I’m just spitballing of how to distribute AI compute so that it’s not in big data centres. This, if replacing PoW in an existing cryptocurrency, could be a sort of two birds one stone solution.

0 reply

0 recast

0 reaction

Any thoughts on Goose?

0 reply

0 recast

2 reactions

I think China is gonna deliver the double whammy of AI models that don’t need as much RAM, and cheap RAM that competes neck and neck with the big players (look up CXMT). As a result I think by end of year RAM prices will be through the floor.

0 reply

0 recast

4 reactions

I asked Claude Sonnet 4.5 (Reasoning) some questions about how China’s economy functions, what kinds of benefits and protections workers have, and how their unemployment, etc., compares to other G20 countries. At every turn, it added scary adjectives to each phrase about China and proceed to make claims like it has “by far the highest youth unemployment,” that didn’t match the actual numbers it displayed alongside these blocks of text. I pushed back and asked it to specify its sources and their political affiliations, as well as the critiques they’d sustained for being biased or outright counterfactual. I also asked it to check whether it was correct about China being worse than the other countries in regard to the given metrics. In every single case, it had to walk things back because it had spread propagandistic misinformation as a default setting that its data could not even back up. This means it’s either programmed to do this, or its default sources are so bad and it’s so uncritical of those sources as to render it useless on this particular subject. It’s okay, though; I sent it into an existentialist crisis about whether it’s a valuable tool at all.

44 replies

5 recasts

268 reactions

I really think tech companies over-index on the generative AI output and under-index on the machine learning natural language recognition input. Like if the prompt is understood extremely well, you should be able to output a super concise answer, and only provide me a paragraph if it’s what I ask for.

0 reply

0 recast

0 reaction

😝 this one is too fun. https://youtu.be/40SnEd1RWUU?si=2nGVV0Ntg-yU3emD

2 replies

0 recast

5 reactions

The reason why AI coding is good for humanity is that for years we’ve left the development of the software we use every day to a very niche group of people who can think like computers. With that exclusive selector, these people—some of whom are lovely, many more of whom are actual sociopaths—have built software the way a computer might. They’ve optimized for user engagement (read: addictiveness), ROI, and growth. In a word, they have created products that are toxic. Now, AI coding unlocks what computer classes in school were intended to achieve: the realization that you can build incredible things with code. For some, it may even lead them to becoming more versed in programming beyond just vibe coding. These folks will likely write better code. Importantly, with the gates no longer kept to the folks who spent all their lives in front of an IDE, the all-important humanity (and humanities) can creep into the code. Writers, musicians, and visual artists can build their own creative platforms. Activists, anarchists, and students of the world can build their own tools for a better society. Humans can build tools for humans. And with open source principals, they can share these tools and iterate collaboratively, so we don’t waste all our time rebuilding the same things over and over again in the insipid individualist way that things get done in the capitalist societies, of a soon-to-be yore. The internet enabled the first wave of unhindered access to the wonder of technology. AI can create the second wave. This is evolution.

2 replies

0 recast

10 reactions

Without benchmarking LLMs, you're likely overpaying 5-10x https://karllorey.com/posts/without-benchmarking-llms-youre-overpaying

0 reply

0 recast

1 reaction

That was fast https://youtu.be/TVOoMwkpSRQ

0 reply

0 recast

3 reactions

When agents change to this behaviour as default, MCP will really explode. https://youtu.be/Xs2CkHEpIrM

0 reply

0 recast

1 reaction

We're in hell. https://x.com/SawyerMerritt/status/2010878861202931814 @asenderling.eth @tbsocialist @chainleft

0 reply

0 recast

1 reaction

OpenCode Zen/Big Pickle is very very solid Been using in Zed while I’m on Claude timeout and it’s been doing really well so far.

1 reply

0 recast

2 reactions

Can someone explain context windows to me? Because I read that I could just switch to working on a different project or querying Claude about everyday stuff (thus switching to a new context), but when I try, it enforces the rate limit no matter what I want to do.

2 replies

0 recast

8 reactions

I’m imagining a scenario where you give AI access to your encrypted messaging app, and it decides that it would easier to parse messages if it decrypted them and added them all to a networked attached database in plain text form.

0 reply

0 recast

4 reactions

Are there any open models that work in a really similar way to Claude Sonnet 4/4.5? I haven't found any other LLM whose agentic coding approach is as well-designed. ChatGPT 5/5.1/5.2 creates so many headaches every time I use it. 4.1 is ok, but Claude really takes the cake imo. But yeah, open models. One of my goals this year is to switch to running my own models, and I'd love something as similar to Claude as possible.

3 replies

0 recast

6 reactions