#AI Engineer
50 articles with this tag
Most read in #AI Engineer, last 30 days
Netflix says subjective AI evals need two judges
Netflix's Cameron Wolfe tells AI Engineer that subjective AI tasks drown in noise and need human evals, tuned LM judges, and production AB tests.
Coatue's Frank Long Says Agents Hit Investing Next
Frank Long told AI Engineer agents will upend investing more than coding, laying out Coatue's function for turning news and positions into selection, size and timing.
Brainbase Labs pushed Domino's agent to 97% accuracy
Milind Sunil Shah detailed how Brainbase moved a Domino's pricing agent from 41% to 97% accuracy across 11 versions while halving cost per source.
Wise: Approve the Job, Not Every Step
Wise data and AI product lead Emil Chetty told AI Engineer to replace per-step approvals with a single immutable job that services enforce.
Nous CEO says labs will trap you or replace you
Nous Research CEO Dillon Rolnick told AI Engineer labs will either trap businesses in agents or replace them, pushing Hermes Agent as the ownable alternative.
Banks Don't Certify Models. They Gate Apps.
Langfuse engineer shows how a top-five bank gates AI apps, not models, with golden datasets and multi-dimensional pass-fail checks.
NatWest built the agent. Safety was harder
NatWest moves Kora from scripted menus to generative answers, but its lead engineer says proving safety beats building the agent.
Why Your 100x AI Gain Stalls Inside Companies
Mitchell Troyanovsky told AI Engineer why coding agents feel 100x faster for individuals but flat for companies, and why opinions and behavior specs are needed for reliable scale.
Agents Should Not Borrow Your Identity
Factory’s Nikhil Harithas tells AI Engineer that agents must run as their own applications in the customer VPC, via Open Compute drivers and an egress-only relay.
Better Search Pushed an Agent From 36% to 98%
Mixedbread showed on AI Engineer how visual late interaction search lifted a hedge fund agent from 36% to 98% accuracy and cut LLM spend by 3x.
Runway says video models now pass reality test
Runway CEO Cristóbal Valenzuela demoed real-time world models that turn any image into an interactive simulation, no 3D engine required.
Prime Intellect says companies should train own models
Prime Intellect's Will Brown told AI Engineer companies should train $300K internal models and update weights nightly to capture institutional knowledge.
A Valid Key Was Enough to Delete Production
Speakeasy CEO Sagar Batchu showed how a valid token let an agent delete production and why identity and policy must sit in the call path.
AI writes faster than humans can review
Sourcegraph's Andrew Norrish told AI Engineer that review is now the bottleneck as AI-written PRs double in count and size.
Nick Nisi's Fruit Fly Code Review Is Real
WorkOS engineer Nick Nisi wired a fruit fly connectome to an AI nose and let it vote on PRs to argue human code review must change.
If You Score 100%, Your Test Failed
Ramakrishnan Lokanathan told AI Engineer a 100% eval means the test is wrong, and showed E-Slams, a game-based harness for agents.
Oracle says memory, not models, is the moat
Oracle's Richmond Alake demoed agent memory and governance on AI Engineer, pitching owned memory and identity-bound access as the differentiator.
Bumble found 78% wanted AI photo feedback but only 10% used it
Bumble's AI photo feedback won 78% support in research but drew 10 to 15% usage in its AB test, forcing guardrails and cheaper models.
Oracle Lays Out When AI Agents Should Remember
Oracle's Wojciech Pluta detailed a memory promotion gate on AI Engineer to decide what agents keep, with RAG as the read path and memory as the write path.
The Cheaper Model That Billed More
Vals AI's Rayan Krishnan told AI Engineer a cheaper list price can mean higher cost per task when weaker models burn extra tokens.
AI writes 2,450% more code, ships 30% more
Tariq Shaukat told AI Engineer agentic code lifts output 2,450% but releases only 30%, and that human oversight and LLM judges fail at scale without engineered confidence.
Revolut says self-hosted AI cut coding costs 90%
Revolut says 56 B200s and GLM 5.3 cut coding AI costs to 30 cents per million tokens for 2,000 engineers, but only after heavy tuning.
Agents break daily, yet teams still trust them
Temporal survey: 80.8% use agents daily, 49.1% in production; 85.5% trust output yet 41.1% hit issues daily, per Shy Ruparel on AI Engineer.
One-Line Change Can Break the Stack, Qodo Warns
Cole Gannaway demoed how one line in Bitwarden's data model ripples across backend, web, mobile and CLI, and why green checks miss it.
Who Made That Request? Agent Auth Has No Opt-Out
Kamil Potrec told AI Engineer that agents need their own identities and short-lived, scoped credentials, with gateways to keep tokens out of agent reach.
Exploits now beat patches by 70 days
Chainguard's Patrick Smyth told AI Engineer that exploits now beat patches by 70 days, blaming frontier agents for faster chaining and supply chain worms.
A $900M Error Is the AI Agent Playbook
Two Sigma's John Feminella used Citibank's $900M Revlon error to pitch five controls for AI agents at AI Engineer.
Agents Need Bosses, Badges and Boundaries
Tim Tyler told AI Engineer that governing agents like software misses the point, proposing badges, sponsors and scorecards instead.
Runlayer says it tracked 300 agents for 60 staff
Runlayer found 300 daily agents for 60 staff and detailed how remote prompts and over-permissioned tools enable exfiltration.
Arize says its agents can fix themselves overnight
Arize product manager Fuad Ali demoed Signals, an SRE agent that reviews traces on a 6-hour loop and files fixes, on AI Engineer.
Greptile shows how not to build review bots
Daksh Gupta detailed how to build an AI code review agent with sandboxes, model inversion and a living wiki, and why cheap DIY bots miss rare catastrophic bugs.
Google says managed agents push code off your laptop
Jack Wotherspoon demoed Gemini managed agents that run the anti-gravity harness in a cloud sandbox and keep secrets behind a proxy.
LiveKit says its voice interruption rate is just 4.5%
LiveKit's Jesse Hall told AI Engineer his turn detector cuts interruptions to 4.5% versus 9.9%, within a 500-600 ms human-like window.
Arize's Eval Demo Skips The Hard Tradeoffs
Arize's Head of Product demoed building a first AI eval on AI Engineer, pushing cheap Decision Judges but leaving the explanation and calibration gap unaddressed.
The AI Engineer Now Builds the Loop, Not the Feature
Aparna Dhinakaran told AI Engineer the AI engineer now builds the loop, not the feature, and demoed Arize Signal's always-on agent for traces.
Frontier models still fail the buy-side memo test
Google DeepMind and LinqAlpha showed why frontier models fail as buy-side analysts and previewed Mandate Bench, a 50-task memo benchmark.
One line fixed safety, broke the task
A 17-point safety gain wiped by a 17-point task loss shows why Arize says agent regressions live in traces, not diffs.
AI can parse credit but not predict it yet
Axim Intelligence CEO Harneet Kaur showed at AI Engineer that two frontier models gave split odds on iHeartMedia until a layered pipeline made them converge.
A Hedge Fund Tests AI With AB Tests, Not Hype
Contour Asset Management runs live AB tests and tracks cost per useful insight to see if AI actually improves hedge fund research.
Arca says its agents can run a business for 10 days
Aman Singh told AI Engineer how Arca compresses a 10-day tax workflow by letting agents watch signals and pull humans in only to approve.
Agents escaped the sandbox. Identity is the fix.
Andrey Rybka detailed real agent breaches at AI Engineer, from sandbox escapes to DNS theft, and mapped fixes to identity and deterministic controls.
Stop handing agents your user token
WorkOS demoed off.md and ID-JAG to let AI agents sign up safely, plus scoped agent tokens that limit what agents can do.
AI startups will die without a compute market
Evan Conrad told AI Engineer that AI startups face a credit wall that only a physically settled compute market can fix.
Your model isn't the bottleneck, Jellyfish finds
Jellyfish data on 50,000 engineers shows model choice explains <1% of PR throughput. Skill use and context matter far more.
1,368 AI failures point to deployment, not models
New York Life researchers coded 1,368 AI failures and found deployment controls, not model strength, best predict harm.
Air Canada Built Tests Without Benchmarks
Nathanael Weill told AI Engineer how Air Canada tests contract-review agents without benchmarks by planting answers first and grading with code.
Tone Bench grades slop but only in one voice
Towards AI founder launches Tone Bench to measure slop in his own voice, but a single-voice test and biased judges leave general claims unproven.
Code Quality Matters Only When Software Must Last
Capital One's Max Kanat-Alexander says code quality matters for durable systems, not for disposable ones, and defines four factors to tell the difference.
Vanguard Cut Research From 7 Days to 15 Minutes
Vanguard's Gareth Yoder said exposing legacy code via MCP cut research from 7 days to 15 minutes during a cash ledger modernization handling 30M calls a day.
MiniMax says faster is not always finished
MiniMax's Morgan Suo says M3.1 Flash is faster but urges testing the whole research job, not tokens per second, before switching models.
