Tencent recently bought 100k-200k H20 to serve R1. [1] I think it's not clear open source will tank nvidia price. And you won't place a lot of bets if the outcome is anywhere from certain.
I think the current US administration significantly undermines the framing of democracy versus authoritarianism. As much as Dario may want to push alternative narratives, most AI practitioners—who have no financial ties to OpenAI or Anthropic—likely view this as an attempt to squash competition from open weight models.
You are really underestimating how many users just forget they have such unused subscriptions, and how much of subscription based company monthly revenue is those that are not used at all.
My vibe question checking suggests otherwise. Even o3-mini-high is not as good as r1, even though it's faster than r1. Considering o3-mini is more expensive per token. It's not clear o3-mini-high is cheaper than r1 either even r1 probably consumes more token per answer.
Americans need to understand that the Chinese are not obsessed with the US. They don't have a saboteur mindset. They want development not because they want US to fail and China to win. It's really sad to look at US state of affairs right now. It used to have a mindset if abundance. It definitely doesn't right now.
Deepseek inference API has positive margin. This however does not take into account R&D like salary and training cost. I believe OpenAI is the same in these aspects, at least before now.
The one your laptop can run does not rival what OpenAI offers for money. Still, the issue is not whether third party can run it, it's just the OpenAI seems not putting API as their main product.
Not necessarily. DeepSeek will probably only threaten the API usage of OpenAI, which could also be banned in the US if it's too sucessful. API usage is not a main revenue for OpenAI (it is for Anthropic last time I checked). The main competitor for R1 is o1, which isn't gnerally available yet.
Brzezinski wrote about US not wanting an integrated Eurasia (the heartland), so that Europe can be more dependant on the US. Germany used to have more integration which make its industry thrive because it can take advantage of Russia's cheap energy and China's industrial goods for its industry. Ukraine war completely changed that, which marks the first time since World War II that Germany has faced two consecutive years of economic decline.
The current copyright system is clearly broken. Copyright itself is not a natural right but a construct that was created relatively recently to incentivize innovation within society. Currently, we still want to train models using copyrighted materials. What we don’t want, however, is for companies to use these materials without giving back, especially to the content creators themselves, such as book authors, publishers, Stack Overflow answer writers, and Stack Overflow as a platform. We need to find a way to properly compensate these contributors to keep book writing, answer writing, and other creative efforts going, preferably while keeping the content publicly accessible. In the future, I hope we can figure out a model that might draw inspiration from music streaming (though perhaps not exactly like it).
This is r1 not r1-lite-preview. Supposedly r1-lite is a much smaller model, where r1 here is the 600+B MoE, same size as Deepseek v3 they released earlier.
It's interesting that nvidia takes somewhat opposite position on this "restricting adversarial nations" perspective. I think OpenAI wants the us government to restrict open source model such as deepseek or qwen from being accessed by US public and preferably other western nation by using leverage on those nations' ability to access AI chips from nvidia. It's pretty cunning strategy to build regulatory capture while using other company's market power behind the usual good vs bad rhetoric.
I don't think this proves that the LLM is just "pattern matcher". Human makes similar mistakes too, especially when under time pressure (similar to non-reasoning model that needs to "use system one" to generate answer on one go). This is further evident that if you specifically ask the models to pay attention to traps, or just ask follow up question "are you sure?", then they usually can get it right.