Learning Library

← Back to Library

GPT‑4 Mini Nears AGI, Costs Skyrocket

Key Points

  • OpenAI has quietly released the new “03” model to a very limited pool of vetted researchers for safety testing, with a public “mini” version slated for January and a full rollout planned for the following year.
  • Early testers say the 03 model is edging toward artificial general intelligence, prompting OpenAI to develop unprecedented alignment and red‑team safety measures before broader deployment.
  • Running the most capable version of 03 on the ARC AGI benchmark costs roughly $1,000 per task (and $100 for the mini), prices most users are unwilling or unable to pay for routine answers.
  • By 2025, raw intelligence will no longer be the primary constraint; societies will match tasks to the appropriate level of AI, reserving the most powerful models only for problems that outstrip human expertise.
  • The current hype cycle of ever‑more intelligent language models will give way to pragmatic, application‑focused use, with a clearer understanding of each system’s true capabilities.

Full Transcript

# GPT‑4 Mini Nears AGI, Costs Skyrocket **Source:** [https://www.youtube.com/watch?v=klDJJFB6HAE](https://www.youtube.com/watch?v=klDJJFB6HAE) **Duration:** 00:04:04 ## Summary - OpenAI has quietly released the new “03” model to a very limited pool of vetted researchers for safety testing, with a public “mini” version slated for January and a full rollout planned for the following year. - Early testers say the 03 model is edging toward artificial general intelligence, prompting OpenAI to develop unprecedented alignment and red‑team safety measures before broader deployment. - Running the most capable version of 03 on the ARC AGI benchmark costs roughly $1,000 per task (and $100 for the mini), prices most users are unwilling or unable to pay for routine answers. - By 2025, raw intelligence will no longer be the primary constraint; societies will match tasks to the appropriate level of AI, reserving the most powerful models only for problems that outstrip human expertise. - The current hype cycle of ever‑more intelligent language models will give way to pragmatic, application‑focused use, with a clearer understanding of each system’s true capabilities. ## Sections - [00:00:00](https://www.youtube.com/watch?v=klDJJFB6HAE&t=0s) **OpenAI's Limited GPT‑03 Release** - OpenAI has handed the highly capable GPT‑03 model to a tiny, vetted researcher pool for safety red‑team work, emphasizing its near‑AGI performance, upcoming public rollout, and the significant economic and cost considerations of its broader adoption. ## Full Transcript
0:00at 10:00 a.m. today December 20th open 0:03AI closed their 12 days of open aai ship 0:06Miss by releasing the 03 model to a very 0:10very very tiny group of researchers you 0:13are not going to get it I am not going 0:14to get it this is for safety researchers 0:17only and you have to apply and you have 0:19to be cleared the reason for that is 0:22that they are going to be releasing this 0:24to the public in the next year they say 0:27in January for the Mini version and then 0:29after that for the the full version of 0:3003 but they need to complete red teaming 0:33they need to complete safety work first 0:34and that safety work has only become 0:36more important as the model has become 0:39more capable in fact open AI said they 0:41had to Pioneer a new kind of alignment 0:44testing and safety work just for the 03 0:46model because it's so capable and that 0:49brings me to sort of the big headline 0:51here 0:5203 according to the people who have 0:54played with it and tested it is 0:57approaching general intelligence 1:00and we should start to think about where 1:04we want to use it and apply it if it is 1:06indeed going to be an autonomous system 1:09capable of doing most economically 1:11valuable work more effectively than 1:13humans which is the the rough rule that 1:16uh open AI keeps for what artificial 1:18general intelligence 1:20means the trick is it's not 1:24cheap so even as we think about what to 1:26do with it we have to think about what 1:28we're willing to spend on it and I think 1:30that for those of us who are listening 1:32to this and saying well there goes my 1:33job it's not exactly that clear I looked 1:37at a graph displaying the costs of 1:40running heavy compute tasks appropriate 1:43for 1:4403 in the arc AGI test 1:48suite and I'll link The Arc AGI test 1:51report for 03 here it is over 1:55,000 per task and I want to ask you are 1:59you you ready to pay $1,000 for an 2:02answer from chat 2:05GPT I not very many of us 2:10are are you ready 2:13to are you ready to pay $100 for an 2:16answer from 03 mini that's the cost that 2:1803 mini was running not very many of us 2:21are and so the reason I call that out is 2:26because we are going to live in a world 2:28in 2025 2:30where 2:32intelligence is not a bottleneck 2:36anymore and instead we have to think 2:38about how to apply the correct level of 2:41intelligence for the task that we want 2:43to accomplish and 99% of the time it is 2:46going to be less intelligence than the 2:48maximum available and we will bring in 2:5103 as a society when we really need to 2:54solve a hard problem like if we want to 2:56solve something that a PhD can't 2:58effectively solve that's fine 3:00we would bring in 03 for that but in 3:03most cases for most of our work we 3:05wouldn't need it and we would have 3:06plenty of cheap capable intelligent 3:09models and so we've been in this Rush 3:13since 3:142022 of getting more and more 3:16intelligent models and sort of being in 3:18awe of how well they do at mimicking 3:20human language and intelligence and 3:22calling out their shortcomings that's 3:23been the story the story in the years to 3:26come is going to be we don't actually 3:29know all the capabilities of these 3:31systems sometimes their intelligence 3:33outstrips us we use them for specific 3:36applications and for most of us 3:38intelligence is just like the air we 3:40breathe it's no longer a bottleneck we 3:42can have all the intelligence we need at 3:44our fingertips it's just about getting 3:46the right intelligence into place that 3:48is a massive change uh and it's going to 3:51take a long time for us to figure out 3:52what it means and 2025 is the year we're 3:54going to start to do that so 03 is here 3:58and we're all going to find out what 3:59that means uh when it starts to come out 4:01next year cheers