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

461–470 of 765 posts

Re: Promising results from DeepSeek R1 for code

#461

Post by brandall10 »

simonw wrote: post #3

Given these initial results, I'm now experimenting with running DeepSeek-R1-Distill-Qwen-32B for some coding tasks on my laptop via Ollama - their version of that needs about 20GB of RAM on my M2. https://www.ollama.com/library/deepseek-r1:32b It's impressive! I'm finding myself running it against a few hundred lines of code mainly to read its chain of thought - it's good for things like refactoring where it will thi…

If you have a bit more memory, use the 6 bit quant, takes up about 26gb and has been shown to be very minimally lossy as opposed to 4bit.

Also serve it as MLX from LMStudio, will speed things up 30% or so so your 6bit will have similar perf to the 4bit.

Getting about 12-13 tok/sec on my M3 Max 48gb.

Re: Promising results from DeepSeek R1 for code

#462

Post by rileymichael »

catigula wrote: post #445
rileymichael wrote: post #425

Earlier quoted context omitted.

gee, I wonder why the guy with an enormous vested interest in pushing this narrative would say that? in general, the people saying this sort of thing are not / have never been engineers and thus have no clue what the job _actually_ involves. seems to be the case here with this person.

Don't you think software engineers have a vested interest in their jobs being relevant, just with less information?

> Don't you think software engineers have a vested interest in their jobs being relevant

virtually everyone has a vested interest in their jobs being relevant

> just with less information

i'm not sure how someone who has no relevant background / experience could possibly have more information on what it entails than folks _actively holding the job_ (and they're not the ones making outlandish claims)

Re: Promising results from DeepSeek R1 for code

#463

Post by kragen »

gmt2027 wrote: post #134
jeswin wrote: post #31

> 99% of the code in this PR [for llama.cpp] is written by DeekSeek-R1 I hope we can put to rest the argument that LLMs are only marginally useful in coding - which are often among the top comments on many threads. I suppose these arguments arise from (a) having used only GH copilot which is the worst tool, or (b) not having spent enough time with the tool/llm, or (c) apprehension. I've given up responding to these.…

If AI increases the productivity of a single engineer between 10-100x over the next decade, there will be a seismic shift in the industry and the tech giants will not walk away unscathed. There are coordination costs to organising large amounts of labour. Costs that scale non-linearly as massive inefficiencies are introduced. This ability to scale, provide capital and defer profitability is a moat for big tech and th…

I'm tempted by this vision, though that in itself makes me suspicious that I'm indulging in wishful thinking. Also lutusp wrote a popular article promoting it about 45 years ago, predicting that no companies like today's Microsoft would come to exist.

A thing to point out is that management is itself a skill, and a difficult one, one where some organizations are more institutionally competent than others. It's reasonable to think of large-organization management as the core competency of surviving large organizations. Possibly the hypothetical atomizing force you describe will create an environment where they are poorly adapted for continuing survival.

Re: Promising results from DeepSeek R1 for code

#464

Post by catigula »

rileymichael wrote: post #462
catigula wrote: post #445

Earlier quoted context omitted.

Don't you think software engineers have a vested interest in their jobs being relevant, just with less information?

> Don't you think software engineers have a vested interest in their jobs being relevant virtually everyone has a vested interest in their jobs being relevant > just with less information i'm not sure how someone who has no relevant background / experience could possibly have more information on what it entails than folks _actively holding the job_ (and they're not the ones making outlandish claims)

Good counter-points!

That being said, I suspect Dario has very skilled engineers advising him.

Re: Promising results from DeepSeek R1 for code

#465

Post by fsndz »

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…

I think the secret of DeepSeek is basically using RL to train a model that will generate high quality synthetic data. You then use the synthetic dataset to fine-tune a pretrained model and the result is just amazing: https://open.substack.com/pub/transitions/p/the-laymans-intr...

Re: Promising results from DeepSeek R1 for code

#466

Post by gauge_field »

amarcheschi wrote: post #317
gauge_field wrote: post #266

Earlier quoted context omitted.

That is what I am struggling to understand about the hype. I regularly use them to generate new simd. Other than a few edge cases (issues around handling of nan values, order of argument for corresponding ops, availability of new avx512f intrinsics), they are pretty good at converting. The names of very intrinsics are very similar from simd to another. The very self-explanatory nature of the intrinsics names and havi…

If I had to guess, it's both the title ggml : x2 speed for WASM by optimizing SIMD and the pr being written by ai

+ Deepseekai recently being in the headlines + Lack of knowledge around simd extension. Modern social media is interesting...

Re: Promising results from DeepSeek R1 for code

#467

Post by monsieurbanana »

catigula wrote: post #440
monsieurbanana wrote: post #428

Earlier quoted context omitted.

That's a huge wall of text. Ctrl+f 2027 or "years" doesn't turn up anything related to what you said. Maybe you can quote something more precise. I mean, 99.99% of engineering disappearing by 2027 is the most unhinged take I've seen for LLMs, so it's actually a good thing for Dario that he hasn't said that.

I think your text search might be broken, or you missed the context. Dario's vision of AI is "smarter than novel prize winners" in 2027.

Sorry, Dario's Claude itself disagrees with you

> The comment about software engineering being “fully automated by 2027” seems to be an oversimplification or misinterpretation of what Dario Amodei actually discusses in the essay. While Amodei envisions a future where powerful AI could drastically accelerate innovation and perform tasks autonomously—potentially outperforming humans in many fields—there are nuances to this idea that the comment does not fully capture.

> The comment’s suggestion that software engineering will be fully automated by 2027 and leave only the “0.01% engineers” is an extreme extrapolation. While AI will undoubtedly reshape the field, it is more likely to complement human engineers than entirely replace them in such a short timeframe. Instead of viewing this as an existential threat, the focus should be on adapting to the changing landscape and learning how to leverage AI as a powerful tool for innovation.

Re: Promising results from DeepSeek R1 for code

#468

Post by svieira »

simonw wrote: post #199
bigpingo wrote: post #198

Earlier quoted context omitted.

I have always had the same line: AI will blow my mind when it solves an unsolved mathematical/physics/scientific problem, i.e: "AI, give me a proof for (or against) the Riemann hypothesis"

That happened back in 2023: https://www.technologyreview.com/2023/12/14/1085318/google-d...

Actually, it happened _long_ before that - 2018 was when I became aware of this technique, but I'm sure there's previous art: https://nullprogram.com/blog/2018/07/31/ (Prospecting for Hash Functions for those who already know).

That said, this is really brute forcing, not what the OP is asking for, which is providing a novel proof as the response to a prompt (this is instead providing the novel proof as one of thousands of responses, each of which could be graded by a function).

Re: Promising results from DeepSeek R1 for code

#469

Post by lukan »

kragen wrote: post #434
simonw wrote: post #217

Earlier quoted context omitted.

My review of 2024 is a good place to catch up on what's changed in the past 12 months: https://simonwillison.net/2024/Dec/31/llms-in-2024/

That post is the best summary I've seen of what happened in LLMs last year, but what's crazy is that it feels like you wrote it so long ago, and it's only been four weeks! So much has changed since then!

Do you mean mainly deepseek, or did I missed something big?

Re: Promising results from DeepSeek R1 for code

#470

Post by kragen »

catigula wrote: post #448
throwu2728828 wrote: post #426

Earlier quoted context omitted.

[flagged]

It's a shame that AI seems to be causing a lot of despair, even prior to its vision being complete. I was forced to implement AI systems that toasted many of our employees.

Toasted? With, like, an oven? Or do you mean with champagne?
Post reply on HN