News for hackers. Stuff that matters.
Re:Grok3 Launch by llm_trw on Tuesday February 18, 2025 @10:37AM (#43090738)
Because the difference between a model that costs 10 million to train and a model that costs 10 billion to train is 6 months.

Deepseek R1 is something that you can run in a garage on hardware that the average software engineer can buy with a months salary and when it came out last month it was better than _every_ other model.

Re:Grok3 Launch by cedws on Tuesday February 18, 2025 @10:55AM (#43090968)
What about third world programmers? They can’t necessarily afford a $5000 GPU. If it weren’t for the “generosity” of tech companies like GitHub granting free LLM usage, they might be locked out entirely. This would put them at a disadvantage, we can argue to what degree but it’s still a disadvantage.

Depending on the curve we’re on, LLMs may grow more resource hungry while becoming closer to human performance in software engineering tasks. It’s not unimaginable this would concentrate productivity in the upper class of software engineers that can afford the hardware and/or licenses.

Re:Grok3 Launch by llm_trw on Tuesday February 18, 2025 @10:59AM (#43091029)
$5000 monthly salary?

My heart goes out to all the oppressed programmers in the EU.

Re:Grok3 Launch by ben_w on Tuesday February 18, 2025 @04:59PM (#43095658)
EU isn't "third world".

Kenya is, and at current exchange rates the range in this citation is $558-$1876/month: https://www.glassdoor.com/Salaries/kenya-software-engineer-s...

Re:Grok3 Launch by llm_trw on Tuesday February 18, 2025 @05:27PM (#43095912)
With salaries like that it's not first world either.

If op is trying to catch up to frontier models locally on a budget 1/5th of what you can get in the west then I can see why she would feel the way she does about Ai.

Re:Grok3 Launch by ben_w on Tuesday February 18, 2025 @06:05PM (#43096282)
What are you even talking about?

You said a month's salary; cedws said "third world" devs can't afford $5000 — and this is correct, third world devs can't afford that. cedws did not say EU, they said third world, at least not here. You said EU, not them.

When you reply to me with "With salaries like that it's not first world either", who is "they"? The country I said explicitly was third world? Because that's a tautology.

Re:Grok3 Launch by Fnoord on Tuesday February 18, 2025 @11:14AM (#43091274)
Deepseek R1 performs well on a 600 EUR Jetson, and a 700 EUR AMD GPU. Both bought during COVID crisis. It is that quick. However, don't ask it about certain sensitive topics.

You can bet your ass Musk is using his AI tools as propaganda tool for his advancement, just like he does with X. We already seen prompt leak of Grok, it wasn't neutral.

Re:Grok3 Launch by mrtesthah on Tuesday February 18, 2025 @12:46PM (#43092750)
This is exactly correct. Grok is already slandering specific journalistic outlets such as “The Information” that publish negative stories on Musk.
Re:Grok3 Launch by mwigdahl on Tuesday February 18, 2025 @11:24AM (#43091471)
No, it was not. It was not better than o1, nor o1-pro. Yes, it was _cheaper_ than those models, and superior in price/performance if the performance was acceptable. But in terms of raw performance it was behind them.
Re:Grok3 Launch by johnfn on Tuesday February 18, 2025 @11:56AM (#43092020)
Chatbot arena leaderboard[1] says otherwise. 4o is ahead now, but that version of 4o was released after r1. r1 was ahead of all version of 4o and o1 at the time of release.

https://huggingface.co/spaces/lmarena-ai/chatbot-arena-leade...

Re:Grok3 Launch by mwigdahl on Tuesday February 18, 2025 @11:38PM (#43098637)
Chatbot arena leaderboard is a good test for vibes and style of response, but not much else. R1's performance in objective benchmarks (coding, etc.) showed very good performance, granted, but inferior to the full o1 and o1-pro models.

It's still a very impressive feat, but it wasn't frontier-pushing.

Re:Grok3 Launch by _heimdall on Tuesday February 18, 2025 @12:47PM (#43092779)
Is it your expectation that as models get cheaper we won't be developing much more powerful models at the higher price range?

Its already worth noting that we already ran into the self hosting at scale problem. People don't want to run a web server and instead accept all the problems that come along with the convenience of social media. Why would LLMs, or any future AI product, be different?