191–200 of 518 posts
Earlier quoted context omitted.
> sustained profitability (see Netflix for another example) What? Netflix is incredibly profitable.
Probably a bad example from my part, but also because of increasing the costs and offering a tier with ads. I was mostly talking about the Netflix as it was originally concieved. "Give access to unlimited content at a flat fee", which didnt scale pretty well.
Earlier quoted context omitted.
-pro models appear to be a best-of-10 sampling of the original full size model
how do you sample it behind the scenes? usually best of X means you generate X outputs and you choose best result. if you could do this automatically, it would be game changer as you could run top 5 best models in parallel and select best answer every time but it's not practical because you are the bottleneck as you have to read all 5 solutions and compare them
remember they have access to the RLHF reward model, against which they can evaluate all N outputs and have the most "rewarded" answer picked and sent
Earlier quoted context omitted.
I used to think the models got worse over time as well but then I checked my chat history and what I noticed isn't that ChatGPT gets worse, it's that my standards and expectations increase over time. When a new model comes out I test the waters a bit with some more ambitious queries and get impressed when it can handle them reasonably well. Over time I take it for granted and then just expect it to be able to handle…
Re-run your historical queries, or queries that are similarly shaped.
how do we know it's not a quantized version of o3? what's stopping these firms from announcing the full model to perform well on the benchmarks and then gradually quantizing it (first at Q8 so no one notices, then Q6, then Q4, ...). I have a suspicion that's how they were able to get gpt-4-turbo so fast. In practice, I found it inferior to the original GPT-4 but the company probably benchmaxxed the hell out of the tu…
> Today, we dropped the price of OpenAI o3 by 80%, bringing the cost down to $2 / 1M input tokens and $8 / 1M output tokens.
> We optimized our inference stack that serves o3—this is the same exact model, just cheaper.
Earlier quoted context omitted.
This is almost certainly what they're doing and rebranding the original o3 model as "o3-pro"
Nope, not what we’re doing. o3 is still o3 (no nerfing) and o3-pro is new and better than o3. If we were lying about this, it would be really easy to catch us - just run evals. (I work at OpenAI.)
Has anyone noticed that OpenAI has become "lazy"? When I ask questions now it will not give me a complete file or fix. Instead it tells me what I should do and I need to ask a second or third time to just do the thing I asked. I don't see this happening with for example deepseek. Is it possible they are saving on resources by having it answer that way?
Yeah, our models are sometimes too lazy. It’s not intentional, and future models will be less lazy. When I worked at Netflix I sometimes heard the same speculation about intentionally bad recommendations, which people theorized would lower streaming and increase profit margins. It made even less sense there as streaming costs are usually less than a penny. In reality, it’s just hard to make perfect products! (I work…
how do we know it's not a quantized version of o3? what's stopping these firms from announcing the full model to perform well on the benchmarks and then gradually quantizing it (first at Q8 so no one notices, then Q6, then Q4, ...). I have a suspicion that's how they were able to get gpt-4-turbo so fast. In practice, I found it inferior to the original GPT-4 but the company probably benchmaxxed the hell out of the tu…
how do we know it's not a quantized version of o3? what's stopping these firms from announcing the full model to perform well on the benchmarks and then gradually quantizing it (first at Q8 so no one notices, then Q6, then Q4, ...). I have a suspicion that's how they were able to get gpt-4-turbo so fast. In practice, I found it inferior to the original GPT-4 but the company probably benchmaxxed the hell out of the tu…
Hw about testing same input vs output with same seed on different dates. If its a different model it will return different output.
OpenAI dropped the price by so much that the server also went down.
Is the price drop really the reason for their recent outage? Or is the price drop an attempt to cover up bad news about the outage with news about the price drop?
This makes no sense. No way a global outage will get less coverage than the price drop.
Also the earliest sign of price drop is this tweet 20 hrs ago (https://x.com/OpenAIDevs/status/1932248668469445002), which is earlier than the earliest outage reports 13hrs ago on https://downdetector.com/status/openai/
Earlier quoted context omitted.
How are LLMs cached? Every prompt would be different so it's not clear how that would work. Unless you're talking about caching the model weights...
This document explains the process very well. It’s a good read: https://platform.openai.com/docs/guides/prompt-caching
> Regardless of whether caching is used, the output generated will be identical. This is because only the prompt itself is cached, while the actual response is computed anew each time based on the cached prompt