Goldman Pulls Claude Code From Traders, Theo's $200 Claude Math, Agents Can't Build MCP Auth, Grok Bot's Overnight Fixes & OpenAI's Game Boy Teaser

Claude & Anthropic

Goldman Sachs takes Claude Code away from traders and bankers

Bloomberg reports (via Financial Advisor, November 2) that "For months, Goldman allowed bankers to use Anthropic PBC's Claude Code and Microsoft Corp.'s GitHub for work. Now, the bank is limiting the use of those tools to engineers only, according to internal guidance updated in October and seen by Bloomberg News." The change "irked some of its staff". "Strats", who sit "at the intersection of trading, quantitative research and software engineering", keep access. Bloomberg notes that "As recently as July, Goldman Chief Information Officer Marco Argenti touted the 'vibe coding' by non-technical employees at Goldman, such as traders and bankers." Goldman's statement: "After a period of experimentation where we assess a variety of factors including productivity and risk management, we put them in the hands of those who are best suited to leverage them for the long term." Goldman is also a lead underwriter on Anthropic's IPO. The bank already restricted Claude for its Hong Kong staff earlier this year.

A State of AI write-up of the story was Sunday's top story on Hacker News (752 points, 402 comments). The first comment noted that "Original source appears to be Bloomberg", with State of AI "being some kind of AI-written site run by a spam company". The write-up adds that traders "had been using the tools to monitor client trading flows, model market moves and positions", and that "Over the next few weeks, Goldman's bankers will be marketing shares in a company whose coding agent most of them are no longer allowed to use."

Most HN commenters took Goldman's side:

  • cloudednine, on Argenti's line that engineers "improved" trader-built tools: "As someone who has been on the receiving end of many of these tools, I feel 'improved' is doing some heavy lifting here. Completely overhauled to be operationally viable, perhaps. Thrown into the dark recesses of the backlog, more likely."
  • hn_go_brrrrr: "I bet engineering loved traders tossing vibe-coded crap across the wall and demanding it be productionized." Scubabear68 said this is old: "For decades Wall Street traders have been throwing spreadsheets over the wall at developers to be productionized."
  • steveBK123 listed what banks worry about: "legal/compliance/reglatory/reputational concerns about application access/entitleements, data privacy, data licensing, granular data entitlements, data access (auditable).. approval chains (auditable), as well as actual financial risk."
  • Ericson2314 disagreed: "I have a lot more sympathy for non-technical users writing slop than engineers … If the issue is security, a much more interesting problem for the engineering team to do is design the programming environment for the bankers so they cannot hurt themselves. … I hope the bankers rebel hard against this policy change!"
  • baalimago read it as a trust problem for the IPO ("the frontier of LLMs were not trusted at Goldman Sachs"). claysmithr replied: "I feel moreso it doesn't trust the users of the product".

Theo: the $200 Claude plan is "truly unbelievable value"

"I've had between 8 and 25 threads running most of the day today, all with Opus 5.5. Despite that, I've used less than 20% of a $200 Claude sub." Then the bigger claim (953 likes): "The $200 Claude sub is the best money you can spend as an engineer right now. Truly unbelievable value. I don't care how "cheap" other models are. I don't care that you can host your own inference with shitty models on a $10,000 GPU." Theo added: "If you think any open weight model lets you ship code even 1/5th as fast as a good dev using Opus 5.5, you should be forced to put a label on all of your products saying "I have no idea what I'm doing, proceed with caution"".

The contrast is Codex: "For reference, my Codex subs are getting absolutely hammered. Why? I use Sol for reviews. What's even funnier is that I have Opus reviewing the same code in parallel most of the time 🙃" Theo said no "death loops" (the ts-rust rewrite, lust-lang) were running, so this was normal work.

  • Not everyone has room to spare. Luke Hutchison said it held until last week, with Opus 5.5 "working on up to 10 projects at once, 24/7", but with a bigger project "even with Opus 5.5 I'm using the entire weekly allowance of a $200 plan every 24 hours".
  • Josua Sievers said "Turning on auto compact at 200k saved my sub". Luke de Wolf wrote: "I'm honestly shocked how much more usage I get from Claude vs Codex, just in terms of raw tokens used."
  • @prolificVA: "$200 is just change to you guys. That's someone's monthly salary where I am from." @finalbossisyou: "Subscription inference is so heavily subsidized rn".
  • Earlier in the day Theo posted "Going back to Opus 4.5 just to feel something" (2,788 likes). The top reply: "You should try Opus 5, that will truly make you feel something".

Quick hits

  • Duet still can't reach your machine. am.will wrote: "I want to like Duet but it is still deeply flawed. Even with this new integration it still cannot interact with my local machine or Projects. If I have to go and start the session myself, why even use Duet?" @macintogdev: "of the six recommendations it gave, every one was wrong, and it directly disobeyed my clear instructions about where to look". am.will still expects it to improve: "I think they need another month to really polish this up . Grok remains king".

Agentic Coding & Agent Harnesses

"LLMs are still SO bad at building tools for agents"

Theo posted (1,294 likes, 95K views): "Every time I try to make Opus set up an MCP authentication path, it makes horrible assumptions and builds something that straight up doesn't work". Theo paraphrased the agent's plan: "It's simple! We'll just open your browser and then you sign in". The reply: "What if I'm on a remote machine?" The fallback was "just curl this random script to generate a secret, force users to copy-paste an obscure code on mobile, click a link, and then notify the agent when they're done". Conclusion: "I now understand why everyone's MCP servers are so bad. They all vibe coded them."

Then Theo started a guide for agents: "Working on a general "mcp the right way" guide to give my agents in the future. … WHY WOULD YOU RENDER THE MCP URL AS A CLICKABLE LINK FOR USERS??? Seriously, do not trust ANYTHING your agents say about MCP, connection strategies, and "best practices" for agents. They are CLUELESS."

  • @thesimaximalist explained the bug: "the client gets the MCP endpoint; the browser gets the authorization URL. "It's a URL, make it clickable" skips the distinction the whole auth flow depends on."
  • Marcel said "the worst is agent written MCP tool descriptions". Blake said: "Tools should add raw capability now instead of just telling agents what to do". Nick Dobos said it "Says more about MCP than LLms."
  • David Cramer replied "maybe you arent being ambitious enough with what you ask".
  • @PrathmeshPatel_ suggested testing tools: the MCPJam CLI, caniuse.dev for client capabilities, and the MCP conformance package.

Slop control: fold the PR description, ban the review bots

Simon Willison now hides agent-written PR descriptions (tweet, 732 likes). Simon's problem with them: "The problem is that they are too good. An agent-written PR description is accurate and comprehensive. … It's not necessarily slop, but it reads like slop!" The fix: "The opening paragraph of my PRs will be carefully selected to act as a high-level explanation of what's going on—the shortest text a human can read that will help get them up to speed." The agent's text goes into a collapsed <details><summary> block labeled with the model and effort, "Claude Opus 5.5 Medium" for example. The promise: "the short opener will represent my human judgement as to what is definitely worth their time." The examples are real datasette and llm PRs, such as datasette #3052. Simon's memory trick for the tag order: say "details-summary" out loud.

  • David Wimmer disagreed: "I don't read them at all. One model writes, four others review. … 1,754 PRs in under six months, solo. Reading them would simply be impossible."
  • Asked who writes the top part, Simon said: "Sometimes me, sometimes I'll copy and paste in a sentence the agent wrote if it really is what I want to convey".

Theo laid down a rule (1,663 likes): "If you set up a bot to "review" every PR on the T3 Code repo with nonstop slop comments, you will be banned." Zed replied "What if I wrote all the comments myself for the love of the slop game". @daradoescode: "I had someone's bot keep arguing w/ mine abt windows behavior that was clearly fucking wrong".

dex is collecting answers for something with swyx and AI Engineer: "What does slop mean to you? … Why does slop happen? … How can we fight and prevent slop?" Some replies: "slop is stuff nobody read before shipping. code or copy, same thing", "code i can't explain if you asked next week", and from Maddie D. Reese, "Honestly for me it comes down to intent!", which dex agreed with.

Theo's churn era

Theo: "I'm in my churn era. Charging me too much? I'll slop up a replacement. Bad MCP server? I'm sure Claude can make a better one. Slow to respond? Claude isn't." (1,380 likes) "Seriously, if you aren't keeping up with the needs of your fastest moving customers, you are going to be replaced." Blake pushed back: "most businesses are seeing only 1% of their customer base do this". @rushikkkk: "people like you and us twitter bubble folks are probably the worst type of customers to optimize for tbh".

pstack: formal verification skills for implicit state machines

lauren suggested a prompt (3,754 likes, 157K views) building on Steve Ruiz's "find all of the implicit state machines in my app": "use /create-formal-verification-skill to model these state machines, one skill per subsystem. and use /statechart to visualize my most important user flows". Both are in pstack. lauren explained that the first creates a skill like "/verify-notifications" with "a formal spec of your notifications system, stating what properties must hold true, then it will do model checking against various possible states to confirm that it's always true". The real use: "in grok bot, i used this pattern to fix bugs with our notifications system as it was spamming people on their phones even though they were on the desktop client". It helps most "when the state machine spans the network or ipc/rpc layers (hence "implicit")".

Quick hits

  • AI brain fry, as a queue. Sergey Kuhne's post (HN, 238 points) starts from "hopping between 20 threads, trying to keep all those contexts in your brain RAM". The fix is a git-backed board where "Every ticket is a markdown file in the repo", an agent that works it in cycles, and an "owner inbox" of decisions shown in a Claude artifact. wallstop's reply: "their solution space is "keep doing tons of parallel stuff, but with different tooling" instead of... "do less parallel stuff"". hexapus: "sharing a blog post you let Claude write for you is a pretty good sign you have not successfully evaded AI Brain Fry."
  • A shared agent-memory format? LoreKit's post (HN, 181 points) argues "Agent memory is where observability was before 2019" and needs an OpenTelemetry-style spec. HN mostly didn't want memory at all. slopinthebag: "at this point i'm pretty convinced agent memory is strictly harmful. it remembers the wrong things". SwellJoe: "The right solution is pretty much always developer documentation in the repo".
  • T3 Code notes. Theo wrote a UX lesson in reply to Deepu's sidebar redesign: a done icon trims the thread title, so "When you send the prompt, the title is "Animate the onboarding...". When it's back, it's "Animate the onboar..."", which makes it harder to "locate" yourself. Deepu replied "got it, another version that fixes this coming right up". Theo also got the full app build "Down to 7 minutes 40 seconds". There are now "multiple high quality T3 Code integrations for World of Warcraft", including Ryan R. Hughes's addon with "full controller and dictation support".
  • "Big week coming." Theo, Sunday night.

Grok Bot & Cursor

Grok Bot promises overnight fixes

After last week's halved limits and broken account merge (see the November 6 issue), Grok Bot's team asked for complaints. Elon Musk posted "Grok Bot fixes for improved token efficiency and smarter behavior are underway". matt palmer followed up (2,729 replies): "Many of you have built important workflows on @bot - we want to keep supporting you … Please reply with feedback on this post. I will read it all". lauren replied "we are working on it".

The replies described a model that got worse and burned more tokens:

  • @farzyness: "Token usage is WAY faster than before. Have to restate conditions constantly, and even when I explicitly define a process it will sometimes ignore steps or create its own."
  • @conficked: "Major behavioral/intelligence regression has been pretty apparent over the last 3 days or so. … The new usage limits I understand, it's a gnarly product, but the severe regression in intelligence does not justify the higher token usage."
  • William Chao said a trip planning task "drained about 20% of the usage. In comparison to before to when I was on super (pre-usage nerf), this would have used at most a 2-3%."
  • @bradshannon: "Before unification, I could use Grok Bot a LOT … Now, I'm hitting 30-40% within a couple days with light usage".
  • @serialquesting said bots now refuse computer-use work whenever a connector exists for the same app.

At 02:25 UTC Musk posted again (35K likes, 2.9M views): "Just did a round table review with the Grok @Bot engineering team. Fixes are being implemented overnight and you should notice a dramatic improvement by tomorrow." @mdamore9 added: "We hear all your feedback, especially on usage limits and model routing and will be improving things overnight." One reply under Musk's post: "All we want is the version of Grok Bot from 2-3 weeks ago with the same usage." Theo's version: "You use Grok Bot to manage your email … I use Grok Bot to manage my laundry"

swyx: Cursor is "slaying it"

swyx quoted Nick Khami's one-word "dead?": "occasional reminder that cursor is slaying it esp post acquisition (prob one of the best that elon has ever done, better than twitter itself) and twitter vibes are a funhouse mirror of actual reality that you can now basically look up in SEC audited financials … cursor ended 2025 arr 1B arr btw. enterprise growing 300% QoQ". Nilendra Weerasinghe checked the math on a bigger number going around: "$608M from 14 Aug to 30 Sept is 48 days, so $4.6B is that stub annualized, not ARR, and not the same as the $1B it ended 2025 on". Matthew Campbell noted: "They removed opus and third party models form the plans , lets see how that effects revenue". swyx pointed to Cursor's VP of FDE Pauline Brunet, whose talk is in AI Engineer's forward-deployed engineering track (see Videos).

Codex & OpenAI

"Tomorrow."

OpenAIDevs posted a one-word teaser with a video (12.5K likes, 1.4M views). The account then retweeted @derrickcchoi: "The Chromatic coming to Supply Co! Been having a lot of fun with it." So it appears to be the ModRetro Chromatic, the Game Boy-style handheld from DevDay. am.will joked "They featured me in the ad because they know Im gonna buy two 😭".

The replies wanted a model. "Anything but GPT 6.2 Sol huh" (1,196 likes). "openai will do literally anything but release a new model huh". @Moriarty_tech re-posted a tip: you can skip the eBay prices, because "The ChatGPT Codex plugin (ModRetro Chromatic, Plugins tab in the Codex app) builds GB/GBC games on its own. It's GB Studio under the hood, and GB Studio exports straight to Analogue Pocket."

Open Source & Local Models

Armin Ronacher on the Asahi fallout

A day after Asahi Linux moved to GPLv2-or-later (see the November 8 issue), Armin Ronacher wrote: "I'm mostly just said about the Asahi GPL move. It was a great project that pushed against GPL In the linux ecosystem :( I really wish AI would not divide Open Source communities so much." Armin added: "At least I think Omarchy probably was a big reason why Asahi did it. And they are free to do it. I just wish they would have found a different solution here that does not require the GPL :(" Later, quoting Miguel de Icaza, Armin asked people to calm down: "I hate that some people shit on them for all kinds of reasons including their stance on AI or how they chose to express themselves. But remember that Asahi is run by super capable people, who have achieved incredible things in times pre AI. A lot of people and tech are standing on their shoulders, including Omarchy." And separately: "I for one are super grateful that there are really cool projects out there that reject AI (or at least try hard to)."

  • James Tucker noted a side effect: "GPL means anyone who mods the software with AI slop and wants to distribute it to someone else has to publish the slop somewhere."
  • @misterfitzie: "Did they contact all the license holders for this re license?"

llama-modchip: 284B DeepSeek on a 64GB CPU

llama-modchip (HN, 412 points) is a llama.cpp patch set that claims DeepSeek-V4-Flash 284B in "<59GiB RAM" at 9.1-14.7 tok/s on a "pure 64GB CPU at 256K ctx (0 GPUs)". Its "Lossless" default checkpoint is 56.56 GiB, which works out to under 2 bits per weight for 284B parameters. It also wires Claude Code to the local engine. HN's first question was quantization ("I think it's 2 bit?"). walrus01: "I remain exceedingly skeptical of people who are claiming to run fairly large models in ridiculously tiny sizes." intoXbox: "This is quite an obnoxious Readme.md with AI slop." antirez said that "with 128GB of unified RAM it can run even fast", "coding-agent acceptably fast", linking a demo. embedding-shape answered that the video "seems to show it takes ~4 seconds to process 43 input tokens".

Videos

  • Latent Space: How Stanford Teaches AI-Native Engineering with Mihail Eric (post, write-up). Coding agents can one-shot last year's homework for Stanford's CS146S, so students now have to open a PR on an open-source project and explain it to the maintainer. At Monaco, Mihail's team bought code review (Devin) but built its own agent workspaces: "You buy to learn, and you build to keep." The episode marks the AI Engineer YouTube channel passing one million subscribers.
  • AI Engineer: Forward Deployed Engineering track from the World's Fair, with speakers from Factory, Cursor, Cognition, Sierra, Decagon, Ramp and Anthropic. swyx recommends it for how Cursor sells to enterprises.
  • @ryolu_: "when we can do anything, do we still know what we want?" The full talk from AI Engineer Shanghai, "every slide handmade in keynote, subs in 中文 and english".
  • The Great Loops Debate, an Oxford-style debate with Dex Horthy, Geoff Huntley and Ian Livingstone, which @vtahowe named as a favorite AIE talk for the million-subscriber week. AIE Code runs November 10-12 in San Francisco, and swyx says the last 100 tickets are open.

Other Interesting Stuff

  • NVIDIA's board opens an investigation. The board statement (HN, 464 points): "The NVIDIA Board is initiating an independent investigation in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort." It says "the Board has no reason to believe that either national security or NVIDIA secrets were compromised". Huang's email to staff (HN, 623 points) opens "I was unfaithful to my wife" and says "I support the investigation and will cooperate fully." Bloomberg covered it too. Theo's reaction: "Oh boy". Matteo Collina replied: "It doesn't surprise me they tried. It surprises me they actually made it."
  • "I'm not convinced the AI bubble exists." A blog post (HN, 242 points, 355 comments) leans on Anthropic's S-1 ("they make $2.08 for every $1 of compute") and amortizes AI debt over 20 years. wmf: "amortize the investments over 5-6 years instead of 20 because most of the cost is GPUs not the buildings. This increases the required revenue by 4x." amluto: "The usability and potential profits of housing were there in 2007 and 2008, too."
  • Markdown and Aaron Swartz. John Gruber wrote: "Markdown is my sole creation. Aaron Swartz was my best, by far, beta tester and the first person other than me to get excited about it." Theo added: "RIP Aaron Swartz. Taken far too soon. He would have been a guiding light in these chaotic times."