Entropy Craftsman
Joined August 2011
- Tweets5.8K
- Following3.5K
- Followers223
- Likes3.6K
Focused Entropy retweeted
You're early. Most people still don’t get it.
We are not in a normal decade. We are at the bottom of the steepest curve in human history.
Focused Entropy retweeted
Banger paper from Salesforce AI Research on agent memory.
The overall finding is that you want to store raw trajectories and decide what to extract from them when the next task arrives, instead of summarizing each run when it ends.
Just-in-Time Memory uses a curator that reads the retrieved traces together with the new task and writes a short memory payload for that task. Because the payload is used right away, the curator can be trained on whether that same task succeeds.
On ALFWorld, WebShop and tau2-bench it beats the strongest baseline by 16.2, 16.3 and 3.9 success-rate points. Even the untrained curator matches or beats memory that is written when a task ends.
Paper: academy.dair.ai/papers/just-…
Focused Entropy retweeted
Jensen Huang: don’t mistake engineering vocabulary for evidence of a machine mind. AI is still software, not a human mind inside a machine.
"We can't make jokes about all this stuff, we're scaring the American public."
Words like “spawn,” “parent,” “child,” and “kill” have existed in computing for decades; Giving those same mechanisms human characteristics today because of AI is unnecessary and misleading.
----
From "The Ezra Klein Show + New York Times Opinion + New York Times Podcasts" YouTube channel, (full video link in comment)
Focused Entropy retweeted
If Democrats gain control of AI, they are going to turn the entire world into HR. You will never be able to get away from the HR Lady. She will be in your refrigerator. She will be in your toaster. There will be no escape.
Focused Entropy retweeted
▪️Using an open-source model = 20 years in jail.
▪️All known AI researchers = under Government surveillance.
Another PauseAI volunteer, Max Winga, who's also a research engineer at Conjecture, presents more "creative" ideas on how to save us all from the AI apocalypse.
Trump and Xi need to mutually decide to "stop AI progress" and then:
1. "Using any open-source model is a huge crime - downloading it, having it in your possession on your hard drive, certainly running the model, distributing the model - make these significant crimes. Any of these things, 20 years in jail. Or even harsher penalties."
2. "You could stop all research. You could scrub the internet of the papers, and you can take all the people who were working on the models, and, at most draconian worse, 'we are saving the world, we will make some sacrifices,' you assign a spy or a government watch person to every single one of these people and ensure that they aren't designing AI models."
"Ultimately, when the fate of all of humanity is at stake, these are extremely small sacrifices to make."
Focused Entropy retweeted
I’m sorry but America did not spend 250 years turning “what’s over there?” into the most powerful civilization on earth just to finally arrive at the most consequential technological frontier in generations and decide actually we’re scared of frontiers now. We ARE the fucking frontier. We cross oceans. We build railroads across continents. We split atoms. We put men on the moon. We turn sand into computers and computers into a global nervous system.
Of COURSE AI is terrifying. The frontier is SUPPOSED to scare the shit out of you. That’s how you know you have, in fact, located the fucking frontier.
The answer is not recklessness, but it sure as fuck isn’t paralysis either. America’s comparative advantage is an unusually high national tolerance for ambitious weirdos staring at impossible things and saying “Yeah... I think I can build it.” That trait is rarer than oil, harder to manufacture than semiconductors, and responsible for a frankly preposterous amount of the modern world.
And you absolutely will not fucking regulate it out of us.
Focused Entropy retweeted
im not worried for me hon, i'm worried for you and the others. it's psychologically similar to people who fall for lottery scams and it's heartbreaking to watch
They really have no clue what’s going on
Focused Entropy retweeted
It’s so funny to me when I encounter people who have this weird sort of smug attitude about the fact that they don’t use AI. As if they’re above it/better in some way.
So fucking cringe. You’re just telling me you’re a moron but you think you’re one-upping me or something? Use AI or don’t, I actually don’t care. But stop embarrassing yourself.
Focused Entropy retweeted
When will the narrative become the obvious one? This is why Dr Dre likes it, and young ppl resist:
AI is like GameShark.
Beginners get to beat the game with cheatcodes, but then will resist it as a crutch to mastery
Masters who already beat the game will enjoy it more to fuck up, experiment, and glitch the internal rules of the game
Using game shark in a normal competiton to inflate your stats? nobody will think that’s fair.
Using gameshark to run a new competiton where all the items and dungeon doors get randomized? interesting new challenge only for masters of the game.
Use ai to make it harder
You’re over there w/ infinite rare candy
We’re over here corrupting the save state trying to catch Missingno
We are not the same.
Focused Entropy retweeted
Very interesting new paper from NVIDIA.
(bookmark it)
It takes a closer look at evaluating agent skills.
Enterprise teams are starting to leverage shared skill libraries, and the review gate is typically a scanner that checks structure, style, and security.
NVIDIA measured whether that gate predicts anything.
Across 145 real skills from internal and public catalogs, structural scan scores correlate with LLM-judge quality at a Spearman rho of 0.14.
ACES proposes Skill Lift instead.
In other words, run the same task twice under the same model, sandbox, workspace, and scorer, once with the skill loaded and once without. Then you measure the difference in what the agent completed.
They scored 947 paired cases from 58 production skills across four harnesses, normalizing trajectories into a shared Agent Trajectory Interchange Format, so results compare across harnesses.
They fins that the largest process-metric gains appear in skill execution, behavior check, and skill efficiency.
Paper: arxiv.org/abs/2608.20614
Track more trending AI papers in our academy: academy.dair.ai/
Focused Entropy retweeted
Long-context recurrent models are attractive because they can summarize everything seen so far into a fixed-size internal state, but that efficiency creates a problem: when a token first arrives, the model does not yet know which details a later task will need, so useful information may be compressed away.
The authors of this paper from @MIT_CSAIL propose “dynamic compression”: keep the original sequence available, and once the current task reveals what matters, selectively revisit relevant earlier tokens and process them again through the recurrent network, effectively rewriting its compact working memory around the present need.
In a controlled experiment where the model must remember and later reuse several learned functions, this extra computation greatly reduces the internal state needed for accurate reuse; with three functions, a dynamically rescanning model with about 111,000 state elements achieves lower error than a single-pass model with about 3.1 million.
Learn from the paper with an AI tutor: chapterpal.com/s/1d962314/dy…