iiinova — where your digital self keeps evolving

dahara1 @dahara1.bsky.social · 08/10/2026

I have been selected as an Alibaba Cloud MVP in recognition of my ongoing activities, such as releasing Japanese-optimized quantized Qwen. I will continue to work on various aspects of AI models without limiting myself to any single specific technology. www.alibabacloud.com/blog/meet-au...

alibabacloud.com

Meet Autumn 2026 Alibaba Cloud MVPs

We are proud to announce the members of the Autumn 2026 Alibaba Cloud MVP.

000

dahara1 @dahara1.bsky.social · 07/10/2026

I seem to recall that the people who invented the steam engine said similar things.

000

dahara1 @dahara1.bsky.social · 12/08/2026

Value That Generative AI Cannot Surpass

000

dahara1 @dahara1.bsky.social · 08/08/2026

暗算はそこそこ自信があったのですが合計ランダム数字とランダム四則演算には苦戦しました、ゲムダンお疲れ様でした！

100

dahara1 @dahara1.bsky.social · 22/06/2026

I've created a ranking page that analyzes the most popular fifa World Cup/soccer videos on YouTube around the world every day. Updated daily during the period! If you're a soccer fan, check it out! There's also a search page! webbigdata-jp.github.io/soccerscope/

webbigdata-jp.github.io

SoccerScope — FIFA World Cup 2026 YouTube Trends & Fan Reactions

Daily FIFA World Cup 2026 statistics on YouTube football videos trending worldwide — views, countries, fan sentiment, teams, and themes. Data & analysis by TubeSaku.

000

dahara1 @dahara1.bsky.social · 18/06/2026

Twitter now allows users to bet money to boost their tweets, but since these aren't treated as advertisements, it's become impossible to gather information from them.

000

dahara1 @dahara1.bsky.social · 11/06/2026

I read on X that when Fable was shown Dario's blog "Policy on the AI Exponential," he calmly pointed out that Dario wasn't playing a hobbit. I like Dario, so I hope he realizes it soon.

100

dahara1 @dahara1.bsky.social · 07/06/2026

This might be considered a dark joke, but I couldn't help but laugh.

000

dahara1 @dahara1.bsky.social · 06/06/2026

The Gemma 4 12B performs a long "Think" operation during inference processing. The benchmark script that worked fine on the E4B timed out and did not complete processing overnight.

000

dahara1 @dahara1.bsky.social · 01/06/2026

Yes, it's gone from my screen too.

010

dahara1 @dahara1.bsky.social · 30/05/2026

Should we refute someone who claims, "You're wrong because chatGPT is saying your opinion is wrong"?

000

dahara1 @dahara1.bsky.social · 27/05/2026

I believe that "human value, not AI value," will be important in the world to come, but YouTube are trying to regulate low-quality content created by AI and wrongly suppressing human-made content.

020

dahara1 @dahara1.bsky.social · 26/05/2026

Japanese professional baseball team managger became enraged his daughters fight and pushed one of them to the ground. The 18 year old daughter receiving advice from chatGPT, reported her father to child protective services and he lost his job. it's first time she had ever fought with her father.

000

dahara1 @dahara1.bsky.social · 25/05/2026

What I realized after starting my own YouTube channel is that data and numbers aren't very popular with the public; instead, methods that appeal to emotions and desires are what become popular. The same might be true in the world of AI.

000

dahara1 @dahara1.bsky.social · 20/05/2026

Gemini 3.5 Flash $1.5/$9 Gemini 3.1 Pro $2/$12 3.0 Flash $0.5/$3 3.1 Flash Lite $0.25/$1.5 Claude Opus 4.7 $5/$25 Sonnet 4.6 $3/$15 Haiku 4.5 $1/$5 GPT 5.5 pro $5/$30 5.4 mini $0.75/$4.5 5.4 nano $0.2/$1.25

010

dahara1 @dahara1.bsky.social · 18/05/2026

deepseek v4 flash $0.14/$0.28 v4 pro ※$0.435/$0.87 ※ until 5/31 limited offer Kimi K2.6 $0.95/$4 K2.5 $0.6/$3 GLM GLM-5.1 $1.4/$4.4 GLM-5 $1/$3.2 GLM-5-Turbo $1.2/$4.0

000

dahara1 @dahara1.bsky.social · 15/05/2026

Recently, many hackathons have focused on scoring based on whether the project attracts attention through videos and other means. While this may be meaningful in encouraging participation from non-engineers, is there really a need for a hackathon to be a competition focused on attracting attention?

000

dahara1 @dahara1.bsky.social · 07/05/2026

I decided to install Linux Mint on a laptop that doesn't support Windows 11 and use it as the main server for my AI agent. There's no need to buy a Mini Mac, right?

000

dahara1 @dahara1.bsky.social · 04/05/2026

ChatGPT Images 2.0 is easy to use because, like the free version of Nano Banana, it doesn't have any visible watermarks on the images.

010

dahara1 @dahara1.bsky.social · 02/05/2026

All the models are good. However, in recent years, with official announcements and past examples becoming less reliable than before, Gemini's approach left a good impression on me.

010

dahara1 @dahara1.bsky.social · 28/04/2026

Pure evil However, in the age of the Agent Web, it seems there are also websites that provide false information when they detect AI.

000

dahara1 @dahara1.bsky.social · 28/04/2026

End of Sora. They probably don't have much attachment to the product.

000

dahara1 @dahara1.bsky.social · 27/04/2026

It's really better to run DeepResearch with three models.

000

dahara1 @dahara1.bsky.social · 18/04/2026

I asked Claude, chatGPT, and Gemini to conduct a deep research on the same topic. chatGPT stuck to official announcements. Claude stuck to facts (even if those facts were old). Surprisingly, Gemini got the closest to the essence of the matter.

130

dahara1 @dahara1.bsky.social · 18/04/2026

Claude and Claude Design have separate weekly usage limits.

010

dahara1 @dahara1.bsky.social · 14/04/2026

llama.cpp now supports voice input for gemma4.

000

dahara1 @dahara1.bsky.social · 10/04/2026

For me, the difference between Opus and Sonnet was that Sonnet asked mostly meaningless questions. Lately, Opus has been asking more questions, and I do feel like its intelligence has decreased.

000

dahara1 @dahara1.bsky.social · 09/04/2026

Hermes Agent worked quite well even with gemma4 E4B (8B). Installation was also easy. youtube.com/watch?v=Ml-5L8CddP0

youtube.com

dahara1/gemma-4-E4B-it-UD-japanese-imatrix with nousresearch/hermes-agent demo

YouTube video by webbigdata

010

dahara1 @dahara1.bsky.social · 04/04/2026

While the performance comparison between Gemma4 and Qwen3.5 is not yet complete, one thing that can be said is that Gemma4 is considerably faster (because the Reason output is not very long).

000

dahara1 @dahara1.bsky.social · 04/04/2026

free version of SORA can no longer generate any videos. It shows 1 credit remaining, but it can't generate any Some people say they can generate videos with the Pro version, while others say the Pro version is also full of errors. How many videos can actually be generated per day?

000

dahara1 @dahara1.bsky.social · 29/03/2026

It's now possible to generate it, thank you. It's really a shame that this model will become unusable. I wish they hadn't decided to put watermarks on paid subscriber videos.

000

dahara1 @dahara1.bsky.social · 29/03/2026

Sora has been stuck for over half a day with an error preventing video creation. This is what happens when a service is shut down.

100

dahara1 @dahara1.bsky.social · 28/03/2026

Sora is incredibly crowded for the first time in a while.

000

dahara1 @dahara1.bsky.social · 27/03/2026

I haven't used Seedance 2.0 so I can't comment on it, but Sora had better expressive capabilities than Veo3, so it's a shame it's being discontinued.

000

dahara1 @dahara1.bsky.social · 22/03/2026

The image support in llama.cpp has evolved dramatically without me realizing it. It can now handle multiple images.

000

dahara1 @dahara1.bsky.social · 18/03/2026

The Gemini app can summarize YouTube videos. This is convenient, but if you don't allow Gemini to use your chat history for learning (turn off [Gemini App Activity]), it will completely lie (hallucinate) without accessing the actual video. This could be a fatal mistake if you don't notice it.

000

dahara1 @dahara1.bsky.social · 15/03/2026

In the past, benchmarks for tools included questions that tested knowledge that could not be answered correctly without using the search tool. However, current LLMs can be answered correctly without using the search tool, and get lower score.

000

dahara1 @dahara1.bsky.social · 15/03/2026

Claude March 2026 usage promotion From March 13, 2026 through March 27, 2026, your five-hour usage is doubled during off-peak hours (outside 8 AM-2 PM ET/5-11 AM PT). Usage remains unchanged from 8 AM-2 PM ET/5-11 AM PT.

001

dahara1 @dahara1.bsky.social · 12/03/2026

Qwen 3.5 is clearly overthought and the line of reasoning seems repetitive, but give it enough space to think and you'll be amazed at the results.

000

dahara1 @dahara1.bsky.social · 11/03/2026

If you think Qwen 3.5's performance is lacking, increasing the context length 30k over may improve it. Qwen 3.5 seems to require very long inference outputs.

000

dahara1 @dahara1.bsky.social · 10/03/2026

insanity is doing the same thing over and over and expecting different results While frequently misattributed to Albert Einstein, the phrase's exact origin is unclear, appearing in literature as early as 1938 and in1980s literature.

000

dahara1 @dahara1.bsky.social · 08/03/2026

I read that GPT-5.4 improved creative writing but worsened hallucination, and that makes sense.

000

dahara1 @dahara1.bsky.social · 05/03/2026

I was actually told that today. Wow! Does that AI cost money?

000

dahara1 @dahara1.bsky.social · 04/03/2026

I didn't see any news on BlueSky about Qwen's technical leader Junyang Lin leaving Alibaba Cloud. Is my feed biased?

000

dahara1 @dahara1.bsky.social · 02/03/2026

The idea of handing control of a mini PC to an AI and having it perform tasks at your command isn't particularly appealing to me. Perhaps my sensibilities are outdated, or perhaps I've already learned with VibeCoding that adding new features and changes that way will quickly become impossible.

000

dahara1 @dahara1.bsky.social · 01/03/2026

It's not widely known, but Google's AI-related hackathons held at DevPost often award participants with $100 in GCP credits. Useful for experimentation. geminiliveagentchallenge.devpost.com/rules

geminiliveagentchallenge.devpost.com

Gemini Live Agent Challenge

Redefining Interaction: From Static Chatbots to Immersive Experiences

000

dahara1 @dahara1.bsky.social · 27/02/2026

I'd like to see Anthropic reconsider its stance on the Open weight AI models, but I renewed my Claude Pro plan yesterday. By the way, I feel like Opus is male and Sonnet is female.

000

dahara1 @dahara1.bsky.social · 26/02/2026

@hf.co huggingface, a company with a slightly NSFW atmosphere.

000