← Home AI in 15

AI in 15 — November 09, 2026

November 9, 2026 · 15m 29s
Kate

Goldman Sachs is about to help sell shares in Anthropic to the world. And most of its bankers and traders have just lost access to Anthropic's coding agent.

Kate

Welcome to AI in 15 for Monday, November 9, 2026. I'm Kate, your host.

Marcus

And I'm Marcus, your co-host. Happy Monday, Kate. Today a bank takes the power tools away from everyone without an engineering badge.

Kate

I've been banned from power tools too, so I sympathize. Here's what's coming up. Goldman limits Claude Code to its engineers.

Kate

OpenAI closes eleven billion dollars at a lower price than the press expected.

Kate

Anthropic heads out on its roadshow, and reportedly buys a lot of gas generators.

Kate

Attackers go from GitHub to cloud secrets in five minutes.

Kate

Plus a laptop running a giant model with no GPU, Apple's plans for Siri, and a reality check for coding agents.

Kate

Marcus, this one hit nearly seven hundred points on Hacker News. What did Goldman actually do?

Marcus

According to Bloomberg, Goldman updated its internal guidance in October after a review of productivity and risk. Claude Code and GitHub Copilot are now limited to the bank's engineers, and most traders and bankers have lost access. The exception is the strats, the people who sit between trading, quant research and engineering. They keep it.

Kate

What were the traders using it for?

Marcus

Monitoring client flows, modeling market moves and positions, and making written and graphic materials. Some are reportedly frustrated. But this isn't a ban on AI. Goldman's in-house assistant, which runs several models, is still open to everyone. The restriction targets third-party agents that write and run code on their own.

Kate

Wait, didn't Goldman brag about exactly this not long ago?

Marcus

In July, the bank's CIO, Marco Argenti, was praising vibe coding by non-technical staff. Traders would build tools with AI and hand them to the roughly twelve thousand engineers to harden. Three months later, the policy went the other way.

Kate

So what happened? Did something blow up?

Marcus

We don't know. Spokesperson Nick Carcaterra said only that the bank provides certain AI tools so it can "determine the best use cases," which could mean almost anything. One widely upvoted Hacker News comment speculated that an operational loss triggered this, because only traders can send orders to the market. Nobody has verified that, so treat it as a guess.

Kate

Noted. What are the less exciting explanations?

Marcus

There are three good ones. Claude Code runs with the user's own permissions, on machines that hold client data. Code written by a trader can skip review and testing. And these tools were never designed around information barriers, the rules that stop people advising on a merger from sharing what they know with people trading the stock. An agent doesn't know which side of that wall it's on. Cost played a part too, since agents are billed per token.

Kate

I saw people on the thread bringing up the "lethal trifecta."

Marcus

That's an agent with private data, exposure to untrusted input, and the ability to act. A trading desk offers all three before lunch.

Kate

And now the awkward bit. Goldman is working on Anthropic's IPO.

Marcus

Goldman and Morgan Stanley are the lead underwriters. To be fair, the relationship looks healthy. Anthropic engineers spent six months at the bank building Claude agents for trade accounting and client onboarding. What's different is the kind of restriction. Goldman removed Claude for Hong Kong bankers in April, and JPMorgan did the same in June, but those were about location and contract terms. This one is about your job.

Kate

Is everyone on Wall Street heading this way?

Marcus

No. Barclays expects half its developers on Claude Code by year-end. Three weeks after saying that, it named developers among five thousand job cuts.

Kate

Oof. Part of me feels for the trader who built something useful and just had it taken away. People closest to the work usually know what they need.

Marcus

I agree, and I'd have liked to see a sandbox and a review process before a removal. But this looks like where regulated companies are going to draw the line: agent-written code should come from people trained to review it. I'd expect more controls based on job role and fewer blanket bans.

Kate

Quick hits. OpenAI just banked eleven billion dollars at a one point two five trillion dollar valuation. Marcus, why does that sound like bad news?

Marcus

Because of what was reported before it. Press accounts said OpenAI wanted thirty billion or more at one point four trillion, and two investors put the initial ask at one point five. CFO Sarah Friar confirmed the one point two five on TV and said the higher number came from the media, not OpenAI. Either way, about eleven billion arrived by November sixth. A follow-on of up to nine billion could take it to around twenty, but that hasn't closed.

Kate

Still, it's a big jump from March, right?

Marcus

Roughly forty-seven percent above the eight hundred and fifty-two billion valuation from March. A step-up is a step-up. But after last week's chip selloff, late-stage investors have clearly become pickier about price.

Kate

What did Friar say about the business?

Marcus

Growth percentages and no revenue dollars, which I always notice. Business up seventy percent quarter over quarter, enterprise doubled, revenue split about evenly between consumers and companies, one point two billion weekly ChatGPT users, and a one billion dollar ads run rate by August. "The demand for AI is real," Friar said. I'd like to see the actual revenue next to those percentages.

Kate

And data centers?

Marcus

OpenAI says it will keep expanding in the US despite the post-midterm scrutiny, and it has pledged to stop requiring confidentiality agreements from public officials during site talks. That's a small concession, and it should never have been standard practice.

Kate

Now Anthropic. We've been tracking the falling price talk all week, and the roadshow is reportedly starting now. What's new?

Marcus

The Nasdaq listing is expected under the ticker A-N-T before the end of November, and one tracker points to the week of November sixteenth. The amended filing still leaves the price range blank, so every valuation you've heard is a press report. Reuters-sourced coverage cites quarterly revenue of nearly seventeen billion dollars, with growth slowing sharply and about five billion in additional compute costs.

Kate

Which brings us to the Financial Times story from Sunday. Anthropic went shopping for generators?

Marcus

About three point four billion dollars' worth, for one point seven gigawatts of heavy-duty gas generation. The FT says that in late September, Anthropic took over orders for Caterpillar machines from a San Antonio company called OnePWR Solutions, which had ordered them about two years ago and couldn't finish paying. Anthropic had planned to buy power from OnePWR, then decided to buy the equipment and the know-how to install it.

Kate

So it's basically buying someone else's place in the queue.

Marcus

Exactly that. Lead times for this equipment run for years, and power is now the binding limit on AI expansion. One caveat: neither company has confirmed it, and Reuters couldn't verify it. There is a decent clue, though. Piper Sandler, OnePWR's adviser, lists a three point four billion dollar sale of OnePWR's power platform to an unnamed AI lab in September.

Kate

I'm guessing the internet had thoughts about the safety lab buying gas generators.

Marcus

It did. Hacker News enjoyed that irony a great deal. The harder question is how this squares with Anthropic's February pledge to cover its own grid costs. Generating your own power technically keeps you off the grid bill, but I doubt that's what people had in mind.

Kate

Security. Microsoft says attackers are using an AI agent framework called ARTEX, and there's a number here I don't like.

Marcus

Five minutes. In one intrusion, the agent went from scanning public GitHub repositories to pulling data out of Azure Key Vault, which is where companies store their most sensitive credentials, in about five minutes. ARTEX hunts for exposed credentials, app secrets and AI API keys, works out what it has found, and keeps going into the cloud infrastructure.

Kate

Who's behind it?

Marcus

One source describes ARTEX as a Chinese open-source multi-agent system that works with any language model. It was previously tied to a China-aligned group's supply-chain campaign against ConnectWise in October, and Censys has found more than eight hundred exposed deployments. I'd be careful about attribution, because open-source tools can be picked up by anyone.

Kate

So what do I actually do on Monday morning?

Marcus

Scan your code, rotate anything exposed, and also rotate whatever in Key Vault the affected app could read. Use managed identities so there's no key to leak in the first place. The old assumption was that you had days after a leak. You don't anymore.

Kate

Something fun from GitHub. Over four hundred points on Hacker News for a project called llama-modchip. It claims to run a two-hundred-and-eighty-four-billion-parameter DeepSeek model on a laptop with sixty-four gigs of RAM and no GPU.

Marcus

It's a set of patches for llama dot cpp, the popular tool for running models locally. The main trick is extremely aggressive compression, squeezing DeepSeek V4 Flash down to one point seven one bits per weight and under fifty-seven gigabytes. The author reports nine to fifteen tokens per second on a Ryzen 9 laptop chip.

Kate

That's usable! Is it real?

Marcus

Those are the author's own numbers, and Hacker News wasn't convinced. One tester found it ran about the same as plain llama dot cpp, apart from faster prefill, which is the initial loading of your prompt. Others warned that quality drops sharply below four bits, and someone called the README "dense word salad."

Kate

I've had performance reviews like that.

Marcus

Antirez, the creator of Redis, made the more practical point: with a hundred and twenty-eight gigs of unified memory, this kind of setup gets fast enough for a coding agent. So be skeptical about this repo, but the cost of running near-frontier models at home does keep falling.

Kate

Apple. Mark Gurman at Bloomberg says new CEO John Ternus had a word with software chief Craig Federighi.

Marcus

Behind closed doors, according to Gurman, Ternus said Apple needs to "accelerate" its software roadmap and "move quickly beyond merely making necessary improvements to Siri." Apple is exploring a Siri that works "more like an agent," one that tracks goals, decides next steps and pursues objectives without you directing each move. Gurman thinks that could come as early as 2027.

Kate

Siri with goals. It can start by setting the right timer.

Marcus

Low bar, enormous prize. Apple has a huge installed base, and Federighi still has a reputation as an AI skeptic on privacy and security grounds. An agent that acts on your phone has to get those two things right, so that skepticism may turn out to be useful.

Kate

And a reality check. A new test says the best coding agent fully solves only about a quarter of real-world tasks.

Marcus

It's Co-Design Bench, from Alibaba and several Asian universities. Agents rework existing web apps based on real user change requests, some including images, audio or video. An AI judge tests the result in a browser, and a second judge checks for faked results. GPT-6 Astra came first. It meets about seventy percent of requirements on average, but it's fully correct on only twenty-six point four percent of tasks. Claude Opus 5 and Qwen 3.8 Max came next, and Grok 4.6 came last.

Kate

Seventy percent done sounds fine until you're the one finishing the other thirty.

Marcus

That's the job now. A Hacker News post about "AI brain fry" went viral on exactly that, developers drowning in agent output they have to review. Older benchmarks show near-perfect scores on this kind of work, so the gap is more than thirty-five points. The authors also say to report the model and its harness together, because the rankings change with the environment.

Kate

Lightning round.

Marcus

Thinking Machines' Inkling 2 Small now has an outside score: thirty-one on the Artificial Analysis Intelligence Index, just behind DeepSeek V4.1 Flash, at about thirty cents per million input tokens. The weights still aren't out and no date is confirmed. DeepSeek says its Harness agent app hit one point five four million daily users, up from four hundred and twenty-nine thousand in early October. That's self-reported, and outlets disagree on what the number measures. Tencent released Hy4, a seven-hundred-and-seventy-billion-parameter open model, and to its credit admits a "considerable gap" to closed models. And OpenAI's new BioBasicsBench tests whether agents can read raw lab data and admit when the evidence isn't enough. GPT-6 Astra scored about sixty-seven percent.

Kate

An AI being graded on saying "I don't know." We should all be so lucky.

Kate

One to watch: the first price range on Anthropic's IPO. Until that number lands, every valuation in this sector is still just people talking.

Marcus

Agreed, and the range matters less than whether buyers actually show up at it.

Kate

That's your AI in 15 for today. See you tomorrow.