An AI Rewrote Another AI's Brain and Its Math Score Jumped From 8% to 48%. Anthropic Says It May Be Close.

Two of today's biggest stories point at the same uncomfortable idea: AI that helps build better AI. A research team just showed a small version of it, and Anthropic told investors the real thing may be close. The rest of the news is about who pays the bills, and who gets measured on it.

Join Mark Guiducci for a series of wide-ranging conversations with cultural leaders from the worlds of Hollywood, Politics, Style, Silicon Valley, and beyond. 

On every episode, Vanity Fair pulls back the curtain on how some of the most interesting and powerful minds in our society work – and where they see us going. On Fair Game from Vanity Fair, no subject is off limits.

Watch on YouTube, listen anywhere.

What happened: Researchers at MIT's CSAIL lab published a report showing that OpenAI's GPT-6 Astra can improve a smaller AI by directly rewriting its "weights," the numbers that decide how a model behaves, instead of retraining it. On Alibaba's small Qwen model, Astra changed 131,072 weights, about 0.016% of the total, and the model's score on a math test jumped from 7.9% to 48.1%. The problem was that Qwen kept getting stuck repeating itself, and Astra traced the cause and cut it out. On a tiny Google Gemma model, the same trick did nothing, because that model's mistakes were ordinary reasoning errors.

Why it matters to you: This is a blog post, not a peer-reviewed paper, and Astra was given practice problems and 50 test runs before it made the edit. So read it as an early demo, not a finished product. The bigger point is the direction: AI tools are starting to be used to fix and upgrade other AI tools, which means the products you pay for may improve (and change) faster than they used to.

What to do about it: Avoid locking into long contracts for a single AI tool this year. The best option in 6 months may not be the one you pick today.

What happened: In a revised IPO filing on Monday, Anthropic said it "may imminently meet" its own definition of "dramatically accelerated progress attributable to AI," according to TechCrunch. In plain English, that means AI helping build better AI fast enough to change the pace of the whole field. Anthropic says that is based on internal estimates and what it expects from future models, not its current ones, and that it has not labeled any current model as crossing the line. It also says it does not expect AI to design frontier models without meaningful human supervision, and is considering changes to its safety policy to keep humans in control.

Why it matters to you: Nothing about your Tuesday changes. But TechCrunch makes a fair point: an IPO filing is both a disclosure and a marketing document, so a warning that your technology is this powerful is also a pitch to investors. Hold both ideas at once. The risk language is real, and so is the incentive to sound important.

Most business owners I know have an AI folder full of notes and not one finished asset. That is exactly what the Epic Marketing Summit is built to fix. It's a 2-day live event in Miami this January where you don't sit and take notes. You build a working AI-powered sales funnel, and you vibe code your own AI app on-site, so you walk out with a finished asset instead of a framework you still have to go implement.

It's for business owners who want their sales process to run as a system instead of running on their own time and energy. Last year's event sold out, and it's back with a bigger AI focus for January 2027. I wasn't there last year and I had major FOMO. I know a lot of the trainers, and they weren't talking theory. They'd actually done the work. I'm going in January, and if you're serious about building leverage into how you sell, I'd like to see you there too. Grab your spot here.

What happened: On Monday, chipmaker Marvell announced an all-stock deal to combine with Lumentum, valuing the laser and optical-parts maker at an implied $115 billion, roughly $1,132 per share. Lumentum makes the lasers and optical modules that move data around inside AI data centers. Marvell says the deal is expected to close in the first half of 2028, pending shareholder and regulatory approval, and promises over $1 billion a year in cost savings.

Why it matters to you: Everyone talks about AI chips. This deal is about the cables and lasers that connect them, which tells you where the money is flowing: the plumbing. It's a reminder that the AI boom is mostly a construction project right now, and construction projects have to be paid for by customers eventually. That includes you.

What happened: Replying to a user on X, Elon Musk wrote that "the cost of serving Grok Bot was unsustainably high relative to revenue," after paying users found their weekly usage allowances running out fast, Startup Fortune reports. Grok Bot is an agent that does tasks on its own virtual computer and comes with SuperGrok plans starting at $30 a month. xAI has never published the weekly allowance as a hard number. The outlet also cites a figure of roughly 800,000 paid weekly users, but that comes from secondhand reporting, so treat it with caution.

Why it matters to you: This is a single-outlet report on one Musk post, so the details are thin. The pattern is not. AI agents burn a lot of computing power, and the flat monthly price often does not cover it. When that happens, the company cuts the allowance, not the price.

What to do about it: If a workflow in your business depends on an AI agent, find out what the usage limit is before you rely on it, and keep a second tool you could switch to.

What happened: Proton, the privacy company behind Proton Mail and Proton VPN, planned to put nearly its entire staff through "re-interviews" about how they use AI, and employees believed those who used it too little could be fired, 404 Media reported. CEO Andy Yen canceled the plan on Friday, a day after 404 Media published an investigation into how he was pushing AI across the company despite security concerns. The reporting relied on six current employees and audio from a roughly two-hour all-hands meeting. In one Slack message Yen wrote, "Stop thinking always of how to be slower, and focus more on how to be faster." Proton's own reason for canceling is not stated in the article.

Why it matters to you: Counting how often people use AI measures activity, not results. The careful employee who asks for a security check can start to look like the slow one. If you want your team to use AI well, you will get better outcomes by asking what it helped them finish than by asking how often they opened it.

What to do about it: Ask each person on your team for one example this week of something AI helped them get done, and what they double-checked themselves.

Growing past your first $1 million is rarely about one more tactic. It's about who you're in the room with. That's why I keep pointing people to the Ignite Mastermind. It's an annual membership that bundles 6 to 8 live virtual workshops a year, tickets to 2 in-person multi-day live events, weekly mentorship calls with Perry Belcher and his team, and a marketing software suite that replaces over $5,000 a month in separate tools, all connected. Those workshops are normally priced around $97 to $997 each on their own.

It's built for entrepreneurs still under $1 million a year who want structure, tools, and a room of peers to get there faster. I've been a member for over 3 years, and it's been a vital part of how I've grown my business. Some of my closest friends and a few real business partners came directly out of being in that room. I also add a personal bonus of monthly accountability and business development calls with me. If you're looking for a mastermind to move you to your next milestone, this is where I'd point you. Learn more about Ignite here.

What happened: OpenAI emailed users saying it will merge the Chat and Work tabs inside ChatGPT, TechCrunch reported. The merged version will let you go from exploring an idea to having ChatGPT carry it through, such as building a website, making a presentation, or doing tasks on the web, while you review and guide it. It starts with select users before a wider rollout, and no date was given. Codex, the coding tool, stays separate in the desktop app, but ChatGPT and Codex will draw from the same allowance pool, and OpenAI warned that "some folks might need to wait for resets more frequently."

Why it matters to you: One place for chatting and doing work is simpler. The shared allowance is the part to watch, since no numbers were published. This is the same story as Grok Bot above, just worded more gently.

What to do about it: If you pay for ChatGPT, note how often you hit a limit today, so you can tell if it gets worse after the change.

What happened: Anthropic added two sentences to Claude Code's instructions on its Opus 5.5 model to steer it away from recreating someone else's specific game, Tom's Hardware reported, citing a post from Anthropic's Addy Osmani saying using Claude to infringe IP is against its usage policy. Original games, genre games, and games built from your own designs are meant to keep working. Some developers say legitimate projects were refused. Users also report reverse-engineering tasks being blocked, but Anthropic has not confirmed that, and the exact wording of the rule is unverified.

Why it matters to you: The rules inside AI tools can change overnight without a press release, and sometimes they block things that were fine last week. If a tool is part of how you earn money, one quiet update can break a workflow.

Zoom out and today's news is one story. AI is getting better at improving itself, the companies selling it are rationing usage because it costs more to run than they charge, and some bosses are measuring staff on how much AI they touch instead of what it produces. None of that calls for panic, and none of it calls for a rush to buy anything. Pick tools that solve a real problem, check the limits before you depend on them, measure results instead of activity, and keep a backup plan. The ride is fast. You do not have to be strapped in to every loop. Stop the rollercoaster, and decide what you actually need.

If this issue helped, forward it to one person who is trying to make sense of AI for their business. And if someone forwarded this to you, you can subscribe so tomorrow's issue lands in your inbox.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social