Seems imminent that more companies will want to score rollouts from deployed agents performing in the real world and feed that back into their training pipelines but it's not immediately obvious whether you should score from explicit user feedback, judge model, metrics?
“Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents”
Long-running agents often forget earlier requirements, failed attempts, or useful discoveries, even when that information is still in context.
This paper adds a separate memory agent that tracks important facts and only reminds the main agent when the information is likely to affect its next action.
This improved pass@1 by 8.3 points on Terminal-Bench and 6.8 points on τ²-Bench, while a trained open-weight memory model also gained 3.5 points.
@space_colonist hate to break it to you dawg but it’s the end of the anthropocene whether today or tomorrow, 2030 or 2050, the brain is a chemical computer and distillable into an electrical one. it’s over and we’re so back
"The Pentagon announced they're doing an LCCM buy from Anduril; our Barracuda-500 class. We will produce 7,000 barracudas in 2027. Deliveries on Tomahawks and Patriots are both less than 1,000 a year, so we will make ten times more Barracuda-500s than Patriots or Tomahawks."
Optimizing your efficiency as a human augmented by AI you either have to leverage your ability to go more broad than AI (you have longer memory and a bigger context window) or go deeper than AI (being an expert in some niche)
The former is more exhausting and less interesting
You essentially advocating with a 20 year old to share his political opinions to the massive audience I have is wildly irresponsible.
Im very conscious of the influence I have, which is why I only speak about topics such as Looksmaxing, which I spent years of my life researching.
More influencers need to humble themselves and realize that they are ignorant politically (or just generally fucking stupid).
I recommend using GPT-5.6 Sol Medium as your daily driver, and for really hard problems to switch to Extra High.
If you want the absolute best and are not afraid to burn usage fast, then Ultra is a beast.
Buy this robot for $20,000. Buy that Nvidia super desktop for $94,000. then for the low price of $114,000 you literally have god level intelligence and a slave to stick it in. Universal prosperity is coming
Talking to 5.6 Sol is the first experience I have had of not wanting to skim through a response to get to the point. Signal:noise ratio is getting better!