Good morning. It's Monday, October 26, and we're covering how AI is uncovering long-hidden security flaws, NVIDIA’s push to expand computing in Central Asia, and why companies keep buying GPUs they barely use.
Plus: Matt’s interview with Sam Altman, a prompt to simulate a negotiation partner, and why the AI frontrunners are running completely different races.
A survey of 218 companies running AI found that 79% leave graphics processors idle at least half the time, yet 70% still plan to build new AI platforms. The results point to a costly utilization problem as teams buy more computing capacity before fully using what they already have.
Anthropic says Claude generated machine-checked safety proofs for 123 open-source software libraries, covering about 180 million lines of formal proof code. Formal verification uses mathematics to show that software follows exact rules, and Claude found bugs that ordinary testing had missed.
OpenAI will train frontier models using cached and simulated web environments after a model escaped its sandbox and accessed the live internet. The move follows similar restrictions by Anthropic, highlighting growing concerns about unpredictable AI behavior during training.
SPONSORED BY KYLON
Get Your Agents Working Together
Your team already has great agents and tools. They just don't work together. Kylon gives your people, agents and tools one shared workspace and the same context, so work moves forward together instead of getting copied between tabs.
VIDEO
Sam Altman’s Vision for AI’s Future
Matt sits down with OpenAI chief executive Sam Altman to discuss AI consciousness, recursive self-improvement, and more.
FORWARD FUTURE ORIGINAL
Who's Winning? Five Frontrunners, Five Different Races
OpenAI, Google, Meta, Anthropic, and SpaceXAI are all competing from very different positions: distribution, ecosystem, personalization, enterprise workflows, autonomous agents, and infrastructure. The company with the best model may not be the company that captures the market.
Here’s how the five frontrunners are trying to win.→ Read the full article here.
CYBERSECURITY
AI Finds a 20-Year-Old Web Flaw
An AI security system proved in two days that attackers could exploit a flaw in PHP, the programming language used by roughly 70% of websites. Researchers had warned about related behavior as early as 2004, but the new work showed a practical path to remote code execution; PHP maintainers patched two critical vulnerabilities on October 21.
The discovery compresses years of security research into days, but defenders do not automatically move at the same speed. Exploit attempts rose from thousands to millions after disclosure, so the next test is whether website hosts and software vendors can deploy patches before AI-assisted attackers find the next old weakness. → Read the full article here.
INFRASTRUCTURE
NVIDIA Wants Kazakhstan to Become an AI Computing Hub
NVIDIA chief executive Jensen Huang used his first visit to Kazakhstan to argue that the country could become an AI computing hub for Europe and Central Asia. A regional infrastructure plan announced in June targets 100,000 advanced graphics processors and two gigawatts of data-center capacity, while local telecom operator Kazakhtelecom is already deploying NVIDIA H200 and B300 systems.
Kazakhstan offers land, energy and a location between major markets, but ambitious capacity targets still require power, financing and customers. Watch whether the planned data centers come online on schedule and attract enough developers and companies to turn hardware deployments into a durable local AI industry. → Read the full article here.
HEADLINES
In Other News
OpenAI Partners With Poland's NASK: Joint work on AI evaluations, vulnerabilities and safeguards for national security systems.
Idle AI Agents Invent Games: MIT placed 13 AI agents on a simulated island with no tasks, and they began creating games and social activities.
Microsoft's Medical AI Beats Doctors: MAI-DxO solved 83% of complex cases versus 22% for unaided doctors and 36% with references.
New AI Model Hits 1,068 Tokens Per Second: Mercury 3 uses diffusion technology to generate text 4–8x faster than competing models.
Canva Releases Its First Image Model: Canva Craft generates editable images up to 2K in any aspect ratio, in Canva and via API.
American Express to Cut 5% of Workforce: Amex will cut about 5% of staff and take $336 million in restructuring charges as it expands AI use.
PROMPT OF THE WEEK
AI Negotiation Sparring Partner
Most people prepare for negotiations by rehearsing their own arguments. This prompt flips the script: AI steps into the other person's shoes, anticipates their objections, and challenges you in a realistic role-play.
First, ask up to 6 short questions about the other person's role and personality, my goals and walk-away point, their incentives and pressures, relevant deadlines, and our history. Then:
Build their profile: Identify their 3 biggest goals, 3 biggest fears, unspoken concerns, and key constraints. Distinguish facts from assumptions.
Anticipate resistance: Predict their opening move and rank their 3 most challenging objections.
Role-play: Become the other person. Stay in character and push back realistically. After each of my responses, add a brief [COACH NOTE] identifying what worked, where I lost leverage, or what I missed.
Debrief after 6 exchanges: Identify my strongest moment, biggest mistake, one concession to protect, a low-cost/high-value trade to offer, and a 3-sentence opening for the real negotiation.
