Wednesday, May 27, 2026
Biology is the next scaling frontier.
May 27 · 10 videos
Alex Rives is scaling biology like LLMs.
BioHub committed $500M to virtual cells.
Sentry found 67% of AI use is comprehension.
OpenCode hit 8M users in months.
YC turned its internal database into an agent OS.
The Bitter Lesson has arrived for proteins.
“67% of my AI usage was comprehension and only 2% code generation.”
Power agents with full context of your experiments and traces with W&B MCP server
Nico · Weights & Biases · 15 min
Watch on YouTube →Weights & Biases launched a hosted Model Context Protocol server to integrate experiment data into agent workflows. This allows coding agents to autonomously analyze training runs and identify regressions.
- The hosted MCP server supports SaaS, Dedicated, and On-Prem deployments with 20 specialized tools.
- Agents can now identify Skilled Silence where a run finishes but yields 0% evaluation metrics.
- Standardized discovery tools allow agents to self-heal when provided with underspecified prompts.
- Integration with Mistral Chat enables mobile workflows for checking project status via natural language.
- The system automates executive report generation by summarizing key performance metrics from W&B projects.
- Context management is identified as the primary bottleneck for effective AI agent performance.
The Bitter Lesson is Coming for Proteins - Alex Rives, BioHub
Alex Rives · Latent Space · 70 min
Watch on YouTube →Alex Rives explains how scaling laws are revolutionizing protein biology through metagenomic datasets. BioHub is investing $500 million to create predictive oracles for cellular physiology.
- ESMFold 2 has resolved 1.1 billion protein structures at atomic resolution from a 6.8 billion sequence atlas.
- Biological structure and function emerge from simple next-token prediction without explicit biological priors.
- The Virtual Biology Initiative includes a $500 million commitment to bridge molecular and cellular modeling.
- Biology is treated as an information processing machine where amino acid contexts determine functional meaning.
- Automated lab-in-the-loop systems are required to overcome the data barrier in open domain experiments.
- Open science models act as catalysts for global research in complex areas like gene editing.
Comprehend First, Code Later: The AI Skill I Rely On Daily
Priscila Andre de Oliveira · AI Engineer · 17 min
Watch on YouTube →A Sentry engineer analyzed her own AI usage to find that comprehension far outweighs code generation in complex systems. She proposes a framework for using AI to navigate legacy codebases.
- Analysis of 116 Claude sessions showed 67% of prompts were for comprehension and only 2% for code generation.
- The Catch Me Up framework explores architecture, conventions, feature traces, syntax, testing, and history.
- Generating code without deep context leads to slop and increased technical debt in high-stakes environments.
- Engineers should transition to agent managers who orchestrate multiple AI tools across several monitors.
- AI acts as the cheapest senior engineer teammate that can fill context gaps from time zone differences.
- The goal is to move from vibe coding to Keynote Code which respects system longevity and stability.
Building OpenCode with Dax Raad
Dax Raad · The Pragmatic Engineer · 81 min
Watch on YouTube →Dax Raad discusses the explosive growth of OpenCode and the economic realities of the AI developer tool market. He warns about the hidden cognitive load and technical debt associated with AI-generated code.
- OpenCode grew from 650,000 to nearly 8 million monthly active users in just three months.
- The OpenCode Zen inference business reached a $50 million run rate within six months.
- AI agents can create a muted prickle where engineers lose the natural friction felt when writing hacky code.
- LLM inference margins are estimated at 80 to 90 percent despite global GPU shortages.
- Taste is defined as the ability to understand how new features interact with existing product cohesion.
- Positioning as a neutral open-source party is a strategic defense against high-resource competitors like OpenAI.
He Raised $70M to Cure Every Disease With AI
Samuel Rodriques · Weights & Biases · 74 min
Watch on YouTube →Samuel Rodriques of Future House discusses building Kosmos, an AI scientist designed to automate drug discovery. He argues that human talent is the primary bottleneck in biological progress.
- The Kosmos agent has made 20,000 to 30,000 novel scientific findings since its launch.
- Kosmos can write 45,000 lines of code in a single 6 to 12 hour run to support its reasoning.
- Future House uses multi-agent systems for high-throughput reasoning to identify potential disease treatments.
- Rodriques advocates for decentralized clinical trials to accelerate the drug approval process.
- Focused Research Organizations (FROs) fill the gap between academic research and for-profit ventures.
- The future of pharma involves lean companies running thousands of drug programs in parallel via AI.
Why Rust is the Ideal Language for Vibe-Coding
Daniel Szoke · AI Engineer · 16 min
Watch on YouTube →Daniel Szoke argues that Rust's strict compiler makes it superior for AI-driven development compared to Python. He posits that deterministic guardrails are necessary to catch non-deterministic AI errors.
- LLMs are non-deterministic systems that require validation by deterministic guardrails like the Rust compiler.
- Rust's constraints allow autonomous agents to enter a tight loop of compile, fail, and fix until code is safe.
- Dynamic languages like Python allow AI to generate runnable code that contains hidden data races.
- The goal of AI coding should be the first provably correct run rather than the first successful execution.
- Alien Intelligence failure modes in LLMs are fundamentally different from human errors and harder to spot.
- Every compile-time error resolved by an agent represents a production bug that never ships.
Inside YC's AI Playbook
Pete Koomen · Y Combinator · 46 min
Watch on YouTube →YC leadership shares their strategy for building an internal agent infrastructure. They argue that AI should be the operating system of an organization rather than just a feature.
- YC provides agents with read-only SQL access to a centralized Postgres database containing all company context.
- The internal registry contains over 350 tools that agents can use to automate workflows for non-technical staff.
- The Dream Cycle allows agents to review daily interactions at night to self-improve their skills.
- Jevons Paradox suggests that reducing the cost of data queries leads to an exponential spike in strategic questions.
- Companies should adopt a trust-default culture where agent conversations are transparent to encourage learning.
- Gary Tan wrote 40,000 lines of code in 3 days for Gbrain using internal agent tools.
Private Markets, Software Repricing and Capital Allocation | Marc Rowan on a16z
Marc Rowan · a16z · 55 min
Watch on YouTube →Apollo CEO Marc Rowan discusses the shift toward private markets and the impact of AI on enterprise software. He predicts significant repricing in the SaaS sector as software production costs drop.
- Private markets are becoming the primary engine for the real economy as public markets concentrate in 10 stocks.
- Apollo manages $1 trillion in assets with 80 percent focused on credit-based investments.
- AI-driven repricing and zero-cost software production may lead to disastrous returns for legacy SaaS private equity.
- Four public tech companies are projected to spend $800 billion in CAPEX this year for AI infrastructure.
- Leadership should value merit adjusted for distance traveled rather than immutable characteristics.
- Enduring cultures must foster intellectual insubordination where the right answer wins regardless of hierarchy.
Protect Your Inner Peace
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Rob Dial uses the loss of his Instagram account to teach emotional regulation and presence. He emphasizes that inner peace is a daily practice rather than a destination.
- Suffering is often caused by imagined realities and catastrophic projections rather than objective truth.
- The Oak Tree Mindset involves remaining rooted regardless of external hurricanes like financial loss.
- Children absorb the emotional regulation of their parents through the concept of borrowed nervous systems.
- Diversifying business dependencies prevents the loss of a single platform from causing an identity collapse.
- Peace is a choice made multiple times a day to break the cycle of emotional instability.
- Rob Dial lost 1.7 million followers and 11,000 posts when his account was temporarily disabled.
The maturity phases of running evals
Phil Hetzel · AI Engineer · 18 min
Watch on YouTube →Phil Hetzel from Braintrust outlines the transition from vibe checks to structured evaluation flywheels. He argues that identifying failure modes is more productive than treating evals like unit tests.
- Evals should focus on identifying specific failure modes rather than trying to cover every possible outcome.
- The four maturity stages of evals range from human vibe checks to complex tool-calling agent assessments.
- Evaluating CRUD-based tools requires capturing the exact state of external systems at the time of the trace.
- A quality flywheel uses production traces to inform offline experimentation and continuous improvement.
- Directional trends in LLM-as-judge scoring are often sufficient for quantifying the impact of agent tweaks.
- Transitioning from proof-of-concept to production is the primary bottleneck for generative AI customers.
References
PeopleAlex Rives (https://x.com/alexrives/status/2059622778945343669) · Dax Raad (https://x.com/thdxr) · Marc Andreessen (https://x.com/pmarca) · Rob Dial (http://coachwithrob.com) · Gary Tan · Samuel Rodriques · Priscila Andre de Oliveira · Daniel Szoke · Phil Hetzel · Pete Koomen
ToolsWeights & Biases MCP · ESMFold 2 · Claude · OpenCode · Sentry · Braintrust · Kosmos · Gbrain
PapersNature (AMD treatment study)