Hacker News Reader

Comment by junon | Hacker News Reader

Not sure what the fuss is. I tried Deepseek earlier today for the first time and it was even worse than o1 when it came to reasoning skills and following my requests for how I wanted to engage with it.

o1 at least gives it to me straight. When I ask it to engage in more back and forth before assuming what I'm after, it tends to follow through. Deepseek seemed immediately eager to (very slowly) feed me a bunch of made up information thinking that's what I wanted.

I feel as though a lot of people get hung up on these sort of "micro benchmarks" whereas trying to get practical work done is severely under tested. I'm not a fan of openai at all but I don't have the spare compute to run anything locally so o1 suffices for now.

Still don't see how this is anything but a win for Nvidia though.

View on HN

The excitement isn't the capabilities of the model, it's how efficiently it was created. One of the major lessons in AI in the last couple of years was that scale mattered - you would want to throw more and more compute at a problem and that has turned into incredible share prices for Nvidia and incredible investments in data centre and energy generation. If it turns out that actually we didn't need quite such incredible scale to get these results and actually we were just missing some really quite basic efficiency optimizations then the entire investment cycle into Nvidia, data centres and energy generation is going to whipsaw in an incredible way.

Essentially, Deepseek is showing that there is a lot of room for improvement with AIs. To paraphrase Orwell, AIs are a lot more like Alarm Clocks and a lot less like Manhattan Projects.

Which doesn't make sense. If anything it has the chance to make AI finally profitable and financially sustainable. Maybe that is the fear!

It can be a win for nvidia but bad for nvidia stock if the win is smaller than the initially expected win.

The business is doing good, the wannabe traders not so much

Deepseek shows you details of the reasoning so you can trust its answers more and you can correct it when it makes a bad turn.

o1 also does this.

o1 does not show the reasoning trace at this point. You may be confusing the final answer for the <think></think> reasoning trace in the middle, it's shown pretty clearly on r1.

Is this different data or different annotation

I wasn't really referring much to the UI as I was the fact that it does it to begin with. The thinking in deepseek trails off into its own nonsense before it answers, whereas I feel openai's is way more structured.

All you get out of o1 is

    Reassessing directives

    Considering alternatives

    Exploring secondary and tertiary aspects

    Revising initial thoughts

    Confirming factual assertions

    Performing math

    Wasting electricity

... and other useless (and generally meaningless) placeholder updates. Nothing like what the <think> output from DeepSeek's model demonstrates.

As Karpathy (among others) has noted, the <think> output shows signs of genuine emergent behavior. Presumably the same thing is going on behind the scenes in the OpenAI omni reasoning models, but we have no way of knowing, because they consider revealing the CoT output to be "unsafe."

o1 does not output the full CoT tokens, they are not comparable.

R1 is the first model I've used that one-shotted a full JavaScript tetris with all the edge-case keyboard handling and scoring. It also one-shotted an AI snake game. With the right prompts I've found it consistently better than o1 and Claude 3.5 Sonnet.

Reply on news.ycombinator.com