Patrick O'Brien

Thinking isn't the issue. The issue is the bloat that LLMs introduce and never clean up. You can have the worlds best implementation plan, but implementation plans are additive. Aside from that, architectures need to grow and change over time. But LLMs do not refactor architecture, they pile on top of it.

@AllTheTokens it forces you to think

At HumanLayer, we’re on a mission to solve the AI slop code problem. In 2025 we open-sourced our Research, Plan, Implement framework, now deployed inside fortune 500s like Block and Uber - places where shipping slop is just not an option And that was just the beginning. Today, we’re opening access to HumanLayer - an Agentic IDE, collaboration platform, and building blocks for your software factory. HumanLayer enables engineers solving hard problems in complex codebases to: > move 2-3x faster across the entire SDLC (not just coding) > maintain rigorous standards for system architecture and program design Hundreds of engineers at companies of all sizes are already using HumanLayer to ship fast without sacrificing quality. I'm excited to invite you to try humanlayer today at humanlayer.com, and I'm even more excited to see what you build. @0xblacklight and I are deeply grateful to our team, our customers who give us so much incredible energy and feedback, our investors who have always been in our corner, and our friends and family who have supported us along this crazy journey if you're a staff or principal engineer trying to make AI coding work at scale for your team, we'd love to hear from you as @swyx likes to say - let's make this the year of no more slop

@jahooma "Steal my codebase"

Announcing 100% free MiniMax M3 with a blazing fast provider. - Up to 200 tokens/sec 🔥 - It's so good we had to crown it "RECOMMENDED"! Try now, it's free: npm i -g freebuff

@herdrdev A plugin for agent-vault would be awesome.

a plugin marketplace is on the way. until then; building something? share it. push it to github, add the herdr-plugin topic, and drop it under this thread so everyone can find it! see the example repo: github.com/ogulcancelik/h… plugin docs: herdr.dev/docs/plugins/

herdr 0.7.0 is out, and it's a major one: it introduces plugins! the idea is simple: herdr stays lean, and everything custom gets extended through plugins. shareable, scoped, built however you want, to fit your own flow. with this release we're also shipping a few examples of what the plugin system can do. first up: a telegram plugin. herdr already controls your agents and knows their status, so the plugin just hooks into agent events and pings telegram the moment one needs you. notification lands → `herdr --remote` or ssh from your phone → straight back to the agent that needs you.

@herdrdev Love this software, and the API backing it is amazing too. I've built up my own agent orchestration system using a fork of Codex CLI backed by Herdr.

still leaving your laptop open so the agent doesn't die? still hand-rolling tmux + ssh + notifications? still can't check on it from your phone? you don't have to. try herdr.dev

Voice prompting is 4x faster than typing. but i wanted more. Multi-agent orchestration with a shared memory system, running on virtual private servers shipping straight to production. All controlled by my voice with nivida parakeet running locally or gpt realtime 2 in the cloud. CNVS is a mac os app built from the ground up in swift for raw performance on apple hardware. agents have bidirectional control, they can spawn and prompt each other, terminals, and browsers. Create loops and even draw diagrams straight on the canvas. nothing comes close, and im just getting started.

@ZackKorman @DanielMiessler The answer is an obvious: Anthropic is terrible. But amazingly, some people still haven't figured this out. They're a company made by a liar, full of liars, with a business culture of lying and deceit. They are genuinely one of the worst companies in the world.

Tonight at 7pm in the UK (2pm ET, 11am PT) I'll be debating @DanielMiessler. The topic is broadly: Is Anthropic good or bad? I will be wearing my UwU Underground OpenFlaw shirt because they haven't made a Dario is bad shirt yet.

Everyone talks about intelligence. Almost nobody talks about trust. For days now, I’ve been posting about serious issues with Z.ai: transparency, reliability, token accounting, limits, and whether users are actually getting what they paid for. And the silence from many AI influencers has been honestly disappointing. These same people are always saying open models will win, that intelligence should be accessible to everyone, that open source is the future. They celebrate benchmarks, post hype threads, compare models to Fable/Mythos, and chase engagement every time a new result looks impressive. But when the issue is reliability, honesty, transparency, billing, limits, and real user experience — suddenly the room goes quiet. This is not just “my personal issue.” Dozens, probably hundreds, of users have reported similar problems under my posts. Z.ai advertised Pro as offering “15x Claude Code” usage. I subscribed. Then, after I exposed that this claim did not match reality and the posts reached nearly half a million views, they changed it. Then I showed Zcode reading less than 2M tokens while the actual usage system counted around 60M tokens and burn through my usage in less than an hour without finishing the task, That suggests cached/repeated context is counted against users in a way that is completely invisible to them — more than 40x higher than what the tool itself shows. That is not a small UX bug. That is a serious trust problem. If the strongest open-source AI lab, with one of the strongest open models in the world, has issues like this, then what should we expect from smaller labs and weaker providers? And let’s be honest: intelligence alone is not enough. What is the value of a brilliant model if I cannot use the service reliably 90% of the time? What is the value of benchmarks if the product experience is opaque, unstable, and users don’t know what is actually being counted? I personally paid for an annual Pro subscription, and after more than a month, I still feel I have barely benefited from it. If I stay silent, and you stay silent, and everyone stays silent, then who exactly is supposed to hold these companies accountable? Who is supposed to make sure users get what they paid for? At this point, I’m genuinely disappointed. Not only in Z.ai, but in the content ecosystem around AI — where hype gets amplified instantly, but accountability is treated like an inconvenience.

@the_codewala Because most developers are mid. The good ones use linux.

Linux runs the world's infrastructure... but macOS runs most developers' desks. Why?

@KTmBoyle Are we supposed to be excited about death machines?

FIGHT UNFAIR 🇺🇸

@rowlsmanthorpe Altman is the only sensible one in the room.

Exclusive: in the room with the AI bosses at the G7 I've been speaking to people familiar with the talks, in particular the messages from Demis, Sam and Dario All said: we need international governance and we need it fast But they sharply disagreed about how Two differences stand out in particular: 1. Liberty vs control Amodei called for a US-led coalition of democratic countries to control access to the technology and "isolate common adversaries" such as China Altman said that once guardrails were in place, "we must err toward human liberty. We want everyone on earth to benefit from this technology and to figure out for themselves how to use it" 2. The role of the labs All three emphasised the need for international standards, but it was clear that some expected the labs to play a bigger role than others Hassabis proposed a "technical standards body that is supported by leading labs" - something he suggested he'd been working on Altman warned that the AI labs should not have too much control "There is a threat more insidious than the technical risks of this technology," he said. "It is the threat that the very real risks that AI poses become the justification for concentrating power in the hands of the few" Many people will write all this off as spin. A lot of truth in that! But these debates matter all the same Like it or not, the world's governments are looking to the AI bosses for suggestions on how to govern AI There is clear agreement on a technical standards led by the US. But right now there's nothing you could call a plan

@AntoineViallon @leyten @Zai_org Not for productive levels of development work. Codex xhigh 5.5 fast is around 60tps, and even that is slow for productivity. Aside from that, who runs only a single agent? Most devs run multiple in parallel.

@AllTheTokens @leyten @Zai_org That's absolutely usable lmfao. For a single request at least.

Wow, it has happened! 30.55 tok/s on GLM-5.2 4-bit (from @Zai_org) ran by six RTX Pro 6000's across the USA scattered over WAN! I can't believe this. It was an insane build, you can read more about it on github.com/leyten/shard

@7NewsBrisbane Good job Australia, wasted a billion dollars on a scam.

Australia could be the first to have a quantum computer that can solve some of the world's most complex problems. Construction started on Thursday in Moreton Bay. The country was chosen because we have one of the largest populations of quantum physicists, despite our small size.

@_inception_ai If it can't compete with frontier labs, or at least frontier open-weight LLMs, there's really little point. It's call and all, but still...

Welcome to the diffusion era. We bet on parallel generation years ago, when it was a contrarian idea. It's great to see the industry arrive. Mercury 2 continues to lead the Pareto frontier for quality, speed, and cost among publicly available diffusion LLMs.

LLM efficiency needs to be solved at a token abstraction level. Why am I predicting the next token, when I could be predicting the next group of tokens as a single symbol? Fine-grained prediction should be a depth complexity and refinement problem over a lossy or fuzzy faster prediction using larger symbols. The whole architecture needs to be rethought. It's backwards. An LLM should pre-complete rough abstractions up-front, and refine those into targeted results. Almost like a diffusion process, but not quite. Refinement should be the "thinking" depth. Surface level results should be weaker but still in the ballpark, and incredibly fast.

@Pirat_Nation Yet they leave Reddit untouched because they use it for propaganda.

4chan’s lawyer says the UK government “doesn’t understand the internet” and has given Ofcom an “impossible task” by trying to enforce British online safety rules on US-based platforms. Preston Byrne, who represents 4chan, argued that the site is protected by the First Amendment and doesn’t have to comply with UK demands. “The UK just isn’t getting the memo that the United States has a different system and we do things in a different way.” He also claimed Parliament misunderstood how the internet works and set Ofcom up with a mission it can’t realistically enforce. Ofcom fined 4chan £520,000 under the Online Safety Act for failing to implement age checks and other safety measures. 4chan has refused to pay. Via: LBC.

Profiting off crisis should be illegal. This is how you get misaligned incentives that result in global corruption and chaos.

@udiWertheimer @AlexFinn He's not wrong. Opus 4.8 sucks.

@AlexFinn > 2 bit quant > better than opus 4.8

I can't believe this is real I have GLM 5.2 running 100% locally on my Mac Studio. 2 bit quant. The results I'm getting are better than Opus 4.8 It's now powering my Hermes Agent and Codex. 100% free, local, private super intelligence on my desk I also have it in a loop coding for me 24/7 now I thought we were at least a year away from this type of event. It happened today. The model takes up about 250gb of memory. So you can technically run it on a Mac Studio with 256gb, but you probably want the 512gb memory version (please tell me you listened to me 5 months ago when these were sitting on store shelves) With Fable gone, I now have Opus 4.8 level intelligence on my desk for free. This is the future. Local, private, secure, personal super intelligence. If you're still writing off local AI as a fad or engagement bait, you are officially delusional