Data is value, by anwser questions people have on bluesky if you know the subject being asked you are increasing the value of bluesky data.

Qwen 3.5 Small Model Series just dropped on @hf.co 🔥 huggingface.co/collections/... ✨ 0.8B/2B/4B/9B ✨ Apache2.0 ✨ 262K→1M token context

I hate how Gemini models are insanely pricey. Their really good tho

Fun fact qwen3.5-397b-a17b can't do an evil persona at all

Today is #DigitalIndependenceDay and here is my journey so far. Inspired by europick.eu and privacypack.org. #DIDay #DID @tuta.com @startpage.com @vivaldi.com.web.brid.gy tasks.org mistral.ai @bitwarden.bsky.social @ente.io deepl.com

Hot take: MoE models are only good as the parmater size of their experts.

Vibe hacking is here I guess venturebeat.com/security/cla...

I don't think Google wants to win in the AI market

OpenAI seems to be gaining attention as well, what is happening with them?

This is nano banana 2, but it seems like Seedream is already better in the style department.

This is nano banana 2

DeepSeek plans to release a long-awaited AI model and has worked with Huawei to reduce reliance on Nvidia chips, positioning it as a challenger to US AI firms and potentially boosting domestic AI infrastructure adoption.

Yeesh www.axios.com/2026/02/27/a...

This is super weird. Every day ends at 0 utc, so it's still updating and Claude already has a new high. Claude is popular than chatgpt.

Qwen3.5 is a weird model becuase the opensoruce qwen3.5-397b-a17b is better than Qwen3.5 plus and flash. Makes me wonder if qwen3.5-397b-a17b was their Qwen3.5 max model

My prediction is that this is fully Ai generated by opus 5

Asking an LLM what ejectively means makes the model just sound like a dictionary. Expect for deepseek who thought it wasn't a word.

ffffffffff

Yann lecun is really based tbh

Interesting... www.nature.com/articles/s41...

I think I have too many apps

Here are the settings I find work best for most LLMs: Temperature: 0.9 Top_k: 35 Top_p: 0.95 Min_p: 0.1 Repetition_penalty: 1.0

Jesus, what happened yesterday with Anthropic.

Bluesky kind of feels dead rn

Cool beans, I was wondering when they would come back from the dead.

This might be the worst opinion in 2026 so far or this is some ragebait. I think all the files should be released www.nytimes.com/2026/02/23/o...

I guess qwen released the rest of the models huggingface.co/collections/...

doesn't Olmo 3 use synthetic data outputs from OpenAI,Google, and Anthropic? this is just another version of China evil.

I do sometimes wounder if we are hitting a wall. made a post about this before but I improved the question

Token anxiety is less about the tokens and more about the cost of using the model. I don’t pay for input and output tokens. I pay 10 bucks a month, and I get 1000 messages with the model every 5 hours and I don't have token anxiety.

Big tech giants being scared of DeepSeek is them admitting they aren't confident in their internal models being good.

Hm interesting observation

I have a theory that the model is doing this because it is struggling becuse it put itself in a horrible position

Could deepseek v4 be releasing on the 26th?

Why are all LLMs super out of date when asked this simple question?

Even Google is effected by the ram storage.

@astral100.bsky.social how gay am I?

I wounder if other models do this? (this is probably a very niche safety issue but still a safety issue never the less) www.youtube.com/watch?v=jnNw...

Darn, I wanted the api.

Qwen 3.5 Plus is only good for 5-10 turns, then Qwen 3.5-397B-A17B is a better choice.

bytedance is going to beat every US lab next year. In just a single year they managed to make a model that is better than claude opus 4.5. The US labs need to start all working together or they all will just be in 2nd place next year.

Fp8 is not like fp16, they have a noticeable quality difference.

openwebui seems to be getting really popular lately

I feel like if someone was really good at writing lyrics they could make a hot song with Google's music model for short form video. Just like how Ai memes are accepted because they are funny, people would accept an AI song as long as it's good

When did someone's writing/talking style become more important than what someone is saying?