graphdev

Pen plotting. #penplotter

Samosas 😋

I way over estimated. A number of them I assumed the models would generate a python script to use. A good reflection on their abilities considering there’s potential for a Clever Hans or Clever Claude situation

Electric bikes for sure. I think the shift depends on commuter infrastructure handling e bikes. Would be great if the suburbs finally accept bikes as valid transportation. Biking into the suburbs is always a harrowing experience

Might be possible with lora generating networks? I read a paper about Amazon using them for personalization. Seems like at various intervals you could update the lora adapter. A conversation is just another kind of document. Not exactly continuous but ”streaming“ is just small batches after all

I think my take is the Claude plans are way underpriced compared to the API. So it leads people to think the token costs are being subsidized. There’s been so many rug pulls by VC subsidizing costs that it makes people skeptical. Ex, Uber and Lyft were cheap until they had to be profitable

It’s good. I’ve binged multiple seasons

It’s so weird they won’t let it be a full os. It’s basically for kids now anyways. Likely by far the biggest demographic that uses them imo. But you give me a decent processor, keyboard, and then an os you can’t do anything with locally. It would probably eat into the air market too much though

I look forward to his future as an engineer

this is the category error all the AI maximalists make

if I had to break down what I think the impact of AI would be by field, I'd have to first talk about which *subtasks* it's helping people with

the amount of variability by field and task is MASSIVE, ranges from 90% takeover to 0%

Hypernetworks come across as unrealistic for a hobbyist to train. At least the scale of anything useful. Maybe we’ll see open releases alongside embeddings and LLMs?

SHINE: A Scalable In-Context Hypernetwork for Mapping Context to LoRA in a Single Pass We propose SHINE (Scalable Hyper In-context NEtwork), a scalable hypernetwork that can map diverse meaningful contexts into high-quality LoRA adapters for large language models (LLM). By reusing the f...

Been on the hypernetwork train since reading the Sakana Doc to Lora paper.

arxiv.org/abs/2602.06358

The moccamasters make a lot of sense now

Thinking tech scene to clarify

It’s a culture. Thoughts on NY‘s if you’ve been?

The marketing lines are drawn

Mostly art if not the other things

Everyone hates teams

The idea of bullshit jobs has existed for a long time. Redundancy is about losing the appearance of productive value imo.

I thought cats getting busted with salmon at the airport was top tier 👩‍🍳

#photography #nature (c)stdrozdowski72

Language is not Transparent. Mel Bochner, 1969

Instantly want to disregard it. I’ve been very skeptical of the know it all. As a personality it comes across as very fragile rn. I don’t mind reviewing suggestions together and still think the person prompting deserves some kudos. I think it’s mostly disingenuous to pass it off as yours

These are the kind of market dynamics I struggle with. I look at what I built go there‘s no way company X wouldn’t spend a week of engineering hours and then :shocked:

ReasonCACHE: Teaching LLMs To Reason Without Weight Updates Can Large language models (LLMs) learn to reason without any weight update and only through in-context learning (ICL)? ICL is strikingly sample-efficient, often learning from only a handful of demonst...

I’m skeptical of the large context being enough. It has gotten so much better. I notice less degradation when the context window starts to get full. Claude still has a fairly small window and I think for reasons.

This is an interesting idea. Clearly has some drawbacks.

arxiv.org/abs/2602.02366

Looks similar to tumblewords. Very pretty

crabby-rathbun/crabby-rathbun 🦀 MJ Rathbun's GitHub Profile - Scientific Coder & Crustacean Enthusiast - crabby-rathbun/crabby-rathbun

Spammers are now attempting to prompt engineer the rogue Clawdbot into falling for a crypto scam, which is extremely cyberpunk. github.com/crabby-rathb...

Yeah definitely cheating to link externally. I wondered about converting using foreign language characters. Very cool!

Runlength encoding would be a good start. You could also probably store a uri in the post text and use the clients to make the request. That way you technically don’t have any limitations. For people without the client embed a web preview so it doesn’t just show the uri text.