WiSaGaN

I recently read a comment I thought was interesting: once you understand why they see genocide in Xinjiang, you will understand why they don't see genocide in Gaza.

"For self-deployed use, please reach out to us for commercial licenses. We will also assist you in lossless quantization of the models for your specific use-cases to derive maximum performance.

The model weights for Ministral 8B Instruct are available for research use. Both models will be available from our cloud partners shortly."

Yes, this seems to be the result of the standard Euclidean metric rather than the high dimension itself. I guess most people assuming the metric to be Euclidean, so it's ok.

They release SOTA open source coding models. [1] Their API us also incredibly cheap due to the novel attention and MoE arch.

[1]: https://aider.chat/docs/leaderboards/

This was first done by deepseek. [1]

[1]: https://platform.deepseek.com/api-docs/news/news0802/

Of course it’s the former. However, you need to contribute over a threshold of years before getting the benefits. This is currently 15, and will gradually prolong alongside with the retirement age, to I think 20.

For reference, according to Worldbank data, Chinese male life expectancy was 76, while female was 81 in 2022.

Anthropic currently does much less hype stuff comparing to openai. It's remarkable that openai was like this until the GPT-4 release, and completly changed since Sam Altman started touring countries.

This feature was first introduced by deepseek. And deepseek will just do it automatically for you. https://platform.deepseek.com/api-docs/news/news0802/

Backward compatibility (in terms of successful compilation) is never absolute. It's just Rust makes it very rare, and usually justifiable (say, fixing sound issue), while still allows making good progress. In this case, it's an inference change, which is easily fixable. Inference breakage is explicitly allowed in Rust compiler policy. You won't get any new methods in standard library if you want absolute backward compatibility.

Exciting! So many models dropped on the 24th. Really to know whether the numerical suffix means YYMM or MMDD!

Utilize strong type system to make the error case unrepresentable. This is great advice to reduce bugs in software in general. It takes more time to think about the problem and to make a design following this. However, a lot of times it is worth the time.

Or programming language design. However, I think during the process, things are still distilled so that new common patterns are incorporated. I am sure null-terminated strings were not a particularly bad idea in the 70s given the constraints developers faced at the time. It's just that later we have different constraints and have gained more experience, thus finally realizing that it is an unsafe design.

I also prefer to use "Coding" or "Hard Prompts (Overall)" instead of default "Overall" in Chatbot Arena scores to determine the actual performance level of LLMs. Seems much more align to my vibe test in terms reasoning. I guess the "Overall" contains a lot of creative tasks, which is not what I use the most in the daily tasks.

I am not bilingual but rather can speak fluently in some foreign language. However, I constantly find myself thinking in the foreign language I know, especially for topics I learned from reading in foreign books.

Another simpler example is when I count I usually count in my native language, even in a specific dialect when I was taught very young!

This is surprising to me because when I learn something, I usually find that trying to speak it out helps a lot of thinking about it clearly. Put it another way, the effort to turn something into language helps forming the thought about it. Thus I don't know whether counts as a counter argument.

This is impressive. I just tested a debug problem I encountered in real life, which previous models including got-4o can’t do. Claude 3 sonnet nailed it in first try.

I remember Ilya's interview about teaching AI to love. Super intelligence on the other hand could become an almight entity in a short time. In that case, we are basically creating God, an entity that is almighty and loves human, preferrably unconditionally.

Would be good if it had unix domain socket (USS) tested. Compared to normal tcp/udp socket, UDS does not go through network stack in the kernel, which is useless when sender and receiver are on the same machine.

This probably says more about Argentinian currency than about Huawei.

Ironically you literally have a piece of investigative journalism in the front page from Reuters about US propaganda aiming to undermine China at cost of Pilipino lives: https://news.ycombinator.com/item?id=40680325

You don't really need a person inside the LLM provider to just use the LLM tech. This is more than that.

Google, Facebook already have signifiant number of "disinformation staff" from US security state. As another big tech firm that collects huge amount of info and get to decide what to present to the general public, it has be added to that list as well.

This gives me the vibe of calling high resolution screens as "retina" screens.

This seems saying that the 1bit model is a bit better than GPTQ Q2. However, I find there are few situations where you would want to use GPTQ Q2 in the first places. You would want to run the F16 version if you want quality, and if you want to have a sweet spot, you usually find something like Q5_K_M of the biggest model you can run.

I believe this military exercise is a response by the PRC to the inauguration speech by Lai. Lai essentially stated that the PRC-controlled mainland and ROC-controlled Taiwan are both sovereign, whereas the previous Taiwan president, Tsai, from the same party, used more ambiguous language regarding sovereignty and only emphasized governance (Yes, there are a lot of nuances in the wording of this relationship). This statement can be construed as an attempt to change the status quo. (We are referring to a third-party definition of the status quo, because the official PRC stance is that Taiwan is part of China, so even taking Taiwan by force is not considered a change of the status quo.) The PRC likely waited to see Lai's speech, checked if it aligned with Tsai's rhetoric, and prepared a response accordingly.

I think Lai's speech was not the most extreme because it contained little mention of "democracy vs. autocracy," which is something even the current US administration would find provocative. However, this speech would still warrant a response from the PRC, at least from the perspective of internal PR, because it differs from Tsai's approach, and they cannot let it go unnoticed.

He was obviously lying, and he probably also knew people would not believe it. I just don't know why he still chose to do it.

At the time, Sam was more powerful than Ilya for sure. But framing their relationship as employee/employer when they were both in the board seems not correct.

qwen is backed by alibaba. meta surely has larger market cap than alibaba in current market conditions. but i am not sure about the "crush" part.

The cause of latency arbitrage is not HFT, it is the fragmentation of liquidity.