will keep putting out things that are broken, and things that are good. Hopefully others would like to help (:

ARC AGI bench has it all wrong. This is the real test.

A lot of people are calling Hermes Agent the end of OpenClaw. BRUH! It's not... Nous Research trains actual models and they built an agent around that expertise. The local model routing is solid, but the part that matters for your business is that your conversations become fine-tuning data. You can train a model on how you actually work. 00:00 The Problem with Local AI Models 00:25 Introduction to Nous Research 01:04 Cross-Platform Agent Capabilities 01:44 Deep Local Model Integration 02:30 Routing Tasks to Different Models 03:06 Conversations as Training Data 03:50 Hermes Agent vs. OpenClaw 04:15 Future Plans and Series Overview

Added a flight recorder to https://t.co/6TyHKaj8lb So now I can replay those human flights of real players later on as kinda "ghost" players (like in race games) Because I also have AI players now but they suck and are kinda boring and not as fun as real humans flying around https://t.co/qfFNhF0MQ1

One definition of intelligence is “learning from error signals” When reality hits back, intelligence learns and adjusts But what about persevering for months, years, decades when you have nothing but error signals: Neural net research pre 2012, Edison trying for the light bulb 100+ times A scientist/mathematician struggling with a theory their entire life Even when there are no subtle positive signals like having nothing to show for it the entire time What is the catch here?

Where intelligence begins: Electricity is useful because it is an exceptionally tractable way to make physical states change quickly, reliably, and at scale. Information is not the charge itself. It is the structured distinctions embodied in a state, abstract in description but, in any actual system, physical in embodiment. Intelligence begins when such distinctions do more than trigger responses, such as supporting models, compare expectation with outcome, and revise action through error. The decisive step is representation. Once learning or design makes a pattern reliably correspond to features of a world beyond itself, a tiny local event can stand for something not here. A neuronal spike or transistor switch can then represent something distant, past, future, or merely possible. Then the system is no longer governed only by what is immediately there; it is also guided by what its states are about. Regular physics still governs every charge. But physics has been organized into a loop in which electrical events carry informational form, and that form constrains the electrical events that follow. The deepest fact is that matter can be arranged so that the absent becomes causally relevant through a pattern present within it. That is the threshold at which representation becomes causally relevant, and intelligence begins.

remember, in agent land the cost function is O(k) agent calls, not O(n) CPU cycles or disk reads this should be obvious, but I've caught opus trying to optimize for the latter at the expense of the former careful out there

This is a great read for anyone interested in understanding how training is done and everything around it. https://t.co/00XZQs9RhD

In a few weeks the most powerful AI model of all time Claude Mythos will release This makes me deeply nervous Not because of cybersecurity risks or anything like that But because it will quite obviously be significantly more expensive which will cause the wealth gap to explode Let me explain First the obvious: tokens aren’t getting cheaper. In fact, they’re getting significantly more expensive Almost every new version of ChatGPT and Claude brings a slight bump in price over the last one And plans haven’t been going down either, they’re only coming out with more expensive ones. ChatGPT Pro plan for $250 a month. Claude Max for $200. GPUs, RAM, CPUs all going up in price. And now Mythos, which the leaked blog post hinted won’t even be included in a plan. It will only be in the API for what will be an astronomical cost. And do you seriously doubt this won’t lead to an upcoming $2,000 a month Ultra plan that every other AI company will immediately copy? It’s one thing to make luxury items more expensive. It’s another thing to make intelligence more expensive. Intelligence that is critical to getting ahead in a crumbling economy. Let’s just call it what it is: using AI gives you an advantage against everyone else. Those with AI are keeping their jobs. Those not using AI are losing their jobs Now a new level of intelligence that will only be accessible to the rich is coming out. Only the rich will be able to use this super intelligence to create more economic value than others. What happens to the people that can’t afford Mythos? Or ChatGPT 6? They are left with a major disadvantage in the economic battlefield. Then on top of that, both OpenAI and Anthropic are going to IPO this year (it’s killing the middle class that this didn’t happen years ago, but that’s another story) They both are heavily incentivized right now to explode revenue as much as they can. They both are incentivized to make these new models as expensive as humanly possible. The middle class is already gutted. A middle class without access to the intelligence that the upper class will have will only gut them further. If a job position is between someone in the middle class with Claude Sonnet, and someone in the upper class with Claude Mythos, the Claude Mythos candidate with 100% get the job. It’s like a ballet dancer getting in a weight lifting competition with someone on insane amounts of steroids. Or say someone with Claude Opus has a genius idea for a business, and someone with Claude Mythos gets the same one. The one with Claude Mythos will release a significantly better product much much faster, crushing the person with Opus. I’m very pro-capitalist. In fact, I might be a radical capitalist. But at the same time this country (and this world) needs a middle class. I don’t know the answers or solution. There probably isn’t one. I honestly don’t even know what I’m trying to achieve with this post. I just have gotten incredibly scared over the last few days thinking about this scenario. I think the best plan of action at the moment is to just create as much economic value as you possibly can right now. (Ethically) earn as much money as possible. Save everything. If you want to compete in the future, you’re going to need to be able to afford the top tier intelligence. It’s critical for you and your family to survive. But in the meantime, don’t let anyone tell you intelligence is going to become “too cheap to meter”.

The upgraded Skill Creator from Claude Code team has built in eval framework that runs tests with child agents to determine effectiveness and drive improvement What you have in the skill is years of researcher experience encoded Very surprised by effectiveness https://t.co/0lrTi6BIpr