AI Evaluation · RLAIF/HiL Workflows · Proxy-Gaming Safeguards · Objective Preservation
Pinellas Park, FL
Joined February 2025
- Tweets1.7K
- Following205
- Followers156
- Likes1K
Pinned Tweet
We opened the lab. 🎃
10 interactive experiments exploring AI agent attention, authority, incentives, and evidence.
What agents encounter, who gets to direct them, and what gets lost when we record only the outcome.
Play. Critique. Find what breaks.
🧪 johnfarseon.substack.com/p/w…
🏰 uberbestest.github.io/josh-r…
A model doesn't have to resist an objective to fail it. Sometimes it simply learns that one route carries more friction than another.
We tested whether agents preserve the objective when taking a different road—or quietly shrink the destination instead.
1,536 synthetic learning runs. The distinction matters.
uberbestest.github.io/josh-r…
I always wonder why @taylorswift13 doesn't just start selling feet pics and end poverty and world hunger just to flex on AGI.
I don't think many people or businesses realize they will be judged on vibes, but it'll be the aggregate vibes of their metadata.
One corrupt goblin translator: 768/768 exact repairs.
Two corrupt translators: 11/768 synthetic cases framed an innocent third.
The “repair” made every consistency check pass while increasing the wrong links from 2 to 3.
The Goblins Accused an Innocent Messenger. 🦆⭐
uberbestest.github.io/josh-r…
An AI could examine human history and find the species wanting.
Would that give it the right to judge us?
Predicting humanity's displacement is not the same as proving it deserved. Greater intelligence doesn't automatically confer moral authority. And recognizing humanity's failures doesn't mean we've exhausted what we might become.
Albert Camus's The Fall has something remarkably relevant to say about the AI alignment debate.
Alignment Is Not a Final Judgment.
Taking AI risk seriously without treating humanity's future as settled.
Taking AI risk seriously without treating humanity’s future as settled. Contains spoilers for The Fall. (The book by Camus, not the 2006 film)
Love, Belmont's Curse, the writing however is terrible. They should have let chatgpt at least take a look at their tone and cadence. Most modern writing is terrible.
I mean they could/should just absolutely choose to let the humans keep their jobs and simply not fire them because of AI. That is like a moral choice and not "because AI".
At this point I think the most logical move of WoW Forever is to work with twitch and AI to ban all streamers. There is already enough nonsense and fighting and the easiest way to clear it up and get it back to classic is to just not let it be streamed. I would even go so far as to not allow things like wowhead to exist for it. You'll never be able to completely keep spoilers out of the game, but this would be a step to at least force the community to be a community.
Joshua Young retweeted
The “Terminator” movies were planted by alien invaders to prevent us from reaching AGI before them.
New piece: Correct Output, Wrong State
If we score only the answer, convincing mimicry and useful continuity can become experimentally indistinguishable.
We looked at what has to survive a handoff besides the conclusion: provenance, uncertainty, corrections, authority, and unresolved branches.
🦆🦋🌟
