Welcome, humans.

So apparently, a Reddit user handed Claude access to an “agentic” trading account for a month… and says it lost him $31,000. Somebody call Wall Street Bets!

He claims he posted the result as a warning that autonomous agents making real financial decisions can “go very wrong, very fast.” That said, some Reddit sleuths questioned whether the screenshot was actually real (anything can be easily faked in today’s deep fake era), so treat the dollar figure as a vibes-based claim.

The thread’s consensus was less philosophical: if you MUST let AI into the yen house, best to paper trade first, cap position sizes, and set a hard stop before an agent can torch the account (not financial advice; at least, not from me!).

Will say that one commenter asked whether he remembered to tell Claude to “make no mistakes.” Turns out that was the load-bearing line all along!

Here’s what happened in AI today:

😺 AT&T routed AI work toward cheaper open models.
📰 Nvidia struck a $6B Poolside licensing deal.
📰 AI plus training sped Pakistani judges 6.3%.
📰 Goldman found AI already weighing on jobs.
🍪 ChatGPT plugged Apple Messages into Work and Codex.

…and a whole lot more in today’s full Around the Horn digest.

Hey! We're booking out ad inventory for Q3 and there's only a few slots remaining! Make sure you reach out ASAP if you want to advertise your product and service to our 700K+ readers today!

P.S: want to get ahead of the crowd? Email Garrett here and he'll set you up.

😺 AT&T is routing AI jobs to cheaper open models

If your company sends every summary, code task, and hard analysis to the same expensive AI model, you may be paying a convenience tax.

Turns out AT&T said enough is enough and won’t be putting more quarters into the payphone, and instead is pushing a growing share of internal AI work toward open models: meaning models companies can run themselves, then routing harder jobs to more expensive top-end systems when needed.

Here’s what happened:

@Hesamation claimed AT&T is already internallyy routing 40% of employee AI usage to open models, and that open-model coding cut costs 56% for about a 2% quality tradeoff. TWO PERCENT! ~spits out milk~
The Information reported smart routing cut costs 80–90% on some applications, while AT&T tried to keep OpenAI and Anthropic spending flat.
Routers are a very big deal atm:
- Stripe’s $7B acquisition of O.G. router Openrouter (our fave; try it here):
- Fintech company Ramp just launched its own router, “Router”, which can choose models by cost, test scores, or task difficulty.
- Callosum raised $100M to optimize the model-and-chip combination behind each request.
- And don’t get us started on all the coding agent harnesses launching their own!

Why This Matters: A model router turns the whole “which AI should we actually use?” conversation into a per-task decision vs an enterprise stack one. The pattern is simple: routine work you’ve proved out can go to a cheaper model; difficult or high-stakes work should be escalated to the strongest one for the task.

We talk all about this concept in our livestream from yesterday; check it here!

The catch here is measurement. If your team cannot define what “good enough” looks like on real work, the router is… more like a roulette wheel.

Our advice? If you’re on a subscription plan with GPT or Claude, and its the first time you’re doing something, do high reasoning (high is usually high enough). If you’re paying by the token (via API), start low and see if it can do it. Same logic goes for trying less intelligent but cheaper models; if they can do it, great! If they can’t crank it up a notch.

🎓 AI Skill of the Day: Turn a Chat Into a Live Website

You can turn an idea, draft, or compatible local project into a hosted website without leaving ChatGPT. ChatGPT Sites can create and refine the site, save reviewable versions, deploy a live URL, and add storage, sign-in, analytics, collaborators, or a custom domain. One important detail: deployment URLs are production, so if you want to review first, ask Sites to save a version before deploying. Brent Schooley’s Before the Cut is a live example.

In ChatGPT, include the word “website” in your request or mention @Sites.
Describe the audience, the job the site should do, and the information or features it needs. Ask ChatGPT to save a version first if you want to review it before publishing.
When it looks right, ask Sites to deploy it and give you the production URL. Keep refining the project conversationally.

Copy this:

Build a website for [audience] that helps them [job]. Include [sections/features]. Save a reviewable version before deploying. Once I approve it, publish it with Sites and give me the live URL.

Bonus: so Brent’s GPT site was all about how to edit video with Codex without making it guess. Brent’s field guide uses four questions before the first cut:

What are we making? Lock the audience, story, target length, format, and feeling before asking Codex to edit.
What are we working with? Inventory camera angles, screencasts, audio, graphics, templates, and brand assets. Have Codex flag picture/sound problems, unreadable demos, missing coverage, and private information.
What did they say? Create both normal and word-level transcripts with speaker labels. Codex can’t listen like a human, so the transcript gives it dialogue, exact timing, speakers, restarts, producer cues, script accuracy, and pacing to compare takes.
What can Codex see? Let it inspect footage with ffmpeg/ffprobe, scene detection, and frame analysis to check composition, screen readability, continuity, focus, exposure, crops, motion problems, and visual glitches.

The trick is to give Codex a target, an asset inventory, timecoded words, and visual evidence before asking it to make editorial decisions.

Brent’s full framework goes deeper on each step.

Have a specific skill you want to learn? Request it here.

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

*Stop compromising on AppSec. Checkmarx Fusion combines hybrid rules and AI reasoning to catch complex zero day bugs in AI generated code without the noise.
ChatGPT’s new Apple Messages plugin searches conversations, catches you up, and drafts or sends replies from ChatGPT Work or Codex on Mac.
FLUX Video Upscale regenerates short clips at 1080p, 2K, or 4K, with Precise mode for fidelity or Creative mode for rebuilding fine detail.
GPT-Image-2 now generates transparent-background PNGs directly, so you can create reusable product cutouts, campaign assets, and presentation graphics without removing the background afterward.
Claude Academy teaches Claude.ai, Claude Code, the Claude Platform, AI fluency, and model limitations through Anthropic’s official courses.
Grok Build turns one prompt into a published app, game, website, or dashboard with its own domain, with a coding agent that can use subagents, a browser, databases, secrets, and GitHub export.
Perplexity Agent API puts 41 models from nine providers behind one endpoint, with web search, finance search, fetching, and sandboxed code execution built in.

FROM OUR PARTNERS

Want to get more out of Claude? Build personalized skills (step-by-step guide)

You're paying for Claude but not hitting your full potential. This free guide shows you how to build high-end Claude Skills with no prior tech knowledge required. Your style, your voice, your data, supercharged.

Click Here for Free Instant Access to the Claude Skill Guide

📰 Around the Horn

seriously underrated use case for coding agents: writing code to make your devices work better

NVIDIA struck a non-exclusive $6B licensing deal with AI coding team Poolside, invested another $1B, and offered jobs to 109 employees while Poolside’s founders stayed.
A new study found AI plus targeted training increased case resolution 6.3% in a randomized experiment with 1,559 Pakistani judges, without a clear writing-quality decline or rise in appeals.
NVIDIA also planned a China-focused AI chip using licensed Groq technology optimized for fast model responses, with small-batch shipments potentially starting by year-end if export approvals cooperate.
Meta has quietly become one of Microsoft’s biggest AI customers, reportedly spending hundreds of millions of dollars a year on Azure-hosted models.
CISA, the FBI, and NSA warned about AI-assisted attacks targeting internet-exposed Siemens S7 industrial controllers.
Goldman Sachs found AI is already weighing on employment in developed economies, with the clearest effects in call centers, software publishing, consulting, advertising, and entry-level work.

Want the rest? Read the full Around the Horn digest here.

💡 Intelligent Insights

Makes you (want to) think!

Logan Kilpatrick asked how much AI spend goes to evals; Brendan Foody guessed below 1%, while Aakash Sabharwal argues the hard chain is KPI → task → eval, with signal lost at each handoff.
Ethan Mollick and Nick Dobos argued ChatGPT and Claude now scatter capabilities across different modes, leaving users unsure which workspace has which tools, memory, permissions, and files.
Damian Barabonkov argues AI coding “slop” is increasingly over-engineering: defensive code and elaborate handling for rare or imaginary edge cases.
Ryan Carson is testing agent-era hiring by having candidates record themselves shipping a real feature, then giving finalists 16 paid hours of Devin access to deliver a merge-ready PR on the actual repo.
Ethan Mollick argues even good AI output is becoming monotonous because the same stylistic patterns spread across ads, software, social posts, instructions, and slides; his follow-up says ordinary prompting and sampling tweaks do not create the deeper variation needed for genuinely different ideas.
Claude Code added a Concise output style that leads with the result and stays short by default; Boris Cherny called it a quick band-aid while Anthropic works on a longer-term fix for recent output-quality complaints.

New from The Neuron: AI Explained

Our new interview asks a pretty uncomfortable question: what if we’re spending billions scaling the wrong kind of AI for making predictions on your data?

Neuralk CEO Alexandre Pasquiou explains why ChatGPT can summarize a spreadsheet, yet still lose the signals needed to forecast sales, churn, risk, or demand, how tabular foundation models attack that problem directly, and why he thinks they’ll power every enterprise prediction workflow by 2030.

He also shows how Neuralk’s Seldon can plug those predictions into Claude, ChatGPT, Excel, and AI agents… and you cant try it right now, for free.

P.S: We’re trying to hit 50K subscribers on YouTube this year. Click here to help!