Eric Florenzano

@ericflo.bsky.social

Building assistance! Used to make AR/VR. Before that made mobile apps. Before that made web things. People first, then products, systems, data. he/him

Man, Dems are really falling on the wrong side of this AI issue, and that's really hard for me. Trump, who I despise, is somehow perfect on this issue. What the heck do I do with that?!

As soon as I posted this, Anthropic, OpenAI, and xAI set aside their differences, formed a cartel, and agreed to slow down under the guise of "safety." I wonder how voluntary this slowdown is, or if we are seeing the bend of the S-curve that I've been feeling, and companies scrambling...

Eric Florenzano@ericflo.bsky.social · 3w ago

This may be contrarian but as good as Fable/Astra are, I think LLM capability has started to plateau and improvements will be narrower from here due to costs, but I think we still have another 10x-100x value to find in the harness, 100x in infra, and probably 1000x to find in products on top of that

I tried giving a software spec for an agent harness to Astra 6, Fable 5.1, Muse Spark 1.3, and Gemini 3.8 with a /goal to build it, and literally none of them got even close. Yet the game I one-shot prompted yesterday is nearly production-ready. It's hard to wrap my head around spiky capabilities.

This may be contrarian but as good as Fable/Astra are, I think LLM capability has started to plateau and improvements will be narrower from here due to costs, but I think we still have another 10x-100x value to find in the harness, 100x in infra, and probably 1000x to find in products on top of that

I hear the fan spinning on my PC in the other room. What could that be? Oh yeah, I forgot, two days ago I gave a local Qwen-3.8-27B agent a /goal to clean up and organize a vibe slop codebase, and it's still going. If this works out, best use of idle PC time ever. Slop on weekends, clean on weekdays

I'm surprised Google hasn't bought Huggingface yet. Google has a tons of cloud storage and bandwidth, seems like strong cultural alignment, and would be Google's AI-era equivalent of Microsoft buying Github.

Qwen 3.8 27B is looking like a fantastic model, but it's also making me really appreciate what Meta did with Glimmer 30B and its KV cache. Glimmer's KV cache is far more manageable in a 24GB VRAM budget. Being able to compare both approaches is so nice! This rocks.

The one thing I hope we don't solve soon is continual learning. Cloud vendor lock-in will be total when that happens. We have to make sure local LLMs are good by then, and your device (physical property possession) is both serving and learning.