Why wouldn’t it? Making the model 10-15% smarter, whatever that actually means, would make many people very happy. I just wouldn’t trust the information that’s currently circulating based on hearsay.
As a enthusiast I am hoping for something good and worthwhile.
Don’t get me wrong above, INCREMENTAL IS GOOD!
If it reduces hallucinations by 90%, then I’m all about it!
I think though the hype train is making it sound like we hit the Singularity or something. And similar to what @qrdl says, show me the money!
But hey, can’t wait for SORA still. Maybe that’s unveiled at dev day?
Intuitively this makes sense, but it goes against The Bitter Lesson, which unfortunately is still valid.
Maybe with “wet-ware” or physical neural networks this will be different and irrelevant, but I (for now) have a hard time seeing how a real-life neural network modeled after our own could reach a level of intelligence higher than humanity. That’s a lot of constraints to make it useful. Going to be crazy when this does happen though 
I won’t believe it until I see ThoughtEmporium have a neural network play Doom! 
If the resources and computations can get better (BitNet when pls), weights become easier to manipulate & understand, and the training data becomes more rich, then who’s gonna stop this train 
Looking forward to SORA and strawberry at Devday!
At the obvious risk of being completely wrong, these are my predictions about Strawberry.
The biggest breakthrough in Strawberry is about planning…long term dynamic planning. Long term planning, that implies amongst other things, involving reacting to data, information and knowledge that a system does not currently have; but may in the future.
Given 1,2 & 3, the formulation of problem statement will become key and Strawberry will perform this “long term” research; as opposed to “good for daily tasks”; till it solves the problem.
Different orgs will have very different goals and each one of those goals will spawn a different “universe” with the unimaginable different meaningful combinations and permutations of the data at the disposal of Strawberry along with the dynamic planning with the ability to react to external changing data, information and knowledge.
I am on the same line, except I would be upset to find out that they solved the “strawberry issue” by just using tools instead of attacking the issue on a fundamental level, which is tokenization. We can already use tools! Applying tools is not anything revolutionary.
Like imagine if when concrete was invented people realized that rebar was needed to support it even further against tensile forces. Then the concrete company was hyping up a new release. “We now sell concrete WITH rebar!!! REVOLUTION!!” 
I am really hoping it’s a smart combination of identifying characters when necessary (letters, numbers, as they are meant to be seen) and then having an internal processing stage where it can go back to it’s own text and reconsider what it has typed out, instead of being forced to finish what it started. Jeez, I use this exact format in half of my posts.
I have (I think) seen this happen with Cursor (the IDE) where it jumps back and forth between lines of code while writing it out. Is very cool to watch.
But isn’t the point that there has been no official communication from OpenAI about “strawberry” whatsoever? I’m comparing this to Q*, which has been hyped as some kind of magical breakthrough based on rumors and unconfirmed leaks.
Don’t get me wrong either, a leap in capabilities would be fantastic. But OpenAI hasn’t announced anything about a “strawberry” that would perform so exceptionally well. Or did I miss something?
In short: Who should we expect to show us the money, and why? What currency are we even talking about?
Not explicitly but they definitely are supporting the hype wagons with cryptic messages.
We, the people are relentless in being entitled! If you show us breakthroughs we now expect breakthroughs always!
OpenAI using the veil of “safety” doesn’t help the case either. Obviously if you drip feed very cryptic information people will invent the most insane crackpot theories.
As some of the researchers on this forum focusing on multi-agents have already discovered, this, too, is completely doable with a different mindset as of today. For example, with graphrag, you can already generate the set of next questions to ask. In other words, I would be disappointed if the quote above is the only thing that they have in “Strawberry”.
If Strawberry is what I have outlined above, it is definitely a security risk.
The currency is tokens and the value is intelligible, actionable sequence of tokens that we have not yet seen.
I am not aware of any cryptic messages 
Anticipation is the greatest joy.
Agreed. I think it’s all fun but can definitely understand why there’s some frustrations from it.
In due time! I’m looking forward to trying it out.
Are you telling me this wasn’t his enthusiasm about fresh growing strawberries!?!
Me and my gardening club will never recover from this
The approach I am thinking with neuromorphic computing is more in line with efficiency gains (less electricity) and more processing power per watt.
Similar to what Intel is up to these days:
This sort of thing would be required to sustain a continually learning strawberry! 
These two Youtube videos from Matthew Bergman answer what it is, how it works, what it means for end-users;
Hope that helps
Unfortunately not a super high quality source, considering sus-r turned out to be grok 
Lots of theory craft and borderline cargo-culting in this space drowning out actual knowledge, were it to exist 
Why wouldn’t it? Making the model 10-15% smarter, whatever that actually means, would make many people very happy.
If Strawberry is just another CoT agent framework,10-15% smarter is super weak. 3rd parties can do much better than that.
A base model that is 10-15% smarter or cheaper would be great. Love the mini! LLama 3.1 is awesome.
But the speculation is that Strawberry is not a base model.
This is super cool, and I’m really excited to see the results. The electricity demand of current AI is 100% unsustainable.
I really hope they can get some control and generate some cool stuff with this. Neural networks are so freaking cool, but are still a completely untamed beast.
This is what I understand as well. It seems like whatever the upcoming model is (Orion, which is a great name, I’m so happy they’re moving away from GPT names) will be trained using whatever Strawberry is. Which doesn’t necessarily mean it’s not a model, but that it’s just not planned to be a general purpose model
When all is said and done, it seems to me this topic isn’t about frustration with technology, disillusionment, or the feasibility of superintelligence.
Rather it appears to question marketing practices that create hype before any concrete developments.
Still, I wouldn’t comment on specific companies’ or organizations’ marketing tactics.
I can see how you might conclude that, but there is a much larger and more pressing issue that is specific to this topic, which is sb 1047, the singularity, and the potential impact on our species.
Rumors of a super powerful AGI agent framework are not just another vaporware joke. It’s a potentially existential development that affects us all.
At least, if it’s true. Personally, I remain skeptical until I see some kind of novel advancement in math or something similar.
If it’s not true, then at most it’s yet another agent framework in a crowded industry of agent frameworks. So, yeah. Not very exciting.
If it’s an improvement on the base model, cool. Every little bit helps.
