Benjamin Gögge-Feiersinger

games + data + humanities - m.a. game dev + research, b.a. cultural sciences | DE: @bggg

[Benjamin Gögge-Feiersinger @bnggge]

@mitsuhiko https://t.co/wUTKSDWOsw — just saying …

[Benjamin Gögge-Feiersinger @bnggge]

@TencentAI_News There is mismatch between “open-sourcing” and the license file …

[Benjamin Gögge-Feiersinger @bnggge]

@badlogicgames You can check for early access of the thinky machines micro sequence model and plug it into the TTY without any tools. One letter for a letter. If the user wants to interfere, they need to send messages to the agent from another terminal session …

[Benjamin Gögge-Feiersinger @bnggge]

@ivanfioravanti @balamuralimenon On the M5 Pro low power pre-fill and token generation goes down to 2/3 of the full power performance using about 40% of the energy in MLX (based on Mactop).

[Benjamin Gögge-Feiersinger @bnggge]

@digitalix And the fan goes “Eeeeh!”.

[Benjamin Gögge-Feiersinger @bnggge]

@deliprao Comparing SVG-output to SVG-output: this is Gemini Pro 3.1. It took more than a minute. Generally vector generation especially for diagrams or animations can be more efficient than generating millions of pixels … What’s the name of the startup?

[Benjamin Gögge-Feiersinger @bnggge]

@OpenAI It would be great to be able to switch the interface language. The translated interface is subpar ...

[Benjamin Gögge-Feiersinger @bnggge]

@simonw I totally agree with the sweet spot of the 3B model, but at least the reasoning version does not produce any usable German with Q8. Based on the great multilingual performance of the Magistral and Mixtral models I was hoping for better results.

[Benjamin Gögge-Feiersinger @bnggge]

@eliebakouch @maximelabonne In the llama.cpp repo are configurations for 3B, 8B, and 14B versions, if I am not mistaken.

[Benjamin Gögge-Feiersinger @bnggge]

@esratech Thanks for the great talk, Esra!

[Benjamin Gögge-Feiersinger @bnggge]

@SimonXinDong Nice work! The approach is intriguing. I enjoyed reading the paper. Any timeline for a HF release?

[Benjamin Gögge-Feiersinger @bnggge]

@ggerganov Thanks Georgi for all the efforts!

[Benjamin Gögge-Feiersinger @bnggge]

@osanseviero a) Gemma3n is doing a good job in following instructions for thinking. b) Allowing to set a thinking budget between 0 and X,000 would be interesting. c) Add thinking as an adapter. d) Optimize for token efficiency during thinking.

[Benjamin Gögge-Feiersinger @bnggge]

@OfficialLoganK @GoogleAIStudio Is there a deprecation policy for the API?

[Benjamin Gögge-Feiersinger @bnggge]

@mamagnus00 The right order of succession after Chat-GPT 3.5 might be humanity’s last exam …

[Benjamin Gögge-Feiersinger @bnggge]

@lateinteraction Nvidia has released Nemotron-Nano-9B-v2 together with most of the training data including sources: https://t.co/WqjqXiniFL

[Benjamin Gögge-Feiersinger @bnggge]

@simonw The only model that somehow worked for me with Ollama was Llama 3,1. But yesterday was its first birthday, so it’s approaching retirement age.

[Benjamin Gögge-Feiersinger @bnggge]

@AidfulAI @jtdavies To factor in the degradation with growing context windows I would propose to add the hour of the night to this matching analogy.

[Benjamin Gögge-Feiersinger @bnggge]

@googleaidevs Is there an availability guarantee for a minimum time frame and expectations on the length of minimum deprecation periods?

[Benjamin Gögge-Feiersinger @bnggge]

@casper_hansen_ Link: https://t.co/AG2ecEYXZb