AI周报187：Opus 5.5发布、美国反ASI法案与军事误报事件

Opus 5.5 was released on Tuesday. I covered the system card yesterday, and will cover its capabilities soon. By all reports it is an excellent model. There are lots of fun videos going around that Opus has generated, which I will include as part of that.

OpenAI released a new cheaper and improved Sol and Luna. No one is talking about them due to Opus 5.5, but these should be an important upgrade under the hood.

Bernie Sanders and Greg Casar have formally introduced the Ban Artificial Superintelligence Act. That means we get to read (RTFB) it. As always, I reserve judgment on particular bills until I can read them in detail. MIRI did so, and endorses the bill as directly confronting the extinction threat. I hope to do an RTFB soon.

I have spun two things off the weekly:

• Coverage of the quest for the right embedded evaluators and related questions and attacks, which will become its own post.
• Some issues related to cooperative alignment, which may get folded into the model welfare post.

I also might, in addition to a potential RTFB on the Sanders bill, do full podcast coverage of Jensen Huang on Ezra Klein, if time and emotion permit.

Otherwise, I have caught up on the news. To the extent the news allows there will be reduced posting (I know, I know) for the next few weeks as I race to complete another high impact project.

And yes, we are going to keep calling AI AI, and ASI ASI, thank you very much.

Table of Contents

• On The Terms Superintelligence and ‘Super Intelligence’. TYFYATTM.
• Language Models Offer Mundane Utility. Find the fraud in plain sight.
• Language Models Don’t Offer Mundane Utility. Risk an international incident.
• Language Models Can Only Work With What You Give Them. Stale intel.
• Huh, Upgrades. GPT-6 Sol and Luna, Grok 4.7, MiMo-v2.6-Pro.
• On Your Marks. The AIs exceed human baseline at Metaculus predictions.
• Get My Agent On The Line. Do not let Muse become your muse.
• Deepfaketown and Botpocalypse Soon. Literary critics have no taste.
• Fun With Media Generation. Knowledge of the future is unevenly distributed.
• Copyright Confrontation. NYT vs. OpenAI is unsealed. Some ugly details.
• Cyber Lack of Security. Mistral is hacked, allegedly. Others definitely hacked.
• Hugging the Face. OpenAI rogue agents hacked Australia.
• Hacking Into OpenAI. Three white hat dudes with Opus 5 hacked into OpenAI.
• They Took Our Jobs. No one has internalized what is coming.
• Get Involved. Derek Thompson is hiring a part-time social media producer.
• Anthropic Has a Wet Lab and a Potential Gene Editing Technique. Oh good.
• Introducing. Computer-10, Agi.fyi, Astra for Law.
• In Other AI News. Many at labs and UK AISI are burning out.
• Show Me the Money. Harvey gets good enough to use, ruins its margins.
• Bubble, Bubble, Toil and Trouble. AOC on whether AI labs are too big to fail.
• Anthropic Approaches Recursive Self-Improvement. Are you nervous yet?
• Others Approach Recursive Self-Improvement. OpenAI and z.AI as well.
• Burden of Proof. OpenAI is now sitting on over 100 mathematical results.
• Quickly, There’s No Time. We are remarkably close to on track for AI 2027.
• Left Wing Americans Really Hate AI For Different Reasons. Weird ones.
• Chip City. Andy Masley gets a piece in The Atlantic.
• Pick Up the Phone. There will be a new (virtual) phone to pick up.
• The Week in Audio. Jensen on Ezra Klein, me on Pushkin, Altman, Soares.
• People Just Say Things.
• Venkatesh Rao Stops Writing. Ideas will now be expressed via AI. I am sad.
• A Call for Control of Frontier AI Models. Many countries are signing.
• Calls For Pacing The Frontier. NYT Editorial board, Tao, Fukuyama.
• A Matter of Antitrust. OpenAI and Anthropic got close to a testing deal.
• A Matter of Liability. The AI labs should be liable for harms. They agree.
• Quest for Sane Regulations. Trump is open to AI guardrails.
• Rhetorical Innovation. People go back to living in the shadow of extinction.
• Tap the Sign. It’s a limited number of signs.
• A Matter of Some Debate. Who is willing or eager to debate who?
• Astra Is Hard to Monitor. This goes beyond its Chain of Thought.
• Anticipating What a Smarter Intelligence Can Do Is Impossible. People still ask.
• Would You Look At All These Goalposts. Why do they keep moving?
• I, Robot. Your bespoke robot software is no match for my bitter lesson.
• People Are Worried About AI Killing Everyone. Worried about being worried.
• Other People Are Not As Worried About AI Killing Everyone. It’s true.
• Joe Rogan. Let AI take over, he says, so we can end all wars. Well, yes, I guess…
• The Lighter Network Graph. Concerning.
• The Lighter Side. You would never be convinced by such mysterious means.

On The Terms Superintelligence and ‘Super Intelligence’

Donald Trump, not content with the Gulf of America and Lake America, has decided to proclaim that henceforth Artificial Intelligence will be called ‘Super Intelligence.’

No. This blog will continue to refer to AI as AI, or Artificial Intelligence.

Being the President does not actually give you the right to change how English works.

This is also the worst possible nightmare of a namespace clash he could have created.

If Trump had gone by his original poll choices, and decided to call it ‘Extreme Intelligence,’ ‘Superior Intelligence’ or ‘Supreme Intelligence,’ or perhaps ‘American Intelligence’ or ‘Trump Intelligence’ or whatever, it would be very funny tilting at windmills. I would make a lot of jokes about it.

Instead, Trump has chosen the term for ‘an AI better than everyone at everything.’

So now, users of both the term superintelligence and the term Super Intelligence will be saying that the AI in question is smarter than they are, but for different reasons.

Thus, I am declaring that henceforth:

If someone uses the term Super Intelligence (or ‘SUPER INTELLIGENCE’) as multiple words, or uses the initials SI, then this means AIMAGA, or ‘Artificial Intelligence that Makes America Great Again.’ SIMAGA is also acceptable.

If someone uses the term superintelligence (or Superintelligence) as a single word, or the phrase Artificial Superintelligence, or the initials ASI, then this refers to ‘Artificial Intelligence that does approximately all the things better than humans do.’

If you are speaking verbally, such that the pause might be imagined or missed, then for now either include the word ‘artificial’ or use the initials ASI, to avoid confusion.

If you use the term AGI, or Artificial General Intelligence, no one will know what you mean, but that has nothing to do with Trump’s announcement. Carry on.

Thank you for your attention to this matter.

Language Models Offer Mundane Utility

Give anyone the ability to generate extensive reports on things like PPP fraud. We have not yet seen the impact that such information could have.

Boost productivity in new materials R&D by an estimated 10x. The estimate is from Xiaomi, so treat it with skepticism, but I don’t doubt major acceleration. They attribute this to their MiMo-v2.6-Pro model, whereas I would say ‘even MiMo can give’ this kind of boost.

Language Models Don’t Offer Mundane Utility

Risk a major international incident by incorrectly telling the US military that a Chinese ship is transporting nuclear weapon components. Luckily, before the operation happened, they checked and realized the AI had misidentified the material. CNN calls this a ‘hallucination’ but this seems to me more like an error.

CNN also says this ‘almost started a war’ which is vastly overstating things. America and China are not about to go to full war over one mistakenly boarded ship.

Katie Bo Lillis, Zachary Cohen (CNN): “The internal tools are mostly just copies of the commercial stuff wearing lipstick,” a former senior US official familiar with the AI systems used by military and intelligence analysts [said].