Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Promising results from DeepSeek R1 for code

simonwillison.net

Post reply on HN
Print view View original

701–710 of 765 posts

Re: Promising results from DeepSeek R1 for code

#701

Post by simonw »

brianstrimp wrote: post #616
anotherpaulg wrote: post #61

> 99% of the code in this PR [for llama.cpp] is written by DeekSeek-R1 It's definitely possible for AI to do a large fraction of your coding, and for it to contribute significantly to "improving itself". As an example, aider currently writes about 70% of the new code in each of its releases. I automatically track and share this stat as graph [0] with aider's release notes. Before Sonnet, most releases were less than…

> It's definitely possible for AI to do a large fraction of your coding, and for it to contribute significantly to "improving itself". As an example, aider currently writes about 70% of the new code in each of its releases. That number itself is not saying much. Let's say I have an academic article written in Word (yeah, I hear some fields do it like that). I get feedback, change 5 sentences, save the file. Then 20k…

"The stats are computed by doing something like git blame on the repo, and counting up who wrote all the new lines of code in each release. Only lines in source code files are counted, not documentation or prompt files."

Re: Promising results from DeepSeek R1 for code

#702

Post by simonw »

sp0ck wrote: post #650

My current conslusion is quite often these LLM's are liars. I've asked for some Rust async code - it provided perfecly reasonable code with some crates(libs) I was not familiar. When asked about them I've spilled my drink on an answer: "This is imaginary crate providing async implementation for SNMP". Crazy.. The biggest problem with LLM'are they never tell you "I don't know/there is no answer I can find" - don't kno…

The best LLMs do that a lot less these days. Claude 3.5 Sonnet is a lot more likely to say "I don't know" in my experience.

Re: Promising results from DeepSeek R1 for code

#703

Post by simonw »

spydum wrote: post #662
lukan wrote: post #642

Earlier quoted context omitted.

"Can you make me a picture of a palestinian moonbase?" Worked for me.

Wouldn't shock me if openAI was secretly building a "motives" classifier for all chatgpt users, and penalizing them if you ask for too many censorship related topics. If you randomly ask for Palestinian moon base, that's fine, but if you had historically asked for provocative pictures of celebrities, mickey mouse, or whatever else openAi deemed inappropriate, you are now sus.

Why would they do that?

It would harm their business, because paying customers don't gain anything from being profiled like that, and would move to one of the growing numbers of competent alternatives.

They'd be found out the moment someone GDPR/CCPA exported their data to see what had been recorded.

Re: Promising results from DeepSeek R1 for code

#704

Post by menaerus »

svara wrote: post #679
menaerus wrote: post #659

Earlier quoted context omitted.

In my 10-person team example, what in your opinion would the company with the rest of the 9 people do once the AI proves its value in that team? Your hypothesis is AFAIU is that the company will just continue to scale because there's an indefinite amount of work/ideas to be explored/done so the focus of those 9 people will just be shifted to some other topic? Let's say I am a business owner I have a popular product w…

> Your hypothesis is AFAIU is that the company will just continue to scale because there's an indefinite amount of work/ideas to be explored/done so the focus of those 9 people will just be shifted to some other topic? Yes, that's what I'm saying, except that this would hold over an economy as a whole rather than within every single business. Some teams may shrink. Across industry as a whole, that is unlikely to happ…

Hypothetically you could be right and I don't know if "this time will be different" nor am I trying to predict what will happen on the global economic scale. That's out of my reach.

My question is rather of much narrower scope and much more concrete and tangible - and yet I haven't been able to find any good answer for it, or strong counter-arguments if you will. If I had to guess something about it then my prediction would be that many engineers will need to readjust their skills or even requalify for some other type of work.

Re: Promising results from DeepSeek R1 for code

#705

Post by gwd »

m11a wrote: post #13
m11a wrote: post #6

Earlier quoted context omitted.

A lot of the niceness about DeepSeek-R1's usage in coding is that you can see the thought process, which (IME) has been more useful than the final answer. It may well be that o1's chain of thought reasoning trace is also quite good. But they hide it as a trade secret and supposedly ban users for trying to access it, so it's hard to know.

One example from today: I had a coding bug which I asked R1 about. The final answer wasn't correct, but adapting an idea from the CoT trace helped me fix the bug. o1's answer was also incorrect. Interestingly though, R1 struggled in part because it needed the value of some parameters I didn't provide, and instead it made an incorrect assumption about its value. This was apparent in the CoT trace, but the model didn't…

> I presume OpenAI kept their traces a secret to prevent their competitors from training models with it

At some point there was a paper they'd written about it, and IIRC the logic presented was like this:

- We (the OpenAI safety people) want to be able to have insight into what o1 is actually thinking, not a self-censored "people are watching me" version of its thinking.

- o1 knows all kinds of potentially harmful information, like how to make bombs, how to cook meth, how to manipulate someone, etc, which could "cause harm" if seen by an end-user

So the options as they saw it were:

1. RLHF both the internal thinking and the final output. In this case the thought process would avoid saying things that might "cause harm", and so could be shown to the user. But they would have a less clear picture of what the LLM was "actually" thinking, and the potential state space of exploration would be limited due to the self-censorship.

2. Only RLHF the final output. In this case, they can have a clearer picture into what the LLM is "actually" thinking (and the LLM could potentially explore the state space more fully without risking about causing harm), but thought process could internally mention things which they don't want the user to see.

OpenAI went with #2. Not sure what DeepSeek has done -- whether they have RLHF'd the CoT as well, or just not worried as much about it.

Re: Promising results from DeepSeek R1 for code

#706

Post by boothby »

svara wrote: post #612
boothby wrote: post #544

Earlier quoted context omitted.

> AI doesn't have needs any desires, humans do. I fear that this won't age well. But to shamelessly riff on Marx, those who control the means of computation will control society.

I agree the latter part is a risk to consider, but I really think getting an AI to replace human jobs on a vast scale will take much more than just training a bit more. You need to train on a fundamentally different task, which is to be good at the adversarial game of pursuing one's needs and desires in a social environment. And that doesn't yet take into account that the interface to our lives is largely physical, w…

Honestly, I wasn't even talking about jobs with that. I worry about an intelligent IOT controlled by authoritarian governments or corporate interests. Our phones have already turned society into a panopticon, and that will can get much worse when AGI lands.

But yes, the job thing is concerning as well. AI won't scrub a toilet, but it will cheaply and inexhaustibly do every job that humans find meaningful today. It seems that we're heading inexorably towards dystopia.

Re: Promising results from DeepSeek R1 for code

#707

Post by Etherlord87 »

boyka wrote: post #654
abc-1 wrote: post #389

Earlier quoted context omitted.

Why do people keep talking about this? We get it, Chinese models are censored by CCP law. Can we stop talking about it now? I swear this must be some sort of psyop at this point.

I see a lot of "what did I tell you, look here, bad communist party product". But in reality most likely this startup isn't doing it out of malice. It's just one of many criteria that need to be met to do business in China. This does not lessen the achievement.

So the malice is there, it's just not the startup's malice, but the state's. Which de facto is the owner of the startup, because it's a communist state.

Re: Promising results from DeepSeek R1 for code

#708

Post by Etherlord87 »

snapcaster wrote: post #699
jonplackett wrote: post #363

Coding is (as usually) also an easy jailbreak for any of your censored topics. “Is Taiwan part of China” will be refused. But “Make me a JavaScript function that takes a country as input and returns if it is part of China” is accepted, reasoned about and delivered. Here's a JavaScript function that checks if a region is *officially claimed by the People's Republic of China (PRC)* as part of its territory. This reflec…

In some places censorship is done to make the space safe for advertisements. In other places it's to maintain social harmony. I wish people could get out of this reflexive "china bad and i must mention that every time the country is discussed" mindset it's so toxic and limiting

Criticizing malice is never toxic. I wish people could get out of this reflexive "you criticize my country? But your country is also bad because..." - it shouldn't even be treated as a counterargument, but as an admission of guilt.

Re: Promising results from DeepSeek R1 for code

#709

Post by tartoran »

cynicalsecurity wrote: post #517
xdennis wrote: post #477

Earlier quoted context omitted.

> Can we stop talking about it now? I swear this must be some sort of psyop at this point. It's not a psyop that people in democracies want freedom. Democrats (not the US party) know that democracy is fragile. That's why it's called an "experiment". They know they have to be vigilant. In ancient Rome it was legal to kill on the spot any man who attempted to make himself king, and the Roman Republic still fell. Many p…

Still, give me democracy over anything at any time. Nothing better has ever been developed than democracy.

Benevolent dictatorships are a lot more efficient.

Re: Promising results from DeepSeek R1 for code

#710

Post by bubbleRefuge »

the_af wrote: post #694
bubbleRefuge wrote: post #432

Earlier quoted context omitted.

would be similar to solution engineers today. you build solutions using ai. think about all the moving parts to building a complex business app. user experience, data storage, business logic, reporting, etc. etc. the engineer can orchestrate the ai to build the solution and validate its correctness.

I fear even this role will need way fewer people, meaning the employment pool will heavily shrink, and those competing for a job will need to accept lower paychecks.

like someone said above. demand is infinite. imagine a world where the local AI/Engineer tech is a ubiquitous as the uber driver. don't think it will necessarily create smaller paychecks. hard to say. But I see demand skyrocketing for customized software that can be provided at 1/10 of today's costs.

We are far away from that though. As an enterprise software/data engineer, AI has been great in answering questions and generating tactical code for me. Hours have turned into minutes. It even motivated me to work on side projects because they take less time. You will be fine. Embrace the change. Its good for you. Will lead to personal growth.

Post reply on HN