|
AI Builders Digest
Monday, November 9, 2026
|
|
Two of today's stories are about AI agents doing science. One measures whether agents can handle a wet lab. The other is already proposing new drug molecules. That gap, from "can it read a gel?" to "here's a novel compound," tells you how fast this is moving, and how much harder it's getting to know what benchmarks are actually measuring.
|
|
---
|
|
01
|
What Stanford's AI engineering course teaches that your company probably doesn't
|
|
|
Latent Space's interview with Mihail Eric, who teaches CS146S at Stanford and leads AI at Monaco, surfaces a genuinely useful tension: when students can one-shot last year's homework with an agent, what do you actually teach them? Eric's answer is to assign open-source contributions where a human maintainer reviews the patch. You can automate the code. You cannot automate the conversation about whether the maintainer wants it, what else it might break, or why the existing behavior exists. His team at Monaco runs the same logic in production, using coding agents to ship and investigate outages while keeping the judgment layer human.
|
Why it matters: Most companies adopting AI coding tools are skipping straight to "ship faster" and not asking what skills atrophy when agents write the first draft every time. If your junior engineers never have to explain their code to a skeptical senior, they won't be able to when the agent gets it wrong. And the agent will get it wrong.
|
|
Source →
|
|
---
|
|
02
|
AI turned cybersecurity's asymmetry upside down, and China may feel it most
|
|
|
ChinaTalk's Chang Che lays out what happened when an OpenAI model found a critical bug in PHP, the language running 70% of the world's websites. Hacking attempts started before the patch was ready. Google's incident response team now clocks the average time-to-exploit for new vulnerabilities at negative seven days. Before AI, finding and exploiting software holes took months of skilled labor. Now the holes are visible to anyone running the right model, and the cost of exploitation has collapsed.
|
Why it matters: If your company runs any PHP-based software, your security team's patching timeline just became a liability it may not know it has. The old assumption, that attackers need time to find and use a vulnerability, no longer holds. Patch windows that used to be measured in weeks now need to be measured in hours.
|
|
Source →
|
|
---
|
|
03
|
Box CEO: long-running AI agents don't need to be near your data center
|
|
|
Aaron Levie made an observation worth sitting with: unlike most software, AI agents aren't sensitive to latency in the way a web app is. If an agent is doing hours of work, it doesn't matter if it's running in a distant data center. Send the task, get the result. The data can move in encrypted form, the work happens remotely, and the user just waits for an output.
|
Why it matters: This quietly undermines the enterprise argument for on-premise AI. If the work takes long enough that round-trip latency doesn't matter, cloud-hosted agents beat in-house infrastructure on cost and capability. The "we need AI on our servers for security" position gets harder to defend when the agent is already designed to run far away and return only a result.
|
|
Source →
|
|
---
|
|
04
|
OpenAI built a benchmark for AI agents in biology labs
|
|
|
OpenAI released BioBasicsBench, a set of 100 tasks testing whether AI agents can interpret raw experimental data from wet-lab biology, the unglamorous work of reading assay results, flagging anomalies, and understanding what an experiment actually produced.
|
Why it matters: Before an AI agent runs experiments autonomously, someone has to prove it can read them correctly. This is OpenAI staking out that it will define what "good enough" looks like in biology, before regulators or competitors do.
|
|
Source →
|
|
---
|
|
05
|
Google DeepMind's Antigravity is already proposing new drug molecules
|
|
|
Google DeepMind published details on Antigravity, an agentic system for drug discovery that can search chemical compounds and propose novel "rigidified macrocycles," a class of molecules relevant to hard-to-treat diseases. It pairs with a set of Cheminformatics Skills that handle the specialized chemistry reasoning.
|
Why it matters: OpenAI is measuring whether agents can read a lab notebook. Google DeepMind's system is already suggesting what to put in one. That is a meaningful gap in ambition, and pharmaceutical companies paying for AI drug discovery tools should be asking which side of that gap their vendor sits on.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|