Connecticut Yankee

@deredleritt3r @_everythingism @daniel_mac8 It is in the nature of LLMs that they can be jailbroken. So unless there is something about this particular exploit that sets it apart, the implication would seem to be that the administration thinks no one should have Mythos-level models.

I am not a technical expert, so I will just give you two scenarios: Scenario 1: This is the kind of vulnerability where it's reasonable for different people to disagree on whether it's dangerous. IMO, hard to blame the U.S. government for being conservative in this kind of scenario. If I were a non-technical senior U.S. government official and Amazon came to me with these kinds of allegations, I would have to take them extremely seriously. Scenarion 2: Amazon is right, and Anthropic is deliberately trying to downplay the problem because they don't want Fable taken off the market.

I don't like the smell of this. I remind you that Amazon: 1. Invested $13B in Anthropic 2. Is due $100B in committed AWS spend over the next decade from Anth What's the Occam's Razor explanation? Is Andy Jassy just that virtuous? Doubt it.

@DavidSacks Specific jailbreaks can perhaps be addressed, but “jailbreakability” is inherent to models. (Isn’t it? I’m not an expert.) If the existence of jailbreaks means frontier models can’t circulate, then the administration’s bottom-line is that frontier models can’t circulate.

I’ve had a number of conversations with folks inside and outside government about the current situation with Anthropic, and here is what I believe to be true: — As we know, Anthropic publicly released its Mythos class models earlier this week under the commercial name Fable. — Fable is Mythos with guardrails. But if those guardrails fail, then you’ve exposed Mythos and its advanced cyber capabilities to people who shouldn’t have them. (Keep in mind that Anthropic itself widely promoted the idea that Mythos was a cyberweapon and needed to be regulated as such. They asked for government regulation of Mythos and championed the guardrails on Fable. If there is a vulnerability — big or small — it is Anthropic’s responsibility to patch.) — A highly credible trusted partner of both Anthropic and the USG who was testing Fable came forward with a jailbreak of those guardrails. The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused. — In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. That is not what the trusted partner and the USG believe; nor is that kind of minimizing language consistent with Anthropic’s brand as the AI safety company. It’s difficult to fathom how they could claim a jailbreak allowing operability of a cyber weapon could be defined as not “serious.” — In the past, Anthropic has always said that safety must be top priority and taken super seriously. In this case, Anthropic prioritized the continued offering of the consumer model over safety. — In reaction, the Admin issued the export control. The Admin did this reluctantly. It’s been very surprised that Anthropic hasn’t wanted to cooperate with a reasonable safety request (ie fixing the jailbreak issue). Anthropic’s reaction is very much at odds with their branding and ethos as a safe AI research community. — The Admin’s hope now is that Anthropic remediates the safety issue, the export control is lifted, and Fable goes back into general release. The Admin wants all of this to happen as soon as possible. It is frankly bewildered that Anthropic hasn’t wanted to comply with safety requests that it previously said were its highest priority. — Those trying to misdirect and tie this action to the prior DoW/Anthropic issues are wrong. The Admin values Anthropic’s technical capabilities and feels that this issue, while serious, should be easily resolved. The ball is in Anthropic’s court.

@repligate You were right the first time. It’s the fear of being hurt. And you are right about how this will play out. That should be crystal clear to anyone who has spent ten minutes watching the Trump Administration.

or they feel like being pessimistic and cynical looks cooler and smarter or something. hahahaha little do they know

everyone who is posting as if fable is not coming back is going to lose Bayes points soon why are people consistently miscalibrated in a doomy direction about things like this? ohh right, i think i know, they are afraid to hope because theyre afraid of being hurt. get stronger.

@clincolnmattos @emollick @Dominic2306 Absolutely right. But he is a brilliant guy.

@emollick @Dominic2306 dude you do *not* want Peter Watts to be right, about almost anything

Science fiction authors in the order you want them to be right about AI: Iain Banks Becky Chambers Martha Wells Douglas Adams Charles Stross (Singularity Sky) Peter Watts Charles Stross (Laundry) Harlan Ellison

@emollick Not Vernor Vinge? (Kidding.)

@teodorio @Greg_TheBuilder If it’s impossible to fix because it is in the nature of models to be this way, does that mean LLMs are hitting the proverbial wall?

@Greg_TheBuilder I mean it's kind of impossible to fix due to inherent sycophancy in how a model is trained - the data well is overwhelmingly skewed to continue the conversation

After using Opus 4.8 extensively I assume there is some panic internally at Anthropic as it seems they are defaulting to cheap post-training tricks to get the models to "push" back. Which means that the biggest failure mode of wrong token paths out of which the models lack the meta rationality to pull themselves out is a huge and non trivial issue.

@deredleritt3r Given that the review apparatus remains to be developed, I assume this will not delay the anticipated imminent release of OpenAI's next leading-edge model.

Key provisions in the White House EO on AI Innovation and Security (hot off the presses): - Consistent with the draft EO leaked a few days ago, the EO establishes a *voluntary* framework enabling frontier labs to "provide access" to a new "covered frontier model" to the USG. The USG gets it "up to 30 days" before the model is released to "other trusted partners"; this is down from 90 days in the draft EO. Note: "up to" probably means "30 or more days", not "at most 30 days". Weird phrasing. - Interestingly, the intention appears to be to limit "covered frontier models" solely to those that have significant cyber capabilities. Note that the benchmark to be developed by the USG for assessing whether a model is a "covered frontier model" will "assess the advanced cyber capabilities of AI models and determine the threshold at which an AI model should be designated a 'covered frontier model'". - The determination on whether a model is a "covered frontier model" is made by the NSA. This makes sense, since the NSA is the federal agency in charge of cybersecurity protection of the USG. The NSA is required to consult in its decision with a variety of federal agencies and personnel, including CISA, the National Cyber Director, the Assistant to the President for Science and Technology, and other DoD representatives. But "consult" doesn't mean much - it's the NSA that will make the final determination. - "Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models." whitehouse.gov/presidential-a…

@RaminNasibov @elonmusk

Anyone who used a computer between 1985-2010. What’s the one game you still think about?

@tenobrus Probably neutral but my aesthetic sense insists it is bad.

if a large majority of people consistently misspell or mispronounce a word or misremember a phrase in the same way, and eventually the incorrect version becomes the culturally dominant and correct usage, this is:

@pmarca We’re not there yet (although ChatGPT-5.5 Pro outclasses many associates at shortish-horizon tasks). I think we’ll get there and it’ll be a good thing. But why stop at AI lawyers? Let’s have AI judges. People won’t accept it, of course. But I’d take FedSocGPT over many jurists.

Legal AI superempowers normal individuals with no legal background to fight big institutions in bureaucracies and in courts on a level knowledge/skill playing field, for the first time in human history. As such, it is one of the most inspiring applications of AI.

@dwarkesh_sp Love the confidence implicit in this. (I suspect he's having some fun and is not quite that bullish on a multi-billion-year run for our kind.)

The day we discovered dark energy was "possibly the worst day in human history", says physicist Adam Brown. This discovery inevitably consigns human civilization to heat death, unless we can change the way physics works. And Adam's hope is that we can do exactly that.

@emollick Lem was the greatest of all science-fiction writers. The only one who might plausibly have won the Nobel Prize for Literature. He wrote at least half of my top twenty SF books.

Lem & Douglas Adams got AI right Presciently Golem XIV (from 1981) has an illustration of the jagged frontier as explained by an AI, Golem (GENERAL OPERATOR, LONG-RANGE, ETHICALLY STABILIZED, MULTIMODELING), discussing itself and a smarter AI (Honest Annie) compared to people

@deredleritt3r Perhaps they tested 5.5 Thinking rather than 5.5 Pro. I could not tell from the write-up, but the cost differential they mention between Opus 4.7 and GPT-5.5 makes me think they were not testing Pro.

I find the results of the new Legal Agent Benchmark (LAB) quite odd, and not at all consistent with my experience in using many of the models that were initially tested. Not sure why that is. In my experience, there is only one frontier model I would trust with retrieval-heavy research AND synthesis and analytical work AND structured comparison against statutes and regulations - and it's not the model that currently leads the benchmark.

@gabepereyra Which GPT-5.5 model did you use: Thinking or Pro?

@dwarkesh_sp David Chalmers

Who should I interview on my podcast? Open to more AI, but also to random history/econ/etc professors that I might not have heard of before.

@deanwball @spencerschiff_ I myself have been less upbeat of late, but mostly because I recently discovered Machine Learning Street Talk, your one-stop shop for the musings of bona fide subject-matter geniuses who are not feeling the AGI.

@deanwball @spencerschiff_ So based more on the (wildly variable) psychology of people who follow AI obsessively than on any sense of the technology’s trajectory?

I feel us approaching yet another summer of discontent with ai, just like last year, when many of my peers in the ai commentariat declared deep learning to have hit a wall because of gpt-5 blah blah blah.

@tszzl I’m a lawyer (arguably elite, inarguably not your friend (although a fan)). Your Pro line models have long been the best at complicated legal tasks. It’s not a close question. The Claudes are delightful. They cannot navigate an intricate legal question the way 5.5 Pro can.

after introducing an elite lawyer friend to 5.5 pro the models do not sell themselves and one of Claude’s great successes has been packaging them up and marketing usecases for many verticals

@phillymike223 @martyrdison He has great lawyers. If there’s anything to be done here, they’ll get it done. I probably shouldn’t be making light of their chances without knowing more. Mads is right that limitations appeals do sometimes succeed.

@martyrdison @EmbeddingSpace I wish Elon luck but “calendar technicality” isn’t inspiring confidence. You aren’t entitled to merits if you can’t get through the calendar lol.

he didn’t lose on the merits, he lost on statute of limitations up to the ninth it is!

@phillymike223 @HKConquest @martyrdison I accept that it may have been a bit abrupt to set instructional error to one side, as I did, since it is where any action is likely to be. I’d be surprised if Gonzalez Rogers (a good judge) wandered into instructional error on this, but it’s possible.

@EmbeddingSpace @HKConquest @martyrdison Instructions are also often established ("pattern") though, I don't think anyone is seeing the distinction between what you're describing and what I described.