Week of October 18, 2026
Greg Brockman's case that security spending should dwarf software spending now that AI makes code nearly free, plus why Anthropic's slowing growth and CoreWeave's distressed bonds mean investors will scrutinize AI bets much harder.
This Week's Top Videos
The Changing Physics of Cybersecurity and Drug Development with OpenAI President Greg Brockman
By No Priors Podcast
Greg Brockman argues security spending should vastly outstrip software spending, because AI makes writing software nearly free while making reliable, trustworthy software scarce. He also points to OpenAI publishing nearly 400 resolved open math problems as proof that AI now generates genuinely new knowledge. For builders, security and verification are becoming the real moat.
- As AI pushes the barrier to entry for software toward zero, the value shifts to software that is reliable, secure, and trustworthy. Brockman claims cyber and security spend should vastly outstrip software spend, and says most people have not yet internalized how much the physics of cybersecurity will change.
- OpenAI published nearly 400 important open math problems that its models resolved, and mathematicians are now combing through them for insights. Brockman says the question of whether AI can produce new knowledge is now definitively answered: yes.
- Brockman draws a clear line on current capability. Models can now crack hundred-year-old math problems but are not yet inventing whole new theories. Founders should calibrate science-AI bets to problem-solving, not paradigm creation.
- Unlike the brute-force, machine-checked four-color theorem proof, these AI proofs contain surprising, human-legible ideas. Domain experts report unexpected applications of existing results, which points toward a new phase of machine discovery that produces real insight.
- The opportunity is in domain-tuned intelligence that amplifies human leverage. The same capability that solves math problems can be applied to biology or to running a small business, putting previously unthinkable results within reach of individuals.
- As models develop better theory of mind, they can guide humans along the path to new understanding rather than handing over opaque answers. This makes AI-powered learning and explanation tools a major underappreciated product category.
Anthropic’s IPO Momentum as CoreWeave Rattles AI Trade
By Bloomberg Technology
Anthropic told prospective IPO investors its Q3 revenue hit $16.8B, but quarter-over-quarter growth slowed to under 50% from closer to 100%. That spooked markets. Meanwhile CoreWeave's 2032 bonds trade at distressed 15% yields on $50B+ of debt. Founders should expect capital markets to scrutinize AI growth durability and debt-funded infrastructure much harder now.
- Anthropic reported $16.8B in Q3 revenue to prospective investors. Quarter-over-quarter growth fell below 50%, down from closer to 100% the prior quarter, and the market read that deceleration as disappointing despite the company's scale.
- Anthropic disclosed an 85% gross margin and 900% net revenue retention, meaning existing customers now spend roughly 9x what they did a year ago. Investors are focused on whether AI labs can generate free cash flow and real return on invested capital.
- An attendee at the investor meeting flagged OpenAI's Codex push as a competitive headwind for Anthropic's coding revenue, though they expect growth to reaccelerate in Q4. Even category leaders face pressure when a rival targets their strongest product line.
- Anthropic ran about five hours of programming for analysts and investors at its headquarters, a standard pre-IPO playbook. Bloomberg's reporting points to a public filing soon, with formal marketing potentially starting the week of November 9.
- CoreWeave's 2032 notes yield around 15% and its credit default swaps trade at distressed levels. The company carries over $50B of debt and has interest payments running ahead of operating income, a sharp fall for a firm that was pursuing an investment grade rating.
- The bigger risk for debt-financed AI buildouts is the structural rise in long-end rates and term premium, not a single near-term Fed hike. A hot CPI print amplified that fear, and the AI trade has shifted from FOMO to skepticism.
I Tested Gemini 4 vs Opus 5.5 vs GPT-6 Astra vs Muse Wonder (5 Use Cases)
By Peter Yang
Opus 5.5 swept Peter Yang's code-only creative tests, winning the Blender castle cinematic and the one-shot Star Fox-style game. Gemini 4 finished a clear second by cramming in extra features like first-person view and battle cruiser summons. Meta's Muse Wonder looked generations behind. Builders choosing a frontier model for visual or creative coding should start with Opus.
- On the Himeji Castle test, Opus 5.5 won. The prompt asked each model to build the castle in Blender and render a 30-second dawn cinematic with music, entirely in code. Opus produced travel-commercial quality shots with accurate architecture. The final ranking was Opus, Gemini 4, GPT-6 Astra, then Muse Wonder.
- Gemini 4 overbuilds to impress. In one shot, its Starcraft-themed rail shooter added a first-person view (V key), battle cruiser summons (B key) and barrel rolls. That made it less polished and more crowded than Opus, but it was the most feature-rich.
- GPT-6 Astra nails mood and atmosphere, such as water reflections and ambient bird sounds. It falls behind on fidelity and depth. Its castle looked flat, and its game had only one enemy type with plain backgrounds.
- Meta's Muse Wonder performed far below the other three on 3D and games, looking like output from a model several generations older. Meta's Chief AI Officer Alexandr Wang (spelled 'Alex Wang' in the transcript) has said it was not trained for 3D. Treat it as a candidate for other use cases, such as personal agents, rather than visual coding.
- Use a personal benchmark prompt you can judge well. Yang reuses a Star Fox clone for every new model and remixes it with domain knowledge he knows deeply, like Starcraft Terrans versus Zerg. A repeatable, taste-driven test reveals model differences faster than public leaderboards.
- For animated storytelling shorts, the winner depends on taste. Opus led on graphics and animation and drew the most emotion from Yang. Gemini 4 was the most artistic. Astra's robot story was charming but had confusing visual moments.
Figma for Edu: Working with typography systems
By Figma
A type scale exists to remove choices, so nobody wastes time debating 64pt versus 65pt. Figma's Edu team shows how variables feed composite text styles (Display, H1 to H5, body, caption) that swap whole brands or breakpoints through Appearance modes. This matters more now that AI generates layouts you need to judge and correct.
- A typographic scale is a set of constraints. Limiting font size options stops teams from debating 64pt versus 65pt and keeps a brand consistent when many people design with it.
- A type system covers more than font size. Define the font family, size, matching line height, weight, purpose of use, and letter spacing or tracking for every style.
- Name styles after their HTML function: Display, H1 through H5, body, and caption. This maps design directly to code and gives the structure room to grow as new screens and modalities appear.
- Bind font properties to variables, then bundle them into composite text styles. In the Appearance panel you can switch modes to swap an entire type scale between brands, such as a serif brand and a sans serif brand.
- The same variable modes can resize type for desktop, tablet, and mobile. One set of styles then handles responsive scaling, so you don't need duplicate component sets for each breakpoint.
- Type literacy matters more as AI builds layouts. Designers need to judge things like measure (the width of a block of body text) and bake typographic rules into systems so generated output stays correct.
Better Work Won’t Get You More Clients. This Will.
By The Futur
Scaling happens with scissors, not addition: cut non-essential, unprofitable offers before adding new ones. Sell clients the six-pack abs, not the nutrition plan, then deliver what they actually need. Map your content to Eugene Schwartz's five awareness levels, and attend one industry conference to learn in 3-4 hours what would take 6 months.
- Sell what clients want and give them what they need. Market the outcome (more leads, customers, profit), then fix the CRM, messaging, website, and hiring gaps once they are working with you.
- Use Eugene Schwartz's awareness levels from Breakthrough Advertising (unaware, problem aware, solution aware, product aware, most aware) to split content by audience stage. Skip the unaware entirely.
- Attend your target industry's conference to compress research. In 3-4 hours at a pharmacy conference in Dallas, Alan Moore learned what would have taken 6 months: their fears, hot-button issues, and where they make and lose money.
- Choose hunting (cold outreach, gifting, targeting 10 CEOs) or fishing and farming (nurturing a market with content) based on honest self-awareness. Most gentle, creative introverts are not hunters, so hunt only during lean seasons.
- Scaling happens with scissors. Before adding low-ticket products, memberships, or masterminds, cut the offers that are neither fun nor profitable and double down on the essentials.
- Deliver quick wins early in an engagement. Clients need a feeling of momentum and progress, because very few people tolerate a long stretch with no payback.
What's new in Grok Bot - Oct 13, 2026
By Replit
Grokbot now gets its own email inbox, a step toward treating AI agents like employees that sign up for services and digest newsletters for you. This week's release also adds free, zero-setup X monitoring with no developer API key, a Shopify integration, a faster higher-resolution agent computer, and five new languages.
- Grokbot can provision its own email inbox when you ask it for a handle. It can then sign up for services, subscribe to newsletters, and summarize purchases and updates you want out of your personal inbox. Handles are first come, first served, so claim one early.
- X search and monitoring now works with zero setup and zero cost. You don't need to connect an X account or get a developer API key, and it no longer burns developer credits.
- The new Shopify integration lets store owners track orders, check store status, and query metrics conversationally through the agent instead of the Shopify dashboard.
- Grokbot is an agent running its own computer, and that computer is now faster with a larger display resolution. Screenshots and visual outputs come back at higher fidelity, which helps with screenshot-heavy and recording workflows.
- Email draft cards now include formatting previews and a sender dropdown. You can check that emails look professional and pick which alias they go out from before sending.
- Smaller updates include in-chat search of bot templates and the marketplace, PR hover previews for engineering work, and a right-click menu for routines. Five new languages were also added: Marathi, Tamil, Telugu, Swahili, and Hausa.