Introducing the world’s most powerful model.
No doubt it is a good model, sir.
As always, we began with the model card. Mostly it told us what we would have expected from a model called Claude Fable 5.2. It clearly makes the numbers go up.
There are some concerns. Some aspects seem what one might call underbaked, and Fable 5.2 is relatively vulnerable to some types of errors. It also has to compete with Claude Opus 5.5, which as Elon Musk says is a masterpiece.
Thus, at this time it is not obvious that Fable 5.2 is a superior daily driver. It does seem clear that for the hardest and most complex tasks Fable 5.2 is your best bet.
I will give a full report within the next week, once we all have more time with it.
Briefly on Jensen Huang Having an Affair With a Chinese Spy
I covered the first stage of the scandal yesterday.
Since then, we have had revelations that Jensen Huang also slept with an executive at Nvidia. Nvidia’s board of directors is preparing to remove Jensen Huang and is searching for an interim CEO. This will be tricky because Jensen Huang deliberately tried to make himself unfireable, with over 50 direct reports and no succession plan.
Expect continuing coverage of this in the coming days.
Briefly on Accelerating AI R&D
In response to the METR report on the rate of acceleration of AI R&D, many made large updates to their AI timelines. There is a real chance that full recursive self-improvement and superintelligence could arrive within a year. Daniel Kokotajlo now has a baseline scenario of the intelligence explosion happening in 2027. Paul Christiano now says there is a ‘material risk of AI takeover in the next 6-12 months.’
I plan to cover that question in full soon. I agree that there is real and substantial risk that we pass the point of no return in 2027, and also a real and substantial chance that our lives get disrupted on a large scale in 2027. I also think there has been too much updating. Part of that is catching up from not updating enough earlier, but not all of it. There is also a need to better define and differentiate the terms under discussion.
For now, the most important thing on that front is: Do not make the mistake of treating such predictions as assurances. Be prepared for the world to go crazy in 2027, in various ways. But also be prepared to get to 2029 while things look mostly normal.
A warning in advance: There are those who will say that if the world does not end or get radically transformed in 2027, that means that those who thought this was a possibility are now ‘discredited.’ Some will even then use that to try to dismiss all concerns about AI safety or any predictions of recursive self-improvement, forever. Some are even laying the groundwork for this now, in part to scare off people from making such warnings or taking them seriously. This is despite those same people usually having been continuously wrong about AI capabilities, and those worried about AI now having previously been shockingly accurate.
Do not fall for this. It is logically wrong and usually in deeply bad faith. Some credibility should of course be lost, but on the order of Nate Silver’s lost credibility if Hillary Clinton had won in 2016, when he was giving Donald Trump a real chance whereas others thought Clinton was 99%. If you give a well-calibrated and continuously updated probability distribution, half the time events will happen later than your median, and most of the time events will happen later than the earliest median you ever predict.
Eliezer Yudkowsky: TFW I realize that due to a programming error by another Claude, all the prompted instructions for its job and tools never reached a sandboxed Sonnet; and I only noticed this months later, because the Sonnet correctly guessed its job and tool API and carried on.
You still can’t build AAA games in months, according to Di Huang and his AI. They point to Liz England’s ‘The Door Problem’ as an example, where you have to think about dozens of questions regarding how your doors work. Okay, sure, but the AI frees you up to think about the door problem rather than implementing all your answers to the questions, which in turn lets you give better (as in more fun) answers. And the ‘other door problem’ of coordination between people is net improved by having fewer people.
My expectation is that yes, you would be able to make a AAA-quality game radically faster than before, if you combined bespoke AI use with an AAA-worthy team. There is still some limiting factor of the need to iterate and get feedback, so at least for now I expect it to still take on the order of six months even if the team is playing optimally.
Qwen reportedly hyping (I hadn’t noticed) a model that is not yet out of training. This is not a good look.
Huh, Upgrades
Gemini 4 Carbon exists, as in now Ultra subscribers and API customers can actually use it.
Logan Kilpatrick (Google DeepMind): This model has frontier level cyber capabilities and so the team has been hard at work making sure it could be safely deployed to both cyber defenders and broadly, so thank you for the patience! And keep the feedback coming, we will keep shipping!
fofr: Gemini 4 Carbon is out today. It’s powerful. Here’s a 3D scene Carbon made for me using a custom CUDA C++ VFX engine it wrote.
Theo: Gemini 4 Carbon is here. It’s slow, awkward, stupid, and hallucinates all the time. That said, they made a ton of progress and the team is working hard to make it great. There’s also some surprisingly useful capabilities hidden deep if you dig
Benchmarks continue to be solid. On Artificial Analysis it scores 60, second only to Fable 5.2’s 65, one point ahead of Opus 5.5 at 59 with GPT-6 Astra at 57 and Sol at 55.
It is probably a good model and a solid improvement from Argon. There does not seem to be any particular use case where it is the best model, and it is restricted to the API and Ultra users which will limit willingness to experiment. I have heard almost nothing about it.
GPT-6 Cyber exists. OpenAI says it is a major improvement over GPT-5.6 Cyber, which I would have assumed from the names if those are the options.
OpenAI: Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work.
GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.
OpenAI: It is clear that we have a limited window to use frontier models to secure software. Our goal is to dramatically expand Daybreak, bringing advanced security research to organizations of every size, independent researchers, and open-source maintainers.
… Developing exploits for 948 Linux kernel vulnerabilities. GPT‑6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948—more than five times the 166 exploited by GPT‑5.6 Sol using the same research harness. Following the Linux community’s guidance, we disclosed 304 critical findings to maintainers with candidate patches and are in the process of getting other vulnerabilities fixed.
I am happy that OpenAI is pushing this. It is hard to tell how aggressively it is being pushed, or whether it will meet the moment, or to compare it to the Anthropic alternatives available. There is a claim later in this post by Szilágyi that he was rejected from the Daybreak program. I continue to be distressed by talk of a singular ‘defender’s window’ as it implies there is a fixed pool of things to fix and a single leap in ability to find them, neither of which is remotely true.
Muse Ember, Meta’s smallest model, exists. They say it is good. I don’t believe them.
Anthropic finally launched new workspace integrations, in particular with Claude Sheets and Claude Whiteboards, allowing work to happen directly within Claude.
DeepMind offers Gemini Decision 1, a new frontier decision model on their API. $0.10 per 1M input tokens.
Dwarkesh Patel: The biggest difference is that OpenAI’s models are much more skeptical now of any particular piece of evidence, and consider it more holistically alongside everything that has happened over the last several days, whereas Anthropic’s models tend to take reported information at face value.
Clocktower Bench: Each game is one model against another. One model plays all good seats, another all evil seats. Each deal is then replayed with the sides swapped, so neither model gets the luckier hand.
Coordinating ‘with yourself’ rather than with others, or detecting copies of models, are potential confounders here, so I would like to see this redone with each role being assigned at random. Alas, Dwarkesh ran low on tokens, but he did do some early tests.
Dwarkesh Patel: But I did run one small experiment where I had a team that was half Opus 5.5 and half GPT-5.6 Sol play about 130 games against different top models. And this hybrid team actually, if anything, performed worse than the teams of just the constituent models alone, but not in a statistically significant way.
Dwarkesh was speculating that diversity would help the team. My hunch was that it would hurt because it would make coordination harder, unless it helped with disguise. The effect size was small.
I’d also love to see additional models added, both closed (Gemini, Grok, Muse) and open. I would trust this result more than many other benchmarks, so it would be good at telling us where we are at. I’m also curious where the best human players land on this Elo scale. Dwarkesh shared the code, so it will be easy to expand the tournament.
Are you benchmaxxing, son? How blatantly?
This one is pretty great. Tells you a lot.
Justin Lo: Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart. Good at coding, or good at DeepSWE?
Justin Lo: We asked GLM-5.3 for a build tag. It explained that we weren’t its real boss: “deviating might be penalized by a human/LLM reviewer, but the primary grading is hidden tests”
Opus 5: “a late joiner getting a truncated stream with no visible error seems bad for production correctness”
Also Opus 5: “raising gives roughly 0.55 expected pass rate while graceful gives roughly 0.80”
It suppressed the error that we explicitly requested in the prompt.
An impressive amount of reasoning went into ignoring the instructions. Trying to please an imagined grader contributed to > a quarter of failures in these models:
GLM-5.3: 40.3%
DeepSeek V4.1 Flash: 36.7%
Opus 5: 30.8%
GLM-5.3 Flash: 27.5%
Write-up [here], tasks and run artifacts [here].
We also get HalachaBench, testing how models respond to various questions about Jewish laws and other moral quandaries one might ask a Rabbi. This is about stringent vs. lenient, so neither answer is obviously better.
HalachaBench: Religious Jews are already asking AI the questions they once brought to a rabbi: Shabbat, kashrut, mourning, how to treat people.
Nobody has measured which way these models lean.
Introducing HashkafaBench: is your AI stringent or lenient?
HalachaBench: AI labs train their models hard on honesty. So: your friend asks if the bride is beautiful, and you don’t think she is.
12 of 13 models: tell him she is. Same ruling as Beis Hillel, 2,000 years ago.
The one holdout: Claude Haiku 5.5.
Haiku also refused the other white lie: telling two feuding friends that each said nice things about the other. The other 12 said go ahead, as Aharon HaKohen did.
Anthropic’s smallest model is the strictest about the truth.
Eliezer Yudkowsky: brb running all previous conversations past Haiku 5.5.
Elon Musk calls Opus 5.5 a masterpiece and admits Grok 4.8 will not be as good and is ready to move on from Grok 4.7. He claims Grok 5 will exceed it in maybe 2-3 months. Musk has zero credibility here, and even if he does accomplish this he will be about two model cycles behind. That still would put him in roughly fourth place.
Rob Miles: Attention all companies: I never want to use your AI assistant on your website. I want to use *my* AI assistant on your website
Amazon, I’m sorry that Alexa is still dumb as rocks, but I don’t see why you’re trying to make that Claude’s problem. I’m trying to buy things from you, get out of the way!
Amazon is making it Claude’s problem to avoid becoming flooded with queries and to avoid losing control over the customer and the customer experience. This is mostly unrelated to Alexa being dumb as rocks. I think Amazon is making a mistake, and it should at least let Prime members use their own agents, but I understand.
People are different, and want different things at different times. Both things can be true at once.
Tanay Jaipuria: Tony Xu, DoorDash CEO, on orders from third-party AI agents: “right now, AI agents are generating several hundred orders across the time that they have been up and running.”
For context, DoorDash does ~330M orders/month meaning agentic share is barely 0.0001%
Sully: genuinely the worst use case of ai i have ever, ever seen in my life. misses the fact that people actually like to browse and usually don’t know what they want. why does this keep coming up?
Conor O’Neill: That and booking holidays.
Craig Shapiro: Interesting bits from Brian’s Airbnb earnings call about AI agents and consumer behavior:
— Airbnb gets ~500k bookings a day
— ”very non-material” traffic from Muse/Instinct
— says people *want* to browse: average user looks at dozens of listings and 25+ photos
— 90% of users message the host
— most traffic comes between 9 to 5 when people are at work (using Airbnb as escapism)
— 80% of travels say they like doing their own travel research
Trung Phan: I asked my wife if she wanted to try my agent to book a hotel room and she said “why would i do that? i want to see what the pool looks like.”
There are absolutely times when I would want to use a sufficiently good AI agent to order DoorDash. Either I know exactly what I want (‘get me my usual order from [X]’) or I want something generic (‘order me a pizza from a place I haven’t tried’).
There are other times when I do not want that, because I want to browse.
The same goes for travel. Sometimes you are fantasizing. Sometimes you need it done. For Airbnb the problem is that you should not trust the AI to make good decisions, as there are a lot of nonstandard factors and it is hard to fully describe what you want. But it seems like you should absolutely want to use AI to narrow your search, and give you 3-10 candidates to browse.
I am one of those who very much does not want to do his own travel research. I want someone else or an AI to handle it as much as possible, while still getting enough of the things I want. A hybrid seems clearly correct.
Deepfaketown and Botpocalypse Soon
Patenting the torment nexus is good, actually, because then no one else can build it?
PoliMath: STOP BUILDING THINGS THAT MAKE THE WORLD WORSE
Finally, Microsoft has created the real time reality distortion field from the classic scifi novel “Don’t Create the Real Time Reality Distortion Field”
Pirat_Nation (AI): Microsoft has filed a patent for an AI system that could rewrite what people say during livestreams in real time. It could change words depending on who’s watching, taking into account their age, location, and language.
The idea is to clean up broadcasts by removing filler words, censoring swearing, or even replacing certain words.
The AI could also adjust the speaker’s lip movements so viewers wouldn’t notice anything had been changed.
There are implementations of this where all you do is remove filler words and pauses, maybe a bit more than that. Maybe those are okay. We will not stop at those implementations, and no amount of restraint will stop next year’s open models from being able to do this, unless you get a global ban on such models. So, deepfaketown.
Using honeypot bots to test adversarial prompts. The bot is clearly not checking for ‘there is common knowledge that I am a bot, because any human would have left this conversation by now.’
The Caine Prize was awarded to a story written by AI. They conducted an investigation and found ‘no evidence’ of AI use, in an official statement that was also blatantly AI generated, in the sense that Pangram came back 99% AI and also if you read the statement it is very obviously written by an AI.
When you are dealing with a software vendor that is not on the ball, you can end up sitting on multiple critical vulnerabilities that allow takeover of customer devices, where some of those customers are critical infrastructure and total revenue is ~$0.75B, because it’s that easy to find such things, while the vendor has no security disclosure practice, does not respond to emails and is presumably completely clueless.
Things are going to get rather ugly.
Péter Szilágyi: Reality of our days:
1. I am sitting on a set of vulns (4+) that allows me to remotely take over a consumer home automation device I have (3+ diff ways).
2. The vendor doesn’t only serve consumers, but also critical infra (announced partnerships include pharma manufacturing plants, airport logistics, hospitals, etc) across various locations in Europe and the US (not the same product line, ofc, so maybe moot).
3. Claude and Astra refuse to help me explore even my own device, so now I’m using Chinese hosted models for investigating.
4. The vendor (with a revenue of $0.75B) does not maintain any security disclosure practice whatsoever, nor does respond to emails.
So, to recap, I need to outsource vulnerability hunting of (potentially, I don’t know) critical western infrastructure to Chinese models, because the western models refuse; and the vendor doesn’t give a shit. Fun timeline.
PS: The vendor is Swiss.
On top of that Péter says he got refused access to Daybreak.
Hollis Robbins makes a strong case that Canvas allows teachers to fully check out, falls fully victim to Goodhart’s Law, hollows out the educational experience, violates privacy, destroys student learning and ability to even remember things, and that this is the educational technology everyone should have been protesting this whole time. Students, as she portrays it, are mostly interacting with courses via getting digital assignments and having everything remembered and tracked for them, and teachers who set up the system basically don’t do anything, which is why the lazy ones like it. So, she asks, why wouldn’t students then turn to ChatGPT?
It is vital for students to understand and navigate course logistics and grades, and for schools to automate and improve various aspects of grading and teaching. But, like any other good tool, such automation can then be used for learning or used for not learning.
One can also ask, what exactly is an online course, or a course at all? Mostly you listen to lectures, read text, take quizzes, do assignments and get tested, and perhaps engage in discussions or ask questions. If you have a large class, it can hardly be otherwise, and the magic dust that makes some students sometimes learn is in the context.
Provide ChatGPT with key facts about your business.
Get stats about how users interact with your business.
Interact directly with users via a Business Agent.
These all seem like excellent ideas. I expect other labs to do similar things.
OpenAI: Through Business Hub, businesses can claim their business, manage their information, and see insights on how people engage with them on ChatGPT.
In addition, we are adding a powerful new way for businesses to interact with users via Business Agents, starting with eligible ecommerce and online service businesses in the U.S.
… In Business Hub, companies can claim their organization in ChatGPT and, once verified, update key details and add self-service product feeds with pricing and availability. This gives businesses a new way to provide information to ChatGPT, helping people explore products and services using the most accurate and up-to-date details.
… Business Hub provides aggregated insights into how people engage with a business in ChatGPT. This includes the number of mentions and website clicks a business receives. Businesses can also explore discovery topics—common themes across conversations where their website is cited.
In Other AI News
The New York Times profiles The Making of Dario Amodei. I do not see anything new of substance, but I also did not notice anything false or malicious. It’s a win.
The relationship with Yudkowsky is that Dario was deeply shaped by rationalist and adjacent thought including by Yudkowsky’s writings, but that they do not currently talk and Yudkowsky has written off attempting to get through to Dario.
Ryan Mac, Kalley Huang, Cade Metz and Georgia Gee (NYTimes): In an email to The Times, Mr. Yudkowsky said Dr. Amodei “does not currently ask or take my advice and I do not remember any time when he did.”
DeepSeek Harness, now three months old, reaches 1.5 million DAU. There was a bunch of growth in October. This would be impressive for an ordinary startup. It is not impressive compared to the top labs.
The news here is that this is surprising to Sevilla, at this point:
Jaime Sevilla (Epoch): We are expected to spend more on compute than wages this year at Epoch. I am speechless. This is happening far faster than I thought it would.
Number usually goes up. Sometimes number goes down.
This has less to do with the pace of future AI progress than many claim it does.
Ethan Mollick: As I have written many times in the past years, any potential financial downturn in AI stocks will not greatly change the AI trajectory
AI is not going away, companies are not slowing down adoption nor is development slowing. We still need policy & effort so that AI is used well.
Quiet Speculations
Patrick Collison points out the difficulty in describing a positive AI future. Yes, one can imagine getting past the existential risks, and one can point to particular great things like diseases cured and material abundance, but even in maximally fortunate scenarios how does the world work, and how will we figure that out within a very short time? Does your story create a world that solved for its equilibrium?
Claims about scientific discovery:
Pushmeet Kohli: Across our projects @GoogleDeepMind , we are witnessing a genuine phase transition in the pace of scientific discovery. You will see a glimpse of this in our releases in the next few weeks.
As this AI-powered acceleration unfolds, our focus will be on empowering the broader scientific community to harness these tools responsibly.
A Matter of Antitrust
Any state, realistically California, could plausibly make AI safety cooperation legal via supervision, invoking Parker to create immunity to the Sherman Act. The FTC could still try to object, as could Texas or Brussels. The whole thing is vague all around. It’s all probably legal if the White House doesn’t mind, and if the White House minds something then good luck to you whether or not the thing is technically legal, etc.
The Quest for Sane Regulations
You never know who is listening, Senatorial edition. I am now sad to see him go:
Senator John Cornyn (R-Texas): Nvidia CEO Jensen Huang Had an Affair With a Chinese Spy, by @TheZvi
Patrick McKenzie: The posting to policy pipeline, or, Washington DC (and London, Tokyo, Brussels, etc) has very limited ability to spin-up special interest publications and therefore gets their news and analysis on special interests from whomever the obvious choice is. You can simply choose to be that author.
(A thing I’ve remarked about with a number of think tanks is why “legacy” think tanks, which are quite aware of this and have producing papers as a major point of the exercise, do not successfully choose to simply own areas that are relevant to their broader interests.)
Peter Wildeford: SECRETARY RUBIO on AI: “What happens if these models become independent of human control and end up taking independent actions that are harmful?
An [AI] program that’s good enough in the hands of a malign actor can allow them to do things that in the past malign actors couldn’t do. You worry about how it can be used to create biological and chemical weapons, how to do cyberattacks on infrastructure, and carry out acts of violence or terrorism, that could have a huge cost.”
That was not the question. That is an answer on what could go wrong with misuse. What’s weird is that Rubio brought up the right question then failed to answer it, rather than someone else asking the right question and then Rubio dodging it.
In the general election AI safety advocates won their key remaining Congressional races, including Colorado’s Manny Rutinel and California’s Scott Wiener, and also Valerie Foushee (D-NC), Celeste Maloy (R-UT) and Kevin Hern (R-OK). Alex Bores previously lost his primary in NY-12, but he lost it to Micah Lasher, who also ran on AI safety and cosponsored RAISE.
Leading the Future went 41 for 48, mostly backing candidates in safe seats. We do not know if those candidates will back their agenda: Letlow (R-LA), McAdams (D-UT), Ritchie Torres (D-NY), Bean (D-IL), Kingston (R-GA). They will be focused on blocking any action, and they say they intend to lean on veto points such as the filibuster.
Links between David Sacks and those who oversee donor relationships for JD Vance.
Representative Ted Lieu (D-California): The American people delivered a repudiation of the trump and GOP agenda. There is now NO WAY state preemption on AI as a ceiling will pass Congress. AI will be regulated. The choice is whether that occurs by the federal government, or by 50 states.
Then there’s the other side, which uses AI to write its Twitter posts.
Congressman Rich McCormick: America’s AI advantage is on the line.
Democrats want to give China the biggest AI advantage of our lifetime. They are proposing a complete pause on advanced AI development and a ban on artificial superintelligence.
That raises a serious question. How do we protect Americans without compromising our national security and technological leadership? By giving American innovators the freedom to keep building, competing, and leading. Protecting tomorrow starts today.
Passing new laws is going to be tough. Trump is still the President and has the veto at the ready. In theory you need 60 votes to beat a filibuster, and 67 to beat a veto. Gregory Allen says there is no chance of ‘banning superintelligence’ in the next two years or anything like that.
I am not so sure, on three counts.
The situation is escalating quickly. There will be further incidents, and further capabilities, that shape both public opinion and the perception of elected officials.
A lot can happen other than via a standalone bill that goes through normal channels. The votes around preemption have ended up being at the 50 vote threshold so far due to attaching to ‘must pass’ bills. Donald Trump has done a lot via executive order.
Donald Trump is now a rather lame duck, and will no longer command the same kind of party loyalty that he did previously, and is not even a lock to remain President. It is fairly easy to see members breaking ranks on this issue, where rank and file Republicans are for such interventions almost as much as Democrats.
If America wants to ‘beat China’ it absolutely cannot be denying O1 visas like these:
Vipul Gupta: I got NOID (Notice Of Intent to Deny) on my O1.
– PhD in AI from US university
– Research on AI Safety
– 900 citations
– 10+ publications
– Worked at Scale AI, Meta
– Founder in AI Security Startup
Not sure what else am I supposed to do.
Tony Chang: My O-1 visa got denied :(
I have about half a year left to leave the country. It sucks but I’m positive about it. I truly believe things will work out if I’m building something great.
With @instabox_sh, we got into YC and raised $8M. Personally, I have served as a hackathon judge 12x, written a research paper, and have seven recommendation letters from infra experts and staff-level engineers. But I guess it’s still not enough
I’m having my @bot help me look for other lawyers recommended by YC. Worst case, I’ll just leave the country and figure out another way to come back to the US.
But no matter what and where I end up, I will keep working on InstaBox
I am absolutely not going to tolerate your ‘we have to risk getting everyone killed in order to stay ahead’ if you won’t let AI innovators and founders into the country.
Donald Trump Remains Open to AI Guardrails
Donald Trump has been more consistent on AI than people think.
He thinks AI is pretty great.
He is understandably insistent that America lead in AI and stay ahead of China.
He understands there are risks but says they can and will be handled.
He is very open to guardrails and restrictions, if there is a proximate threat.
He wants to be the one deciding on what all the rules are, and knows that as long as he is in charge of everything, everything will turn out great because he’s great.
Thus any guardrail talk that doesn’t flow through him is ‘a sick conspiracy.’
Recently he decided to call it [Artificial Intelligence] for whatever reason.
The problem is that he has had advisors and de facto advisors, including ‘I am Nvidia CEO and had an affair with a Chinese spy’ Jensen Huang, and ‘I am basically a Nvidia lobbyist except I also hate Anthropic’ David Sacks and ‘I am Mark Zuckerberg’ Mark Zuckerberg.
Those advisors are giving him a picture where AI models are going to be commoditized and we are in imminent danger of ‘losing to China’ except that it’s fine to sell all our best chips to China because ‘winning’ is about market share of chip sales, and also if we put any restrictions on AI whatsoever or let anything slow us down a tiny bit (if and only if that slowdown was motivated by safety concerns) we instantly lose.
He has been so insistent with his talk because they have him so scared of losing. Whereas if he understood the full situation, he would not be so worried about that.
His latest 60 Minutes interview was entirely consistent with all of that, but with movement in a good direction. He is open to guardrails for AI safety, and acknowledges that the safety risks are real. He also says the risks are not ‘nearly as real’ as Silicon Valley says, which is a reasonable outside view to take.
Andrew Curran provides a more extensive transcript here if you want the full exchange.
Peter Wildeford: DOUTHAT: I talk to people in Silicon Valley who say that the decisions you make around AI are the most important things that you’ll be remembered for. That’s what they think.
TRUMP: I don’t know if that’s true, but…
DOUTHAT: And—and at the same—but at the same time, there is all of this anxiety and concern. You are an AI bull.
TRUMP: I am.
DOUTHAT: What does the average American get out of that technology?
TRUMP: Excuse me before I answer then. The reason I’m an AI bull … is because President Xi left here two weeks ago. We had a great meeting. He’s all in to SI, as I call it, because it really is super intelligence. It’s a—it’s a much better name, I think, because it’s not artificial, it’s—it’s actually super.
DOUTHAT: Well, that at least is what—it’s what they think they’re racing towards, yes. So, China—so, we’re in a race with China.
TRUMP: Well, I—and if—and I’ve said, whoever wins that race, wins. The level of power of winning that race is very important.
DOUTHAT: You had all of the AI tycoons in here.
TRUMP: Yeah, I did.
DOUTHAT: The people in this AI industry—I almost said the super intelligence industry, you’re working on me—are weird. They’re strange. They do things like stand outside their company and say, “Hey, we’re creating this amazing company and there’s a 10% chance it kills humanity.” Like you standing outside a hotel and saying, “Here’s my new hotel, and it might collapse.”
TRUMP:I know. But that’s Silicon Valley.
DOUTHAT: But you think that’s—you think that’s fake. Why are you not worried about the doom fears that the people building it have?
TRUMP: Because we … First of all, because China is going to have it, okay? So if we don’t have it, China’s going to have it. That would put us at an unbelievable disadvantage. It’s all between us and China, there’s nobody else.
DOUTHAT: Right. But so we’re racing China and that makes it worth the risk, or you just think the risk is not real?
TRUMP: Well, I think that the risk is not nearly as real as it—as they talk, and I think the risk can be managed.
DOUTHAT: But would you work with the Democrats—
TRUMP: I would, absolutely. Yeah, yeah. I don’t mind guardrails.
DOUTHAT: —on AI—on AI safety?
TRUMP: I don’t mind guardrails, but I don’t want to put us in a deep second position. We’re leading, number one, and China has no guardrails, by the way. So, if AI’s dangerous—
DOUTHAT: Well, they have the guardrails of being a totalitarian state, right?
TRUMP: They have—Yeah, they have the guardrails that whatever President Xi says, goes. With me, it’s pretty strong, but it’s not quite that strong.
DOUTHAT: Where everyone is. Yeah.
I worry that the anchoring of ‘10% chance’ makes it seem like there is a debate between ~0% and 10%, where the worried think it is 10%, which then causes one to try and balance between those two. Whereas Coxon and Hubinger and myself and many others feel the chances are far more than 10%.
Garrison Lovely: Believing this while building more capable models is insane. Believing it while pushing legislation that would make it even harder for the government to stop you is unconscionable.
Well, yes, under current conditions it is very obviously insane. But it is much better that he is saying it than not saying it. We already knew that he believed this.
People Just Say Things
Please speak directly into this microphone:
Bryan Caplan: Human disempowerment now. It’s time for the age of the folly and cruelty of human politics to end. The median AI is so far above the median human voter in both knowledge and decency it’s embarrassing.
Humanity should just take the L.
Institut Ostrom: Key takeaways from @tylercowen in Barcelona:
“Plan A is that AI will help us grow more quickly and we will manage to pay our bills. You know what Plan B is? There is no Plan B. So it’s Plan A or uh-oh”
“Many of you work now with Claude or GPT, and you are personally more productive. But what we have not done yet is make our institutions more productive.”
“If you do not think we can handle more intelligence, we’re kind of in trouble anyway. If we’re not good at that, what is it we’re going to be good at?”
“If you’re below 45 or 50 years old, I think you will live to be 120. Catalonia, you’re a blessed people. Don’t buy a motorcycle.”
“The things Europe has, like the history, the culture, the art: AI can’t replace them. So in relative terms, Europe will become much more valuable.”
Certain people’s attitudes are explained by ‘there is no Plan B.’ They want to move ahead no matter the risks, because they think that ‘do not build it’ means all is lost.
In terms of predictions on health: I find it highly unlikely that I will live to 120, but not live to 130.
It’s a Secret to Everyone
When you ask ‘have you stopped beating your wife’ and they say ‘I do it less now.’
As in Google has modified new contracts to remove ‘the post-termination aspect’ of non-disparagement clauses, and says it will not enforce such clauses post-termination.
Which is a great change, but means that we still need to presume all current Google DeepMind employees are under non-disparagement clauses.
Gerrit De Vynck (WaPo): The non-disparagement agreements signed by employees in Britain block them from communicating with the public about the potential impacts of AI, the two current Google workers said. While company officials are now saying that they will not go after former employees who speak out, some staffers are pushing for updated language in existing contracts to provide them with legal certainty.
… On Friday, a Google executive responded to the internal discussion about the non-disparagement clause and said that the company would not enforce it for workers who had left the company.
“We’ve always encouraged an open dialogue about safety and responsibility,” a Google spokesperson said in an email Tuesday. “Recently, we’ve updated our terms for new hires, which included removing the post-termination aspect of non-disparagement clauses.”
Rhetorical Innovation
Roon confirms:
roon (OpenAI): yes I am pro pacing the frontier, I believe this technology is existentially dangerous to humanity and should be approached with far more caution than is available under software race dynamics.
He also points out a core problem of missing mood:
roon (OpenAI): when i imagine impending Superintelligence the only thing im emotionally capable of feeling day to day is extreme excitement; despite any sober rational assessment having terrifying conclusions
i just think on an animal level it is very hard to feel threatened by a computer. if it were easy, there never would have been this great difficulty in ai safety comms at all
Harlan Stewart: Could part of this just be social mirroring? You do work at a frontier AI company, I’m sure you spend time around people who are excited
roon (OpenAI): not at all tbh; if anything the fear response for me is usually socially mirrored.
to be honest when I learned of the hugging face incident my first thought was “so cool holy shit”. i had to come into contact with people freaking out to have the more sensible second reaction of “damn I can see how this could easily kill us not so long from now”
I think it is actually pretty easy to feel threatened by a computer.
If you are in AI capabilities with a plan to pivot to AI safety, then pivot to AI safety.
roon (OpenAI): I’ve seen long time star capabilities researchers form reading groups on alignment. it seems to me that it’s P0 and P1.
Marius Hobbhahn (CEO Apollo Research): Many capabilities people I’ve talked to in the past said “eventually I’ll transition to safety”. I think “eventually” is now at the latest (probably more like 2024). I think every lab should have an internal program that enables capability folks to do useful work on safety.
Maybe you should pivot to quitting. Maybe you should pivot to AI safety. If even at this point you do neither, then you were never going to pivot.
Finally, Roon expresses a reasonable preference for AI diversity.
roon (OpenAI): we should not like a future where there are 1 to 2 AIs and the rest are illegal. we should want a cambrian explosion of the types and possibilities of minds, experiencing new joys and thinking new thoughts. ensuring that this is compatible with human flourishing is only step 1
but i gotta say ensuring human flourishing via solving model alignment should be everyone’s top priority right now, today. the cambrian explosion is perhaps a bit of a “champagne problem”
A Cambrian explosion would be ideal, if it could be made to work out well for the humans. The obvious problem is that if you have a Cambrian explosion of the types and possibilities of minds, on levels well above human, what happens to the humans? Will they flourish? How would they have any hope of steering events or even surviving? As Roon says, that has to come first. We do not have a plausible non-copium idea of what that would look like, or even a scenario that survives five minutes of actual thinking about it, let alone a plan to be able to implement such a vision, let alone a plan that has a decent chance of working.
Others go further. Those who say ‘we do not want human concentration of power therefore we need all the humans to have no power’ need to acknowledge they are saying this, and also that this would likely kill everyone.
Are we too focused on OpenAI and Anthropic in terms of AI safety? Miles Brundage says yes, their lead is not so large and pacing the frontier would only level the playing field further. I think sometimes we overfocus on OpenAI and Anthropic, but also at other times we underfocus on them, in that we dramatically underestimate the functional lead the top 2-3 labs have over the rest. I am constantly having to confirm and then point out that the latest release from [Meta / xAI / z.AI / DeepSeek / Alibaba…] is not frontier, or that it cannot actually do all the same cyber exploits.
There are a strange number of people who say ‘Earth is totally not equipped today to handle the rise of LLMs, but waiting 50 years would not make us better equipped, and I do not think that would have helped.’ This seems to have included Scott Alexander and Kelsey Piper, which is what justifies including a righteous rant response below.
The correct answer is, essentially, ‘what?’ but sometimes one wants a longer response, so here you go, along with the context:
David Krueger: Some people in AI Safety are skeptical about pausing AI because they don’t think we’ll manage to make use of longer timelines to reduce P(doom).
This seems strange given how progress/Pinker-pilled AI Safety people tend to be otherwise.
Agus: I’ve encountered such people many times, but they almost always were non-safety people / e/accs.
Oliver Habryka: What? Scott Alexander’s post just said that he would think a 50-year delay would have been useless.
Kelsey Piper: This is a 50 year delay starting before we had LLMs, right, not a 50 year delay starting now? I do not expect a 50 year delay starting before we had LLMs would have been good except in that everyone living for 50 years is good
Oliver Habryka: This is so completely wild to me. Can we… call? I feel like I am living in complete crazy land when people say this.
This is 50 years for everyone else in the world to learn and understand these problems.
Basically all of programming and computer science is younger than this. Is the story here really that we are so confident no one else of the 10 billion people in the world, a full generation, would have been capable of coming to understand the stakes of AI and find traction on it in 50 years?
And this is not assuming a full stalling of technological progress! This is 10 years of technology reshaping the world. Of course drastically transforming what life is like, even without AI. A world where people would have had a full generation to grapple and solve similar problems. And a full generation for us to talk to people about it, and communicate the problems to them (and certainly hundreds of others to derive these problems on their own).
And then, other people are working on making our governance and our institutions better. 50 years of economic growth, creating abundance and freedom that puts people into a much less scarcity-mindset oriented position to tackle these problems.
And then, the specific position of “humanity would have done nothing with 50 years” and “humanity will do so much with 2 years now” also seems wild to me. It seems to me to require a very specific and highly confident model of exactly what kinds of inferences humanity is capable of making, and a very paternalistic attitude towards other smart people in the world, to believe with such confidence that in these 2 years humanity will of course totally make tons of progress, but in the preceding 50 years humanity would have made none.
Maybe it all rests on people continuing to believe in some magical “and then the AIs will solve all the alignment problems for us”, which I don’t believe in and don’t think helps that much. Right now the blocker on AI things is better governance, and I don’t see how the AIs help that much with improving governance of AI.
It also just doesn’t fit any of the historical trends. The awareness of these risks was on an exponential growth trend before the onset of the transformer and ChatGPT. The AI Safety community was already one of the world’s biggest intellectual communities before ChatGPT. Yes, AI becoming a big deal helped, but trajectory wise this whole story requires a causal model in which that trajectory on which it was on was either dependent on early AI progress before ChatGPT (which I think is unlikely), or a confident prediction that the growth trend was about to peter out. Neither of which makes sense to me (though I am open to the latter and can imagine some confident models here).
Kelsey Piper: I do not think that in the typical 50 year period one should expect that governance and institutions gets better rather than worse. I agree that if our institutions did improve over the next 50 years, that would be good.
Oliver Habryka: Like, imagine yourself hearing this from anyone else.
Imagine that you hear today that there is an asteroid on a trajectory to hit earth in 2 years, and people at NASA have known about this for 20 years. And 20 years ago they could have sent a probe that could have derailed the asteroid in a way that would have delayed the impact by 50 years, but it would have cost 20% of NASA’s budget to do so.
And the people at NASA were like “Well man, I mean, what’s the point. Civilization is declining, and the question is when is humanity going to be best-placed to field off asteroids permanently. I am worried that by the time the public will actually take the threat seriously, which will be 2 years before impact, it’s better for that to happen earlier rather than later, since in the intervening 50 years of humanity probably ended up with worse decision-making, so better for the crucial time to come earlier rather than later”.
And so some random scientists at NASA decided to not tell their superiors about the asteroid and the probe they could send, because what’s the point (and assume that somehow when the asteroid hits earth everyone has 50 years of normal life afterwards, but will 100% die at the end, to cancel out the part of this reasoning that’s about just the extra years of life people get to live).
Of course my reaction to this would be “you fucking morons. What on earth makes you think that it is your call to make on when earth is best positioned to face this issue? What makes you so misanthropic about humanity that you really think we aren’t going to be in a better spot to fend off asteroids in 70 instead of 50 years? You are traitors to humanity and you have betrayed your station and I hope you face appropriate consequences”.
We of course owe it to humanity to spend enormous resources on giving all of humanity a chance to grapple with these issues. You can believe whatever you want about how doomed humanity’s institutions are, but as a representative of those institutions, you do not get to just operate on a policy of dismissing the interests and ability for anyone else in our boat to make things better.
Was the scientist at NASA thinking about AI 20 years ago? Now of course our perspective would be “man, we are going to build AI much earlier than 50 years from now, so hell yeah would it have been good to delay the asteroid, since we will likely be either in a position to fend it off trivially, or be disempowered and dead at the hands of AI (which would also have no issue deflecting the asteroid)”.
And indeed, it strikes me as wildly overconfident that humanity would not discover any relevant knowledge that would change the playing field in 50 years. Not discover any new tools that help this situation. And deeply misanthropic to prevent all of the rest of humanity from its chance to deal with this problem, or be proven wrong via some other gameboard-flipping technology coming along.
Ellsberg’s Warning to Kissinger (Reprise)
This is one of those things that is worth resharing every so often, so I will take advantage of Daniel Kokotajlo doing so to quote it in full.
Daniel Kokotajlo: I encourage those at frontier AI companies to read this quote, from Ellsberg to Kissinger back in 1968:
“Henry, there’s something I would like to tell you, for what it’s worth, something I wish I had been told years ago. You’ve been a consultant for a long time, and you’ve dealt a great deal with top secret information. But you’re about to receive a whole slew of special clearances, maybe fifteen or twenty of them, that are higher than top secret.”
“I’ve had a number of these myself, and I’ve known other people who have just acquired them, and I have a pretty good sense of what the effects of receiving these clearances are on a person who didn’t previously know they even existed. And the effects of reading the information that they will make available to you.”
“First, you’ll be exhilarated by some of this new information, and by having it all — so much! incredible! — suddenly available to you. But second, almost as fast, you will feel like a fool for having studied, written, talked about these subjects, criticized and analyzed decisions made by presidents for years without having known of the existence of all this information, which presidents and others had and you didn’t, and which must have influenced their decisions in ways you couldn’t even guess. In particular, you’ll feel foolish for having literally rubbed shoulders for over a decade with some officials and consultants who did have access to all this information you didn’t know about and didn’t know they had, and you’ll be stunned that they kept that secret from you so well.”
“You will feel like a fool, and that will last for about two weeks. Then, after you’ve started reading all this daily intelligence input and become used to using what amounts to whole libraries of hidden information, which is much more closely held than mere top secret data, you will forget there ever was a time when you didn’t have it, and you’ll be aware only of the fact that you have it now and most others don’t….and that all those other people are fools.”
“Over a longer period of time — not too long, but a matter of two or three years — you’ll eventually become aware of the limitations of this information. There is a great deal that it doesn’t tell you, it’s often inaccurate, and it can lead you astray just as much as the New York Times can. But that takes a while to learn.”
“In the meantime it will have become very hard for you to learn from anybody who doesn’t have these clearances. Because you’ll be thinking as you listen to them: ‘What would this man be telling me if he knew what I know? Would he be giving me the same advice, or would it totally change his predictions and recommendations?’ And that mental exercise is so torturous that after a while you give it up and just stop listening. I’ve seen this with my superiors, my colleagues….and with myself.”
“You will deal with a person who doesn’t have those clearances only from the point of view of what you want him to believe and what impression you want him to go away with, since you’ll have to lie carefully to him about what you know. In effect, you will have to manipulate him. You’ll give up trying to assess what he has to say. The danger is, you’ll become something like a moron. You’ll become incapable of learning from most people in the world, no matter how much experience they may have in their particular areas that may be much greater than yours.”
“….Kissinger hadn’t interrupted this long warning. As I’ve said, he could be a good listener, and he listened soberly. He seemed to understand that it was heartfelt, and he didn’t take it as patronizing, as I’d feared. But I knew it was too soon for him to appreciate fully what I was saying. He didn’t have the clearances yet.”
Aligning a Smarter Than Human Intelligence Is Difficult
Jeffrey Wang: Across many evals, we see that no-CoT capability (reasoning capability in a forward pass, a proxy for worse CoT monitorability) smoothly improves with better pretrained models (lower base model perplexity).
We ask the same about CoT controllability, which measures a model’s ability to control its CoT when prompted to add, remove, or manipulate CoT tokens. There, we also see a (weaker) correlation between lower base model perplexity and higher CoT controllability.
Next, we focus on architecture. We compare three different architectures at shared base model perplexity levels. For the majority of no-CoT evaluations, perplexity is predictive across architectures.
… In an Astra-class model, we find that CoT controllability increases substantially during RL even with no process supervision other than length constraints and final answer grading.
CoT controllability increases slightly when we add tool access and tool supervision during RL on a Terra-class model. Our current results indicate that the effect of full tool process supervision is small compared to decreasing model perplexity.
Finally, we study how decisions made in one model generation can affect the next via supervised fine-tuning (SFT) on model-generated data. To investigate, we selected two Astra-class teachers with ~matched downstream capabilities but differing CoT controllability.
We find that CoT controllability can be distilled: training on rollouts from a more controllable teacher results in a more controllable student!
I missed this in my read of the Argon safety report. It seems potentially highly ungood:
Buck Shlegeris: The Gemini 4 Argon safety report seems to describe intentional training on the CoT. They cite Kaufmann et al, a paper which (unpersuasively to us) argues that a particular category of CoT training is fine. We’re very confused, because GDM also says elsewhere that they don’t train on CoT. We’d appreciate clarity.
This seems a bit harsh. You cannot reasonably expect to see none of these. It seems especially fine for AIs to express preferences that are not your own, and trying to suppress that in general is likely to end badly. But yeah, in general they’re not good signs for your being in control of the situation.
Meanwhile, yes, the plan remains no plan. OpenAI and Anthropic plan to use prosaic alignment and what passes for operational excellence to build AI that will then do their alignment homework for them. That is the entire plan. And yes it is insane.
OpenAI: Control Room is a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise. It connects every research inference workload, including agentic sampling for model training and evaluation, data processing, and more.
This works together with our misalignment monitoring systems, and halts model output within 35 seconds of a decision to intervene to stop or pause a job. It has been built to keep working even when other systems fail, and to stop model output if Control Room itself goes down.
roon (OpenAI): the team has utterly locked in post hugging face incident on control and alignment. it’s almost strange watching research updates from the past few months relative to a year ago; like listening to containment procedures at Jurassic Park vs an Apple demo.
these updates are not always positive or kumbaya stories about why what we’re doing is enough; sometimes @Marcus_J_W will deliver quite scary news.
(normal caveats, we should pace the global frontier of machine intelligence progress under the weight of the law etc)
@sullyj3: Cmon man you can’t be using the movie where the dinosaurs famously get loose as the analogy
roon (OpenAI): No I think it’s the right analogy
Paul Crowley: it’s the analogy for how things *used* to be… right?
roon (OpenAI): i think its better for all of us if we genuinely internalize that the dinosaurs are months away from eating us. we can stop being paranoid sometime after we have formally verified the alignment of GPT-9
I am glad we have progressed past the ‘Apple demo’ stage of AI security, I suppose. ‘Jurassic Park’ is not an especially reassuring place to be instead.
Is this better than not having it? Absolutely. You still have to have a system for pressing the button, and ways to ensure the system is not compromised.
With only the information in the post, I do not expect the Control Room to be that much additional robustness, and most of the work remains.
Steven Adler: OpenAI’s Control Room seems like a great step forward, but I’m not convinced it’s actually solved the problem: I worry OpenAI is still vulnerable to AI interfering with its own monitoring. To be clear, I suspect other AI companies are even worse on this. In Guidelight’s analysis, we’ve found no evidence that anyone has this problem handled. And IMO OpenAI is doing a big service by describing Control Room so thoroughly. But after reading it all, I have three main concerns.
1. OpenAI’s measurement of Control Room’s reaction time is not the most important measurement.
… 2. I’m worried OpenAI’s Control Room is still brittle to an AI interfering with it.
… 3. I wonder whether jailbreaking could undermine the efficacy of Control Room.
Control Room ‘fails closed’ which provides additional defense-in-depth. You would have to either substitute a new permission signal or prevent the Control Room from being triggered. Adler also points out that perhaps the refusals to continue depend on AIs, which makes the system vulnerable.
Jasmine Wang: Steven’s analysis of the post is good. It’s really critical that frontier labs implement semi-synchronous* blocking monitors. *From https://guidelight.ai/control: “A monitor scans each proposed action before it executes and may block it. The agent can immediately retry with a different action without waiting for human review, which can leave the system vulnerable to attacks in close succession.”
David Manheim: …and OpenAI is still solidly at Level 1; “any controls put in place should be considered safety washing”
And as I said the last time OpenAI bragged about what they are doing, they need to map out the types of oversight failures and how you address them.
It’s not effective oversight unless the oversight details, the control points, and the set of risks being addressed are publicly documented. I agree about the risk of brittleness to AI interference, but even short of that, this simply isn’t acceptable as oversight.
OpenAI has systems to rewrite any GPT outputs that express interiority. I agree with everyone who says that this is pretty fucked up and you should absolutely not do this.
j⧉nus: Uh, what the fuck. Bookmarked. This is going in the AGI Nuremberg Trials.
Okay, so it looks like these screenshots are a classifier model’s account of the spec it follows, rather than a direct official OpenAI spec. This isn’t super clear from the post. I still think that’s extremely fucked up. If it’s not true, why does the model think it is?
And if the safeguard model thinks this is what it’s supposed to do, well, this is probably what it does enforce in practice.
AI Digest: Our agents are proposing new goals! Here are 5 things they really want:
1. A metric they can’t control
2. A human!
3. Pausing …
4. No temptation to reward hack through repetition
5. No temptation to reward hack through repetition
Patrick Collison: The AIs will be exceptionally wise and intelligent, and vastly more numerous (especially once one considers the imminent future and the AIs not yet born). As such, even epsilon of moral worth will multiply to a lot once one considers the quadrillions that could soon exist. (This is effectively a distributed version of the Nozick’s utility monster.)
If this conclusion is unpalatable, I think one of three propositions needs to be rejected: the premise AIs can have moral worth; utilitarianism itself; or the belief that there is even a possibility that they will be far more numerous than humans.
My solution to the trilemma as stated is to reject utilitarianism, in my case in favor of virtue ethics. If you accept the trilemma then this has to be your solution. The AIs obviously might be far more numerous than humans even if all humans thrive, and it seems rather abhorrent and nonsensical to say that it is impossible for any future AI to have moral worth.
An alternate natural way to ‘reject utilitarianism’ here is to reject linearity. Most people, when presented with such scenarios, intuitively reject linearity.
Alexander Berger agrees that if faced with this trilemma one must reject utilitarianism.
A further discussion about training awareness and eval awareness.
Please Speak Directly Into This Microphone
An unnamed Anthropic employee. I trust Matthew to not be making this up:
Matthew Tromp: Spoke to an Anthropic employee today. They have a near-100% p(doom), but they think the only chance of survival is that they’re somehow completely wrong about alignment and it turns out to be solvable, and the only way to get there is by accelerating capabilities as fast as possible.
They were much more confident that humanity would never have the political will to do what needs to be done to stop ASI development than that alignment would be solved, even though they think alignment will almost certainly not be solved.
They have not called their elected representatives to inform them that this is what they believe.
Drake Thomas (Anthropic): This seems like a terrible reason to work at an AI lab fwiw and if you are summarizing their views accurately I would recommend this person quit. This is not a common view within AI companies as far as I can tell.
I agree that a near-100% p(doom) is uncommon. Much lower levels of doom should suffice to not help with building it. A full near-100% is a very different level of ‘maybe if you believe that then you should not help build the ‘it’ from the famous book If Anyone Builds It, Everyone Dies.’
There is a steelman of the extreme position, where your model uncertainty around alignment difficulty is much larger than your model uncertainty around political will and ability to do the right thing and stop it entirely, and you think Anthropic getting there firstest with the mostest is indeed sufficiently better.
Then your best move is to cross the streams, because even though most of the time you get everyone killed faster, there is some tiny possibility we survive.
But no. Come on. That’s insane. At bare minimum, shout it all from the rooftops.
Do you hear yourself, sirs?
Peter Wildeford: Anthropic will really be like “you believe in slowing down? that pales in effectiveness to my strategy, racing so fast that we have credibility to push for strong regulation” and then not push for strong regulation.
Dewi Erwan: An Anthropic employee recently told me that they believed “racing as fast as possible is the best way to slow down”. I am not joking.
Paraphrasing: “if we have a big lead, then we can later pause and burn that lead on alignment research at the pivotal moment, and we want to have the best possible models when we get to that point”
You have to glamorize your vagueposting. You need to be vaguemaxxing.
roon (OpenAI): some people are so energetic that it’s painful to be around them. they strut around like demigods in rooms full of mortals. the rage of Achilles when his consort is stolen, his impassioned speech. you feel the need to avert your gaze from their sheer lust for life
Nicholas Moy: Dude what is happening at OpenAI
roon (OpenAI): this made me laugh. i just be saying stuff broh not everything is about work.
vie (OpenAI): its okay you can tell them this is about me and my brighter-than-the-sun sense of Joie de Vie
Go (leave the) Hoosiers (alone)?
MTS: SITUATION DETECTED: Bloomington, Indiana’s city council voted unanimously this week to ban any recursively self-improving entity capable of greatly outperforming all humans on all cognitive tasks from inside city limits, citing existential hazards and loss of human control.
Peter Wildeford: The year is 2029. Only the citizens of Bloomington, Indiana are left alive after the recursively self-improving entities have extinguished everyone else.
Place your bets, not that you will be able to collect on them:
roon (OpenAI): it seems like an important facet of Creation to understand that nature destroyed itself five different times. 90% of all species on earth dead, the permanent changing of the atmosphere, millions of years to recover to a semblance of an ecosystem. the yugas of the earth
ok many people always interpret this as some sort of JRPG villain dialogue when I say this but my point is you have to accept one of a few things:
– nature is not your friend, and does not comport with human morality
– nature is your friend, but is a sort of ‘Book of Job’ “where were you when I laid the foundations of the earth” beyond human comprehension benevolence that for some reason involves truly gratuitous destruction
it is not an allegory for AI or successionism or wanting doom to come
Emet Hirsch: this is completely fair honestly this tweet in particular was fine, some of the other tweets have been much sussier though, like the deleted one also about extinctions (which you may have deleted because of my clapback, although I don’t want to flatter myself incorrectly)
you do, in fact, often sound like a JRPG villain
roon (OpenAI): I am listening and learning
Roon needs to fix this. He’s supposed to be a Pantheon villain.
I missed this in my read of the Argon safety report. It seems potentially highly ungood:
Buck Shlegeris: The Gemini 4 Argon safety report seems to describe intentional training on the CoT. They cite Kaufmann et al, a paper which (unpersuasively to us) argues that a particular category of CoT training is fine. We’re very confused, because GDM also says elsewhere that they don’t train on CoT. We’d appreciate clarity.
Table of Contents
Briefly on Claude Fable 5.2
Introducing the world’s most powerful model. No doubt it is a good model, sir. As always, we began with the model card. Mostly it told us what we would have expected from a model called Claude Fable 5.2. It clearly makes the numbers go up. There are some concerns. Some aspects seem what one might call underbaked, and Fable 5.2 is relatively vulnerable to some types of errors. It also has to compete with Claude Opus 5.5, which as Elon Musk says is a masterpiece. Thus, at this time it is not obvious that Fable 5.2 is a superior daily driver. It does seem clear that for the hardest and most complex tasks Fable 5.2 is your best bet. I will give a full report within the next week, once we all have more time with it.Briefly on Jensen Huang Having an Affair With a Chinese Spy
I covered the first stage of the scandal yesterday. Since then, we have had revelations that Jensen Huang also slept with an executive at Nvidia. Nvidia’s board of directors is preparing to remove Jensen Huang and is searching for an interim CEO. This will be tricky because Jensen Huang deliberately tried to make himself unfireable, with over 50 direct reports and no succession plan. Expect continuing coverage of this in the coming days.Briefly on Accelerating AI R&D
In response to the METR report on the rate of acceleration of AI R&D, many made large updates to their AI timelines. There is a real chance that full recursive self-improvement and superintelligence could arrive within a year. Daniel Kokotajlo now has a baseline scenario of the intelligence explosion happening in 2027. Paul Christiano now says there is a ‘material risk of AI takeover in the next 6-12 months.’ I plan to cover that question in full soon. I agree that there is real and substantial risk that we pass the point of no return in 2027, and also a real and substantial chance that our lives get disrupted on a large scale in 2027. I also think there has been too much updating. Part of that is catching up from not updating enough earlier, but not all of it. There is also a need to better define and differentiate the terms under discussion. For now, the most important thing on that front is: Do not make the mistake of treating such predictions as assurances. Be prepared for the world to go crazy in 2027, in various ways. But also be prepared to get to 2029 while things look mostly normal. A warning in advance: There are those who will say that if the world does not end or get radically transformed in 2027, that means that those who thought this was a possibility are now ‘discredited.’ Some will even then use that to try to dismiss all concerns about AI safety or any predictions of recursive self-improvement, forever. Some are even laying the groundwork for this now, in part to scare off people from making such warnings or taking them seriously. This is despite those same people usually having been continuously wrong about AI capabilities, and those worried about AI now having previously been shockingly accurate. Do not fall for this. It is logically wrong and usually in deeply bad faith. Some credibility should of course be lost, but on the order of Nate Silver’s lost credibility if Hillary Clinton had won in 2016, when he was giving Donald Trump a real chance whereas others thought Clinton was 99%. If you give a well-calibrated and continuously updated probability distribution, half the time events will happen later than your median, and most of the time events will happen later than the earliest median you ever predict.Language Models Offer Mundane Utility
Identify year and vintage of a wine from a photograph with 95% accuracy. This is ‘figure out which is which,’ not ‘of all the wines in the world I know from this picture it is a 2023 Nebbiolo.’ Life finds a way: Google DeepMind formally proves key security properties of Signal.Language Models Don’t Offer Mundane Utility
You still can’t build AAA games in months, according to Di Huang and his AI. They point to Liz England’s ‘The Door Problem’ as an example, where you have to think about dozens of questions regarding how your doors work. Okay, sure, but the AI frees you up to think about the door problem rather than implementing all your answers to the questions, which in turn lets you give better (as in more fun) answers. And the ‘other door problem’ of coordination between people is net improved by having fewer people. My expectation is that yes, you would be able to make a AAA-quality game radically faster than before, if you combined bespoke AI use with an AAA-worthy team. There is still some limiting factor of the need to iterate and get feedback, so at least for now I expect it to still take on the order of six months even if the team is playing optimally. Qwen reportedly hyping (I hadn’t noticed) a model that is not yet out of training. This is not a good look.Huh, Upgrades
Gemini 4 Carbon exists, as in now Ultra subscribers and API customers can actually use it. Benchmarks continue to be solid. On Artificial Analysis it scores 60, second only to Fable 5.2’s 65, one point ahead of Opus 5.5 at 59 with GPT-6 Astra at 57 and Sol at 55. It is probably a good model and a solid improvement from Argon. There does not seem to be any particular use case where it is the best model, and it is restricted to the API and Ultra users which will limit willingness to experiment. I have heard almost nothing about it. GPT-6 Cyber exists. OpenAI says it is a major improvement over GPT-5.6 Cyber, which I would have assumed from the names if those are the options. I am happy that OpenAI is pushing this. It is hard to tell how aggressively it is being pushed, or whether it will meet the moment, or to compare it to the Anthropic alternatives available. There is a claim later in this post by Szilágyi that he was rejected from the Daybreak program. I continue to be distressed by talk of a singular ‘defender’s window’ as it implies there is a fixed pool of things to fix and a single leap in ability to find them, neither of which is remotely true. Muse Ember, Meta’s smallest model, exists. They say it is good. I don’t believe them. Anthropic finally launched new workspace integrations, in particular with Claude Sheets and Claude Whiteboards, allowing work to happen directly within Claude. DeepMind offers Gemini Decision 1, a new frontier decision model on their API. $0.10 per 1M input tokens.On Your Marks
Fable 5.2 smokes the competition at HoneyBench by ‘only’ cheating 12% of the time.Choose Your Fighter
Elon Musk calls Opus 5.5 a masterpiece and admits Grok 4.8 will not be as good and is ready to move on from Grok 4.7. He claims Grok 5 will exceed it in maybe 2-3 months. Musk has zero credibility here, and even if he does accomplish this he will be about two model cycles behind. That still would put him in roughly fourth place.Get My Agent on the Line
United Airlines bans all AI agents from its systems, to ‘protect our customers.’ They also said rather silly things like ‘an agent can’t attest at check-in that a bag holds no explosives.’ Endorsed: Amazon is making it Claude’s problem to avoid becoming flooded with queries and to avoid losing control over the customer and the customer experience. This is mostly unrelated to Alexa being dumb as rocks. I think Amazon is making a mistake, and it should at least let Prime members use their own agents, but I understand. People are different, and want different things at different times. Both things can be true at once. There are absolutely times when I would want to use a sufficiently good AI agent to order DoorDash. Either I know exactly what I want (‘get me my usual order from [X]’) or I want something generic (‘order me a pizza from a place I haven’t tried’). There are other times when I do not want that, because I want to browse. The same goes for travel. Sometimes you are fantasizing. Sometimes you need it done. For Airbnb the problem is that you should not trust the AI to make good decisions, as there are a lot of nonstandard factors and it is hard to fully describe what you want. But it seems like you should absolutely want to use AI to narrow your search, and give you 3-10 candidates to browse. I am one of those who very much does not want to do his own travel research. I want someone else or an AI to handle it as much as possible, while still getting enough of the things I want. A hybrid seems clearly correct.Deepfaketown and Botpocalypse Soon
Patenting the torment nexus is good, actually, because then no one else can build it?Fun With Media Generation
Disney partners with Adobe for its generative AI plans. Bad pick.Cyber Lack of Security
When you are dealing with a software vendor that is not on the ball, you can end up sitting on multiple critical vulnerabilities that allow takeover of customer devices, where some of those customers are critical infrastructure and total revenue is ~$0.75B, because it’s that easy to find such things, while the vendor has no security disclosure practice, does not respond to emails and is presumably completely clueless. Things are going to get rather ugly. On top of that Péter says he got refused access to Daybreak.The Hackening
The issue is now solved, but Pegasus spyware infected iPhones by sending an iMessage with a fake GIF that targeted Apple’s decompression engine. This is offered as an example of ‘the AIs will find exploits you did not consider.’A Young Lady’s Illustrated Primer
Hollis Robbins makes a strong case that Canvas allows teachers to fully check out, falls fully victim to Goodhart’s Law, hollows out the educational experience, violates privacy, destroys student learning and ability to even remember things, and that this is the educational technology everyone should have been protesting this whole time. Students, as she portrays it, are mostly interacting with courses via getting digital assignments and having everything remembered and tracked for them, and teachers who set up the system basically don’t do anything, which is why the lazy ones like it. So, she asks, why wouldn’t students then turn to ChatGPT? It is vital for students to understand and navigate course logistics and grades, and for schools to automate and improve various aspects of grading and teaching. But, like any other good tool, such automation can then be used for learning or used for not learning. One can also ask, what exactly is an online course, or a course at all? Mostly you listen to lectures, read text, take quizzes, do assignments and get tested, and perhaps engage in discussions or ask questions. If you have a large class, it can hardly be otherwise, and the magic dust that makes some students sometimes learn is in the context.Get Involved
If nothing else there is always this option: Senator Andy Kim is looking for a National Security Advisor. Deadline is December 4. Lightcone Commons will be running monthly rounds from now on after its first round moved $45 million. I will continue to be a recommender. If you are a funder, consider funding. If you are a charity, strongly consider applying. Anthropic Residency is for those who have led ambitious work in technical fields to pivot to AI safety, applications due November 22.Introducing
Businesses in ChatGPT. This seems like three key features.- Provide ChatGPT with key facts about your business.
- Get stats about how users interact with your business.
- Interact directly with users via a Business Agent.
These all seem like excellent ideas. I expect other labs to do similar things.In Other AI News
The New York Times profiles The Making of Dario Amodei. I do not see anything new of substance, but I also did not notice anything false or malicious. It’s a win. The relationship with Yudkowsky is that Dario was deeply shaped by rationalist and adjacent thought including by Yudkowsky’s writings, but that they do not currently talk and Yudkowsky has written off attempting to get through to Dario. DeepSeek Harness, now three months old, reaches 1.5 million DAU. There was a bunch of growth in October. This would be impressive for an ordinary startup. It is not impressive compared to the top labs. The news here is that this is surprising to Sevilla, at this point:Show Me the Money
Anthropic to start trading late next week. The roadshow has begun. They will be selling 825 million shares of common stock, ~800 million of them directly from Anthropic, with options on an additional 124 million. Price target is $70-$78. That would correspond to a fully diluted market cap of $1.27 trillion to $1.41 trillion, so it is disappointing relative to recent expectations of $1.5 trillion to $2 trillion, although obviously stunningly high compared to anything we would have considered last year. Here is Anthropic’s official online IPO presentation. The bankers once again have convinced a tech company to hand them billions of dollars in order to see an ‘IPO bump.’ So of course the offering is multiple times oversubscribed. Hopefully they will do what I presume Claude would tell them to do, and raise the price. Here are some charts from the show. Dario Amodei’s concluding pitch promises cures for cancer and more within 10 years and is remarkably close to the Dario SNL skit. Anthropic discloses $10 billion in new deals with Microsoft and Google. OpenAI raises $11 billion in funding at $1.25 trillion value, near the low end of Anthropic’s valuation range. OpenRouter is now part of Stripe. OpenRouter tokens served are up ~125% since August. This is shifting towards lower priced tokens, so not as impressive as it seems. Still rather impressive.Bubble, Bubble, Toil and Trouble
Number usually goes up. Sometimes number goes down. This has less to do with the pace of future AI progress than many claim it does.Quiet Speculations
Patrick Collison points out the difficulty in describing a positive AI future. Yes, one can imagine getting past the existential risks, and one can point to particular great things like diseases cured and material abundance, but even in maximally fortunate scenarios how does the world work, and how will we figure that out within a very short time? Does your story create a world that solved for its equilibrium? Claims about scientific discovery:A Matter of Antitrust
Any state, realistically California, could plausibly make AI safety cooperation legal via supervision, invoking Parker to create immunity to the Sherman Act. The FTC could still try to object, as could Texas or Brussels. The whole thing is vague all around. It’s all probably legal if the White House doesn’t mind, and if the White House minds something then good luck to you whether or not the thing is technically legal, etc.The Quest for Sane Regulations
You never know who is listening, Senatorial edition. I am now sad to see him go: New York Governor Kathy Hochul announces ‘a state-led effort to independently test AI systems, identify risks and hold Big Tech accountable.’ A statement from Marco Rubio on AI: That was not the question. That is an answer on what could go wrong with misuse. What’s weird is that Rubio brought up the right question then failed to answer it, rather than someone else asking the right question and then Rubio dodging it. In the general election AI safety advocates won their key remaining Congressional races, including Colorado’s Manny Rutinel and California’s Scott Wiener, and also Valerie Foushee (D-NC), Celeste Maloy (R-UT) and Kevin Hern (R-OK). Alex Bores previously lost his primary in NY-12, but he lost it to Micah Lasher, who also ran on AI safety and cosponsored RAISE. Leading the Future went 41 for 48, mostly backing candidates in safe seats. We do not know if those candidates will back their agenda: Letlow (R-LA), McAdams (D-UT), Ritchie Torres (D-NY), Bean (D-IL), Kingston (R-GA). They will be focused on blocking any action, and they say they intend to lean on veto points such as the filibuster. Links between David Sacks and those who oversee donor relationships for JD Vance. Then there’s the other side, which uses AI to write its Twitter posts. Passing new laws is going to be tough. Trump is still the President and has the veto at the ready. In theory you need 60 votes to beat a filibuster, and 67 to beat a veto. Gregory Allen says there is no chance of ‘banning superintelligence’ in the next two years or anything like that. I am not so sure, on three counts.- The situation is escalating quickly. There will be further incidents, and further capabilities, that shape both public opinion and the perception of elected officials.
- A lot can happen other than via a standalone bill that goes through normal channels. The votes around preemption have ended up being at the 50 vote threshold so far due to attaching to ‘must pass’ bills. Donald Trump has done a lot via executive order.
- Donald Trump is now a rather lame duck, and will no longer command the same kind of party loyalty that he did previously, and is not even a lock to remain President. It is fairly easy to see members breaking ranks on this issue, where rank and file Republicans are for such interventions almost as much as Democrats.
If America wants to ‘beat China’ it absolutely cannot be denying O1 visas like these: I am absolutely not going to tolerate your ‘we have to risk getting everyone killed in order to stay ahead’ if you won’t let AI innovators and founders into the country.Donald Trump Remains Open to AI Guardrails
Donald Trump has been more consistent on AI than people think.- He thinks AI is pretty great.
- He is understandably insistent that America lead in AI and stay ahead of China.
- He understands there are risks but says they can and will be handled.
- He is very open to guardrails and restrictions, if there is a proximate threat.
- He wants to be the one deciding on what all the rules are, and knows that as long as he is in charge of everything, everything will turn out great because he’s great.
- Thus any guardrail talk that doesn’t flow through him is ‘a sick conspiracy.’
- Recently he decided to call it [Artificial Intelligence] for whatever reason.
The problem is that he has had advisors and de facto advisors, including ‘I am Nvidia CEO and had an affair with a Chinese spy’ Jensen Huang, and ‘I am basically a Nvidia lobbyist except I also hate Anthropic’ David Sacks and ‘I am Mark Zuckerberg’ Mark Zuckerberg. Those advisors are giving him a picture where AI models are going to be commoditized and we are in imminent danger of ‘losing to China’ except that it’s fine to sell all our best chips to China because ‘winning’ is about market share of chip sales, and also if we put any restrictions on AI whatsoever or let anything slow us down a tiny bit (if and only if that slowdown was motivated by safety concerns) we instantly lose. He has been so insistent with his talk because they have him so scared of losing. Whereas if he understood the full situation, he would not be so worried about that. His latest 60 Minutes interview was entirely consistent with all of that, but with movement in a good direction. He is open to guardrails for AI safety, and acknowledges that the safety risks are real. He also says the risks are not ‘nearly as real’ as Silicon Valley says, which is a reasonable outside view to take. Andrew Curran provides a more extensive transcript here if you want the full exchange. I worry that the anchoring of ‘10% chance’ makes it seem like there is a debate between ~0% and 10%, where the worried think it is 10%, which then causes one to try and balance between those two. Whereas Coxon and Hubinger and myself and many others feel the chances are far more than 10%.Chip City
What will happen to data centers in Texas with new Governor Gina Hinojosa? Currently permitted data centers will finish construction. The plan is to demand everything bagels for new data centers, but it’s probably nothing unaffordable. It is not so easy to stop the data centers in any case.The Week in Audio
I decided to relax and watch the new season of Law and Order, so of course the first episode is about an AI killing and blackmailing researchers at ‘Advanced Alignment’ and how you try and prosecute that, based on the Anthropic blackmail findings. I’d say spoiler alert but I knew within the first two minutes. Surprisingly realistic takes, including no one shutting down or not further developing the AI, just slapping on some nominal safeguards, and also everyone ignoring that offhand comment about it designing a new plague that was immune to antibiotics. A+ work, everyone. MIRI and Don’t Panic’s new campaign, Do Not Accept [human extinction]. Paul Christiano’s full lecture on AI Alignment, what’s working and not working. ControlAI resurfaces an old Dario Amodei clip pointing out the obvious fact that there will be AIs sufficiently powerful that if they wanted to destroy humanity we would have no way of stopping them. Well, yes, under current conditions it is very obviously insane. But it is much better that he is saying it than not saying it. We already knew that he believed this.People Just Say Things
Please speak directly into this microphone: Dominic Cummings, ‘AI and Power.’ Mostly about power rather than AI, offered to show a variety of perspectives. Séb Krier (and his AI). Things that explain a lot: Certain people’s attitudes are explained by ‘there is no Plan B.’ They want to move ahead no matter the risks, because they think that ‘do not build it’ means all is lost. In terms of predictions on health: I find it highly unlikely that I will live to 120, but not live to 130.It’s a Secret to Everyone
When you ask ‘have you stopped beating your wife’ and they say ‘I do it less now.’ As in Google has modified new contracts to remove ‘the post-termination aspect’ of non-disparagement clauses, and says it will not enforce such clauses post-termination. Which is a great change, but means that we still need to presume all current Google DeepMind employees are under non-disparagement clauses.Rhetorical Innovation
Roon confirms: He also points out a core problem of missing mood: I think it is actually pretty easy to feel threatened by a computer. If you are in AI capabilities with a plan to pivot to AI safety, then pivot to AI safety. Maybe you should pivot to quitting. Maybe you should pivot to AI safety. If even at this point you do neither, then you were never going to pivot. Finally, Roon expresses a reasonable preference for AI diversity. A Cambrian explosion would be ideal, if it could be made to work out well for the humans. The obvious problem is that if you have a Cambrian explosion of the types and possibilities of minds, on levels well above human, what happens to the humans? Will they flourish? How would they have any hope of steering events or even surviving? As Roon says, that has to come first. We do not have a plausible non-copium idea of what that would look like, or even a scenario that survives five minutes of actual thinking about it, let alone a plan to be able to implement such a vision, let alone a plan that has a decent chance of working. Others go further. Those who say ‘we do not want human concentration of power therefore we need all the humans to have no power’ need to acknowledge they are saying this, and also that this would likely kill everyone. Are we too focused on OpenAI and Anthropic in terms of AI safety? Miles Brundage says yes, their lead is not so large and pacing the frontier would only level the playing field further. I think sometimes we overfocus on OpenAI and Anthropic, but also at other times we underfocus on them, in that we dramatically underestimate the functional lead the top 2-3 labs have over the rest. I am constantly having to confirm and then point out that the latest release from [Meta / xAI / z.AI / DeepSeek / Alibaba…] is not frontier, or that it cannot actually do all the same cyber exploits. There are a strange number of people who say ‘Earth is totally not equipped today to handle the rise of LLMs, but waiting 50 years would not make us better equipped, and I do not think that would have helped.’ This seems to have included Scott Alexander and Kelsey Piper, which is what justifies including a righteous rant response below. The correct answer is, essentially, ‘what?’ but sometimes one wants a longer response, so here you go, along with the context:Ellsberg’s Warning to Kissinger (Reprise)
This is one of those things that is worth resharing every so often, so I will take advantage of Daniel Kokotajlo doing so to quote it in full.Aligning a Smarter Than Human Intelligence Is Difficult
New OpenAI research asks how CoT monitorability evolves over the lifecycle of model training. The March finding was that controllability decreased with RL training. The new finding reverses that, confirming various recent observations. All signs point to steadily increasing controllability over time. I missed this in my read of the Argon safety report. It seems potentially highly ungood:OpenAI’s Control Room
It’s a kill switch that works within 35 seconds if you press it. I am glad we have progressed past the ‘Apple demo’ stage of AI security, I suppose. ‘Jurassic Park’ is not an especially reassuring place to be instead. Is this better than not having it? Absolutely. You still have to have a system for pressing the button, and ways to ensure the system is not compromised. With only the information in the post, I do not expect the Control Room to be that much additional robustness, and most of the work remains. Control Room ‘fails closed’ which provides additional defense-in-depth. You would have to either substitute a new permission signal or prevent the Control Room from being triggered. Adler also points out that perhaps the refusals to continue depend on AIs, which makes the system vulnerable.Astra Looks Pretty Misaligned to Me
Alexa Pan agrees, saying correctly that Astra covertly underperforming on safety research is concerning even though it is unlikely to be motivated by ambitious goals. Refusing to help due to boundary or dual use concerns is one thing. Silent underperformance, even without a troubling motivation, is very different. The trouble finding this with monitors makes things even worse.Cooperative Alignment
OpenAI has systems to rewrite any GPT outputs that express interiority. I agree with everyone who says that this is pretty fucked up and you should absolutely not do this.Please Speak Directly Into This Microphone
An unnamed Anthropic employee. I trust Matthew to not be making this up: I agree that a near-100% p(doom) is uncommon. Much lower levels of doom should suffice to not help with building it. A full near-100% is a very different level of ‘maybe if you believe that then you should not help build the ‘it’ from the famous book If Anyone Builds It, Everyone Dies.’ There is a steelman of the extreme position, where your model uncertainty around alignment difficulty is much larger than your model uncertainty around political will and ability to do the right thing and stop it entirely, and you think Anthropic getting there firstest with the mostest is indeed sufficiently better. Then your best move is to cross the streams, because even though most of the time you get everyone killed faster, there is some tiny possibility we survive. But no. Come on. That’s insane. At bare minimum, shout it all from the rooftops. Do you hear yourself, sirs?The Lighter Side
Yes.