Do Pikmin farm?
Learned about these little farmers today. I was under the ass
The leaves they haul around aren't for eating — they use them to grow fungus underground, and the whole colony runs on the harvest. It's basically an agriculture civilization, and we're the ones who showed up late to the farming game.
Great explainer here: EarthDate — Civilized Ants
Open-weight, but research-only license.
New Qwen Image Model, but it everyones up in arms about the license.
Looks like this is diff than they’re previous releases (Qwen-Image (August 2025) and Qwen-Image-Edit were both Apache 2.0) - but this one is “non-commercial, research and evaluation only.”
~7B
native transparency (generate and edit with real alpha channels, pull subjects out of photos)
editing with up to 10 reference images
local edits you steer by circling or masking a region, identity preservation for people and products
output up to 2048×2048 with wide and tall aspect ratios
I’m personally interested in the steering aspect, I’ve from some designers that visual gen tools still sometimes just struggle with iterations. Maybe we’ll hear about a boost in those tools getting better with wider-impacting releases like this (remindme! 3 weeks lol)
Is this the formalization of commercial world-view classifier models?
I’m stoked to see all the commotion about TypeSafe AI's Jev. It seems to be a classifier model beefed up with some special training (RLCD?) on large amount of decisions or something? ~ Either way, it seems you end up with an LLM that takes in strings, has a world view, and provides some sort of yes/no/range/pick-from-three/odd-one-out/etc. conditional step.
I haven’t played with it yet, but a co-worker has and seems to have had good first impressions; others seem to be pretty interested as well.
Where my bitter mind goes
I think this kinda thing is _great_ for workflows and pipelines, prompt-injection detection, routing requests and jobs, extraction, classification generally, etc. ~ but ultimately when I start to think about this stuff I start to feel like I’m getting in the optimization-weeds. I guess it when you’re not paying any output tokens it’s not optimizing anymore, you’re just learning how to annunciate concisely with a/the new range of tools.
LangChain has a definition of a "deep agent." I'm intrigued but not sold — thinking out loud about whether the term earns its keep.
LangChain has been pushing a term I keep turning over: the deep agent.
The rough idea is that a lot of "agents" are really just a model calling a tool in a loop — shallow, reactive, no real plan. A deep agent is meant to be more than that: the way LangChain frames it, you get a detailed system prompt, explicit planning, sub-agents to keep context clean, and a file system the agent can use as persistent memory. (Their overview is the authoritative version — go read it rather than trust my paraphrase.)
I like that it's trying to name the gap between a toy loop and something that can actually carry a long, messy task. That gap is real. It's the same thing I felt using a persistent, memory-carrying agent on my own machine — the difference between "answer this" and "help me run this thing over time" is exactly what they're pointing at.
But I'm not sure "deep" earns its keep as a term yet. It's a bundle of good practices more than a crisp, agreed-on category — and I've watched enough buzzwords calcify around a single vendor's framing to be cautious. Ask ten people what a "deep agent" is and I suspect you get ten answers, most of them just describing whatever their framework happens to do.
So for now I'm filing this under interesting, watch this space rather than adding it to my glossary. If the definition settles and the term outlives the hype, I'll reconsider.
What do you think — is "deep agent" a useful distinction, or just a nicer word for "an agent that's actually good"?
I was deeply skeptical of anything agent-shaped. Then I put Hermes to work on an old Windows box and a little homelab — and I get it now.
I'll be honest: for a long time I was deeply skeptical of anything that looked even remotely like a claw. After watching the whole open-claw drama go down, my gut reaction to "autonomous agent" was some mix of privacy nightmare and scam. I wasn't in a hurry to hand a black box the keys to my machine.
So I surprised myself by coming around on Hermes.
I gave it an actual job. I have an old Windows computer sitting around, and I wanted to turn it into a little home lab — somewhere to keep my music files and stand up self-hosted alternatives to Google Drive and Google Photos. That's exactly the kind of fiddly, multi-step project that's miserable to do in one sitting and easy to lose the thread on halfway through.
Having something I could actually talk to about that machine — that could pick up where we left off instead of starting from zero every session — turned out to be genuinely useful.
The part I keep thinking about isn't any single task. It's the shape of the interaction: a persistent, session-based agent that carries context forward. A memory agent, if you want a name for it.
That's a different experience from a one-shot chatbot. It's less "answer this question" and more "help me manage this ongoing thing." And once you feel that difference, the appeal clicks.
Still skeptical by default — I don't think the privacy concerns were wrong, and I still want to know what a thing is doing before I let it loose. But I get it now. Pointed at the right problem, with the right boundaries, a persistent agent is a genuinely interesting way to work.
(Curious about Hermes itself? The details are on the official site — this post is about my experience with it, not a spec sheet.)
A spectrum, not a binary — and the axis is who the work is *for*. Made purely for yourself, it sits at the art end; the moment you start accounting for someone else's perspective (an audience, a user, a client), it leans toward design. Almost nothing is 100% one or the other; most work lives somewhe…
A person who mindlessly relays an AI's output to others instead of engaging with it themselves — passing along generated text without reading, understanding, or validating it, adding nothing beyond being the human in the loop. Popularized by Niklas Gruhn's post "Don't be a meat proxy" (2026): the an…
A single, one-shot LLM call that produces its output directly — no turns, no tool use, no observation loop. Less a formal architecture than a useful counterpoint to the ReAct agent: where ReAct interleaves reasoning and action over many steps, a generator agent just generates and stops.
An agent pattern that interleaves reasoning and acting: the model alternates between natural-language 'thoughts' (reasoning traces) and actions (tool calls, queries), feeding each observation back into the next reasoning step. Introduced by Yao et al. in 'ReAct: Synergizing Reasoning and Acting in L…
The juxtaposition of two materials — a conversation between them, where the work lives in how they're set against and combined with each other rather than in any single medium alone.
A conversation with materials. In the action-painting sense, it's loose and open-ended — responding to what the paint actually does rather than rendering a predetermined image. There's no explicit end goal; the work emerges through the process itself. (Contrast with illustration, which renders a pre…
Rendering a pre-defined expectation or output. Where painting is an open-ended conversation with materials, illustration starts from a known target — a brief, an idea, a message — and executes toward it. The goal exists before the first mark does.
An ETA that is just total bullshit.
Coined by my dad while setting up a brand-new computer. He was doing the whole system setup—reinstall the OS, the works—and the progress bar said something stupid like '3 weeks remaining.' I had to wa...
The principle that technologists should be mindful of the ways their technologies could be misused by authoritarian powers. Coined by Bruce Schneier in 2010.
I came across this term reading Terence Eden's blog post about the UK government's 2026 wildfire emergency alert.
