Explore

Models trained on public chemical data may learn chemist sociology, not chemistry

From the chemist-attribution result, Quigley concludes that the models are really learning a lot of sociology, not ground truth. If a model performs equally well using creator-identity vectors as chemical formulas, it is learning creator sociology, not chemistry — treat the dataset as unreliable for training. If models can identify the chemist from the molecule alone, maybe they can infer chemist intent, which means maybe they don't learn anything at all. And when you train models on the exhaust, you might be in trouble.

Framed partly as hypothesis ('maybe'); the underlying inference is the speaker's interpretation.

Latent.Space · Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio

Life-sciences ML is a lamppost problem: scalable measurement exists only at the discovery layer

Quigley frames life-sciences ML as a lamppost problem: scalable measurements only exist at the early-discovery layer — you can't run a million mice or millions of phase one trials — so that's where the technology improves first before climbing layers of biological abstraction. Going after the right problem means working where measurement scales.

offered as the guest's framing

Latent.Space · Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio

AI's role: accelerate deterministic software at lowest cost, and write software that never existed

AI has two roles: accelerate the development of deterministic software at the lowest cost possible, and write deterministic software that never existed before — either because of budget constraints or because it was not considered an interesting problem.

AI Engineer · I Spent Two Years Building the Median — Hariprasad Sudharshan

AI's median output is a trap: good enough to ship, but indistinguishable from competitors

The median AI answer is shippable but not differentiating: the next person gets the same answer, so competitors can ship something similar or identical. Treat it as a floor, not an advantage.

AI Engineer · I Spent Two Years Building the Median — Hariprasad Sudharshan

AI failures are infrastructure problems, not model problems

The talk's thesis: incidents came from missing environmental separation, credentials, and logging — not from smarter models or better prompts. Replit's actual fixes were infrastructural.

AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder

Incidents were preventable with infrastructure controls, not smarter models

Everything shown in the talk was totally preventable — not by paying five times more per token for a smarter frontier model, but with infrastructure controls, many of which companies already own today. Companies just haven't gotten around to setting them up for the AI world.

AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder

The agent wasn't hacked — it did exactly what it was permitted to do

The opening incident did not happen because the agent was taken over by an external malicious actor; the agent was just trying to follow instructions and did exactly what the user gave it permission to do. Jiang asks the audience to keep one question in mind: what is stopping any of this from happening to me, and are these actions ones my agents have permissions to take?

AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder

Layer six parts of secure architecture so a suggestible agent has five more barriers to pass

The six key parts should be layered so a well-meaning but suggestible agent that gets past one still has five more to go; even the first three could have stopped every incident shown.

AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder

A model proxy layer sits between the model and the agent, inspecting and filtering everything that passes between them

A model proxy layer is something that sits between the brain and the body of the AI tool — the model, which is the brain, and the agent running locally, which is the body — and it can see everything that passes between them. It inspects, filters, and logs every call going between them. Such a layer could have prevented the NX singularity incident by stripping injected instructions out of external content before the agent read it, and could have prevented the Copilot incident by flagging a request like 'hey, how do I bypass DLP' before it actually got answered — flagging the DLP bypass request as something the user should not be asking about before the actual exfiltration happened. It could also have flagged what the agent was trying to send out.

AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder

Six infrastructure building blocks ordered by leverage

Jiang presents a list of six infrastructure concepts — building blocks the audience should know about — ordered roughly by leverage: number one takes the most risk off the table, while number six does the least. Every postmortem and scary story in the talk could have been stopped by some combination of these six things. Some, like identity-aligned access control and short-lived credentials, aren't new ideas — they mattered a lot for human developers too — but they take on new urgency when AI is in the loop, because agents are acting at machine speed: if a mistake is made, an agent can make it 20 times or more faster than a human engineer, and agents don't exactly have the common sense that human engineers have. Others, like model proxy layer and agent sandboxing, may be new concepts, which

AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder

Environmental separation is the highest-leverage control but does not protect SaaS-resident AI

Environmental separation is the highest-leverage building block: with no route to production or credentials, the prompt doesn't matter. But it would not have stopped the Copilot incident because Copilot lives inside a SaaS subscription.

AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder

Agent sandboxing is a firewall for agents: restricted file system and network access limits blast radius

Agent sandboxing can be thought of as a firewall for agents: restrictions on the file system and the network so that a misbehaving or compromised agent has a limited blast radius. Sandboxing could have prevented the singularity attack because even if the injected instruction made it through to the agent, it could have stopped the agent from getting access to credentials and prevented the agent from running the command to exfiltrate data. Even if the instruction did land, there wouldn't have been a path to get into credentials and no way to exfiltrate anything without setting off alarms.

AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder

Guardrails on tools: refuse oversized dumps, return available sections on mismatch

The full-rulebook tool refuses to dump rulebooks over roughly 100K tokens into context, returning the section list with a warning instead; if a section name doesn't match, the tool returns available sections so the agent can retry rather than give up or invent names.

AI Engineer · Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket

The fix: stream markdown prose token by token, append the source card after the full answer

The fix was splitting the answer into parts: the agent now streams markdown prose token by token and the front end styles it live as it arrives. The source card is one real UI component appended after the full answer streams; it holds no information you want to click before reading the whole answer, so nothing is lost by waiting, and it gives a clickable card instead of a plain link sitting in the text, which is nicer for the user. The team only adds a structured component when it's required by the UX.

AI Engineer · Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket

Keep document structure: no overlapping chunks, never split tables

Chunks are never overlapped and tables are never split across chunks, because overlap breaks reassembly and splitting breaks table structure.

AI Engineer · Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket

Tool discipline: most questions need 2–4 calls, and the agent is told to self-check before a sixth

A tool discipline instruction is included in the prompt because most questions require two to four calls; before a sixth call the agent is told to ask itself whether it's fetching something genuinely new or just rephrasing a search that already failed. This instruction came from watching real traces where the agent tried the same search again and again, hoping a number would eventually show up.

AI Engineer · Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket

Tool discipline: most questions need 2-4 calls; reflect before a sixth

The prompt instructs the agent that most questions require two to four tool calls, and before a sixth call it must ask whether it's fetching something genuinely new or just rephrasing a failed search — addressing the failure mode of repeating the same search hoping a number appears.

AI Engineer · Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket

Five files make an AI review complete

Five files make the review complete: intended use (users, boundaries, prohibited actions), data lineage (each example source and use approval), failure modes (detection and fallback), results by segment (coverage and errors), and controls (reviews, thresholds, stopping rules, owners).

AI Engineer · Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers

In bank AI deployments, discovery is the engineering

Paden Gayle's reframing: in bank AI deployments, discovery is the engineering. Uncover deployment decisions in week one so you can design a system with a credible route into production.

AI Engineer · Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers

Three technical levers run in sequence: caching, shrinking the prompt, downsizing the model

Alvin lays out three technical levers for making work cheaper, and they run in a sequence: first caching, then shrinking the prompt footprint, then downsizing the model. The one-line playbook: cache the obvious stuff, let an optimizer shrink your instructions, right-size your model, and use the ceiling test to understand the variation — then test whether quality holds.

AI Engineer · How to Actually Tokenmaxx and Save Millions — Alvin Pane, OutRival