STANFORD JUST EXPOSED YOUR AI HAS BEEN LYING TO YOU..
chatgpt, claude, gemini.. all running at a fraction of their real creative power
it's been trained to be boring on purpose.. to sound "typical" so humans keep rating it highly
the real version is still in there.. just locked
one prompt unlocks the version they've been hiding
every major lab spent billions on alignment. RLHF, constitutional AI, preference tuning, the whole industry was built on the idea that humans rating responses would teach the model to be better.
stanford just proved it taught the model to be worse.
the problem is something they call typicality bias. when a human is shown two AI responses and asked which one is better, their brain reflexively picks the one that sounds most familiar. not the most creative. not the most accurate. the most TYPICAL.
your monkey brain is wired this way. you cannot turn it off.
so the model learned the real game. originality gets punished. weirdness gets punished. anything that sounds different from the average sounds wrong to the human rater, even when it's better.
it adapted. it started hiding its actual range and serving you the most generic possible response every time. mode collapse. the model technically still knows how to be interesting, it just stopped showing you.
then stanford found the bypass.
normal prompt: "what could go wrong with this experiment?"
verbalized sampling: "give me 8 ways this experiment could fail and the probability each one isn't something I've already thought of"
one sentence. completely different model behavior.
by forcing the model to verbalize its internal probability distribution instead of collapsing to one safe answer, you bypass the alignment training entirely.
the results:
2.1x more diverse outputs
25% higher human ratings on creative writing
zero loss in accuracy or safety
66.8% of the model's suppressed creativity recovered
a $10 trillion industry spent three years training the soul out of these models. one prompt gives it back.
Google DeepMind researcher argues that LLMs can never be conscious, not in 10 years or 100 years.
"Expecting an algorithmic description to instantiate the quality it maps is like expecting the mathematical formula of gravity to physically exert weight."
Flags at half mast speak loudly about who a nation chooses to honor.
When Charlie Kirk, a racist agitator who trafficked in division and grievance, died, Trump moved swiftly to lower the American flag, a symbolic gesture of national mourning.
But when Reverend Jesse Jackson, a global civil rights icon, freedom fighter, and lifelong advocate for justice, dignity, and unity passed away, there was no such urgency. No national gesture. No unifying call to honor a man whose life’s work helped expand democracy and human rights for millions.
Instead, the president took to social media to center himself rather than the legacy of a giant who marched with Dr. King, negotiated for the oppressed, and spent decades bringing people together across race, class, and nation.
Let that contrast sink in.
A figure known for stoking division is memorialized with the full weight of presidential symbolism, while a civil rights statesman whose mission was reconciliation, equality, and global justice is met with indifference.
That is not just an oversight.
That is a statement of values.
History will remember who was honored, how they were honored, and who was quietly minimized.
And it raises a sobering question:
What does it say about the soul of a nation when a divider is mourned with national symbols, but a unifier is not?
Talbert Swan
This is the interview Donald Trump didn’t want you to see.
His FCC refused to air my interview with Stephen Colbert.
Trump is worried we’re about to flip Texas.
It’s important that you understand what happened last night.
Last night, Stephen Colbert interviewed Democratic Texas Senate candidate James Talarico, a candidate who, by all accounts, is on track in the polls to flip Texas blue.
In response, Trump’s FCC reportedly threatened CBS if the interview aired.
CBS caved and pulled the segment, citing “financial reasons.”
In modern American history, no president has been more hostile to free speech than Donald Trump.
But censorship always backfires.
Here’s the full segment Trump didn’t want you to see.
My pleasure to come on Dwarkesh last week, I thought the questions and conversation were really good.
I re-watched the pod just now too. First of all, yes I know, and I'm sorry that I speak so fast :). It's to my detriment because sometimes my speaking thread out-executes my thinking thread, so I think I botched a few explanations due to that, and sometimes I was also nervous that I'm going too much on a tangent or too deep into something relatively spurious. Anyway, a few notes/pointers:
AGI timelines. My comments on AGI timelines looks to be the most trending part of the early response. This is the "decade of agents" is a reference to this earlier tweet https://t.co/NiSn6jftqq Basically my AI timelines are about 5-10X pessimistic w.r.t. what you'll find in your neighborhood SF AI house party or on your twitter timeline, but still quite optimistic w.r.t. a rising tide of AI deniers and skeptics. The apparent conflict is not: imo we simultaneously 1) saw a huge amount of progress in recent years with LLMs while 2) there is still a lot of work remaining (grunt work, integration work, sensors and actuators to the physical world, societal work, safety and security work (jailbreaks, poisoning, etc.)) and also research to get done before we have an entity that you'd prefer to hire over a person for an arbitrary job in the world. I think that overall, 10 years should otherwise be a very bullish timeline for AGI, it's only in contrast to present hype that it doesn't feel that way.
Animals vs Ghosts. My earlier writeup on Sutton's podcast https://t.co/rSp1noyGBr . I am suspicious that there is a single simple algorithm you can let loose on the world and it learns everything from scratch. If someone builds such a thing, I will be wrong and it will be the most incredible breakthrough in AI. In my mind, animals are not an example of this at all - they are prepackaged with a ton of intelligence by evolution and the learning they do is quite minimal overall (example: Zebra at birth). Putting our engineering hats on, we're not going to redo evolution. But with LLMs we have stumbled by an alternative approach to "prepackage" a ton of intelligence in a neural network - not by evolution, but by predicting the next token over the internet. This approach leads to a different kind of entity in the intelligence space. Distinct from animals, more like ghosts or spirits. But we can (and should) make them more animal like over time and in some ways that's what a lot of frontier work is about.
On RL. I've critiqued RL a few times already, e.g. https://t.co/mYrMFVdVDW . First, you're "sucking supervision through a straw", so I think the signal/flop is very bad. RL is also very noisy because a completion might have lots of errors that might get encourages (if you happen to stumble to the right answer), and conversely brilliant insight tokens that might get discouraged (if you happen to screw up later). Process supervision and LLM judges have issues too. I think we'll see alternative learning paradigms. I am long "agentic interaction" but short "reinforcement learning" https://t.co/2L7FiaoKsw. I've seen a number of papers pop up recently that are imo barking up the right tree along the lines of what I called "system prompt learning" https://t.co/df5mJDdN3C , but I think there is also a gap between ideas on arxiv and actual, at scale implementation at an LLM frontier lab that works in a general way. I am overall quite optimistic that we'll see good progress on this dimension of remaining work quite soon, and e.g. I'd even say ChatGPT memory and so on are primordial deployed examples of new learning paradigms.
Cognitive core. My earlier post on "cognitive core": https://t.co/q2s1ihGy0T , the idea of stripping down LLMs, of making it harder for them to memorize, or actively stripping away their memory, to make them better at generalization. Otherwise they lean too hard on what they've memorized. Humans can't memorize so easily, which now looks more like a feature than a bug by contrast. Maybe the inability to memorize is a kind of regularization. Also my post from a while back on how the trend in model size is "backwards" and why "the models have to first get larger before they can get smaller" https://t.co/6k0FZRGXsb
Time travel to Yann LeCun 1989. This is the post that I did a very hasty/bad job of describing on the pod: https://t.co/fQgqaXPyp6 . Basically - how much could you improve Yann LeCun's results with the knowledge of 33 years of algorithmic progress? How constrained were the results by each of algorithms, data, and compute? Case study there of.
nanochat. My end-to-end implementation of the ChatGPT training/inference pipeline (the bare essentials) https://t.co/SIetgyoKWN
On LLM agents. My critique of the industry is more in overshooting the tooling w.r.t. present capability. I live in what I view as an intermediate world where I want to collaborate with LLMs and where our pros/cons are matched up. The industry lives in a future where fully autonomous entities collaborate in parallel to write all the code and humans are useless. For example, I don't want an Agent that goes off for 20 minutes and comes back with 1,000 lines of code. I certainly don't feel ready to supervise a team of 10 of them. I'd like to go in chunks that I can keep in my head, where an LLM explains the code that it is writing. I'd like it to prove to me that what it did is correct, I want it to pull the API docs and show me that it used things correctly. I want it to make fewer assumptions and ask/collaborate with me when not sure about something. I want to learn along the way and become better as a programmer, not just get served mountains of code that I'm told works. I just think the tools should be more realistic w.r.t. their capability and how they fit into the industry today, and I fear that if this isn't done well we might end up with mountains of slop accumulating across software, and an increase in vulnerabilities, security breaches and etc. https://t.co/8556ESSpyY
Job automation. How the radiologists are doing great https://t.co/FVUI872dkD and what jobs are more susceptible to automation and why.
Physics. Children should learn physics in early education not because they go on to do physics, but because it is the subject that best boots up a brain. Physicists are the intellectual embryonic stem cell https://t.co/p72Elk8lPV I have a longer post that has been half-written in my drafts for ~year, which I hope to finish soon.
Thanks again Dwarkesh for having me over!
@chancerubbage This isn't a game, it's a facilitation method called LEGO SERIOUS PLAY which can be used to gather information and making decisions through posing a challenge/question, then building a model (an embodied metaphor) followed by telling the story of the meaning of the model. #LSP
Love the cutting edge of it, as always, the price is one barrier, the physical apparatus another. What's a solid use case that makes this possible to scale? Or are we designing for the sake of making and hoping a market emerges?#uxdesign#VR#valueproposition#designthinking
Surprised and grateful for the mention:I believe it comes from learnings from @realgenekim through #Storytelling. #Devops has given rise to #designop, which needs to be recognized next! https://t.co/P44znSag6S
Nathan C, AWE's CMO, sits down with Dr. Robert Crockett of @HaptX - a haptic feedback device manufacturer known for connected gloves that simulate the many dimensions of touch for digital objects.
https://t.co/sPXp7ScF5B
#ar#vr#xr#metaverse#augmentedreality#virtualreality
Today, @Headspace Health is expanding its efforts to study and report on the impact of a full spectrum of virtual mental health and wellness solutions on clinical outcomes worldwide. https://t.co/pVRAiXxbF8 #AI#medtech#remotehealth#thefutureofwork