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

Open models by OpenAI

openai.com

Post reply on HN
Print view View original

171–180 of 909 posts

Re: Open models by OpenAI

#171

Post by NitpickLawyer »

organsnyder wrote: post #162
NitpickLawyer wrote: post #137

Earlier quoted context omitted.

It's apache2.0, so by definition it's open source. Stop pushing for training data, it'll never happen, and there's literally 0 reason for it to happen (both theoretical and practical). Apache2.0 IS opensource.

What is the source that's open? Aren't the models themselves more akin to compiled code than to source code?

No, not compiled code. Weights are hardcoded values. Code is the combination of model architecture + config + inferencing engine. You run inference based on the architecture (what and when to compute), using some hardcoded values (weights).

Re: Open models by OpenAI

#172

Post by atonse »

a_wild_dandan wrote: post #113
asdev wrote: post #67

Earlier quoted context omitted.

Meta is so cooked, I think most enterprises will opt for OpenAI or Anthropic and others will host OSS models themselves or on AWS/infra providers.

I'll accept Meta's frontier AI demise if they're in their current position a year from now. People killed Google prematurely too (remember Bard?), because we severely underestimate the catch-up power bought with ungodly piles of cash.

And boy, with the $250m offers to people, Meta is definitely throwing ungodly piles of cash at the problem.

But Apple is waking up too. So is Google. It's absolutely insane, the amount of money being thrown around.

Re: Open models by OpenAI

#174

Post by Scene_Cast2 »

shpongled wrote: post #143
m_ke wrote: post #104

Earlier quoted context omitted.

No they’re usually done at each attention layer.

Do you know when this was introduced (or which paper)? AFAIK it's not that way in the original transformer paper, or BERT/GPT-2

Should be in the RoPE paper. The OG transformers used multiplicative sinusoidal embeddings, while RoPE does a pairwise rotation.

There's also NoPE, I think SmolLM3 "uses NoPE" (aka doesn't use any positional stuff) every fourth layer.

Re: Open models by OpenAI

#175

Post by a_wild_dandan »

4b6442477b1280b wrote: post #99
a_wild_dandan wrote: post #62

Earlier quoted context omitted.

Right? I still remember the safety outrage of releasing Llama. Now? My 96 GB of (V)RAM MacBook will be running a 120B parameter frontier lab model. So excited to get my hands on the MLX quants and see how it feels compared to GLM-4.5-air.

in that era, OpenAI and Anthropic were still deluding themselves into thinking they would be the "stewards" of generative AI, and the last US administration was very keen on regoolating everything under the sun, so "safety" was just an angle for regulatory capture. God bless China.

Oh absolutely, AI labs certainly talk their books, including any safety angles. The controversy/outrage extended far beyond those incentivized companies too. Many people had good faith worries about Llama. Open-weight models are now vastly more powerful than Llama-1, yet the sky hasn't fallen. It's just fascinating to me how apocalyptic people are.

I just feel lucky to be around in what's likely the most important decade in human history. Shit odds on that, so I'm basically a lotto winner. Wild times.

Re: Open models by OpenAI

#176

Post by rfoo »

foundry27 wrote: post #154

Model cards, for the people interested in the guts: https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7... In my mind, I’m comparing the model architecture they describe to what the leading open-weights models (Deepseek, Qwen, GLM, Kimi) have been doing. Honestly, it just seems “ok” at a technical level: - both models use standard Grouped-Query Attention (64 query heads, 8 KV heads). The card talks about how…

Or, you can say, OpenAI has some real technical advancements on stuff besides attn architecture. GQA8, alternating SWA 128 / full attn do all seem conventional. Basically they are showing us that "no secret sauce in model arch you guys just sucks at mid/post-training", or they want us to believe this.

The model is pretty sparse tho, 32:1.

Re: Open models by OpenAI

#178

Post by logicchains »

foundry27 wrote: post #154

Model cards, for the people interested in the guts: https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7... In my mind, I’m comparing the model architecture they describe to what the leading open-weights models (Deepseek, Qwen, GLM, Kimi) have been doing. Honestly, it just seems “ok” at a technical level: - both models use standard Grouped-Query Attention (64 query heads, 8 KV heads). The card talks about how…

>They did something to quantize >90% of the model parameters to the MXFP4 format (4.25 bits/parameter) to let the 120B model to fit on a single 80GB GPU, which is pretty cool

They said it was native FP4, suggesting that they actually trained it like that; it's not post-training quantisation.

Re: Open models by OpenAI

#179

Post by echelon »

tedivm wrote: post #60
x187463 wrote: post #23

Running a model comparable to o3 on a 24GB Mac Mini is absolutely wild. Seems like yesterday the idea of running frontier (at the time) models locally or on a mobile device was 5+ years out. At this rate, we'll be running such models in the next phone cycle.

It only seems like that if you haven't been following other open source efforts. Models like Qwen perform ridiculously well and do so on very restricted hardware. I'm looking forward to seeing benchmarks to see how these new open source models compare.

This might mean there's no moat for anything.

Kind of a P=NP, but for software deliverability.

Re: Open models by OpenAI

#180

Post by rrrrrrrrrrrryan »

thimabi wrote: post #15
coreyh14444 wrote: post #4

Earlier quoted context omitted.

Because GPT-5 comes out later this week?

It could be, but there’s so much hype surrounding the GPT-5 release that I’m not sure whether their internal models will live up to it. For GPT-5 to dwarf these just-released models in importance, it would have to be a huge step forward, and I’m still doubting about OpenAI’s capabilities and infrastructure to handle demand at the moment.

It seems like a big part of GPT-5 will be that it will be able to intelligently route your request to the appropriate model variant.
Post reply on HN