Why would reasoning have plateau’d?

I think reasoning ability is not the largest bottleneck for improvement in usefulness right now. Cost is a bigger one IMO.

Running these models as agents is hella expensive, and agents or agent-like recurrent reasoning (like humans do) is the key to improved performance if you look at any type of human intelligence.

Single-shot performance only gets you so far.

For example- If it can write code 90% of the way, and then debug in a loop, it’d be much more performant than any single shot algorithm.

And OpenAI has these huge models in their basement probably. But they might not be much more useful than GPT-4 when used as single-shot. I mean, what could it do what we can’t do today with gpt-4?

It’s agents and recurrent reasoning we need for more usefulness.

At least- That’s my humble opinion as an amateur neuroscientist that plays around with these models.

> Running these models as agents is hella expensive

Because they are dumb so you need to over compute so many things to get anything useful. Smarter models would solve this problem. Making the current model cheaper is like trying to solve Go by scaling up Deep Blue, it doesn't work to just hardcode dumb pieces together, the model needs to get smarter.

You mean like our dumb ass brains? Theirs a reason "saying the first thing out of your mind" is a bad fucking idea, thats what AI's currently do, they don't take a moment think about the answer and then formulate a response, they spit out their first "thought" thats why multi-shot works so much better, just like our own dumb brains.

My brain can navigate a computer interface without using word tokens, since I have tokens for navigating OS and browsers and tabs etc. That way I don't have to read a million tokens of text to figure out where buttons are or how to navigate to places, since my brain is smart enough to not use words for it.

ChatGPT doesn't have that sort of thing currently, and until it does it will always be really bad at that sort of thing.

You are using a hand to hammer a nail, that will never go well, the solution isn't to use more hands the solution is to wield a hammer.

WTF are you even talking about, we're talking about understanding and communication not taking actions, navigating an OS and browser, tabs etc are actions, not thoughts or communication. This model isn't taking actions there is no nail to hammer lol, and if their was you'd be smashing a brain into a nail for some reason.

The topic is agents, the AI acting on your behalf, that needs more than text. What are you talking about?

Your brain is doing this without you realising, just because you aren't verbalising it doesn't mean it's not iterating through its own output and selecting the best answer. Some of our thinking is purely illusion. Models can also do this, some models like Gemini allow you to see this in action where you get to select the best answer, there's also behind the scene methods such as q-learning.

When your muscle memory is instinctively clicking that start button, it's more akin to a very strong weighting after many sessions of reinforcement learning. Our brains may still be dumb but we can quickly say things like 1+1=...2 because we used reinforcement learning to strengthen the weighting back in primary school. We're not sitting visualising an abacus moving in our minds.

It's good in any prompt to give the AI a chance to think without impact (without executing anything, maybe without making anything specifically appropriate for the user to read). That works somewhat similarly to "taking a moment to think."

At that point they'll still have a tendency to use a stereotyped response and stick with it, instead of a thoughtful response, but you can try to address that in prompting too by asking for multiple proposals before choosing one.

Disagree. Even going by your example, AlphaGo uses many iterations of a "dumb" model in order to achieve incredible performance. If it had to single shot the solution with a model 100x bigger, it would perform worse. All that matters is the frontier of intelligence vs cost, and larger foundation models aren't necessarily going to push that frontier forward. AlphaCode hints at that.

I’d be pretty happy if they could just make ChatGPT4 10x faster and cheaper. It’d be fine for basically all of my use cases.

> For example- If it can write code 90% of the way, and then debug in a loop, it’d be much more performant than any single shot algorithm.

OOC, Would this make the academics including algorithms as more or less important in their curriculum? That's a bad win for soceity if it's true.

Reliable agents in diverse domains need better reasoning ability and fewer hallucinations. If the rumored GPT-5 and Q* capabilities are true, such agents could become available soon after it’s launched.

Sam has been pretty clear on denying GPT-5 rumors, so I don't think it will come anytime soon.

Sam mentioned on several occasions that GPT-5 will be much smarter than GPT-4. On Lex Fridman’s podcast, he even said the gap between GPT-5 and 4 will be as wide as GPT-4 and 3 (not 3.5).

He did remain silent on when it’s going to be launched.

OpenAI has been open about their ability to predict model performance prior to training. When Sam talks about GPT-5 he could very easily be talking about the hypothetical performance of a model given their internal projections. I think it’s very unlikely a fully trained GPT-5 exists yet.

Sam has stated that he knows the month GPT-5 will be released.

Given the amount of time and uncertainty involved in training and red-teaming these models, we can assume GPT-5 exists if we take Altman at his word.

It’s going to be launched this year. My buddy’s company had a private demo of gpt5

Ben Horowitz and Andreesen just complimented Sam on their podcast on how smart he is. They then went onto compliment how adept he is at competitive strategy. I wouldn’t trust a word he says about when they will arrive at agi or milestones along the way.

This isn't really new tech, it's just an async agent in front of a multimodal model. It seems from the demo that the improvements have been in response latency and audio generation. Still, it looks like they're building a solid product, which has been their big issue so far.

Its 200-300ms for a multimodal response, thats REALLY a big step forward, especially given it's doing it with full voice response, not just text.

It has to be a separate model doing the voice output right? I can’t imagine they’ve solved true multimodal output from a single model, they’d be bragging about it.

They’re probably predicting tone of voice tokens. Feed that into an audio transformer along with some speculative decoding to keep latency low.

The old voice mode was but everyone including gdb is saying that this one is natively multimodal once it’s fully rolled out, audio in audio out. This has been in the works for a while, you can look up papers on things like OCR-free document understanding and the like but the basic idea is you just train it and evaluate it on whatever media you want it to understand. As long as you can tokenize it it’ll work.

It’s definitely multimodal input. Passing Clip embeddings to an LLM is nothing new, and that’s really all you need for document understanding. It’s almost certainly the same thing for audio. They would have trained a dual encoder that maps both audio and text to a shared embedding space.

What’s not at all clear to me is if they’re doing something special for output. Are you saying OpenAI has moved beyond next token prediction and just hasn’t bothered to mention it?

why does this imply theyve moved beyond next token? is it a throughput issue?

I assume so, you can’t really tokenize audio, at least not high fidelity audio. Audio models like Bark don’t output logits from what I understand. For true multimodal output you’d need a model that can output both logits and audio embeddings.

No, audio is fed directly into the model. There is no text to speech transformer in front of it like there was with chatgpt-4.

Ya so sad that OpenAI isn't more Open imagine if OpenAI was still sharing their thought processes and papers with the overall commity, really wish we saw collaborations between OpenAI and Meta for instance to really have helped push the open source arena further ahead, i love that their latest models are so great but the fact they aren't helping the Open source arena to progress is sad. Imagine how far we'd be if OpenAI was still as open as they once were and we saw collaborations betweeen Meta, OpenAI and Anthropic all working and sharing growth and tech to reduce double work and help each other not go down failed paths.