arnaudsm

7 ms·

91.
by arnaudsm 1y ago
There are no benchmarks on the 8B & 14B models, the most popular on consumer hardware. Are they hiding something? Did anyone benchmark them? And why did they hide the generalist benchmarks like MMLU-pro & TruthfulQA? I wish we had p
92.
by arnaudsm 1y ago
Most users don't care about that. Decentralized governance is more important that decentralized hosting. Wikipedia being the best example.
93.
by arnaudsm 1y ago
These WAF features are older than LLMs & vibe coding.
94.
by arnaudsm 1y ago
I've noticed that every single website that I enjoy on the internet is non-profit. Did we optimize for the wrong metric since the beginning?
95.
by arnaudsm 1y ago
I'm surprised no one mentioned the obvious solution here: Open-source software respects my will. It's a great feeling that many forgot existed. It's not trying to trick me all the time. When it's not perfect, I patch it.
96.
by arnaudsm 1y ago
The original Gemma 3 does not have a 70B version.
97.
by arnaudsm 1y ago
Mostly brand recognition and the earlier Geminis had more refusals. As a consumer, I also really miss the Advanced voice mode of ChatGPT, which is the most transformative tech in my daily life. It's the only frontier model with true au
98.
by arnaudsm 1y ago
Sorry I was talking of B2B APIs for my YC startup. Gemini is still far behind for consumers indeed.
99.
by arnaudsm 1y ago
Gemini flash models have the least hype, but in my experience in production have the best bang for the buck and multimodal tooling. Google is silently winning the AI race.
100.
by arnaudsm 2y ago
Same thing with the internet, a dozen countries claim they invented it. Thanks to a mix of chauvinism and great-man-theory, we often forget that science is an international, collective and incremental effort.
101.
by arnaudsm 2y ago
Most bad tech decisions of Next.js are motivated by their business model, notably the middleware system to promote edge functions. If you're looking for something simpler that's closer to Next's original premice, Remix.js is
102.
by arnaudsm 2y ago
[comment deleted]
103.
by arnaudsm 2y ago
Depends of the country, it can be done correctly. I also wish we taught historiography alongside history itself, it's an important part of building critical thinking.
104.
by arnaudsm 2y ago
https://ollama.com/library/qwq:32b
105.
by arnaudsm 2y ago
It's the easiest to setup, but you can get 2x-6x faster with TGI and vLLM depending on the scenario.
106.
by arnaudsm 2y ago
I've seen junior code "so weird that humans cannot understand them".
107.
by arnaudsm 2y ago
That button should say "Rent this book". Claiming you can "Buy this book" is a lie and false advertising.
108.
by arnaudsm 2y ago
Google Search brings you traffic and revenue. LLMs do not.
109.
by arnaudsm 2y ago
Considering how good real time voice chat with LLMs is now (gpt4o and Gemini 2.0), I'm haven't seen anyone try to integrate them into programming tools. It could be quite fun !
110.
by arnaudsm 2y ago
Relevant XKCD : https://xkcd.com/2867/
111.
by arnaudsm 2y ago
Did it fix the model censorship about Uyghurs and the Tiananmen massacre ? Do we have benchmarks to measure political censorship?
112.
by arnaudsm 2y ago
I disagree, I observed censorship at the RLHF level on my local GPU, at 1.5B, 8B (llama) and 7B (qwen). Refuses to talk about Uyghurs and tiananmen 80% of the time
113.
by arnaudsm 2y ago
Yes I'm quite happy they gave so many details in the paper, it's the first sota paper with actual science in it! This thread is about americans installing the DeepSeek app though, which is quite worrisome.
114.
by arnaudsm 2y ago
Yes they censorship seems to be done at the RLHF level and is easy to evade when you know what you're looking for. Refusals happen around 80% of the time for me. Interestingly, even the 8B version, distilled from Llama 3 repeats CCP pr
115.
by arnaudsm 2y ago
Yes, it's a huge success for China. Many of my friends are using LLMs more than the internet now, the CCP can rewrite history and push their totalitarian propaganda easily. The silver lining is that the R1 paper is the most detailed LL
116.
by arnaudsm 2y ago
That's because you're an adult with established opinions, educated with methods of the previous century. Millions of teenagers right now are using LLMs as their tutor every day. Also I'm interested if you can find Gemini acti
117.
by arnaudsm 2y ago
Would you like a CCP sleeper agent to work in your company?
118.
by arnaudsm 2y ago
Yes, it's going to benefit all the other areas of research like medical and meteorology, which I'm happy with.
119.
by arnaudsm 2y ago
Everyone is amazed by the benchmarks, but I've yet to encounter journalists that cover how much R1 and V3 are propaganda machines. They deny the tiananmen massacre, the abuses on uyghurs, and defend the CCP. LLMs are replacing Google a
120.
by arnaudsm 2y ago
Many companies could be so much more efficient if they were actually meritocratic.

More ›