📅 ThursdAI - Apr 30 - DeepSeek V4 (1.6T MoE), Cursor SDK Wins WolfBench, Mayo's REDMOD Saves Lives, Stripe Gives Agents a Wallet & more

Hey everyone, Alex here 👋

Tomorrow is May. May! I genuinely cannot believe we’re four months into 2026 already, and the AI news cycle is showing zero signs of slowing down. This week’s show was a wild one! We opened with what is genuinely one of the most important AI stories I’ve ever covered (Mayo Clinic AI detecting pancreatic cancer THREE YEARS before human radiologists), we covered the return of the Chinese whale with DeepSeek V4, OpenAI got caught in their own system prompt begging GPT-5.5 to please stop talking about goblins, and I literally gave my coding agent a credit card and asked it to buy my fiancée a wedding gift with the new Strip Link skill and CLI!

Oh yeah, I’m getting married next Tuesday! 💍 So next week’s show will be a little different. I’ll be back the week after to catch you up on whatever drops in my absence (almost certainly something major, knowing this industry).

Lots to get through, so let’s dive in. (also, in the end I have a full month recap of every major launch, don’t miss)

ThursdAI - Highest signal weekly AI news show is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Mayo Clinic’s REDMOD: AI Detects Pancreatic Cancer 3 Years Early 🔥 (X, Blog, Announcement)

I know we usually cover Models, Parameter sizes, MoEs and big copmanies. But this is important. This is the use case that justifies the entire AI revolution, the GPU burns, the buildouts. I want humans to WIN, and Cancer to be fixed!

Mayo Clinic just published a study in Gut (BMJ) validating an AI model called REDMOD that detects pancreatic cancer on routine CT scans up to three years before clinical diagnosis. The numbers are jaw-dropping: They show 73% sensitivity for catching prediagnostic cancers, compared to 39% for experienced human radiologists (while looking at the same exact CT scans).

And maybe the most important bit, at scans taken more than 2 years before diagnosis, the AI catches nearly 3x as many cases as specialists

For context: pancreatic cancer has less than 15% five-year survival specifically because 85% of patients are diagnosed after the disease has already spread. This is the cancer that took Steve Jobs. Imagine if Jobs had access to this AI three years before his diagnosis. That’s the impact we’re talking about.

As Dr. Ajit Goenka from Mayo Clinic put it, the greatest barrier to saving lives from pancreatic cancer has been the inability to see the disease when it’s still curable. This AI can now identify the signature of cancer from a normal-appearing pancreas.

Even better: it runs on CT scans people are already getting for other reasons. No extra screening protocol, no new imaging required. Just smarter analysis of existing data. The model also showed remarkably stable performance across institutions, imaging systems, and protocols, with 90-92% test-retest concordance over serial scans.

Mayo Clinic is now moving this into prospective clinical testing through a study called AI-PACED (Artificial Intelligence for Pancreatic Cancer Early Detection).

When we say “lets f*****g go” that’s what we mean. Yeah getting more intelligence is cool, but I want a world without decease! Let’s F*****g go mayo clinic!

Agentic Commerce - Giving OpenClaw my credit card - safely!

Stripe Link Wallet and Infrastructure CLI (X, Announcement, Blog, Announcement)

Ok, give an LLM your credit card, what can go wrong.. right? Well, it’s clear that this, increasingly, is the future of commerce. Agents will be shopping for us, and we need solutions here. Well, this week at Stripe Sessions (Stripe’s annual product lineup conference) just delivered.

Link Wallet, is a new ... API? CLI? Skill? Definitely a skill, for your agents, to connect with your Stripe Link (the thing that stores your credit cards safely) and then giving your agent a budget, it can go and make purchases in your behalf. Now the trick here, is, every purchase, you get a notification to approve, and the agent never sees your actual credit card number! This I think is the biggest win here.

To test it out , first, I showed Wolfred the install instructions, which are literally this:

Read link.com/skill.md and get me set up with Link

And then I asked Wolfred my OpenClaw assistant to buy me a present of its choice for my upcoming wedding, and that I don’t want to know what the present is, but I can approve the spend!

OpenClaw installed this, sent me a link to connect to my Link.com account, I also downloaded the Link app to receive notifications (and had to enable them by hand, it was a bit annoying to discover, but they said they will fix the onboarding) and .. voila, my agent can now go spend my money, and I get these approval notifications:

The kicker? The present Wolfred sent us is due to arrive like 2 months after the wedding 😂 But hey, it’s still something! My agent went, chose a wedding gift in budget, asked for my approval to puchase, and filled out the details (asked me for a few of them) and voila, first agentic purchase that did not require my credit card exposed!

Stripe announced a whole bunch of other Agentic Commerce Suite features, like Shared Payment Tokens, which are scoped to seller and protected by Radar, MPP (machine payment protocol) and streaming payments using stable coins that are pretty slick and a bunch of other interesting things. This is where the world is moving to, and Stripe is innovating hard here, definitely worth keeping an eye out on what they are

Speaking of agents and stripe, they also opened up the waitlist for projects.dev - which is a way for agents to provision accounts fully on their own, get API keys, and set everhing up from scratch. I think it’s a wonderful addition to the agentic tools and agentic internet! Your agent just runs something like stripe projects add cloudflare/workers abd boom, you have a workers deployment, with credentials synced, no dashboard clicking or API creation!

Big Companies & APIs

GPT-5.5 Goblin Mode: The Funniest Bug Report in AI History (X, Blog)

Someone on X noticed that Codex system message for GPT 5.5 that launched last week has this interesting addition: “Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query” and it has it two times!

This created a bunch of memes, questions and wonderings about ... why would OpenAI care so much about Goblins. And they finally posted a long writeup on why:

the TL;DR there is, GPT 5.5 absolutely LOVES talking about Goblins, trolls and other nerdy creatures. This is a result of them favoring the “nerdy” personality archetype and reinforcing this reward via RL. OpenAI admitted that “Unfortunately, 5.5 started training before we found the root cause of goblins” and so, now, we get 5.5 that LOVES to talk about goblins, can’t stop talking about goblins (unless they are asked to stop by a system prompt)

OpenAI also posted the exact instructions of how to “unleash“ the goblin mode on the blog, which I find hilarious, a company that leans into the meme is a company to be celebrated 👏

GPT 5.5 is as good as Claude Mythos on CyberSecurity

According to the AI Security institute, GPT 5.5 (not the GPT 5.5 - Cyber version that was announced), the one you have access to, is as good as Claude Mythos on vulnerability finding. We previously reported that Anthropic deemed Claude Mythos as “too dangerous to release publicly” and it turns out that that was either a marketing “Myth”, or Anthropic’s inability to server this huge model like they server Opus.

OpenAI Ends Microsoft Azure Exclusivity

This piece of news sent quite a lock of shock throughout the industry, somehow, Sam Altman and OpenAI have been able to negotiate through the very strict deal with MIcrosoft and now are available in AWS as well as Microsoft Azure! Apparently the AGI clause is now gone as well!

For many startups who are locked into AWS and Bedrock ,this is great news, they are not a…