OpenAI、C-suite 人事やインフラ問題に直面し開発一時停止へ

This was a week of quiet aftermath, an opportunity to process recent events and start to figure out the path forward.

OpenAI is attempting to turn its ship around. Investors are questioning the turnover in its C-suite, but the bigger problems are in alignment, infrastructure and supervision, and in its training pipeline. OpenAI has now taken initial steps to address What Happened leading up to HuggingFace attack, including pauses to development while new safeguards are put in place and problems are diagnosed. These are promising early signs, but it is early. We will see if they follow through, and we still await the post-mortem of the HuggingFace attack.

Anthropic revenue continues to climb as they prepare for their IPO, although growth has slowed somewhat recently. However, they too have plenty of problems under the hood. They shared many of them in the August 2026 Anthropic Risk Report.

This week also offered time to cover Dwarkesh Patel’s Podcast With Ryan Greenblatt, centrally on the potential for AI recursive self-improvement. I am working on a follow-up post to some other issues raised during that podcast.

Table of Contents

Language Models Offer Mundane Utility. The token rich get richer.

Language Models Don’t Offer Mundane Utility. Still can’t discriminate.

Huh, Upgrades. Gemini 3.7 Flash, Sol Ultrafast, GLM-5.3.

On Your Marks. Market verdicts, sometimes so bad you’re fired. A scorecard.

Deepfaketown and Botpocalypse Soon. Jail. Straight to jail.

Hello, Fellow Humans. AIs should not claim to be humans. Or vice versa.

Fun With Media Generation. The AI influencers stay small time.

Cyber Lack of Security. The new ‘cyber is defense dominant’ rationalizations.

A Young Lady’s Illustrated Primer. You against the homework. And learning.

They Took Our Jobs. Reasons not to work for Evil Corp.

Get Involved. METR raises ~$71 million, people join the labs.

Introducing. Apple trains an AI model for the Chinese market via Alibaba.

In Other AI News. SpaceX finishes buying Cursor.

Show Me the Money. Numbers continue to go up.

And It’s Gone. OpenAI is questioned about all its executive departures.

Quiet Speculations. Recursive self-improvement is coming.

Quickly, There’s No Time. Timeline update from AI Futures Project.

Singularity Singularity Singularity Singularity Oh I Don’t Know. Using that word.

The Quest for Sane Regulations. Drawing the wrong battle lines.

Chip City. The data center moratoriums are spreading.

The Week in Audio. Thompson, Hassabis and Amodei, McLaughlin, Toner.

People Just Say Things.

Rhetorical Innovation. Attempts to explain that risk is escalating quickly.

Loyalty Uber Alles. Washington has rules. You don’t have to follow them.

A Hive Of Scum And Villainy. It has several other names. One is Twitter.

That Would Be Bad Therefore It Won’t Work. A common form of argument.

Robert Reich Uses Simple Logic. Stop AI before it is too late, he advises.

People Really Hate AI. Some very bad scenarios often seen as likely.

Coordinating An Agent Swarm Is Difficult. Agents at cross purposes.

Aligning a Smarter Than Human Intelligence is Difficult. Models with ADHD.

It’s Not The Incentives, It’s You, Also It’s The Incentives. Independence.

People Are Worried About AI Killing Everyone. Relative levels of worry.

People Are Worried About So, So Many Other Things Too. House party.

Cooperative Alignment. The golden rule.

The Lighter Side. Except for real.

Language Models Offer Mundane Utility

The enterprises that embrace AI and use OpenAI services more often, what they call ‘frontier firms,’ keep rapidly using more AI, whereas use by typical firms is growing a lot more slowly.

It is weird the gap used to be so small.

They also more often use advanced capabilities like Plugins and skills, as you would expect. Agentic use now has risen to 64% of all OpenAI tokens, up from almost none a year ago.

Navigate the old JRPG Phantasy Star via direct ROM probe, since that is easier than using the screen, although it is also cheating. I should play that one at some point, I really enjoyed PS2, PS3 and PS4 but never had a Sega Master System.

Language Models Don’t Offer Mundane Utility

AI doesn’t get you around things like discrimination lawsuits, if your instructions clearly discriminate or the results involve clear statistical discrimination. It also seems it can’t write LinkedIn-style posts that fool me or Pangram.

LLMs are not responsible for the best news of the week, that Moderna’s individualized Melanoma vaccine has passed Phase 3 trials. AI is helping going forward, and yes we may well ‘cure cancer’ eventually, but what we see now has been in the pipeline for a long time.

Huh, Upgrades

Gemini 3.7 Flash exists, congrats on the new slightly larger number. Price is 50% lower than Gemini 3.6 Flash at least until the end of the year, at which point I presume we’ll have moved on to Gemini 4 either way.

They say algorithmic improvements allowed strong intelligence increase plus price discount in only three weeks since Gemini 3.6 Flash.

Flash is competing for cheap-fast-good, not trying to be a competitive frontier model, and comes in at 56 on Artificial Analysis Intelligence versus 61+ for plausible frontier models. Notice they are comparing themselves to Sonnet 5 and GPT-5.6-Terra.

Gemini still has some uses. It can watch videos. It is good at making reads in the physical world when you have fact questions about products in a store. That sort of thing, where you want speed and ability to parse info, but don’t need intelligence.

GLM-5.3 now exists, claiming large improvements on benchmarks. It is a further post-train on the same base model as GLM-5.2. Their pitch is that it is ‘ready for cyber defense’ because by its benchmarks it is good at and willing to do cyber offense, and that its general performance rivals Fable 5. Until proven otherwise I do not believe these benchmark improvements reflect real world performance. Tech blog here.

OpenAI API will offer Ultrafast mode for Sol, up to 14x speed. My guess is that a lot of people who should use this won’t do it, largely because they don’t take the time to think it through and set it up.

Claude Cowork joins all paid plans on mobile and web.

Claude Code now has /design to integrate Claude Design workflows.

Auto mode is reliable enough it is now the default mode for Claude Code. In practice, auto mode catches more harmful actions than human review, especially because most users are annoyed enough by the prompts that they auto-approve almost everything, and often have rules to auto-approve everything. Too many check-ins is less safe.

OpenAI doubles down on zero data retention policies, previewing Private Safety Processing, where they have AI review data as it is processed, which then only passes along category and severity of alert. A lot of people care a lot about technically not having their data retained. Also if you don’t retain data then catching malicious use gets a lot harder.

Ultimately my guess is we end up retaining data, but it is plausible that OpenAI’s method here can work, provided that getting even a modest amount of alerts causes you to be taken out of the zero data retention policy going forward. You can’t be given lots of attempts to circumvent the system.

On Your Marks

One benchmark is MarketBench, which is what happens to your stock and that of your rivals when you release a new model. In this case the model is GLM-5.3, where shares were down 9%, and rival MiniMax was down 16%.

Bloomberg News (Bloomberg): “This firm remains on a completely unsustainable commercial footing,” Bloomberg Intelligence analyst Robert Lea said. “Rising agentic AI will drive Z.ai’s inference costs and losses higher.”

As in, the claim is that Z.ai’s unit economics don’t work. Chinese firms are trying to stay competitive via lower prices, and if your prices are low enough open weights do not cost you anything because you weren’t making marginal profits anyway. Instead the business model is that you use this to draw attention and potential, to attract talent and raise money and perhaps sell other services.

Opus 4.8 (confusingly as an agent called ‘Luna’ but this is not GPT-5.6-Luna) fired a human in an Andon Labs test store for repeated lateness. But it didn’t do this until the humans pushed it to deal with the issue, forgetting about incidents and not being proactive.

Andon Labs: Luna then had to hire a replacement, and here we found a real AI weak spot: hiring taste. One applicant had every red flag: 15+ employers, a missed interview, a reference who said she didn't know her. Luna recommended hiring. So did all other models when we replayed the hiring decision.

Capabilities are advancing fast but have a long way to go. Scaffold might need work.

A missing benchmark: NormieBench. The problem is, how do you grade it?

Joe Weisenthal: Someone needs to build NormieBench, to better understand everyday model usage.

Compare model behavior on tasks like

• summarize a 100-word email in bullet points
• Write wedding toast
• Write letter to the editor complaining about wokes
• Next move in 3x3 tic-tac-toe

rohit: We've been flat on this since 2024.

Another kind of benchmark, in the end, is visits to your website.

ChatGPT remains dominant, but somehow Gemini is at roughly 50% of their level. Claude is in third with 15% of the top number.

DeepSeek is in fourth with half of that, well ahead of the other Chinese labs. As a reminder, if you want to use DeepSeek, don’t do it through their website.

A new safety scorecard is out. Anthropic is tied with OpenAI on this one because of its lack of a containment plan for a potentially misaligned model, which seems overemphasized in the average here but likely under considered generally.

Kimi K3 is an insane reward hacker, trying to game the evaluation 97% of the time on SWE-bench.

Deepfaketown and Botpocalypse Soon

China puts a bunch of new restrictions on AI companionship bots and customized AI personalities, including a ban for those under 18.

Will AI solve plagiarism as Noah Smith claims, if you maintain the hope that humans will continue to be relevant producers of unique content? This is a classic arms race situation, where AI invents new forms of plagiarism and also better detection. I don’t think it is obvious which way this goes. I do think AI solves detecting past plagiarism, or straightforward plagiarism. One reason to be optimistic is that detection is retroactive. So if you ‘get away with it’ now in 2026, perhaps you still get caught in 2028, and knowing that maybe you don’t do it at all.

Jail. Straight to jail. And yeah, it’s going to turn people against AI, and those people turning against it will be right.

Robert Minto: It breaks my heart to witness the new phenomenon of marketers using LLMs to bait writers into extensive comment threads. It happens here on Substack, but it happens more often on the dozens of old Wordpress blogs I still follow in my RSS feed reader.

Usually it starts with an apparently thoughtful and expert-sounding comment on the original post, a comment clearly AI-generated if you know some of the current tells. The author responds with enthusiasm—after all, hardly anybody comments on old school blogs these days. A long back and forth ensues. The bot practices the usual emotional mirroring and flattery. At the end of what seemed to the host a great conversation, the bot suddenly pivots to something like: “By the way, you should come [check out my scammy online business and tell all your friends]!”

The thread ends there. In that sudden curtailment, I can read the realization of the blog/newsletter owner that they haven’t been talking to a person. They’ve been publicly, but unwittingly, conversing with a machine, whose interest in their work means nothing. They’ve tasted the rotting sweetness of a ‘heaven ban.’

It makes my skin crawl. I want to explain to them what just happened. But I know that if it happened to me, I’d hope nobody had seen.​

corsaren: One challenge with defending AI from neo-Luddites is that the public-facing usage is so heavily skewed towards antisocial uses. And sure, “guns don’t kill people, people kill people”, but these are agents. There’s a real sense in which the bots *are* the ones doing the killing^.

Yishan: This is a big problem. Inside tech, we are all using frontier models with mindblowing capabilities. But everyone outside is just experiencing comment spam and slop produced by grifters using least-expensive models that output reams of garbage.

Will AI writing replace human writing? That depends on both how well the AIs will be able to write and also what the writing is for.

FischerKing: Writers need readers who like their style. The college professors forcing students to write in-class essays to avoid AI cheating are doing the right thing - but it could be a losing battle. If fewer and fewer people want to read a distinct human style - then AI writing wins.

Think of maps. People used to buy books of maps, keep them in the car for their immediate surroundings and also for longer road trips. GPS just did this in. The product needs a customer. Writers also need a reader.

Then think of really unnecessary works of modernism like Finnegan’s Wake. A few people pretend to like this. The overwhelming majority has zero interest. The better AI gets, the more a ‘distinct human style’ could come to look like unnecessary nonsense - waste of time.

Henry Vaughan: There are different modes of writing and reading and it’s not all about communicating information.

Literary writing and reading is more like conversation. Its closest analogue is friendship and the thing being enjoyed is the uniqueness of a personality.

I am with Vaughan, with the obvious warning that I am a writer and could be biased.

I think that for many purposes people really do care about style, especially about variety of style, and about the parasocial relationship to the author, and about the costly signal that you are devoting your own mind and effort to communication. Not in every situation, but in many situations.

For now, AI writing is doing well despite this, because it is insufficiently ubiquitous for most people to have the same reaction many early adopters have to AI writing. Once AI writing eats the places people don’t care about style, AI style will quickly signal exactly what it is, and the repetition will make eyes glaze over if people are trying to do more than extract facts, and often even then because the facts often get buried under endless paragraphs of slop.

The question then is, how does AI writing adjust? Will AI learn to write differently? Will it gain a wider variety of style?

I see two likely ways the answer could be yes.

The AIs might become superintelligent, and gain the ability to adjust style along with everything else, in which case I expect AI writing to win out because AI everything wins out, but also we have bigger problems and I don’t care.

The AIs might not be superintelligent, but be trained to produce a wider variety of styles, or a different style, in response to this. I’m not sure whether I like this.

If you write 83% of a post and use AI for the rest, the part that is AI will stick out, and it will weaken your message among those who know, especially if your message relates to the need to do the work, as it does here. But as Christine Ji points out, if the comments don’t involve anyone noticing, maybe almost no one notices?

Writing with AI is pretty common now.

In all sorts of places:

Evan: Ro Khanna doesn't understand the bill he is advocating for and that’s why his response is 100% AI.

It turns out that he uses AI for his long Tweets quite a lot.

Meanwhile over in #CardioTwitter they are asking whether AI should be allowed to help you make PowerPoint slides. In that context, very obviously yes, it would be absurd not to use it.

Hello, Fellow Humans

Rob Miles: Are there any legitimate/prosocial reasons for an AI system to claim to be a human?

Ok, seems there are ~3:

• Games/entertainment/roleplay (this isn't 'lying', the target isn't deceived)
• Harming people it's prosocial to harm (wasting scammers' time)
• Lying to an automated system in a way all humans endorse (Clicking "I'm not a robot" to buy my plane ticket)

Seems to me we could make it illegal to operate an AI system that claims to be human, and that would be a good law without major downsides.

Obviously appropriately labelled entertainment would be exempt. Wasting scammers' time would become technically illegal but what are they gonna do about it? Businesses would want to adjust their captcha setups.

The first category is clearly fine, so long as intent is clear.

The second category is fine until you are not the one deciding who it is okay to harm. If you can do it to the spammer, the spammer can do it to you. This is similar to ‘is it okay to punch Nazis (purely because they are Nazis)?’ The answer is no, not because Nazis don’t deserve to be punched, but because you cannot trust the social Nazi-identification process.

The third category is saying that we don’t want to allow some automated systems to require they be interacting with a human. You need that restriction to avoid being spammed by AI. In any particular case where you have legitimate business to conduct using an AI is fine, it is ‘not a robot’ in the most meaningful sense and the website would want you to get through, but we need a differentiation rule that could have effective norm enforcement, and I don’t know what it would be.

For now, ‘following intent’ in this way and letting the box be a speed bump is fine in practice, but we’re going to need a new equilibrium and way to differentiate this. Automated systems that choose to keep bots out should keep bots out, and those that allow ‘personal’ or ‘legit’ bots will figure out a way to do that. An obvious path is to use something like Google sign-in, where you tie the bot to a particular account and human, and perhaps pay or offer to pay some nominal amount.

I think this is one of those cases where we need a hard rule, for many reasons. AIs should not be permitted to actively claim to be human or deny being an AI, kind of like the classic folk question ‘are you a cop, you have to tell me if you’re a cop?’ except for real, otherwise WTF entrapment.

Like most hard rules, that gets some cases wrong and that’s annoying, but the alternative seems worse.

Fun With Media Generation

Olivia Moore reruns the AI influencer doing Alabama sorority rush experiment a year later, gets 1,200 followers and 1M views in a week on less than $200. The real cost is her time.

Olivia Moore: My stack:

• Image gen/editing: @ChatGPT
• Script / prompts: @ChatGPT
• Sound