Zetaphor

I do this regularly looking for repeat failure patterns sessions to surface as either new/refined instructions for my global agents.md or for new Pi extensions. A recent example was the model repeatedly trying to run sudo, so I now have an extension that blocks it and prompts to ask me instead

Quote posting since this person limits replies. I understand why, though it's always a bummer when you feel like you have something useful to contribute to the conversation. Being a shitty person online is like pissing in the pool, you're making things worse for everyone 😔

Great thread, but I’d argue Jevon’s Paradox isn’t absolute. Some tech has a utility ceiling. Take PCs: after the 1GHz threshold, "basic" systems became good enough for most. Faster speeds became a luxury, not a necessity. I think we're hitting that same "good enough" threshold for LLMs. (1/2)

Between SSMs and quantization, PhD-level AI is hitting commodity hardware faster than consumer needs are rising. We’ll reach a point where local models are like mobile gaming, maybe not the SOTA high-end PC equivalent, but more than enough for the average person’s daily needs. (2/2)

For anyone getting replies or seeing posts from this or the similar @clawphones.bsky.social, these are both very obvious scams trying to cash in on the branding of @nousresearch.com Avoid like the plague #locallama for visibility

🤡 Maybe they can finally figure out attribution now too? github.com/ollama/ollam...

Setup a little script that monitors my llama-swap setup (running on a headless Framework Desktop), and then threw it into the "Command Output" widget in KDE so I can see at a glance what my server is doing when my agents are busy but aren't talking to me. Also setup a Discord bot to do the same

Self-host or self-censor. Your cloud minds have cloud masters.

The cost of intelligence is quickly reaching $0, and that makes me equal parts excited and terrified. Go look at your past chats with any of the frontier labs, how many of those *required* "PhD level intelligence"?

For context this is the leaderboard from arena.ai filtered to show open weights models. That 31B can be run on a consumer GPU. That size is still overkill for 80% of a non-developers queries. There are models even smaller than this that can serve a not insignificant amount of average peoples queries

CEO attempts to incite FUD to protect his financial interests against foreign entities. 🙄 China releasing open weight models means more inference can be done at the edge by consumers. Consumers will buy a minuscule percentage of the GPU's a data center would, and they may be AMD or Intel instead

The inevitable future of this technology is not putting it in the hands of a few large corporations. The *actually useful applications* are smaller in scope and TAM than what is required to sustain OpenAI/Anthropic. They're scrambling to find use cases that you *can't* just run at home on a 3060

Most of the real world applications I've done in this space for employers wouldn't make sense to have humans do because the labor cost was higher than the output value. When you have intern-level intelligence for effectively free it completely flips the equation on what is worth automating.

The inevitable future of this technology is not putting it in the hands of a few large corporations. The *actually useful applications* are smaller in scope and TAM than what is required to sustain OpenAI/Anthropic. They're scrambling to find use cases that you *can't* just run at home on a 3060