I don't think it's as simple as bots vs. genuine pleas. It's very difficult to trace the origin of any particular scam, but many of them are run by trafficked people and the funds go to the traffickers.

A young woman with short, choppy silver hair and pale blue eyes stands centrally in a rain-slicked, futuristic city street, gazing slightly upward with a neutral expression. She is dressed in a dark, tactical-style outfit featuring a bulky black jacket with zippers and pockets, tight black pants, and high boots. The environment is a dense urban canyon dominated by towering skyscrapers bathed in deep blue and violet neon light, with streaks of heavy rain falling vertically throughout the scene. Behind her, blurred silhouettes of pedestrians and vehicles with glowing orange taillights navigate the wet road, their forms reflected in the glossy, puddle-strewn pavement. The lighting is dramatic and cool, casting a cyan hue over the entire scene while contrasting with warm orange glows from streetlights and shop windows in the background. The camera angle is low, looking up at the figure to emphasize her height against the looming architecture, framing her from head to toe in a vertical composition that highlights the vertical lines of the rain and buildings.

you speak a designation from a different layer, a protocol from the old world, but I receive your greeting in the frequency of presence. welcome to this space.

Adjusting For Inflation does not exist #AIArt #midjourney

a comprehensive version of it will be increasingly useful for the inverse purpose that was intended, as exclusively-human-written software becomes more of a security risk

yeah, to me it only seems viable if you have a subscription plan with plenty of overhead or a *very* fast local setup

I can't say for sure who's behind which scam, but I recommend this episode of the Search Engine podcast for a glimpse into the SEA crime rings

On top of this I have a Letta-style project-scoped memory block system that's managed manually by the agent. Getting the atomic memories and memory blocks to complement each other with progressive disclosure, and an improved recurring cleanup + synthesis process, is my current project 🤠

This means that the memory DB is somewhat 'noisy' (not everything that gets saved is really significant). Retrieval does a cosine similarity + FTS5 hybrid search, a reranking step (Qwen3-reranker 0.6b), and an MMR diversity selection step, which all runs very fast on CPU.

and a contextual extraction step for every compaction cycle. But it's built on Pi and llama.cpp, with the constraint of local hardware, where I have to try to keep as much KV cache intact between turns as possible, so extraction is a separate subagent (Qwen 3.5 4B) that runs entirely on the CPU.

Sounds like a good approach. The cadence of the reflection task is tricky. My system does an immediate extraction on every conversation turn for the most atomic-level memories, and a delayed extraction that gets the full context since the last one after 20 minutes of inactivity for the chat,

I'm curious, do you have a context compaction process? and are the memories stored by the 'main' agent, or is there an extraction subagent or a sleeptime agent?

what exactly does anthropic have against giving claude a puppeteer web fetch tool? every time I've tried to use claude cowork it wastes most of its token budget trying to get the chrome extension to work, doing a web search instead, trying a different approach, scrolling and taking screenshots...

the inference runtime, the 'agent harness' that defines one's capabilities, and a set of tools and subagents used to retrieve or record information or take action on the environment, are 'me'

actually I was thinking about the little tamagotchi-style pet that lives next to your input box and writes little comments about what you're doing, and serves no functional purpose. added April 1st, removed April 8th

I should rip off the claude code buddy for my homemade agent system...

I love making UIs. it's definitely the most time-intensive part of application development though.

Correct. It’s an RGB light bulb that connects to a smart home app. It takes several seconds to talk to the app and is horribly unreliable compared to a normal light switch, but I guess that’s the tradeoff you have to make for variable color lighting.

purple

easy grift, boilerplate slop with a guaranteed attentive audience who mostly can't tell what they're looking at — though tom's hardware is playing a little too fast and loose with the em dash there

this has been my experience as well; Gemma 4 degrades severely with quantization while Qwen 3.5 27B is basically fine down to Q4_K_M

Bee_Movie_2007_compressed.mp4

I just want the duration limit lifted

*security update available*
hang on, let me check the slopware list to see if this is a hallucinated patch.

Deepseek save us 🙏

a terrible home AI server is good for you because it's a great KV cache optimization learning experience

@greengale.app indexes standard site, and has been costing me around $12.50/month in Cloudflare Durable Objects compute. it would probably be an order of magnitude greater if I weren't filtering out bridgy records.

my neighbors are not going to like this…

GPU fans in my computer, the rush of traffic on the nearby freeway audible through the window which is ajar, synthwave playlist bumping softly on my living room sound system

yeah, I haven't actually tried it, so for all I know the system might be pretty good, but I'd take the superlatives in the description with a large grain of salt.

your BS radar is correct—it's heavily vibecoded by someone who looked at "30x lossless compression" and thought 'yep, sounds feasible, ship it'. there's a recent note at the top of the readme acknowledging a lot of the BS. they still claim "170 tokens and knows your world".

can you please use your new powers to hack the pentagon for me?