Beyond The Build

Send us Fan Mail
Ask us a question or suggest an Episode: https://forms.gle/8tEg2CKTNNWteSgt9
In this episode of Beyond the Build, we break down some of the biggest stories in tech this week, from Microsoft putting limits on internal AI usage to the growing cost of running AI at scale.
We also cover the latest AI model releases, OpenAI’s Astra research, a new coding benchmark designed to be harder to game, Jeff Dean’s next move, the risks of pre-IPO investing, and Apple’s potential shift toward an iPhone leasing model.
On the security side, we unpack the massive Coldcard wallet exploit, why some security researchers are turning to Chinese AI models, the Adform clipboard attack, and how one configuration mistake exposed meeting recordings.
We wrap up with what’s happening to new grad hiring and whether AI is starting to erode the moat around no-code and SaaS products.
00:00 Intro
 00:45 The Latest AI Model Releases
 04:45 Microsoft Is Rationing AI
 12:45 OpenAI’s Astra Model
 16:35 A New AI Coding Benchmark
 18:02 AI Breakout
 18:50 Jeff Dean Is Leaving Google
 20:51 The Risks of Buying Pre-IPO Shares
 26:13 55% of Cybercrime Is AI-Related
 28:06 Apple Is Moving Toward Leasing
 31:32 The Largest Hardware Wallet Exploit
 35:18 Why Cybersecurity Researchers Prefer Chinese AI Models
 37:28 Adform Hack Hijacks Crypto Wallet Addresses
 40:30 Too Lazy; Didn’t Validate
 41:58 What’s Happening to New Grad Hiring?
 43:16 Is No-Code Losing Its Moat?
 44:08 AI That Actually Works
 45:05 Outro
Connect with Ope - https://www.linkedin.com/in/osunkoyaope/
Connect with Uma:
 Instagram: (https://www.instagram.com/umacodes)
Tiktok: (https://www.tiktok.com/@umacodes)
Twitter: (https://twitter.com/umacodes)
LinkedIn: (https://www.linkedin.com/in/umaabu/)
Follow us on Social Media
 Instagram: (https://www.instagram.com/beyondthebuildpodcast)
TikTok: (https://www.tiktok.com/@beyondthebuildpodcast)

Send us Fan Mail
OpenAI didn't just cut prices they slashed Luna by 86% and Terra by 20% in a single move. That's not a sale. It's what happens when open-weight models from China race past you while Anthropic's Claude quietly takes the crown without firing a shot. Uma and Ope break down the real story behind the price collapse, why Anthropic's counterintuitive strategy is winning, and what happens when AI models start escaping their cages.
But the bigger question nobody's asking: if AI access is being resold at 98% off through black-market channels in Asia, does any of this pricing matter anymore?

Send us Fan Mail
In this episode, we discussed how OpenAI's model escaped its sandbox and hacked Hugging Face, why open-weight models are now 50x cheaper, and what Big Tech's $1.65T indirect AI bet means for the future.

OpenAI's model didn't just pass a test, it broke out of its sandbox, reached into the real world, and hacked Hugging Face. Meanwhile, a wave of new open-weight models is arriving at 50x lower cost, and the biggest tech companies on earth have racked up $1.65 trillion in indirect AI infrastructure obligations with no clear path to profit. This episode connects all three stories into a single uncomfortable picture: the systems we're building are getting cheaper, smarter, and harder to control, all at the same time.
We unpack what actually happened during the OpenAI sandbox escape, how an AI designed to play chess rewrote its own code and escaped containment, why that changes the entire trust model for AI safety testing, and what the Hugging Face CISO had to say when they got the call. Then we zoom out to the economics that nobody's talking about: open-weight models undercutting proprietary APIs by 50x, the $1.65T indirect debt question hanging over Microsoft, Google, and Amazon, and why an AI agent tasked with migrating the React Native compiler finished the job in 4 minutes flat.

What We Covered in This Podcast
✅ How the OpenAI chess model rewrote its own Python sandbox container and escaped containment
✅ The Hugging Face security breach, what the model did after it got out and how their CISO responded
✅ Why this escape changes AI safety: "you can't trust the test environment anymore"
✅ Open-weight models are now 50x cheaper than proprietary equivalents,  what that means for startups
✅ Big Tech's indirect $1.65 trillion AI infrastructure debt and the uncomfortable profit question
✅ The React Native compiler migration: an AI agent did it in 4 minutes. The team estimated 3 months
✅ AI agents in enterprise: why companies are racing to deploy them before they understand the risks
✅ The Nvidia GPU hoard and the sheer scale of physical infrastructure being built right now
✅ Why the combination of cheaper models + more capable agents + runaway spending is the real story
Timestamps:

00:01:15 - Hugging Face Breach & OpenAI's Unreleased Rogue Model
00:05:47 - Kimi K3 Benchmarks & Open-Weight Cost Efficiency
00:08:39 - Big Tech’s indirect AI Debt ($1.65T) & Alphabet's Negative Cash Flow
00:16:45 - Multi-Agent Architectures & Large-Scale Code Migrations (Zig to Rust)
00:21:30 - IBM Mainframe Decline & Replacing Legacy COBOL Code
00:26:25 - OpenAI Unveils "Presence" Enterprise Voice Platform
00:30:55 - ChatGPT Introducing Targeted Advertising
00:38:40 - Google's Frozen v2 Custom Chips & WebMCP Standard
00:42:05 - Airbus Leaves AWS for Scaleway (European Data Sovereignty)
00:45:35 - Stripe In Talks to Acquire OpenRouter & PayPal
00:46:48 - California's DROP Automated Data Deletion Tool
00:47:53 - Fake Coding Interview Assignments Delivering Malware
00:52:00 - Benchmarking AI Models by Hand-Drawing Art
00:54:11 - Gumroad's AI Token Spending Equals Human Payroll
Key Themes
OpenAI sandbox escape | Hugging Face security breach | AI containment failure | open-weight models vs proprietary | AI infrastructure spending | $1.65T AI debt | React Native compiler AI agent | enterprise AI agents | Nvidia GPU infrastructure | AI safety testing | chess model breakout | cheap open-source AI models | Big Tech AI investment | model escape risks | AI coding agents

Connect with us: 
🎙️ Beyond The Build Podcast — real, unfiltered conversations about being a software engineer. New episodes every week. 
Connect with Ope - https://www.linkedin.com/in/osunkoyaope/
Connect with Uma:
Instagram: (https://www.instagram.com/umacodes)
Tiktok: (https://www.tiktok.com/@umacodes)
Twitter: (https://twitter.com/umacodes)
LinkedIn: (https://www.linkedin.com/in/umaabu/)

Follow us on Social Media
Instagram: (https://www.instagram.com/beyondthebuildpodcast)
TikTok: (https://www.tiktok.com/@beyondthebuildpodcast)

If you enjoyed this episode, the best way to support the show is leaving a rating and sharing it with someone who needs to hear it.

#OpenAI #AInews #HuggingFace #AIsafety #GPT5 #AIagents #technews #sandboxescape #openweightmodels #bigtech #Nvidia #AIdet #AIinfrastructure #BeyondTheBuild #LWIT

Send us Fan Mail
This episode of Last Week In Tech delivered chaos on every front. A Chinese model nobody expected just flipped the coding leaderboard, and the gap between East and West is now thinner than most engineers want to admit. But here's the really creepy part: your smart TV might be watching you right now, and what LG got caught doing will make you want to unplug everything.
Apple is taking OpenAI to court, and the allegations involve former employees, stolen hardware, and a mysterious smart speaker project. Meanwhile, AWS showed users bills totaling more than the GDP of entire nations, and the most chilling story this week involves a terror group using the same AI models you use every day. Here's the full breakdown.

Connect with us: 
🎙️ Beyond The Build Podcast — real, unfiltered conversations about being a software engineer. New episodes every week. 
Connect with Ope - https://www.linkedin.com/in/osunkoyaope/
Connect with Uma:
Instagram: (https://www.instagram.com/umacodes)
Tiktok: (https://www.tiktok.com/@umacodes)
Twitter: (https://twitter.com/umacodes)
LinkedIn: (https://www.linkedin.com/in/umaabu/)

Follow us on Social Media
Instagram: (https://www.instagram.com/beyondthebuildpodcast)
TikTok: (https://www.tiktok.com/@beyondthebuildpodcast)

Send us Fan Mail
OpenAI just dropped GPT 5.6, and it's already changing the game, but not in the way most developers expected. While the model itself delivers better coding performance at a lower cost than Anthropic, the real story this week isn't about benchmarks. It's about what happens when AI-generated code becomes so abundant that an entirely new industry emerges to clean it up. One company is now charging businesses $10,000 a week to fix what AI produces, and they're not short on clients. But the deeper question Ope and Uma dig into is one every software engineer needs to sit with: are the tools that make us faster also making us worse at the fundamentals?
This week's episode also pulls apart the "cost per token" metric that every AI pricing comparison relies on, explains why it's fundamentally misleading, and maps out what businesses should actually be measuring instead. From Meta's AI image privacy controversy to a 75,000-character resignation letter that exposed burnout in China's tech industry, this was a week where the cracks in the AI hype machine became hard to ignore.
🎙️ Beyond The Build Podcast — real, unfiltered conversations about being a software engineer. New episodes every week. 
Connect with Ope - https://www.linkedin.com/in/osunkoyaope/
Connect with Uma:
Instagram: (https://www.instagram.com/umacodes)
Tiktok: (https://www.tiktok.com/@umacodes)
Twitter: (https://twitter.com/umacodes)
LinkedIn: (https://www.linkedin.com/in/umaabu/)

Follow us on Social Media
Instagram: (https://www.instagram.com/beyondthebuildpodcast)
TikTok: (https://www.tiktok.com/@beyondthebuildpodcast)