Wednesday, July 1, 2026
The bottleneck shifted from chips to power and trust.
July 1 · 8 videos
Grid lead times hit five years.
GLM 5.2 matches frontier models at 750B parameters.
GPT 5.6 reaches 750 tokens per second.
Kent Beck says empathy is a technical skill.
The era of the Software Factory has arrived.
“The human part is the hardest part in software engineering.”
How Kent Beck shapes the software engineering industry
Kent Beck · The Pragmatic Engineer · 148 min
Watch on YouTube →Kent Beck discusses the critical trust gap in software engineering. He argues that AI creates code faster than humans can understand or trust it.
- Empathy is a technical skill that gates your ability to affect change.
- The 3X framework defines business phases as Explore, Expand, and Extract.
- AI returns the industry to the Explore phase where rapid experimentation is key.
- Leaders must be tree shakers who initiate change during technological shifts.
- Productive discomfort is necessary for growth and should be induced by coaches.
- Product market fit often manifests as people tugging on your shirt for a solution.
The weird history of the internet...
Fireship · Fireship · 8 min
Watch on YouTube →A look at the chaotic evolution of the internet from military projects to AI generated content. It explores the tension between open source utility and corporate monetization.
- ARPANET began as a decentralized framework designed to survive nuclear strikes.
- Tim Berners-Lee created foundational value by releasing the World Wide Web for free.
- Microsoft used its OS dominance to win the browser wars against Netscape.
- PageRank proved that recursive hyperlink analysis is the most profitable way to organize data.
- Modern internet faces a brain rot phase with massive JavaScript payloads and AI slop.
- Technological progress often moves from desk bound utility to pocket sized distractions.
AI Is Running Out of Power
Ejaaz · Limitless Podcast · 32 min
Watch on YouTube →The AI investment thesis is shifting from GPU supply to electrical grid infrastructure. Lead times for power connections are now the primary bottleneck for data centers.
- US data center power demand is projected to grow from 80 to 150 gigawatts by 2028.
- Traditional grid power connections now have a five year lead time.
- Bloom Energy fuel cells offer a 90 day deployment alternative to the grid.
- Value is migrating to Layer 5 infrastructure including power, wiring, and thermal management.
- The CME is setting up compute futures trading indices to financialize GPUs.
- Complexity in AI manufacturing is reaching atomic levels where five atoms determine success.
CoreWeave ARIA: The autoresearch loop for continuous improvement
Zubin Isola · Weights & Biases · 5 min
Watch on YouTube →CoreWeave ARIA introduces autonomous agents to manage the machine learning experimental loop. This shifts the researcher's role from manual tuning to high level problem definition.
- ARIA inspects prior run states to form hypotheses about learning rates and dropout.
- The agent triggers training jobs independently via Weights & Biases Launch.
- Parallel ARIA agents allow researchers to explore divergent research paths simultaneously.
- Small validation loss improvements like 0.0005 are analyzed against randomness jitter.
- AI agents are evolving from simple chatbots into active agents of action.
- Automating the hypothesis to trial loop allows senior talent to focus on what to solve.
Why Everyone Is Suddenly Talking About AI Factories
Kaushik Shirhatti · The TWIML AI Podcast with Sam Charrington · 15 min
Watch on YouTube →The AI Factory model transforms data and energy into business outcomes. It represents a shift from experimental projects to industrial scale intelligence production.
- Enterprise AI ROI increased from 5 percent to 30 percent over the last year.
- Sovereign AI is driving national and enterprise investment for strategic independence.
- Success requires unifying traditional IT silos like storage, compute, and data teams.
- Real value comes from reimagining workflows rather than just optimizing legacy processes.
- The industry is moving toward universal liquid cooling and physical AI robotics.
- Successful enterprises pick 4 to 5 core projects and go deep rather than spreading resources.
When You're Tired Of Everything, Watch This
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Chronic exhaustion is framed as a signal of mental and spiritual misalignment rather than just physical fatigue. Rob Dial explores the metabolic cost of overthinking.
- Professional chess players can burn 6,000 calories a day through intense thinking.
- Exhaustion acts as a protection mechanism to prevent engagement with misaligned environments.
- Radical honesty requires answering what you are pretending not to know.
- Decision fatigue over trivial matters indicates a mentally overtaxed system.
- The energy to tolerate a hated job is higher than the energy needed to build a business.
- Curiosity should replace judgment when analyzing personal energy drains.
AI Just Entered A New Era
Dr. Károly Zsolnai-Fehér · Two Minute Papers · 7 min
Watch on YouTube →GLM 5.2 shows that open weight models are rapidly closing the gap with proprietary systems. It offers a high performance alternative that can be owned and run locally.
- GLM 5.2 is a 750 billion parameter model that matches frontier systems in coding.
- The model uses Preference Optimization and multi-token prediction for long horizon tasks.
- Weight ownership is the only way to ensure long term access and avoid platform bans.
- Running a 750B model locally requires tens of thousands of dollars in hardware.
- The development cycle from version 5.1 to 5.2 took less than three months.
- API pricing must account for high token consumption in dense models using multi-token prediction.
WF2026: Software Factories & Keynotes ft. Microsoft, OpenAI, OpenClaw, Z.ai (GLM), MiniMax, HF
Swix · AI Engineer · 516 min
Watch on YouTube →The 2026 World's Fair highlights the transition to Software Factories and autonomous agentic systems. Engineering is shifting from code production to loop orchestration.
- GPT 5.6 SUL and Terra models offer speeds of 750 tokens per second.
- MiniMax provides 1 million token native multimodal contexts.
- Value Maxing is replacing Token Maxing as the primary focus for enterprise AI.
- Engineers are becoming factory managers who oversee architectural slot free zones.
- Agentic security and formal verification are becoming essential for safe AI deployment.
- Maintain model agnosticism to retain leverage and avoid being priced out by suppliers.
References
PeopleKent Beck · Ward Cunningham · Martin Fowler · Eric Gamma · Larry Tesler · Grady Booch · Paul Baran · Tim Berners-Lee · Marc Andreessen · Vint Cerf · Larry Page · Jensen Huang · Leopold Aschenbrenner · John Carmack · Josh (@JoshKale) · Ejaaz (@cryptopunk7213) · Zubin Isola · Andrej Karpathy · Thierry Piennar · Kaushik Shirhatti · Sam Charrington · Rob Dial (http://coachwithrob.com) · Dr. Károly Zsolnai-Fehér · Swix (@swyx) · Pablo Castro · Peter Steinberger · Vibhav
Tools3X Framework · CRC Cards · ARPANET · TCP/IP · PageRank · Ajax · Bloom Energy · CoreWeave ARIA · Weights & Biases · GLM 5.2 · GPT 5.6 · MiniMax · Codex · Claude