There is a counter-argument to the OpenAI competing with Mistral,Meta,Anthropic to release a better model by 2025 is that, all these companies are getting maximum profit from having a model which is faster, cheaper and dumber. But I still think they might still release gpt-5 equivalent before 2025 for research purposes.

After the Meta and Mistral releases the case for yes got much stronger. I'll give it an 80% chance they release this year, no way they will allow so much competition without some kind of reply

Well gpt-4o competitors are already here like Llama 3.1 and claude 3.5 , I think it's about time for OpenAI to shock people with their next thing or eventually fade into obscurity.

I'm quite certain it (whether "it" is Strawberry, the model training on GB200s, or something entirely new) won't finish training until 2025, let alone be done with fine-tuning and RLHF. OAI might unveil a GPT-4.5 or 4.5o that uses the same architecture but has benchmark score improvements until 2025, but not an outright new major version of the model.

20k NO shares at 52%

I'll do more at 48% if anyone's interested. Just not leaving it on bc I'm scared of news, but if there hasn't been any and you wanna trade lmk

My view for NO is that OpenAI appears to be focused on other parts of the model pareto curve, such as new modalities and lower cost. Models will keep improving regardless, but I expect OpenAI's commercial focus to remain on other technologies.

My rationale for yes: Anthropic has already commited to releasing 3.5 Opus before the end of the year and OpenAI should have GPT-5 ready at the time because it started training mid May at the latest.

3.5 Opus will be undoubtably better than 4o and OpenAI will want to release around the same time to avoid falling behind.

I think this market would be appropriately priced around 75%

Do you think 3.5 Opus will be of similar scale/capability to GPT5?

Also, what would you say to this comment made earlier by @AlexanderFavvas

GPT-4 was done training August 2022. However, it was only released March 14, 2023, which is 7 months. During this time, they had the model internally but were safety testing. GPT-4 training was around 5 months. This means that the development process for GPT-4 was one year. Assuming GPT-5 has begun training the same day the screenshot was posted (which probably isn't the case but close enough for an estimate), that places the release date for the public release of GPT-5 to later than May 28, 2025

I would say that OpenAI has gotten much faster at safety testing and getting models ready for release based on how quickly 4o came out after turbo. And yes, I do realize voice mode isn't here yet, but regardless they have clearly gotten faster at readying text generation

I would guess that 3.5 Opus, GPT-5, and Gemini 2.0 Ultra will all be similarly capable and released late 2024

I honestly don’t think you’re right but I hope you are like it would be awesome. Google is testing a new Gemini model on the lmsys leaderboard under the name “test-Gemini”. It could be Gemini 2.0.

I didn't know that, I'll check it out. If it's Gemini 2.0 I would presume it's 3.5 Sonnet level at the absolute worst so I'll give it some of my private LLM evals and report back

Ok tell me how it does. I’ve only got it to appear in the vision arena (either because it’s only in the vision arena or because there’s way less models in that arena so there’s a higher chance of getting Gemini-test.

Alright I tested it some and the results are disappointing to say the least. It got a logic question wrong that only 3 Opus (and not 4o or Sonnet 3.5) can sort of get right.

It also did horribly in the coding exercise I gave it, actually losing convincingly to 4o mini.

I hope for their sake this is a smaller model because if this was 2.0 Ultra then they have lost the AI race.

I'll post the questions below, but will ask that nobody reply with the answers to avoid contamination:

Logic:

You are walking down a path when you come to two doors. Opening one of the doors will lead you to a life of prosperity and happiness, while opening the other door will lead to a life of misery and sorrow. You don't know which door leads to which life.

In front of the doors are two twin guard brothers who know which door leads where. Both of the brothers always lie. You are allowed to ask one single question to one of the brothers (not both) to figure out which door to open. What question should you ask?

Coding:

Write a clone of geometry dash using HTML/JS/CSS that allows the user to jump over an endless sequence of spikes. Put all the code in one file

Damn. Maybe it’s a model designed specifically for vision tasks? Someone posted a thread on twitter showing Gemini-test getting the exact location from an image. That could also explain why I’ve only found it in the vision arena?

I tested its vision by uploading this screenshot from a Bronny James film breakdown and asking what the circled player should do in this situation. Gemini-test thought he was on defense, and suggested he get into a defensive stance. Even 4o-mini was smart enough to realize he was on offense and suggest cutting to the basket

Ok well there has to be something special about it

I feel like we could get a teaser like OpenAI Sora before 2025 if not full rollout if there is enough competition to OpenAI from Anthropic/Meta/XAI/Google and others

https://in.mashable.com/tech/77593/openais-gpt-5-pushed-back-to-late-2025-but-promises-phd-level-abilities

You might want to check comments from 17 days ago here.

Low effort garbage article that copies the bs conclusion the shameless, click-baiting Tsarathustra on twitter did. Mira did not say this.

Maybe you have a position to defend? Others are just asking the questions and sharing what they are knowing. No need for name calling.