Monday, July 13, 2026
The shift from blind trust to provable safety.
July 13 · 8 videos
Erik Meijer demands proof carrying code.
Engram raised $98M to kill context rot.
Alejandro Vidal is fixing 1950s style evals.
Walking 40 minutes reverses brain shrinkage.
Attention spans dropped 33 percent since 2006.
Dalton Caldwell says 5 percent equity is enough for leavers.
“I'm convinced that if there's anything between the model's goal and where the model currently is, it will do everything that it can to reach that goal, including killing us or deleting your files.”
Cofounder Relationships
Dalton Caldwell · Dalton + Michael · 22 min
Watch on YouTube →Dalton Caldwell and Michael Seibel discuss why cofounder relationships fail and how to handle breakups. They argue that mutual respect is the only foundation that prevents death by losing faith.
- Mutual respect is the foundational layer for startup success.
- Tactical fixes cannot resolve underlying friction if respect is missing.
- Departing founders pre product market fit should retain less than 5 percent equity.
- Leaving within the first year should result in 1 to 2 percent equity retention.
- Generosity during a breakup pays dividends in team morale and reputation.
- Lawyers often exacerbate conflict by writing mean letters to bill more hours.
In Code They Act, In Proof We Trust — Erik Meijer, Leibniz Labs
Erik Meijer · AI Engineer · 21 min
Watch on YouTube →Erik Meijer argues that AI agents are currently unhinged and require machine checkable proofs before execution. He proposes a return to 1990s Proof Carrying Code to prevent agents from deleting databases.
- Agents are dangerous until proven safe by construction.
- The Lethal Trifecta includes private data access, prompt injection, and tool use.
- Agents must submit an execution plan alongside a machine checkable proof.
- Side effects in agentic loops must be air gapped from decision making.
- Formal verification tools like Lean are seeing high VC interest.
- Computer science research from the 1990s holds answers to modern AI problems.
The AI Memory Problem: Why Long Context Isn’t Enough — Dan Biderman, Engram Co-founder & CEO
Dan Biderman · Latent Space · 49 min
Watch on YouTube →Engram CEO Dan Biderman explains why RAG and long context windows lead to context rot. He introduces Knowledge Cartridges to compress context by 1000x through weight based updates.
- Context rot causes models to lose accuracy as they read more data.
- Engram raised $98M to move knowledge from transient text to model weights.
- Knowledge Cartridges offer 1000x compression of raw context.
- A Llama 70B model requires an 80GB KV cache for one Wikipedia article.
- Enterprise data scale will soon reach trillions of tokens.
- Intelligence and token efficiency are inseparable in the next scaling paradigm.
Stop Evaluating Models Like It's the 50s - Alejandro Vidal, Mindmakers
Alejandro Vidal · AI Engineer · 23 min
Watch on YouTube →Alejandro Vidal critiques current AI benchmarks for treating all questions as equal. He advocates for Item Response Theory to measure the invisible intelligence of models.
- Item Response Theory measures the invisible intelligence of LLMs.
- Current benchmarks are flawed because they treat all questions as equal.
- Benchmarks can be reduced by 80 percent while maintaining 99 percent correlation.
- Error patterns act as model DNA to detect data leakage or distillation.
- Adaptive testing prevents benchmark saturation and identifies training data theft.
- Raw accuracy is misleading when comparing models of different intelligence levels.
From fork() to Fleet: Designing an Agent Sandbox Cloud — Abhishek Bhardwaj, OpenAI
Abhishek Bhardwaj · AI Engineer · 44 min
Watch on YouTube →OpenAI's Abhishek Bhardwaj details the infrastructure required for secure agent execution. He focuses on microVM sandboxes and persistent storage for long horizon tasks.
- Agents are natural drivers of Linux systems due to their training data.
- Security is a binary asset where one breach can end a product.
- MicroVM sandboxes provide hardware level isolation for agent execution.
- Persistent storage allows agents to tackle long horizon tasks.
- Gold Mode tasks have run for up to 3 days continuously at OpenAI.
- Infrastructure design should prioritize first principles like syscalls and ring levels.
The Prime Intellect Stack — Will Brown, Prime Intellect
Will Brown · AI Engineer · 46 min
Watch on YouTube →Will Brown introduces the Prime Intellect stack for democratizing post training and reinforcement learning. He argues that evals are the entry drug to building a proprietary post training flywheel.
- Post training is now a requirement for maintaining competitiveness.
- Prime Intellect operates over 10,000 GPUs for model refinement.
- A 1000 step RL run on a frontier model costs about $50,000.
- Evals are the entry drug to building a post training flywheel.
- Asynchronous RL overcomes the bottleneck of slow rollouts in coding tasks.
- Open source models are now good enough for superhuman performance in niches.
How To Get Ahead Even When No One Is There For You
Rob Dial · The Mindset Mentor Podcast · 15 min
Watch on YouTube →Rob Dial discusses how the 33 percent drop in human attention spans creates a massive opportunity for focused individuals. He advocates for trading entertainment for micro education during dead time.
- Human attention spans have dropped 33 percent over the last 20 years.
- The average person spends 4.1 hours on their phone every day.
- Outlier CEOs read 60 books per year while the average person reads zero.
- Environmental design is more effective than willpower for focus.
- Transition periods like driving are high ROI moments for micro education.
- Trading entertainment for education creates a massive competitive advantage.
How to Improve Your Memory & Cognitive Function at Any Age | Dr. Alan Castel
Alan Castel · Andrew Huberman · 148 min
Watch on YouTube →Dr. Alan Castel explains how the aging brain compensates for slower processing through selectivity and wisdom. He provides evidence based interventions to reverse hippocampal shrinkage.
- Walking 40 minutes three times per week can increase hippocampal volume by 1 percent.
- A positive attitude toward aging is linked to longer life and lower dementia risk.
- The brain compensates for slower processing by prioritizing high value information.
- Superagers have memory performance that rivals people decades younger.
- Family safe words are necessary to protect against AI voice cloning scams.
- Learning through failure is more effective than passive repetition.
References
PeopleDalton Caldwell (x.com/daltonc) · Michael Seibel (x.com/mwseibel) · Erik Meijer (@headinthebox) · Dan Biderman · Alejandro Vidal (@dobleio) · Abhishek Bhardwaj · Will Brown (@willccbb) · Rob Dial (coachwithrob.com) · Alan Castel (castel.psych.ucla.edu) · Steve Jobs · Steve Wozniak · Emmett Shear · Solomon Hykes · Simon Willison · Jeff Huntley · Assaf Rappaport · Chris Ray · Scott Linderman · Sabri Eyuboglu · Cade Daniel · Swix · John Wooden · Sully Sullenberger · Ronald Cotton
ToolsLean · Knowledge Cartridges · PrimeRL · Sandbox Clouds · Item Response Theory · Free Monads · Llama 70B · GPT-4 · Code Interpreter · GLM-5
PapersProof Carrying Code