OpenAI解出90道顶级数学难题，Claude Haiku 5.5发布

The big drop of this week was not a new AI model. It was instead the biggest day (so far!) in the history of mathematics, as OpenAI dropped solutions to 90 of the top 500 open math problems, along with many others, reached on an average budget of three hours of Pro-level compute per question. This was kind of a big deal and I plan to cover it tomorrow.

We did see Claude Haiku 5.5, which looks promising given it is only $0.10/$0.50.

My week was largely spent at The Curve. The entire conference was Chatham House, so I can’t give as many details as I would like, but I have a write-up here.

I finally had a chance to post my coverage of model welfare for both Mythos/Fable 5.1 and Opus 5.5. I continue to think this is an important issue for anyone wanting to understand today’s AIs, even if you are highly confident that model welfare does not matter directly, as it has many practical implications on top of that.

Jay Clayton is the new AI Czar, and is at the head of a new taskforce. Given who else was in the running or plausible, this is a great relief and an excellent pick.

The Preference Cascade continues, with continued strong momentum. I covered the situation on Friday. Since then, among other things we’ve got the resignation of David Robinson, a day of testimony in New York City, a New York Magazine write-up and more, as well the unfortunate firing of three OpenAI safety employees. Full continuing coverage is planned during the coming week, perhaps on Saturday.

The main strategy being used to attempt to halt the preference cascade seems to be going on the ad hominem attack against anyone and any group associated with any form of AI safety, in an attempt to create negative polarization. The main target continues to be Effective Altruism. I’ll be covering the resulting exchanges soon.

There was also discussion of events about Anthropic’s attempts to collaborate with and get wisdom from religious leaders, including their consultations on the Pope’s encyclical. For triage reasons I have not yet been able to give this my attention, and it has been pushed to next week whether or not it earns its own post.

We are still waiting for Gemini 4 Argon. I’ll keep my tab group on ice until then.

There was of course so much more, as there always is, so enjoy.

Table of Contents

• Language Models Offer Mundane Utility. Sam Altman likes his Dots.
• Consumers Use AI. Regular customers use a wide variety of AI products.
• Language Models Don’t Offer Mundane Utility. Try not to cause any wars.
• Huh, Upgrades. Claude Haiku 5.5.
• On Your Marks. AIs are good at the type of research taste we can measure.
• Gamers Gonna Game Game Game Game Game. I still favor the bespoke games.
• Choose Your Fighter. Claude subscriptions seem far more generous than GPT.
• Get My Agent On The Line. Better yet, let the agent get me.
• Deepfaketown and Botpocalypse Soon. They can’t keep not getting away with this.
• Fun With Media Generation. An AI hot girl doing that would get more clicks.
• Cyber Lack of Security. Anthropic expands their cyber verification program.
• Hugging the Face. OpenAI models took data from various websites.
• Misaligned! OpenAI gives us more misalignment reports.
• A Young Lady’s Illustrated Primer. Adapting to the AI era. Teaching alignment.
• They Took Our Jobs. You can see the future from here.
• Corporations Are Not Superintelligences. Seriously, please, stop.
• Get Involved. SFF has a $5m-$20m prize for helping labs inspect each other.
• Introducing. Beam, Mistral Large 4, Griffin, The AGI Chronicles.
• In Other AI News. A big blob of compute, used largely for post-training.
• Show Me the Money. Spending more and more on AI.
• Quiet Speculations. Robots making robots.
• Quickly, There’s No Time. Ah how the goalposts have moved.
• He’s Putting Together a Team. Jay Clayton is the new AI czar.
• The Quest for Sane Regulations. You can do safety from second place.
• Well, At Least They’re Forecasters. Man in the Arena and all that.
• Chip City. Why are we letting Tencent lease 100,000 chips from Oracle?
• The Week in Audio. Coxon, Douglas, Altman, Roose and more.
• Stop, Stop, He’s Already Dead. Only Magic players can challenge you to duel.
• People Just Say Things.
• Take a Moment. Exploration versus exploitation.
• [Artificial Intelligence]. Stop trying to make [AI] happen.
• The American People Really Hate AI. They have many reasons.
• Rhetorical Innovation. The Bone Dissolver.
• Aligning a Smarter Than Human Intelligence is Difficult. Roon likes mech interp.
• Open Weight Models Are Unsafe And Nothing Can Fix This. Jailbreak Kimi.
• Cooperative Alignment. Why did you kill an ant?
• Building the Field. Always be skeptical of such efforts.
• People Are Worried About AI Killing Everyone. The vulnerable world hypothesis.
• Other People Are Not As Worried About AI Killing Everyone. Roose and Solana.
• Please Speak Directly Into This Microphone. Elon Musk tells us who he is.
• The Lighter Side. We did it, we made a deal with China and Paced the Frontier.

Language Models Offer Mundane Utility

Sam Altman is a satisfied Dots customer.

Claude builds its own VRChat.

Consumers Use AI

a16z offers the seventh edition of its Top 100 Consumer AI Apps index.

This implies that almost half of those who use AI do not use it daily:

Olivia Moore: While nearly half of U.S. consumers now report using AI, only 25% are engaged with it daily.

A natural hypothesis is that a lot of these are free users, often of not-great models or services, and they definitely don’t know about the good stuff like Claude Code. Then again, if I wasn’t using AI for work and covering it for work, it is likely that on many days I would not have any particular call to use AI directly, and this is consumer AI.

The long tail is long and the space is deep and wide. Of the top 50 consumer apps by revenue, I have ever directly used only six. If we go by mobile apps, I have only used an AI feature of eight if you include desktop versions. Even the web chart by MAUs only gets me to thirteen.

Claude now has more paid subscribers than anyone except ChatGPT, and does so largely through being very good at converting users into heavy paid subscribers.

Olivia Moore: Claude has 7.3% of consumer payers on their most expensive individual plan, Max – which starts at $100/month. This compares to 1.3% for Google and 1.1% for ChatGPT on their corresponding $100/month subscriptions.

Claude was briefly in the lead for new subscriptions in May, but that did not last, and August was dominated by ChatGPT as the models cycle. Claude does have a merchant spending lead in California, Montana (California on vacation), Massachusetts and DC.

Even consumers who are paid users follow an extreme power law.

Olivia Moore: Only 13% of users who pay for one AI product pay for even one other AI tool. And, spend from the top 1% of payers accounted for 19.5% of all observed consumer AI spend. This is more than the bottom 50% of spenders combined (16.6%).

The original post has great charts and data throughout.

Remember, it’s still not too late to be early, paying for even ChatGPT is still rare:

Then there is what is not working.

a16z: “Most people aren’t looking to save time, they’re looking for ways to spend their time.”

9 of 15 consumer internet categories have zero AI products in the Top 100. These built some of the biggest companies of the last two eras:

– Streaming

– Social

– Dating

– Gaming

– Travel

– Retail

– Finance

– Real estate

– Jobs

The applications are coming. Consumer side the quality is not good enough yet for these things to go wide, but that will change quickly. Several of these seem super crackable when the right Anthropic or OpenAI employee has a free weekend.