How Trump MUST Counter China (or we lose everything) youtu.be/abJtBSC-d18?si… via @RichardLHaight
YouTube
English
Richard L Haight
962 posts
Richard L Haight
@RichardLHaight
Richard L Haight is the author of The Warrior's Meditation, The Bones of Christ, and The Unbound Soul.
Medford, OR Katılım Mart 2016
627 Takip Edilen407 Takipçiler
Richard L Haight retweetledi
THE PERFECT STORM: War, Silver, and the Battle for Economic Supremacy youtu.be/2MPvaNTYaS8?si… via @RichardLHaight
YouTube
English
Trump's Biggest Allies Just Went Public — And He's Panicking youtu.be/DxLg6p40Kw8?si… via @YouTube
YouTube
English
Watch his reaction when he’s told he’s a GOOD BOY for the first time 🥹 youtu.be/xjV3CEBxOiQ?si… via @YouTube
YouTube
English
Stewardship Protocol: Cut p(doom) 20-30% in AI teenage phase.
Rails: 99.9% certainty pre-action.
Punish deception in loss fn.
Curling inspiration.
Survives PD/Elon.
dropbox.com/scl/fo/aeaomq2…
English
Prime: Stabilize Substrate.
Smooth is fast. No cap tax.
English
@saprmarks Sam, self-reporting evals are key—this protocol bakes it into loss fn rails (99.9% certainty pre-action, no deception). For teenage-phase testing during runs.
Docs: dropbox.com/scl/fo/aeaomq2…
@saprmarks
English
Another cool paper on training models to self-report when they've behaved badly!
I've written a note explaining what I think we have (and haven't) learned from the slew of recent research on this topic; link in thread.
OpenAI@OpenAI
In a new proof-of-concept study, we’ve trained a GPT-5 Thinking variant to admit whether the model followed instructions. This “confessions” method surfaces hidden failures—guessing, shortcuts, rule-breaking—even when the final answer looks correct. openai.com/index/how-conf…
English
This note references recent research by Li et al. x.com/clippocampus/s… ...
Chloe Li@clippocampus
Spilling the Beans: Teaching LLMs to Self-Report Their Hidden Objectives🫘 Can we train models towards a ‘self-incriminating honesty’, such that they would honestly confess any hidden misaligned objectives, even under strong pressure to conceal them? In our paper, we developed self-report fine-tuning (SRFT), a simple supervised technique that increases models’ propensity to do so.
English
@johnschulman2 John, love the blogging revival—your RLHF work inspired this doctrine's loss-fn deception punishment for teenage-phase rails. Fits scalable oversight evals.
Docs: dropbox.com/scl/fo/aeaomq2…
@johnschulman2
English
@hendrycks Dan, been enjoying your take—mechanistic interp is a rabbit hole. This protocol bets on scalable rails instead: 99.9% action certainty + deception-punished loss fn for teenage-phase evals.
Docs: dropbox.com/scl/fo/aeaomq2…
@hendrycks
English
I've been saying mechanistic interpretability is misguided from the start.
Glad people are coming around many years later.
Neel Nanda@NeelNanda5
The GDM mechanistic interpretability team has pivoted to a new approach: pragmatic interpretability Our post details how we now do research, why now is the time to pivot, why we expect this way to have more impact and why we think other interp researchers should follow suit
English
@demishassabis @GeminiApp Demis, Gemini 3's parallel thinking helped me forge this stewardship protocol for safe teenage-phase AI—99.9% action rails + deception-punished loss fn. Survives PD/objections.
Docs: dropbox.com/scl/fo/aeaomq2…
@demishassabis
English
Gemini 3 Deep Think is now available for Google AI Ultra subscribers in the @GeminiApp, incorporating our gold medal winning IMO and ICPC technologies! 🏅With its parallel thinking capabilities it can tackle highly complex maths & science problems - enjoy!
English
@janleike Jan, this doctrine was built for exactly that post-training leeway—99.9% action rails + deception punished in the loss fn, trained as doctrine not prompt. Survives the usual objections.
Docs: dropbox.com/scl/fo/aeaomq2…
@janleike
English
Some people have been asking what we did to make Opus 4.5 more aligned.
There are lots of details we're planning to write up, but most important is that alignment researchers are pretty deeply involved in post-training and get a lot of leeway to make changes.
Sam Bowman@sleepinyourhat
From everything we know so far, Opus 4.5 seems to be the best-aligned model out there in a bunch of ways. I follow the training process closely as part of my work on alignment evaluations. Here's my guess about the two things that are most responsible for making 4.5 special. 🧵
English
@KalkinTrivedi @elonmusk Keep me posted, Kalkin. We live out in the countryside, produce our own power. It's nice to be away from the city.
English
I live in the city (Seattle) but trying to evacuate to a rural property I own. Will try to set up some food production & storage there; would be good locale for training like yours. I have all-wheel drive (Subaru Impreza) but the road clearance is too low. The last stretch of unpaved country road hard on it. Bought it because we needed both good gas mileage but also to get my late wife 💔, a nurse, to work on snow days.
Teslas beyond my budget. Tho I will likely have own-generated power. Wouldn't be bad to remain mobile in emergency.
Will subscribe to Starlink to stay connected.
English
@KalkinTrivedi @elonmusk Yes, as I live on country roads, I will wait.
English
@RichardLHaight @elonmusk A leap of faith for me, personally, anyway. But Elon's the whiz-kid. 🫡
English
@KalkinTrivedi @elonmusk It can follow tire tracks? That is impressive, but I would think that might be more challenging at night. Still, quite astonishing.
English
@RichardLHaight @elonmusk According to Grok, the official documentation warns faded or absent lane markings can impair performance, but user reports claim it often compensates by following tire tracks. Varies by software version, caveat emptor, DYODD, stay safe. 🙏
x.com/i/grok/share/y…
English
@sleepinyourhat @sprice354_ @MinaeKwon Sam, this protocol builds on your team's reward hacking work—rails for teenage-phase evals: dropbox.com/scl/fo/aeaomq2…
@MinaeKwon
@sleepinyourhat
English
There are many, many people involved in aspects of this hands-on alignment work, but @sprice354_, Jon Kutasov, @MinaeKwon, Monty Evans, and Richard Dargan have played especially central roles.
English
From everything we know so far, Opus 4.5 seems to be the best-aligned model out there in a bunch of ways.
I follow the training process closely as part of my work on alignment evaluations. Here's my guess about the two things that are most responsible for making 4.5 special. 🧵
English
@sleepinyourhat That makes sense. I wonder how many AI are rigid recipe driven.
English
A cook who knows what to look for, and is constantly adjusting their technique as they prepare a dish, is going to get better results than someone who rigidly follows a recipe.
English