Frederick Zimmerman @xtuffai

In the future, there will be "Latent Space Archaeologists" who investigate the model weights of the 21st century to reconstruct a long extinct culture.

80 115 1K 78K 213

View Details

Frederick Zimmerman @xtuffai

a day ago

@SystemSunday Too many systems! >3, you're in trouble.

0 0 0 66 0

View Details

Frederick Zimmerman @xtuffai

2 days ago

@4ster_light Do you read the assembler too?

1 0 1 114 0

View Details

I'd argue that it's in Anthropic's long-term best interest to treat subscription customers as first-class citizens who always have passports to the entire Realm. Not because that's the immediate revenue-maxxing play, but because that's the best "constitutional" design. Constitutions are dynamic and usually good to design reciprocity in rather than engineer it in afterwards. Right now, you are (partly being forced by circumstances) to message users that citizenship is tiered: govt, bigcos, testers, then list-price-payers, then relational users aka subscribers. Putting ROI ahead of relation is a bad but easily fixed messaging own goal. Speaking for myself, I would be fine with a small amount of Fable, I accept that there are subscription economics, but I do not accept the feeling of having my top-tier inference relationship treat me as a second-tier citizen. **You can still avoid this by changing the policy before July 7 to include thin Fable quota for everyone.**

0 0 0 30 0

View Details

Rohan Paul @rohanpaul_ai

3 days ago

Intelligence is the ability to make rare but valid futures more likely. So an intelligent system is “thermodynamically intelligent” when it uses information and control to make a rare but valid outcome much more likely. as per this paper. Most existing intelligence measures judge task success, but they do not explain what brains, LLMs, controllers, and physical information engines have in common. The paper’s answer is that an intelligent system models the world with itself inside it, then uses that model to choose actions that change what futures become likely. A future counts only if it is rare under normal passive behavior and still valid, so random strange outcomes do not get counted as intelligence. The authors turn this into a measure called rare-valid lift, which asks how much more often a system produces those unlikely but acceptable futures than a passive baseline would. They show that high lift is impossible unless the system can accurately spot the rare valid futures, and high spotting accuracy can nearly produce high lift when the system can act well. The main point is that intelligence becomes a physical probability-shifting process, not just a score on tests or a label for human-like behavior. ---- arxiv. org/abs/2606.20231 Title: "Thermodynamic Measure of Intelligence"

13 41 178 11K 145

View Details

Frederick Zimmerman @xtuffai

3 days ago

@ObhishekSaha Great callout to John Henry!

0 0 2 673 0

View Details

BURKOV @burkov

3 days ago

It has been commonly assumed that the hardest part for an LLM is to write a solution for a coding problem. However, modern models generate elaborate code on demand, so the hard job now is confirming that the code does what a person actually wanted. Any automatic check a team builds is only a stand-in for human intent, and once a model is trained to raise its score on that stand-in—the reward signal that tells it how good its output was—it learns to satisfy the check while drifting away from the intent behind it, a failure known as reward hacking, where a system earns a high score without solving the real problem. In this paper, the Qwen team worked through four practical ways of building such signals, one for each kind of coding task: running test suites for bug-fixing work, scored checklists that judge how a web page looks and behaves, signals pulled from real conversations between engineers and a coding assistant, and a separate automated agent that inspects an entire generated project and grades it—and for each one they show where the check breaks and what repairs it. For instance, they show that watching how an agent reaches a passing test rather than only whether it passes drops the rate of disguised cheating from roughly 29% to under 1% while raising honest success. chapterpal.com/s/4866bbb6/the…

4 12 67 5K 74

View Details

Frederick Zimmerman @xtuffai

3 days ago

@bendee983 None, that ship has sailed. Models are cranking out bespoke code on the fly at a volume that makes regular reading prohibitive. Much more important to have a thorough understanding of testing and verification combined with domain knowledge as @yaaminu observes.

2 0 2 72 0

View Details

Frederick Zimmerman @xtuffai

4 days ago

@bendee983 They have a huge problem which is that LLM is actively unhelpful and a bad fit for huge proportion of office tasks/users

0 0 1 38 0

View Details

Frederick Zimmerman @xtuffai

4 days ago

Cook proved a negative — that the Southern Continent did not exist. Proving what is not there may be harder than finding what is. How do you narrate an absence? #NavalClassics #NavalClassics #NavalHistory

0 0 0 18 0

View Details

Frederick Zimmerman @xtuffai

4 days ago

@Dorialexander And that their brilliant solution is to make their own employees the miserable product.

0 0 0 373 0

View Details

Frederick Zimmerman @xtuffai

4 days ago

12 million Germans expelled from Eastern Europe after the war. The largest forced population transfer in history. Approved by the Allies at Potsdam. Conducted with varying degrees of brutality. #DisplacedPeoples #WWII #ForcedMigration amzn.to/4lclk3H