00:00 - Shocking AI leap
01:16 - Benchmark battle begins
    See More ps://www.youtube.com/watch?v=ByeX368H3GQ&t=159">02:39 - Model costs revealed
04:26 - Polyglot dominance explained
06:07 - AGI inference cost
08:20 - Price war intensifies
09:48 - Benchmark flaws exposed
11:15 - Human test issues
13:10 - Surpassing massive models
15:04 - Tiny model power
16:24 - Phone-ready AI
17:21 - Ban threat rises
20:00 - Model delay hints
22:18 - Legal chip mess
24:31 - Hardware bottleneck issues
26:05 - Deep implications loom
28:30 - Will it open-source?    See Less

Just because Chinese companies keep using the same "R1" name doesn’t mean they’re ahead or behind US firms

There’s no universal standard for semantic versioning in AI     See More ompanies just pick whatever names or version numbers they want. So using the name or number alone to compare model updates doesn’t tell you much at all    See Less

If DeepSeek begins to build the most powerful models, and the US bans them, who do you think that's going to hurt? After all, very soon it will be the most powerful models who are buildi     See More t powerful models. Also, it seems naive to believe that the CIA isn't monitoring all of the models throughout the world.    See Less

Wait until these cheap Chinese inference chips will be widely available (and have a good yield and sufficient stability). Then Chinese AI tech will rule the world despite all the restriction     See More ogle will have a realistic chance to survive this.    See Less

Didn't ARC AGI price performance figures previously imply deekseek R! to be underwhelming on a price per raw IQ performance metric against openAI O models, even though superficially it h     See More r or 2 magnitude of cheap tokens to throw at inference to catch up? How much better is this? Do any of these bench marks give token to token performance comparisons or are they just gamed by omission of the the inference compute scaling?    See Less

Here is another problem with AI:

When everyone in the World are getting advice/information/structure etc., from the same 2-3-4 AI models ... Everyone will make the same kind of decisi     See More ments/products/politics/laws etc.

This is how everything-bubbles are made. This is how entire civilizations are driven into a corner.    See Less

Those SEAL results seem to indicate DeepSeek R1 v2 has been trained specifically to score high on common public benchmarks. Private benchmarks (like SEAL) show it scoring much lower compared     See More frontier LLMs. I usually only use the absolute best models for coding work, so that means Claude Sonnet 4 and o4 mini high right now. I’d guess DeepSeek R2 will still be less capable than these models when and if it is released this summer. We shall see.    See Less

but NVidia chips are used for inference as well not just training     See Less

there's so much biased info😅
they spent a lot of money. $ 6m is just a lie and OpenAI hasn't released a new model in weeks and there's Gemini 2.5pro Deepthink that's a     See More r than 2.5pro. also deepmind have so many different kinds of Gen AI. deepseek is not as close as you think it is    See Less

So, would you agree with me that all jobs will be be gone within 2 years?

Just adding my 2 cents of paranoia into this conversation.

Love from Kolkata, India     See Less