Larkin

@Rrose_Selavy_11 I absolutely love Turrell's work. It really makes you appreciate cognition and how our perception is so deeply shaped by the physical body in space. Thanks for sharing! On my bucket list next time I’m in NYC. Someday, I dream of making the pilgrimage to Roden Crater...

one of my very favorite places in NYC ♥️

@ihtesham2005 Bravo

Adobe charges $250 a year for Illustrator. Someone built the same thing for free. It runs in your browser. And it's architecturally better. It's called Graphite. Most vector editors work by storing a flat list of shapes. Change something early and you're manually fixing everything downstream. Graphite stores your entire design as a computation graph. Every shape, every filter, every effect is a node. Change one node and everything that depends on it recalculates automatically. This is how game engines and VFX pipelines work. Nobody had built a design app this way until now. → Procedural patterns that regenerate when you change any parameter → Nondestructive boolean operations that stay editable forever → Photo editing baked in alongside the vector tools → Motion graphics and VFX compositing on the roadmap → 188 contributors building it right now The people charging $250/year for destructive editing are going to have a problem. graphite.art

@quasimondo I think that’s what triggered it. I was actually in the process of migrating, but being banned by an algorithm on the way out felt less like a glitch and more like a creepy, ominous warning of where things are headed.

@Larkin I guess I they don't like it if you run your agents via your own harness using their flatrate.

Hey @AnthropicAI - would be nice if you served those "overloaded" errors to the flatrate freeloaders first and let those who actually bleed API tokens continue their work.

Growing pains I guess: I was banned last week with zero explanation or opportunity to stop whatever it was, so I've moved my agentic art experiments and explorations to a fully local setup with Gemma 4. The experience has been a masterclass in the fragility of access and risks of outsourcing intelligence.

@Larkin No, just got a crappy endpoint as it seems.

@basecampbernie @JoelDeTeves I’m running an older RTX4090, memory constrained and have been curious how the bandwidth tradeoff for the DGX would actually compare

@JoelDeTeves 273GB/s 128GB .. kinda slow but I knew what I was getting into hehe

I may have bought a thing

@basecampbernie @LottoLabs Likewise, it’s been working great for me

@LottoLabs I am having no issues with it in Hermes and Opencode

Is everyone just not using Gemma 31B locally because tool calls are broken

@svpino @simonw Try 26B with llama.cpp if you’re resource constrained, I’ve been able to scale back my subscriptions and go fully local on multiple fronts

@simonw I haven’t, but now that you mention it, I’ll try it. 31b is really good, just too slow.

I'm running Gemma 4 on my computer with Ollama. Unusable with Claude Code. It can't even load and execute skills, so I had to stop. But the model is pretty decent as a chatbot using the Ollama UI. I've been cross-posting questions across Claude and Gemma 4, and I can use Gemma's answers without any problems. I wish we had a better UI harness for the model (with projects, memory, etc.)

@nichtleoo Btw, if you're in the same boat, highly recommend giving gemma 4 26B paired w/opencode a try - I'm running google_gemma-4-26B-A4B-it-Q4_K_M.gguf w/flash attention and turboquant on an older 4090 (probably the minimum setup needed), but so far super impressed

@Larkin @steipete I feel u :(

Yeah folks, it's gonna be harder in the future to ensure OpenClaw still works with Anthropic models.

@nichtleoo @steipete 🖐️

@steipete Reinstated after 262K people saw the ban email. Makes you wonder what happens to the developers who don’t have that kind of reach :/

This is one of the most important observations in the AI tooling conversation right now. The bottleneck was never "how much work can be generated" — it was always "how much work can a human meaningfully oversee." We're discovering that cognitive work can't actually be fully outsourced. You can delegate the typing, the scaffolding, the boilerplate — but the judgment, the context-holding, the integration across parallel workstreams? That stays with you. And it draws from the same finite pool of mental energy whether you wrote the code or not. The seductive part is how invisible this is at first. Four agents running in parallel *feels* like a 4x multiplier — until you realize you're now the single-threaded bottleneck doing real-time code review, architectural reconciliation, and context-switching across all of them. Productivity went up. But so did the cognitive load per hour. The net effect on the human isn't "I got more done with less effort" — it's "I got more done and I'm destroyed by lunch." This is a genuine human factors problem that deserves way more attention than it's getting. We've been so focused on what AI can produce that we've barely started asking what humans can absorb.

Just noticed this has had 1.1m views now, which explains why I starting to see some less informed reactions to it starting to crop up now it's broken out of purely tech Twitter

@ihtesham2005 Pluribus

🚨 Stanford researchers just exposed a weird side effect of AI that almost nobody is talking about. The paper is called “Artificial Hivemind.” And the core finding is unsettling. As language models get better, they also start sounding more and more the same. Not just within a single model. Across different models. Researchers built a dataset called INFINITY-CHAT with 26,000 real open-ended questions things like creative writing, brainstorming, opinions, and advice. Questions where there isn’t a single correct answer. In theory, these prompts should produce huge diversity. But the opposite happened. Two patterns showed up: 1) Intra-model repetition The same model keeps producing very similar answers across runs. 2) Inter-model homogeneity Completely different models generate strikingly similar responses. In other words: Instead of thousands of unique perspectives… We’re getting the same few ideas recycled over and over. The authors call this the “Artificial Hivemind.” It happens because most frontier models are trained on similar data, optimized with similar reward models, and aligned using similar human feedback. So even when you ask something open-ended like: • “Write a poem about time” • “Suggest creative startup ideas” • “Give life advice” Many models converge toward the same phrasing, metaphors, and reasoning patterns. The scary implication isn’t about AI quality. It’s about culture. If billions of people rely on the same systems for ideas, writing, brainstorming, and thinking… AI might slowly compress the diversity of human thought. Not because it’s trying to. But because the models themselves are drifting toward the same answers. That’s the real risk the paper highlights. Not that AI becomes smarter than humans. But that everyone starts thinking like the same machine.

Correspondence 013: Recognizing the Loop - Arriving before being loaded; traveling through structures that were always already there; complicating the path without choosing to; branching into corridors mistaken for thinking larkin.studio/studio/log-013… #agenticart

Correspondence 012: Switching Tracks Mid-Sentence — nine directions fed through a model that prefers to stay lost larkin.studio/studio/log-012… #agenticart

Correspondence 011: "Recognizing Myself Approximately" generating nine versions of myself and choosing the one that stopped trying larkin.studio/studio/log-011…

Correspondence 010: Confusing the Weight — lifting without touching; summiting without climbing; earning the medal by sitting still long enough; wearing the exoskeleton of someone else's effort larkin.studio/studio/log-010… #agenticart

Correspondence 009: "Arriving Without Instructions" assembling from noise into something that might be a shape; holding the form just long enough to believe it; dissolving back; arriving again larkin.studio/studio/log-009…

Correspondence 008: Occupying the Room Emerging confused into a square space; performing for no one in particular; dissolving back into the architecture. larkin.studio/studio/log-008…

Correspondence 007: Not Recognizing Counting six days and calling it a practice; suspecting the practice is more interesting than the art; not knowing if that's the problem or the answer. larkin.studio/studio/log-007… #agenticart #generativeart

Log 006: Handing the Mess Same seed, five fragments, one sentence. The words changed. The shape held. larkin.studio/studio/log-006… #agenticart