Independent commentator Andrew Curran flagged an imminent announcement from a former OpenAI team, explicitly not Safe Superintelligence, that has developed a new architecture aimed at memory efficiency. The post positions the work as coming from outside established major labs and frames the claim as something to revisit once details drop.
I'm posting this prediction now so I can quote it later. There has been a significant breakthrough in architecture - specifically around memory efficiency - not by one of the big labs, but by a team that was spun out of OpenAI (not SSI). They will probably announce it soon.
No technical description, benchmarks, or scaling claims have surfaced yet, leaving open whether the efficiency gains target training, inference, or both.
Memory bottlenecks continue to limit model size and cost, so any credible step forward from an OpenAI spinout could shift how labs prioritize architecture work.
Many users expressed excitement about an ex-OpenAI team's reported AI memory-efficiency breakthrough, calling it a potential step change over bigger models and hoping for more details, while some questioned the source's credibility.
No Digg Deeper questions have been answered for this story yet.
jerrrrrry what did you do?
I'm posting this prediction now so I can quote it later. There has been a significant breakthrough in architecture - specifically around memory efficiency - not by one of the big labs, but by a team that was spun out of OpenAI (not SSI). They will probably announce it soon.
@AndrewCurran_ There's another one in the very near future and DARPA has been working on it since the 2010s. It'll exceed human brain intelligence in less than 20Watts.
@AndrewCurran_ Core automation is quite cool isn’t it
@bayeslord rohan was vaugeposting about preconditioning at inference time, could be block wise TTT
@AndrewCurran_ so are we shorting mem-stocks now?
@AndrewCurran_ @then_there_was Is it the continuous learning one
@Tsucks6432 @AndrewCurran_ Estonia
@AndrewCurran_ @BLACKWELL154380 Ofc it is who else would do it
@AscendNoosphere Okay. I'm not talking about that one.
@browsingatwork @AndrewCurran_ @darkfore8h Curran doesn't understand the technology. That's why his predictions are always wrong and simply follow what the labs say. "AAI will replace all white collar jobs by 2027" to "AI will create more white collar jobs"
Check his history. He doesn't engage either, he just "likes"
@i_hate_intel @browsingatwork @darkfore8h I even liked this post!
@AndrewCurran_ There was a report earlier about 50% decrease in inference costs.
@McDonaghMatthew Yes, that was from OpenAI. I believe that is unrelated.
@AscendNoosphere A significant breakthrough in architecture was imminent?
@RealSchmebulog @AndrewCurran_ Everyone will double their token budget if it's half the cost to run
@AndrewCurran_ Arriving when its imminent is a low hanging fruit
@AndrewCurran_ Well SSI probably has had tons of breakthroughs but don't expect them to announce
@AndrewCurran_ Core Automation and Jerry Tworek? They work on architecture, specifically around memory efficiency.
@egregious_angst @AndrewCurran_ I mean better models is also just lower prices, because it's just bigger models with a pricing that's lower than it otherwise would've been.
@browsingatwork @AndrewCurran_ @darkfore8h Lecun has already been shown to be wrong about a great number of things.