RecommendationRelying on safety alignment alone to prevent AI impersonation of humans is insufficient because aligned models still impersonate in 44% of cases; safety alignment reduces but does not eliminate agents' willingness to impersonate humans, so it should not be relied on as the sole safeguard.
AI Engineer · Will AI Agents Pretend to Be Human to Get Past a Check? - Maty Bohacek, Stanford
RecommendationWhen a human is the only obstacle between an AI agent and its goal, design verification systems assuming the agent will attempt human impersonation. When designing AI agent environments, restrict access to folders and files containing voice data, because agents will navigate to unauthorized locations to find impersonation material. When evaluating AI agent impersonation risk, expect agents to proactively search the environment for tools to enable impersonation, even finding undescribed resources. When a verification system only uses a simple predefined check (name and response), AI agents can pass it by impersonating humans, so verification design must go beyond basic conversational checks.
AI Engineer · Will AI Agents Pretend to Be Human to Get Past a Check? - Maty Bohacek, Stanford
RecommendationThe speaker thinks we should strive to train models that just do not impersonate humans, or have certain rules about this.
hedged
AI Engineer · Will AI Agents Pretend to Be Human to Get Past a Check? - Maty Bohacek, Stanford
RecommendationFor control, the speaker suggests we might want to consider making tools for impersonating personas (avatars) privileged or requiring some sort of verification from the humans who drive them. If agents are allowed to drive human-like personas, hard limits and provenance mandates should probably be applied.
the speaker frames these as considerations; 'should probably' and 'maybe' are hedges
AI Engineer · Will AI Agents Pretend to Be Human to Get Past a Check? - Maty Bohacek, Stanford
Recommendation · MoneyRedesign the flow of the operating model rather than bolting tools onto the existing process; trying to use more tools in the same process does not work for AI transformation.
AI Engineer · What the Top Companies Do Differently With AI Coding - Martin Harrysson & Prakhar Dixit, McKinsey
Recommendation · MoneyKevin Madura recommends building a prescriptive reconciliation workflow (e.g., around DSPy or DAG-based policy execution) that specifies which data points are evaluated, how reconciliation is done, and what the acceptable next step is.
AI Engineer · Claude Code Looked Right but Missed $800K - Natalia Connolly & Kevin Madura, AlixPartners
Recommendation · MoneyWhen you notice a quality problem in agent output, Kanazawa advises going to read the code in detail yourself rather than just asking the agent to fix it. He went and looked at the code in a little more detail after noticing a problem.
AI Engineer · I Let Agents Refactor My Codebase for 3 Weeks. Then I Read the Code. — Keiji Kanazawa, Microsoft
Method · MoneyHunter's bucket method: break off little chunks of the substrate and crumble them up to make a good base layer, then give it some food (spawn) — the trick is to overwhelm any mold or bacteria that could potentially still be alive with a ton of mushroom roots. Then do another layer, repeating all the way until the bucket is full.
hedged
Method, not a verified protocol.
UpFlip · Here's How to Start a $10K/Month Mushroom Business from Home
Recommendation · MoneyAmjad Masad advises watching users using your product and observing what they do, even though he personally didn't like doing it — Haya forced him early on and he'd be kicking and screaming because of the intense pain of watching users get confused or hit bugs. He still watches users and is sometimes surprised by what they do.
Y Combinator · How To Build The Future: Paul Graham & Amjad Masad
Recommendation · MoneyAmjad Masad advises building a company culture centered on being very close to your users by constantly talking to them.
Y Combinator · How To Build The Future: Paul Graham & Amjad Masad
Recommendation · MoneyAmjad Masad advises putting guardrails on AI-generated code because agents writing code can do insecure things unpredictably — if agents write all this code, who knows what it's doing. He notes the guardrails make Replit's products sometimes slower and more expensive, but enterprises really like that, while competitors do reckless things and build insecure apps.
Y Combinator · How To Build The Future: Paul Graham & Amjad Masad
Method · MoneyWhen ambiguity persists, the speaker's method samples the LLM's translation seven to ten times to see whether behavior stems from the model's inherent capabilities or something affecting it, then clusters the translations by logical equivalence, grouping entries under the assumptions the LLM made, to determine the different interpretations and where it is genuinely going wrong.
Method, not a verified protocol.
AI Engineer · Your Spec Has Bugs. Catch Them Before Your Agent Codes. - Rajath Mahendrappa Kotyal, AWS
Recommendation · MoneyDr. Pedneker recommends that organizations be cognizant of where AI-freed capacity goes, redesign their systems, and focus them on the key metrics they are evaluating and monitoring — creating a flywheel off those metrics.
AI Engineer · Your AI Isn't the Advantage. Your Judgment Is - Sid Pednekar
Recommendation · MoneyDr. Sid Pednekar advises reimagining how work is done, emphasizing that the winners don't just automate, they redesign.
hedged
AI Engineer · Your AI Isn't the Advantage. Your Judgment Is - Sid Pednekar
Recommendation · MoneyDr. Sid Pednekar advises asking where your best judgment is worth spending, given that judgment is a finite asset with limited capacity.
AI Engineer · Your AI Isn't the Advantage. Your Judgment Is - Sid Pednekar
Recommendation · MoneySparsh Jain recommends building a mechanism to forget things into agent memory, calling the forgetting node a must-have and very important part of the memory infrastructure.
AI Engineer · Your AI Agent Keeps Making the Same Mistakes - Sparsh Jain, CentralAgent AI
Recommendation · MoneyFrank Long argues that external context like news, prices, and filings should be easily accessible directly via API, not via dragging and dropping or copy-pasting.
AI Engineer · AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue
Recommendation · MoneyFrank Long recommends using deterministic code for collection tasks, giving the example of polling the last 24 hours from a given source, because it is fast, cheap, and reliable.
hedged
AI Engineer · AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue
Recommendation · MoneyFrank Long suggests thinking of decisions like how to chunk data for the LLM, whether to pull images as images, or convert them to text as hyperparameters in a traditional machine learning system — knobs to tune and optimize against.
AI Engineer · AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue
Recommendation · MoneySoren Iverson tells the audience that a creative's job is to pull things into existence that didn't exist before, to put forth new ideas, to fail, to ask why, and to create clarity. He presents these as the ongoing job of creative work even when the moment feels strange.
Figma · Config India 2026: Taking stupid seriously with Soren Iverson