Beyond The Build

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)

Send us Fan Mail
This week on Last Week in Tech episode, we discussed: Anthropic just re-released a model the US government told them to pull. OpenAI dropped GPT 5 but only to 20 companies. In this episode, Uma and Ope break down what these staggered launches reveal about government influence over Claude AI releases, and why both hosts believe these restrictions could quietly hand the AI lead to Chinese labs. The conversation then shifts into a much bigger question: if AI coding tools now write 41% of all new code globally, what happens when nobody reviews any of it?
You've probably heard the narrative that AI is coming for engineering jobs. The data tells a different story, one about rehiring sprees, AI burnout among senior devs, and a hiring wall that's crushing junior engineers specifically. Uma and Ope also unpack why NVIDIA GPU supply constraints are forcing Google to throttle Meta's Gemini access, the agent jacking attack achieving an 85% success rate across popular coding tools, and why the curl maintainer just halted bug bounty submissions entirely after an avalanche of AI-generated noise. The episode closes on a topic that hits every creator and consumer: do you actually own your digital media, or are you just renting it until the license expires?

In this episode, we covered:
✅ Why Anthropic's re-release and OpenAI's gated GPT 5 launch signal deeper government entanglement in model releases 
✅ The routing-layer strategy companies are adopting as AI tokens become more expensive 
✅ How AI coding tools are accelerating technical debt and why code refactoring has dropped sharply 
✅ The real story on engineering jobs: rehiring, AI burnout, and the 20% hiring drop for developers aged 22–25 
✅ Why Google is limiting Meta's access to Gemini and what the NVIDIA GPU supply bottleneck means for the industry 
✅ The agent jacking vulnerability and what it exposes about AI security in coding agents 
✅ How AI-generated submissions broke the open source software bug bounty system and what the curl maintainer did about it 
✅ Google reCAPTCHA's camera-based verification test and the privacy concerns it raises 
✅ The uncomfortable truth about digital media ownership and why your purchased movies and books aren't really yours

📌 Timestamps:
00:01:15 — Anthropic re-releases Fable — and what it reveals about government control 
00:05:13 — Anti-abuse checks, distillation wars, and Alibaba banning Claude AI code 
00:06:47 — The 41% stat — AI coding tools, technical debt, and the death of code review 
00:11:43 — Engineering headcounts, AI burnout, and the junior developer hiring freeze 
00:13:43 — Google throttles Meta's Gemini access — inside the NVIDIA GPU supply crunch 
00:18:30 — Agent jacking explained — the 85% success-rate attack on coding agents 
00:20:17 — Curl maintainer halts bug bounties — AI-generated submissions overwhelm open source software 
00:23:13 — Google reCAPTCHA wants your camera — privacy, bots, and 57% of internet traffic 
00:30:24 — AI companies vs. traditional companies — who controls the systems? 
00:32:52 — Ladrillos.js, RJ the Dev's interview prep app, and digital media ownership

Send us Fan Mail
Ask us a question or suggest an Episode: https://forms.gle/8tEg2CKTNNWteSgt9

This week on Last Week in Tech: a $100M fintech disaster that tricked millions of Americans, Polymarket's fake gambling website exposed by college influencers, and why China's new open-weight AI model is spooking Silicon Valley. Plus, your smart TV might be selling your IP address without you knowing.
What to expect in this episode: Ope and Uma break down the Synapse bankruptcy where $100M vanished from fintech accounts, the hidden residential proxy scandal running through smart TVs, Claude's new government ID requirement, and the escalating AI arms race between the US and China. If you're a software engineer trying to make sense of where tech is heading, this is your weekly must-watch.

📌 Timestamps:
01:14 — Welcome to Last Week in Tech
01:40 — The $100M Fintech Scam: What Happened to Synapse & Yotta
11:47 — Smart TVs Are Selling Your IP Address (Residential Proxies)
18:26 — Polymarket's Fake Gambling Website Exposed
23:46 — GLM 5.2: China's Open-Weight AI Challenges US Dominance
27:28 — The Identity Verification Trap: AI, Privacy, and Global Regulations
32:56 — Flock Cameras & The Mass Surveillance Ecosystem
39:28 — Meta Tracked Employee Keystrokes to Train Internal AI
41:59 — OpenAI Codex Record & Replay Feature
43:38 — Why "The Optimal Amount of Slop Is Nonzero" in Software
45:48 — Inside OpenAI's Delayed IPO
48:18 — Anthropic Accuses Alibaba of Massive AI Distillation Attack
51:38 — Web Dev Update: HTTP Introduces the New "QUERY" Method
52:37 — Outro & Upcoming Live Show Announcements

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)

💬 Drop a comment: Which story hit hardest for you this week; the Synapse scam, the smart TV proxies, or Claude asking for your ID?