Not to mention the competition: chinese open-weight models and open-source harnesses. Qwen3.6-(27B and 35B) have proven to be worthy and capable of running locally. I am confident more SMEs would look into this as a solution given the ballooning costs of API usage. You get a decent setup with an RTX 6000 Pro.
> * That can still yield useful "discoveries" in certain fields, absent the discovery of new mechanics that exist outside said training data
One can argue, new knowledge is just restructured data.
I think the main concerns about LLMs is the inherent "generative" aspects leading to hallucinations as a biproduct, because that's what produces the noi. Joint Embedding approaches are rather an interesting alternative that try to overcome this, but that's still in research phase.
> Qwen3.6 35b a3b is still my local champion but I may use this for auto complete and small tasks.
I second this! Using the Unsloth Q6 (I forgot the exact name). Currently using it with forgecode (with zsh), on my Strix Halo, and it's suprisingly really good. I would say slightly Similar to Haiku 4.5, plus additional privacy, minus speed. It's surprisingly really fast for the hardware, given the speculative decoding, still PP is on the slow side.
They patched the "non-existent" issue it seems. And totally denied it happened in the first place. Honestly, someone should do a dump of redacted client documents to teach them a lesson. Short of a class action lawsuit would be an understatement. This is really huge.
Yeah, I ended up using an old mac mini for my Home Assistant needs. It draws a whopping 7W from the wall at idle (and it's near always idle), but the price of a new RPi is the same as 13k hours of electric usage for this.
Using whatever compute you have sitting in a drawer usually makes the most sense (including an old phone).
> So, the lessons for all other countries in the world is pretty clear: grow yourselves some mountains, dig yourselves a big river, and dam, baby, dam !!
You're forgetting corruption. Many countries can easily go 100% renewable, but there is no profit for dictators/politicians to do so. Most of africa, or the middle east, yet you still have many regions without electricity or water, so that people worry about food for tomorrow instead of better governance in the future.
Sorry but no. There are several major issues with that if you want your power to stay on all the time. Storage would be needed which even for the smallest countries on this list would require over a years worth of worldwide battery production. And grid stabilization would be almost impossible and that's just for starters. All 9 of these countries are mostly hydro. The renewables in this case are almost incidental. Also these dams were built decades ago for reasons that have nothing to do with the environment.
Benchmarking has been already known to be far from a signal of quality for LLMs, but it's the "best" standardized way so far. Few exists like the food truck and the svg test. At the end of the day, there is only 1 way: having your own benchmark for your own application.
> Dr. B is the king of slop, with 84 extensions published, all of them vibe coded.
> How do I know? Most of their extensions has a README.md in them describing their process of getting these through addon review, and mention Grok 3. Also, not a single one of them have icons or screenshots.
> Personally, I’m shocked this number is this low. I expected to see some developers with hundreds!
This is really surprising. Either because Firefox is not that popular ir mozilla has an automatic filter?
While its current performance is not competitive, there are currently interesting options. I got the orange pi riscv version, mainly to test riscv while it's slow compared to other arm socs, it's still better than I expected. There are even risc v TPUs now.
reply